text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We present new results for ground-state candidate energies of Mg-rich olivine (MRO) clusters and use the binding energies of these clusters to determine their nucleation rates in stellar outflows, with particular interest in the environments of core-collapse supernovae (CCSNe). Low-lying structures of clusters [(Mg$_2$SiO$_4$)$_{n}$]{} $2 \le n \le 13$ are determined from a modified minima hopping algorithm using an empirical silicate potential in the Buckingham form. These configurations are further refined and optimized using the density functional theory code Quantum Espresso. Utilizing atomistic nucleation theory, we determine the critical size and nucleation rates of these clusters. We find that configurations and binding energies in this regime are very dissimilar from those of the bulk lattice. Clusters grow with SiO$_4$-MgO layering and exhibit only global, rather than local, symmetries. When compared to classical nucleation theory we find suppressed nucleation rates at most temperatures and pressures, with enhanced nucleation rates at very large pressures. This implies a slower progression of silicate dust formation in stellar environments than previously assumed.'
address:
- '[email protected]'
- '[email protected]'
author:
- 'Christopher M. Mauney'
- Davide Lazzati
bibliography:
- 'bib\_elsevier.bib'
title: 'The formation of astrophysical Mg-rich silicate dust'
---
Introduction
============
Cosmic dust presents an intriguing laboratory to the physicist. The cycle of dust, it’s crucial role in a great galactic recycling and processing of material, necessitates a broad view. Yet any approach to the study of dust that does not incorporate the details of small-scale chemistry and kinetics, up to and including quantum effects, will fail as a predictive model. This is especially true when examining dust formation, when the physics is in the most dynamic phase of the cycle. Silicate dust is a major component of dust present in the ISM. Absorption features at 9.7 $\mu$ m and 18 $\mu$ m are associated with the Si-O stretching and O-Si-O bending modes in silicates [@gibb2004interstellar]. Similar spectral features have been observed in other galaxies [@teplitz2006silicate; @roche2007silicate]. These lines are strong and broad, indicating that in the diffuse ISM silicates are structurally amorphous [@kemper2004absence]. However, crystalline silicate features have been observed around AGB stars and stars with disks [@molster2002crystalline; @olofsson2009c2d], and in comets [@wooden1999silicate]. These observations indicate that a substantial fraction of silicate dust grains are in a crystal structure before being injected into the ISM. It is possible that crystallization occurs directly from the vapor, and subsequent processing by grain-grain collisions, shock sputtering, and thermal annealing leads to amorphization [@henning2010cosmic]. However, the reverse is also possible, where amorphous silicate is formed from the vapor and later processed into crystalline structure.
Determining the formation pathway of silicate dust grains is necessary for making more accurate predictions of dust properties. In this paper we continue the approach set forth in our previous paper on carbon dust precursors [@Mauney15] with the Mg-rich olivine (MRO) clusters [(Mg$_2$SiO$_4$)$_{n}$]{}, for $2 < n < 13$. In principle, to exactly determine nucleation of multi-species molecules, all pathways through a high-dimensional Gibbs energy surface are necessary. For the reasons given in the next section, we instead follow a fixed stochiometric ratio into larger molecules.
This paper is organized as follows: in section \[sec:methods\] we detail our methods for cluster ground state configuration and density functional theory energy calculations. Results of these calculations are presented in section \[sec:results\], showing binding energies of clusters, critical sizes, and nucleation rates. We summarize these results and discuss their impact in section \[sec:conclusion\].
Methods {#sec:methods}
=======
Selecting the nucleation pathway
--------------------------------
In principle the formation pathway of silicates can take many directions through precursor molecules that do not necessarily maintain a fixed stoichiometry. It is to be expected, however, that the stoichiometry of the precursor molecule will eventually approach the one of the crystal. Since it is computationally impractical to compute the properties of all possible precursor molecules, in this study we only investigate molecules of the form [(Mg$_2$SiO$_4$)$_{n}$]{}. While this is a limitation of this work, it is worth noting that it is still a significant improvement over the capillary approximation used in other investigations of silicate nucleation (see, e.g., [@Todini01; @Nozawa03]). Not only the fixed stoichiometry is implicitly assumed, but also the precursor molecules are taken to be spherical and possessing the physical properties - such as surface energy - of the bulk material. While some studies [@Goumans12; @Bromley16] have attempted a full study of precursor molecules of free stoichiometry, they were limited to small molecules ($n < 10$) (the larger molecules in those studies did follow fixed stoichiometries).
Determining ground state configuration
--------------------------------------
A molecular configuration approximating the ground state of the molecule must be supplied to the DFT calculation. We use an implementation of the global minima search algorithm minima-hopping (MH) developed by Goedecker [@Goedecker04]. Minima hopping does not generate new configurations based on random moves, as other global minima techniques such as basin hopping, but rather smoothly follows the energy surface by applying molecular dynamics to the system. Starting from an initial state, the system is given a kinetic energy and allowed to evolve according to the equations of motion. The stopping criterion of the molecular dynamics algorithm is passing over one or two hills on the potential energy surface (that is, the system goes over from increasing to deceasing energy).
New configurations are subjected to a minimization after the molecular dynamics step. The minimized configuration is compared to the previous (beginning) one. If the two configurations are determined to be the same (in this case, comparing inter-atomic distances of atoms), the algorithm returns to performing the molecular dynamics step with the an increased kinetic energy. If the system escapes the current energy well into a new unique one (that is, a new configuration is found), the kinetic energy is reduced, and the process is started again. Details on the specifics of implementation of the algorithm can be found in [@Goedecker04]. All the parameters we used in this article can be found in Table \[tbl:mh\_parameters\_0\].
We use a modified version of minima hopping that implements an atom-swap. This algorithm defines a small (1-5%) probability $\alpha$ that the next cycle will swap positions of two randomly selected atoms in the molecule rather than do a molecular dynamics run. This use of non-local transformations has been applied to basin-hopping and has been found to increase the efficiency of such searches [@rondina2013revised]. In our approach, a temperature $T_s$ is used in a simple Metropolis condition $\exp(-\Delta E / T_s)$ after the swap to determine if the swapped positions should be accepted as the new state of the search, where $\Delta E$ represents the relative change in energy.
The potential function used in this study is the Bees-Kramer-van Santen (BKS) model [@van1990force]. The fuctional form is given as a combination of Coulomb and Buckingham potential terms
$$\begin{split}
\label{eq:buckingham_potential_0}
U(\mathbf{r}) &= \sum_{i < j} U_{ij} + \sum_{i < j < k} U_{ijk}\\
U_{ij} &= \frac{q_i q_j}{r_{ij}} + A_{ij}\exp (-r_{ij}/B_{ij}) - C_{ij} r_{ij}^{-6}\\
U_{ijk}&= K_{ijk} \left(\theta_{ijk} - \theta_{0,ijk}\right)^2
\end{split}$$
where the variable $r_{ij}$ is the inter-atomic distance between atoms $i$ and $j$ and $q_i$ is the charge on atom $i$. The pair-wise parameters $A_{ij}$ and $B_{ij}$ specify a short range repulsive force and $C_{ij}$ a long range attractive force. The three-body parameter $K_{ijk}$ is a force constant, $\theta_{ijk}$ is the angle formed from atoms $i,j,k$, and $\theta_0$ is the equilibrium angle. Values for these parameters are determined from ab-inito studies and reliably reproduce the properties of crystalline and large amorphous silicates. The values used in this study are given in Tables \[tbl:charge\_potential\_parameters\_0\], \[tbl:pair\_potential\_parameters\_0\], & \[tbl:angular\_potential\_parameters\_0\] [@hassanali2007model; @roberts2001investigation; @flikkema2003new].
Table \[tbl:mh\_parameters\_0\] gives the values used for our minima hopping parameters. Verlet integration is used to evolve the system during the molecular dynamics step of the minima hopping algorithm. A Broyden-Fletcher-Goldfarb-Shanno (BFGS) [@shanno1985broyden] algorithm is used for the minimization step.
[ll]{} Mg & +2.0\
O & -2.0\
Si & +4.0\
\[tbl:charge\_potential\_parameters\_0\]
[llll]{} O-O & 22764.0 & 1/0.149 & 27.88\
Si-O & 1283.907 & 1/0.32052 & 10.66158\
Si-Si & 79502.113 & 1/0.201 & 446.780\
Mg-O & 821.6 & 1/0.3242 & 0.0\
\[tbl:pair\_potential\_parameters\_0\]
[llll]{} O-Si-O & 2.097 & 109.47$^{\circ}$\
O-Mg-O & 2.097 & 90.0$^{\circ}$\
\[tbl:angular\_potential\_parameters\_0\]
[lllllll]{} 0.95 & 1.05 & 1.05 & 1.05 & 0.95 & 0.5 & 1000 \[tbl:mh\_parameters\_0\]
Our atom-swap method results in most cases in faster searches. Figure \[fig:mh\_as\_compare\] shows a comparison between traditional minima hopping and minima hopping with atom-swap for a $n=4$ and $n=6$ test case. In both cases, the atom-swap method yields lower energies. Figure \[fig:mh\_as\_param\_compare\] compares atom-swap algorithms with different parameters. The algorithm is not very dependent on the selection of parameters, but appears to work best at $\alpha=5 \%$ and $T=300$. We select these values for this article.
![Performance of traditional minima-hopping algorithms (dashed) compared with minima-hopping with atom-swap included (solid), using the same initial starting configuration.[]{data-label="fig:mh_as_compare"}](figures/compare_atom_swap_mh_0){width="100.00000%"}
![Comparison of different atom-swap parameters (swap % and swap temperature) for the same initial starting configuration.[]{data-label="fig:mh_as_param_compare"}](figures/compare_atom_swap_mh_1){width="100.00000%"}
The number of discoverable minima increases exponentially with the size of the molecule. It is therefore advantageous, especially for larger silicate clusters, to precondition the input to the minima-hopping algorithm. For larger molecules ($n > 5$), we borrow techniques from genetic algorithms used in structure prediction [@Deaven95], where previously determined clusters can act as sub-units in the seeds of larger clusters. Inputs are constructed as combinations of smaller molecules found in previous runs. An example is provided in Figure \[fig:mh\_combine\_0\], where we use the optimized state of [(Mg$_2$SiO$_4$)$_{3}$]{} to generate the initial search state of the [(Mg$_2$SiO$_4$)$_{6}$]{} molecule.
![Example procedure of generating starting configurations of larger cluster searches.[]{data-label="fig:mh_combine_0"}](figures/combine_A){width="100.00000%"}
It is possible that the lowest-lying state produced using an empirical potential will not be the lowest in the DFT calculation. To address this, we select 8 lowest-lying candidate configurations from the global search. These are screened using a low-resolution DFT relaxation, and the candidate with the lowest reported energy is selected for a full DFT calculation.
Many of our larger [(Mg$_2$SiO$_4$)$_{n}$]{} molecules appear amorphous and lack clear symmetries. The potential given by Eq.(\[eq:buckingham\_potential\_0\]) can be dominated by electrostatic forces between the Mg$^{2+}$ anions and SiO$_4^{4-}$ cations, limiting the manifestation of large-scale symmetries arising from chemical bonding. However, the thoroughness of our search and robustness of the MH algorithm gives us confidence that our configurations are energetically low-lying states.
Quantum chemistry calculations of [(Mg$_2$SiO$_4$)$_{n}$]{} clusters
--------------------------------------------------------------------
We use the DFT software Quantum ESPRESSO v5.3 (QE) [@giannozzi2009quantum] to calculate the binding energies of the [(Mg$_2$SiO$_4$)$_{n}$]{} clusters. What follows is an overview of DFT and the plane-wave approach of QE uses to solve the coupled Kohn-Sham (KS) equations. For a more thorough exploration of the theory, we refer the reader to [@Mauney15; @MartinTB]. An initial guess of the many-body electron density $n(\{\mathbf{r_i}\})$, as a function of the electron positions $\{\mathbf{r_i}\} = (\mathbf{r_1},\mathbf{r_2},...,\mathbf{r_N})$, is made and is used to construct an effective potential $V_{eff}([n])$. An updated density $n'(\{\mathbf{r_i}\})$ is found by solving the Kohn-Sham eigenvalue equations
$$\left[\frac{-\hbar^2}{2 m_e} \nabla_i^2 + V_{eff} \right]\phi_i = \epsilon_i \phi_i$$
where $\hbar$ is Planks reduced constant, $m_e$ is the electron mass, $\phi_i$ are the KS orbital basis functions and $\epsilon_i$ are the eigenvalues. The new density is recovered as
$$n'(\{\mathbf{r_i}\}) = \sum_i | \phi_i |^2$$
This iterative procedure is done until self-consistency is achieved, that is when $|n - n'| \le \delta$, with $\delta$ a convergence parameter close to zero. We will be subjecting our clusters to local minimization using BFGS, which can be very computationally expensive. Therefore we set $\delta = 10^{-6}$ to allow for an efficient minimization and still maintain good accuracy of the energy.
The binding energy $E_b$ is given as
$$E_b = E_t - n E_1
\label{eq:binding_energy_0}$$
where $E_t$ is the total energy and $E_1$ is the single monomer total energy. The difference represents all energy released from the system following cluster formation.
Free energy and nucleation
--------------------------
To calculate nucleation rates, we first need to determine the change in free energy when clusters grow. We can construct the free energy of a cluster of size $n$ as
$$G(n) = G_V(n) + G_S(n)
\label{eqn:gfe_0}$$
where $G_V(n)$ is referred to as the volume term, and $G_S(n)$ as the surface term. These represent, respectively, the energy release when moving vapor monomers to the new phase, and the energy barrier necessary to overcome when doing so. The volume term comes from familiar thermodynamics
$$G_V(n) = -n k T \ln S
\label{eqn:gfe_V}$$
with supersaturation ratio (hereafter saturation) $S = p/p_e$ where $p_e$ represents the equilibrium pressure, and temperature $T$. Following [@Kozasa87], we take the pressure as the partial pressure of the key species in the formation of the clusters. The key species represents the constituent element of the vapor with the lowest collision rate, and to good approximation the rate of nucleation is controlled by the density of the key species. For MRO nucleation in astrophysical conditions we find Mg to be the limiting element, in most cases. The equilibrium pressure is given as function of temperature $T$ as
$$\ln (p_e) = -A/T + B$$
where $A$,$B$ are fitted thermodynamic constants. We use the values by [@Nozawa03], $A=18.62 \times 10^4 K$ and $B=52.4336$.
In the classical case of CNT, the capillary approximation is used to represent the surface term. In the atomistic case, we use
$$G_S(n) = \lambda n - E_n
\label{eqn:gfe_S}$$
where $E_n$ is the binding energy of a cluster of size $n$ and $\lambda n$ represents the binding energy of $n$ monomers in the *bulk* solid phase (e.g. the infinite lattice).
For a full atomistic formulation, we go from $n$ continuous to $n$ discrete, and from Eq. (\[eqn:gfe\_0\] -\[eqn:gfe\_S\]) construct the work of cluster formation $W_n$ $$W_n = -n k T \ln S - (E_n - \lambda n)
\label{eqn:wcf}$$ The maximum value of $W_n$ represents the critical cluster size; values at the critical size are denoted with a $*$, so that the critical size is $n^*$, the WCF is $W^*$, ect. To very good approximation the stationary nucleation rate $J_s$ is only a function of critical values
$$J_s = z f^* C^*
\label{eq:jrate}$$
where $z$ is the Zel’dovich factor, $f^*$ is the attachment rate onto critical clusters and $C^*$ is the concentration of critical clusters. The Zel’dovich factor accounts for the possibility that critical clusters will spontaneously lose a monomer and decay into smaller clusters, rather than grow into the new phase. For the purposes of this work we take $z=1$ (that is, all critical clusters will grow into the new phase).
Determination of $E_1$ and $\lambda$
------------------------------------
The properties of the molecule [(Mg$_2$SiO$_4$)$_{1}$]{}, particularly the binding energy $E_1$ and the bulk lattice cohesive energy $\lambda$ of a forsterite lattice with respect to [(Mg$_2$SiO$_4$)$_{1}$]{}, are necessary for our calculations. There exists robust literature on the chemistry of forsterite crystals, but little of it explores possible values of $E_1$ and there is disagreement on the value $\lambda$.
To determine the structure of $E_1$, we take a single [(Mg$_2$SiO$_4$)$_{1}$]{} from the bulk lattice consisting of the octahedral sites M1, M2, and the SiO$_4$ belonging to sites O2 and O3. This molecule then undergoes ionic relaxation in DFT. The ground states of the constituent atoms Mg, Si, and O are then calculated. The binding energy $E_1$ is given by
$$E_1 = E_{T} - 2 E_{Mg} - E_{Si} - 4 E_{O}
\label{eq:e_1}$$
The calculation of $\lambda$ is similar. To set up the lattice, four forsterite molecules are arranged in a supercell with an orthohombic lattice, and then relaxed in DFT as above. The energy $\lambda$ is then $$\lambda = (E_{lattice} - 4 E_1) / 4
\label{eq:lambda_0}$$
Because in the bulk silicates may be either crystalline or amorphous it would be better to use a $\lambda$ reduced from the bulk crystalline to be able to explore amorphous growth. Studies of amorphous and crystalline material properties demonstrate that amorphous material has a lower bulk binding energy than the crystalline form (e.g. [@thogersen2008experimental; @gonccalves2016molecular]). Given the configurations of the MRO clusters in Figure \[fig:fu\_table\], it is reasonable to conclude that MRO grains grow amorphously. Therefore we select value that is lower from our crystalline lattice calculation but still within the range of other studies. In the appendix we explore how different values of $\lambda$ impact our results.
$E_1$ -389.6 Ry
----------------------- -----------
$\lambda$ (QE) 9.5
$\lambda$ (ref)
$\lambda$ (this work) 8.0
: $E_1$ and $\lambda$.[]{data-label="tbl:e1_lambda_values"}
Results {#sec:results}
=======
Ground state configurations of (Mg$_2$SiO$_4$)$_n$ clusters
-----------------------------------------------------------
Our clusters are shown in Fig. \[fig:fu\_table\]. Small molecules exhibit distinct symmetries, whereas large molecules become more amorphous, but with an underlying layering of MgO and SiO$_4$. We find no discernible tendency towards a bulk forsterite lattice structure, and expect clusters consisting of much larger numbers of monomers are necessary for a recognizable lattice structure to form. See the work of [@Horbach96; @Noritake14] for a more systematic overview of size effects in silicate dynamics.
Cohesive energies ($E_b / n$) are plotted in Fig. (\[fig:ebnd\_0\]), along with the selected value of $\lambda$. These plots suggest convergence of these values at large monomer numbers. Edge and surface effects are still prominent in this regime (Horbach et all 1996), leading to non-monotonic growth of the cohesive energies.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu2 "fig:"){width="6em"} n = 2]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu3 "fig:"){width="6em"} n = 3]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu4 "fig:"){width="6em"} n = 4]{}
[![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu5 "fig:"){width="6em"} n = 5]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu6 "fig:"){width="6em"} n = 6]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu7 "fig:"){width="6em"} n = 7]{}
[![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu8 "fig:"){width="6em"} n = 8]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu9 "fig:"){width="6em"} n = 9]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu10 "fig:"){width="6em"} n = 10]{}
[![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu11 "fig:"){width="6em"} n = 11]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu12 "fig:"){width="6em"} n = 12]{} [![Clusters configurations found after DFT optimization. Colors represent atomic species: oxygen(red, small), silicon(tan, medium), magnesium(green, large).[]{data-label="fig:fu_table"}](figures/fus/fu13 "fig:"){width="6em"} n = 13]{}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Cohesive energies (=$E_b/n$) of [(Mg$_2$SiO$_4$)$_{n}$]{} clusters. The constant plots c-MRO and a-MRO represent the bulk cohesive energy of crystal and amorphous [(Mg$_2$SiO$_4$)$_{n}$]{}, respectively.[]{data-label="fig:ebnd_0"}](figures/ebnd_0){width="100.00000%"}
Critical sizes and nucleation rates
-----------------------------------
With binding energies at hand, we can find the free energy values and determine critical sizes and nucleation rates. Fig (\[fig:wcf\_s1000\]) is an example of free-energy curves at selected value of saturation across a range of temperatures. At low temperature values the WCF has no maximum in the range of monomer numbers, indicating a critical size will be larger than the $n*=13$, the largest cluster we studied. For the sake of completeness we would like to investigate clusters in the regime of $n>13$, these clusters are too large to be efficiently computed with our minima search and DFT calculations. Our size limit of $n=13$ is equivalent to 91 atoms, already stretching the limits of our computational techniques. Physically, moreover, molecules of this size have a very slow formation rate, and their contribution to dust creation can be taken to be negligible. Conversely, at large values of $(T,S)$ there is little to no free energy barrier and nucleation begins quickly with small clusters $n \approx 2$.
The WCF plots constructed, we can locate the maximum value and determine the critical cluster size $n^*$. These critical sizes across a range of temperatures and saturations are plotted in Fig. (\[fig:ncrit\_l8-0\]), along with results from CNT for comparison. As expected $n^*=2, 3$ for extreme environments, where density and temperature reduce the free energy barrier to the new phase and nucleation begins quickly. $n^*=7$ is prominent in the middle regions. Comparatively colder and sparser regions of the vapor have larger free energy barriers, leading to larger critical sizes $n* \ge 10$.
With critical sizes determined, nucleation rates follow from Eq. (\[eq:jrate\]), and are plotted in Fig. (\[fig:jrate\_l8-0\]) along with the results from CNT for comparison. Nucleation is suppressed by several magnitudes at all but the largest temperatures and saturations compared to the classical case. Our results imply that nucleation does not take place at significant rates until at least several tens of saturation. This result is consistent with the critical sizes at low saturations being very large.
![Work of Cluster Formation for $S=1000$ plotted across a range of temperatures. At high temperatures the curve is flattened and critical clusters will form at small $n$. For lower temperatures the WCF is more jagged, leading to higher critical sizes.[]{data-label="fig:wcf_s1000"}](figures/wcf_s1000){width="100.00000%"}
![Critical size of [(Mg$_2$SiO$_4$)$_{n}$]{} clusters as a function of saturation at constant temperatures. Dashed lines are the CNT result, solid lines are the results of this article. Note that we impose a minimum critical size of $n=2$.[]{data-label="fig:ncrit_l8-0"}](figures/ncrit_l8-0_A){width="100.00000%"}
![Nucleation rates of [(Mg$_2$SiO$_4$)$_{n}$]{} clusters as a function of saturation at constant temperatures. Dashed lines are the CNT result, solid lines are the results of this article.[]{data-label="fig:jrate_l8-0"}](figures/jrate_l8-0_A){width="100.00000%"}
Conclusions and Discussion {#sec:conclusion}
==========================
We find new results for the ground state configuration of [(Mg$_2$SiO$_4$)$_{n}$]{} clusters and determine the rates of nucleation of MRO clusters in stellar outflows. Using DFT we have determined precise values of binding energies for these clusters, and used these binding energies in an atomistic formulation of nucleation theory to produce nucleation rates for silicates. These values have been compared to CNT and have been found to be significantly different than the classical case. Except for environments of large temperatures and saturations, out ANT approach finds lower formation rates of critical clusters when compared to CNT.
While we expect this trend to hold for regions of low temperature and saturations, the critical sizes that would be expected to be found at these environments are too large for our methods to efficiently determine. It is possible that at low temperatures and at saturations of approximately unity, ANT will predict enhanced nucleation over that of CNT. However, even in CNT nucleation in these environments is negligibly small (see, for instance, the low saturation regions of Figure \[fig:jrate\_l8-0\]).
Locating the ground state of large multi-component systems is difficult. Further limiting research into these clusters is the use of the empirical BKS model in studying nanoscale clusters. This potential form is known to perform poorly when describing the surface chemistry of silicates[@ceresoli2000two; @Flikkeme03]. Ideally an *ab-initio* ground-state search would be preferable. However, global minima techniques require many million energy evaluations, and at present this approach would too computationally prohibitive.
Except for the small clusters $n=2$, $n=3$, and $n=4$ the cluster configurations found do not exhibit strong symmetries or growth patterns. Larger clusters are amorphous and lack any well-structured ground state. This can be seen in the binding energy of the clusters (Figure \[fig:ebnd\_0\]), which stops growing monotonically. In the specific cases of $n=7$ and $n=10$ there is a significant drop in stability. While it is possible that there are configurations of clusters of comparable energies, the relative energies of the clusters will remain the same (for instance, $n=7$ and $n=10$ showing lowered stability.)
The drop in stability at $n=7$ and $n=10$ can be explained by examining the mean coordination (number of bonds) in these clusters. As seen in Figure \[fig:coordination\_0\], there is an increase in Si coordination at $n=7$, indicating that the Si atoms are being weakly bonded to fifth O atom and weakening the stability of the SiO$_4$ tetrahedra. A similar, albeit smaller, effect is also noticeable at $n=10$. It is not surprising, then, that these clusters turn out to be prominent critical sizes. Adding a monomer from the vapor to these clusters will release more free energy by lowering the mean coordination of the SiO$_4$ tetrahedra. At larger cluster sizes there is an overall trend in increasing Si coordination as a result of MgO-SiO$_4$ layering (see below).
![Average coordination of each atomic species plotted against cluster size. Increased Si coordination is seen at $n=7$ and $n=10$.[]{data-label="fig:coordination_0"}](figures/coordination_0){width="100.00000%"}
The layering present in our results is explained by Noritake [@Noritake14]. Briefly, Si-O bonds are much stronger than Mg-O bonds, leading to Si atoms preferentially sharing an oxygen bond as Si-O-Si. Thus SiO$_4$ tetrahedra cluster together and have overlapping O sites. These shared O atoms are not available to bond to Mg atoms, and the Mg will gather around excess O atoms. Our results suggest that during nucleation MROs begin as a melt, and that the emergence of Mg-Si layering explains the location of MRO critical cluster sizes.
Our results support the model that MRO dust precursor molecules form amorphously, when following a fixed stoichiometry formation path. Evidence from simulations of silicate glasses [@Horbach96] suggests that the transition from silicate melt to crystal lattice occurs in ensembles of several hundred to thousands of atoms. Crystalline silicates in late-stage stellar environments will therefore form after a period of processing and annealing. The presence of OH and Fe in the surrounding vapor will enhance the formation of crystalline MROs [@jager1998steps], and their inclusion in nucleation studies may result in more crystalline-like ground states.
Critical sizes for MRO clusters are large in regions of low pressure and temperature. In Fig. \[fig:ncrit\_l8-0\] it can be seen that regions of low temperature and pressure show no critical size in the range of sizes we considered. In the absence of a critical size it is not possible to determine the nucleation rates in ANT. However, as can be seen in Fig. \[fig:jrate\_l8-0\], nucleation rates in these regions will be near negligible, and nucleation in this regime will not be a significant contributor to dust creation. This does not mean that MRO nucleation at low temperature is impossible. Rather, the gas must achieve a very large saturation before nucleation of MRO clusters becomes efficient. At lower temperatures, nucleation of smaller silicate clusters (MgO, SiO) is more prominent [@Jeong03; @Goumans12]. However, at this stage most of the condensable monomers will have nucleated into MRO clusters.
Nucleation rates are suppressed compared to CNT for all conditions save for the largest saturations we studied. In the classical theory, the value of the surface tension allows for clusters to quickly grow to the critical cluster size and begin nucleation. However, our results indicate that this value of surface tension is too low in the classical case (for instance the value given in [@Boni56] $\sigma=425$ erg cm$^-2$), and does not accurately represent the surface physics of nanoscale silicate clusters.
Silicate nucleation rates are suppressed at low temperatures and saturations. This suggests ’staged’ nucleation periods, with carbon forming and growing quickly while silicate clusters take longer to form. Interestingly, the chemical kinetics approach reverses the priority of this staging, with carbon dust forming later than silicate clusters [@Sarangi15]. More detailed observations of the dust formation histories of CCNSe outflows will be useful in resolving this disagreement between models.
This work was supported in part by NSF grants 1150365-AST and 1461362-AST (DL & CM). We also would like to thank Professors Goumans and Bromley for sharing their research and providing insights into silicate cluster growth.
Effect of $\lambda$ values on nucleation rates
==============================================
The parameter $\lambda$ represents the work required to remove a monomer from the bulk solid phase. This value is used in the atomistic formulation of nucleation rates to determine the surface energy of a cluster, as in Eq. (\[eqn:gfe\_S\]). The complex nature of amorphous silicate growth prevents simple evaluation of this parameter. In section \[sec:methods\] we present our approach to selection of this value for this work. In this appendix we present results of nucleation rates with lower and higher values of $\lambda$.
Lower values of $\lambda$ imply a lower energy barrier between phases, and this will induce faster nucleation. This can be seen in Fig. \[fig:jrate\_l7-5\]. Nucleation rates are still suppressed compared to classical nucleation, but lower saturations are required to begin nucleation. The corresponds physically to a mostly free formation pathway. While our results indicate that silicate growth is amorphous, there is indeed some underlying constraints on growth (for instance, the MgO-SiO$_4$ layering) that would argue against using such a low value for $\lambda$.
![Nucleation rates at $\lambda = 7.5$ eV. Dashed lines are the CNT result, solid lines are the results of this article. Temperatures given in Kelvins.[]{data-label="fig:jrate_l7-5"}](figures/jrate_l7-5_2D){width="100.00000%"}
Fig. \[fig:jrate\_l8-5\] represents are larger value of $\lambda$, and shows significantly less nucleation. This corresponds to a much more constrained formation pathway, i.e. crystal growth. As we do not see evidence of crystal structure in our low-lying cluster configurations, it seems unlikely that such a large value of $\lambda$ is applicable to astrophysical dust nucleation.
![Nucleation rates at $\lambda = 8.5$ eV. Dashed lines are the CNT result, solid lines are the results of this article. Temperatures given in Kelvins.[]{data-label="fig:jrate_l8-5"}](figures/jrate_l8-5_2D){width="100.00000%"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The process $e^+e^-\to\pi^+\pi^-$ has been studied with the SND detector at VEPP-2M $e^+e^-$ collider in the vicinity of $\phi(1020)$ resonance. From the analysis of the energy dependence of measured cross section the branching ratio $B(\phi\to\pi^+\pi^-)=(7.1\pm 1.1\pm 0.9)\cdot 10^{-5}$ and the phase $\psi_\pi=-(34\pm 4\pm 3)^\circ$ of interference with the non-resonant $\pi^+\pi^-$ production amplitude were obtained.\
\
[*PACS:*]{} 13.25.-k; 13.65.+i; 14.40.-n\
[*Keywords:*]{} $e^+e^-$ collisions; Vector meson; Hadronic decay; Detector
author:
- |
M.N.Achasov, K.I.Beloborodov, A.V.Berdyugin, A.V.Bozhenok, A.D.Bukin,\
D.A.Bukin, , T.V.Dimova, V.P.Druzhinin, M.S.Dubrovin,\
D.I.Ganyushin, I.A.Gaponenko, V.B.Golubev, V.N.Ivanchenko,\
P.M.Ivanov, A.A.Korol, S.V.Koshuba, E.V.Pakhtusova,\
E.A.Perevedentsev, A.A.Salnikov, S.I.Serednyakov, V.V.Shary,\
Yu.M.Shatunov, V.A.Sidorov, Z.K.Silagadze, A.A.Valishev,\
A.V.Vasiljev\
\
[*Budker Institute of Nuclear Physics and Novosibirsk State University*]{}\
[*630090, Novosibirsk,* ]{}\
[*Russia* ]{}
title: '**Decay $\phi\to\pi^+\pi^-$**'
---
Introduction
============
The decay $\phi\to \pi^+\pi^-$ reveals itself as an interference pattern in the energy dependence of the cross section of the process $e^+e^-\to\pi^+\pi^-$ in the region close to $\phi$ peak. The $\phi\to\pi^+\pi^-$ decay was previously studied at VEPP-2M collider [@Olya; @ND] and current PDG value $B(\phi\to\pi^+\pi^-)=(8^{+5}_{-4})\cdot
10^{-5}$ [@PDG] is based on these results.
The decay $\phi\to\pi^+\pi^-$ violates both OZI rule and G-parity conservation. The decay amplitude in Vector Dominance Model (VDM) was calculated in [@Karn]. The main contribution into the amplitude of the $\phi\to\pi^+\pi^-$ decay in this work comes from the electromagnetic $\phi-\rho$ mixing. The contribution of the $\phi-\rho$ transitions through the $\omega$ meson and other intermediate states such as $K\bar{K}$, $\eta\gamma$, etc. is estimated to be $\sim 20\%$ of electromagnetic $\phi-\rho$ mixing. The value of the branching ratio of the decay $\phi\to\pi^+\pi^-$ calculated from the decay amplitude obtained in the work [@Karn] is almost 2 times higher than current PDG value [@PDG]. Different $\phi-\omega$ mixing models were scrutinized in respect to this decay in [@Ach1]. The branching ratio calculated in this work is lower than that in [@Karn], but discrepancy between the experimental results, especially [@ND], and the theoretical prediction [@Ach1] still exists. Possible mechanisms, which could decrease the theoretical branching ratio, are discussed in [@Ach1]. One of them is the existence of direct decay $\phi\to\pi^+\pi^-$.
Experiment
==========
The experiments with SND detector (Fig. \[SNDtrans\]) at VEPP-2M $e^+e^-$ collider are being conducted since 1995. SND is a general purpose non-magnetic detector [@SND]. The main part of the SND is a 3-layer spherical electromagnetic calorimeter, consisting of 1632 NaI(Tl) crystals [@Lisbon]. The solid angle of the calorimeter is $\sim 90\%$ of $4\pi$ steradian. The angles of charged particles are measured by two cylindrical drift chambers covering 95% of full solid angle. The important part of the detector for the process under study is the outer muon system, consisting of streamer tubes and plastic scintillation counters.
![ Detector SND — section across the beam; 1 — beam pipe, 2 — drift chambers, 3 — inner scintillation counters, 4 — NaI(Tl) counters, 5 — vacuum phototriodes, 6 — iron absorber, 7 — streamer tubes, 8 — outer scintillation counters []{data-label="SNDtrans"}](p10.eps){width="8cm"}
The 1998 experiment was carried out in the energy range $2\mathrm{E}_b$=984–1060 MeV in 16 energy points and consisted of 2 data taking runs [@Prep.98]: PHI\_9801, PHI\_9802. The total integrated luminosity $\Delta L=8.6$ pb$^{-1}$ collected in these runs corresponds to $13.2\cdot 10^{6}$ produced $\phi$ mesons. The integrated luminosity was measured using $e^+e^-\to e^+e^-$ events selected in the same acceptance angle as the events of the process under study. The interference term in the $e^+e^-\to e^+e^-$ cross section due to $\phi\to e^+e^-$ decay was also taken into account. The systematic error of the luminosity measurement was estimated to be 2%.
Event selection
===============
The energy dependence of the cross section of the process $$e^{+}e^{-} \to\pi^+\pi^-
\label{pipi}$$ was studied in the vicinity of $\phi$ meson. Events containing two collinear charged particles and no photons were selected for analysis. The following cuts on angles of acollinearity of the charged particles in azimuthal and polar directions were imposed: $\mid\Delta\varphi\mid<10^\circ$, $\mid\Delta\theta\mid<25^\circ$. To suppress the beam background the production point of charged particles was required to be within $0.5$ cm from the interaction point in the azimuthal plane and $\pm 7.5$ cm along the beam direction (the longitudinal size of the interaction region $\sigma_{z}$ is about $2$ cm). The polar angles of the charged particles were required to be in the range $45^\circ<\theta<135^\circ$, determined by acceptance angle of the muon system.
The main sources of background are cosmic muons and the following processes: $$\begin{aligned}
e^{+}e^{-} \to e^{+}e^{-},
\label{ee} \\
e^{+}e^{-} \to\mu^+\mu^-,
\label{mumu} \\
e^{+}e^{-} \to \pi^{+}\pi^{-}\pi^0,
\label{ppp} \\
e^{+}e^{-} \to K_{S}K_{L}.
\label{KsKl} \end{aligned}$$ To suppress the background from the process (\[ee\]) a procedure of $e/\pi$ separation was used. It utilizes the difference in the longitudinal energy deposition profiles in the calorimeter for electrons and pions. The separation parameter was calculated for each charged particle in an event: $$K=\log\left(\frac{\mathcal{P}_e(E_1,E_2,E_3,E_e)}
{\mathcal{P}_\pi(E_1,E_2,E_3,E_\pi)}\right),
\label{epif}$$ where $\mathcal{P}_{e(\pi)}$ — the probability for an electron (pion) with the energy $E_{e(\pi)}$ to deposit the energy $E_i$ in the $i$-th calorimeter layer. $E_{e(\pi)}$ in our case is equal to the beam energy. The separation parameters distribution for both particles in collinear events with no hits in the muon system, is shown in Fig. \[epi\]. This distribution is asymmetric because the particles are ordered according to their energy depositions in the calorimeter. To select the events of the process (\[pipi\]) the cut $K_1+K_2<0$ was imposed. The background from the process (\[ee\]) was suppressed by a factor of $\sim 3000$, while only 7% of the events of the process under study were lost. Remaining background from the process $e^+e^-\to e^+e^-$ was about 1.5%.
![ Distribution of the parameters $K_1$ and $K_2$ for two electrons and pions from the processes (\[ee\]) and (\[pipi\]). Electrons occupy the top right corner while pions concentrate in the bottom left one. []{data-label="epi"}](epi.eps){width="8cm"}
The events of the process (\[mumu\]) and cosmic muons can be efficiently suppressed by the muon system. We required no hits in the scintillation counters of the muon system. The efficiency of these counters was estimated using cosmic muons selected by special cuts. Due to possible admixture of beam events which actually produce no hits in the muon counters only the lower boundary of the efficiency was obtained: 99.8%. Thus estimated contribution of cosmic muons does not exceed 0.7% of the total number of events of the process (\[pipi\]) and was neglected.
The energy dependences of the probabilities for muons and pions to produce hits in outer scintillation counters were obtained from the experimental data. With energy increasing from 492 MeV up to 530 MeV these probabilities rise from 84% up to 94% for muons and from 0.5% to 11% for pions. In the final selection of the process (\[pipi\]) the background from the process (\[mumu\]) was about 15%.
To suppress the resonant background from the processes (\[ppp\]) and (\[KsKl\]) the following cuts on energy depositions in the calorimeter were applied:
1. the energy deposition in the first calorimeter layer of the most energetic particle in an event is less than 75 MeV;
2. the energy deposition in the third calorimeter layer of the least energetic particle in an event is more than 50 MeV.
In the events of the process (\[ppp\]), which satisfy the geometrical cuts, the energetic photon from $\pi^0$ decay propagates along the direction of a charged pion producing unusually large energy deposition in the first calorimeter layer for this pion. Such events are suppressed by the first cut. The pions from process (\[KsKl\]) are relatively soft with a maximum energy of about 300 MeV and low probability of significant energy deposition in the third calorimeter layer. The second cut is crucial for the rejection of the process (\[KsKl\]). The residual cross sections of resonant background processes were estimated by Monte Carlo (MC) simulation: 0.06 nb for the process (\[ppp\]), 0.09 nb for the process (\[KsKl\]).
To determine the remaining resonant background more accurately the selected events were divided into two data samples using the parameter $\Delta\varphi$: $\mid\Delta\varphi\mid<5^\circ$ and $\mid\Delta\varphi\mid>5^\circ$. The resolution in $\Delta\varphi$ is about $1^\circ$. The main part of the events of the process (\[pipi\]) is contained in the first sample. Due to the emission of hard photons by initial or final particles and errors in the reconstruction of the particle angles some events of the process (\[pipi\]) can migrate into the second sample. The level of the resonant cross section $\sigma^{res}_2$, determined in the second sample, was used to estimate the resonant background in the first sample: $\sigma_1^{res}=k\sigma_2^{res}$. The coefficient $k=1.5\pm0.3$ was obtained by MC simulation of the processes (\[ppp\]) and (\[KsKl\]), its error is determined by accuracy of simulation of energy depositions of pions in the calorimeter. Because the level of the resonant background is low, the error in $k$ does not give significant contribution into the errors of the interference parameters. The cut $\mid\Delta\varphi\mid<5^\circ$ reduces the level of resonant background down to as low as 0.09 nb. This value is less than 1% of the process (\[pipi\]) detection cross section.
The pion polar angle distribution for the process (\[pipi\]) at beam energy higher than 520 MeV is shown in Fig. \[teta2\]. At this energy the cross sections of the resonant processes (\[ppp\]) and (\[KsKl\]) are small. The additional cut on the total energy deposition in the calorimeter $E_{tot}>400$ MeV rejects the events of the process (\[mumu\]). A good agreement between experimental distribution and the simulation of the process (\[pipi\]) shows that selected pion sample is quite pure and the level of QED background is low.
![ The pion polar angle distributions for the experimental events (points with errors) and simulation (histogram) for the process $e^+e^-\to\pi^+\pi^-$. []{data-label="teta2"}](rtf1.eps){width="8cm"}
Data analysis
=============
The fitting of the detection cross sections for the first and the second samples were performed simultaneously (Fig. \[pipitot\], \[restot\]). To describe the cross sections the following formulae were used: $$\begin{aligned}
\label{cross1}
\sigma^{vis}_1(E) & = & \sigma^{vis}_{\pi\pi}(E)
+\sigma^{vis}_{\mu\mu}(E) \nonumber \\
& & +\sigma^{vis}_{ee}(E)+\sigma^{res}_1(E), \\
\label{cross2}
\sigma^{vis}_2(E) & = & C+D\cdot(E-m_\phi) \nonumber \\
& & +\sigma^{res}_2(E), \\
\sigma^{res}_1(E) & = & k\sigma^{res}_2(E), \nonumber \\
\label{rezfon}
\sigma^{res}_2(E) & = & \varepsilon_{res} \cdot
(0.39\cdot\sigma_{\pi^+\pi^-\pi^0}(E) \nonumber \\
& & + 0.61\cdot\sigma_{K_S K_L}(E)), \nonumber\end{aligned}$$ where $E$ is the CM energy; $\sigma^{vis}_{\pi\pi}(E)$ — the detection cross section of the process (\[pipi\]); $\sigma^{vis}_{\mu\mu}(E)$ — the contribution of the process (\[mumu\]) (this process was studied in our work [@mymumu]); $\sigma^{vis}_{ee}(E)=0.2(nb)\cdot (m_\phi/E)^2$ — the contribution of the process (\[ee\]). The ratio 0.39:0.61 between the processes (\[ppp\]) and (\[KsKl\]) was taken from the simulation. The coefficients $C$, $D$ and $\varepsilon_{res}$ were free fit parameters.
The following expression was used for $\sigma^{vis}_{\pi\pi}$: $$\begin{aligned}
\sigma^{vis}_{\pi\pi}(E) & = &\sigma_0(E)\cdot R(E)\left| 1 -
Z_\pi\frac{m_{\phi}\Gamma_{\phi}}{\Delta_{\phi}(E)}\right|^2, \nonumber \\
\label{crosspi}
\sigma_0(E) & = & \frac{\pi\alpha^2\beta^3(E)\mid F_\pi(E) \mid^2}{3\cdot
E^2}, \end{aligned}$$ where $\alpha$ is the fine structure constant; $\beta(E)=(1-4\cdot m^2_\pi/E^2)^{1/2}$; $m_\phi$, $\Gamma_{\phi}$, $\Delta_{\phi}(E)
= m_{\phi}^{2} - E^2 - iE\Gamma(E)$ — $\phi$-meson mass, width and propagator respectively; $\sigma_0(E)$ — the Born cross section of the process $e^+e^-\to\pi^+\pi^-$; $Z_\pi$ — complex parameter characterizing strength of the interference. Two representations of $Z_\pi$ are used in different works: $Z_\pi=Q_\pi e^{i\psi_\pi}=\mathbf{Re} Z_\pi+i\mathbf{Im} Z_\pi$. $F_\pi(E)$ is the pion form factor without $\phi$-meson contribution: $$\label{piform}
\mid F_\pi(E) \mid^2=\mid F^\phi_\pi\mid^2\cdot
(1+A\cdot(E-m_\phi)+B\cdot(E-m_\phi)^2),$$ with $F^\phi_\pi$ as the pion form factor at the maximum of $\phi$ resonance. $Q_\pi$, $\psi_\pi$, $A$, $B$ and $\mid F^\phi_\pi\mid^2$ are free fitting parameters. $R(E)$ is a factor taking into account detection efficiency and radiative corrections: $$R(E)=\varepsilon_\pi\frac{\sigma_{\pi\pi}(E)}{\sigma'_0(E)
\left| 1 - Q'_\pi
e^{i\psi'_\pi}\frac{m_{\phi}\Gamma_{\phi}}{\Delta_{\phi}(E)} \right| ^2}.
\label{Rfac}$$ $\sigma_{\pi\pi}$ is the result of MC integration of differential cross section of the process (\[pipi\]) with all geometrical restrictions [@theor]. Since the probability for pions to hit the outer scintillation counters depends on energy, it was taken into account during $\sigma_{\pi\pi}$ calculation. The remaining contributions into the detection efficiency do not depend on CM energy and pions energies and were included into $\varepsilon_\pi$. The value $\varepsilon_\pi=0.234$ was obtained using MC simulation and experimental data. It is mainly determined by the cuts on energy depositions. Its independence of the pions energy was checked in the range 430 – 530 MeV using the pions from the process (\[ppp\]) with energies up to 450 MeV and pions from the process (\[pipi\]) at the beam energy 530 MeV. The geometrical cuts and the requirement on no hits in the outer scintillation counters led to 50% efficiency losses, so the total detection efficiency of the process (\[pipi\]) was approximately 12% at $E=m_\phi$.
![ The detection cross section in the first data sample . []{data-label="pipitot"}](pidat1.eps){width="8cm"}
![ The detection cross section in the second data sample . []{data-label="restot"}](pidat2.eps){width="8cm"}
The $R(E)$ was calculated by iteration method. As a first approximation the interference parameters $Q'_\pi$ and $\psi'_\pi$ from [@ND] were used. The pion form factor was taken from [@formpi] while calculating $\sigma'_0(E)$. After fitting $R(E)$ was recalculated with corrected $Q'_\pi$ and $\psi'_\pi$. This procedure was repeated until convergence was reached.
The branching ratio $B(\phi\to\pi^+\pi^-)$ is related to the interference parameters by the following formula: $$B(\phi\to\pi^+\pi^-)=\frac{Q_\pi^2\alpha^2\beta^3(m_\phi) \mid
F^\phi_\pi\mid^2}
{36\cdot B(\phi\to e^+ e^-)},
\label{Bpipi}$$ where $B(\phi\to e^+ e^-)=(2.99\pm 0.08)\cdot 10^{-4}$ [@PDG].
The fitting has been performed for each experimental run separately. The results are listed in Table \[tabres\]. The fit parameters for two runs are in statistical agreement, therefore combined fit was performed to obtain the final results also listed in Table \[tabres\]. The observed level of resonant background 0.07 nb is in a good agreement with the MC estimation of 0.09 nb. The fitted values of the coefficients $A$ and $B$ from the equation (\[piform\]) are $A=-(8.5\pm0.3)\cdot 10^{-3}\
MeV^{-1}$ and $B=(4.9\pm 1.0)\cdot 10^{-5}\ MeV^{-2}$.
To check the accuracy of the process (\[mumu\]) background subtraction, the fit to the data with more stringent event selection cuts has been done. The additional requirement that the total energy deposition in the calorimeter is higher than 400 MeV significantly reduced the muon background. The obtained interference parameters: $$\begin{aligned}
Q_\pi = 0.073\pm0.006, \nonumber \\
\psi_\pi = -(32\pm 5)^\circ \nonumber \end{aligned}$$ agrees well with the results from the Table \[tabres\].
The additional contribution into the shape of interference pattern may come from the process $$e^+e^-\to\phi\to f_0\gamma\to \pi^+\pi^-\gamma.
\label{f0g}$$ The process (\[f0g\]) interferes with the process (\[pipi\]) when soft photon is emitted by pions. This contribution estimated using CMD-2 analysis of the process $e^+e^-\to\pi^+\pi^-\gamma$ in the vicinity of $\phi$ resonance [@CMDppg] does not exceed 1.5% of the interference under study. This value was included into the systematic error.
Parameter PHI\_9801 PHI\_9802 Combined
-------------------------- ----------------------------- ----------------------------- -----------------------------
$\chi^2/NDF$ $15.3/26$ $ 27.5/26$ $48.4/58$
$\mid F^\phi_\pi \mid^2$ $2.96\pm 0.03$ $3.01\pm 0.03$ $2.98\pm 0.02$
$Q_\pi$ $0.069 \pm 0.008$ $0.078\pm 0.007$ $0.073\pm 0.005$
$\psi_\pi$ $ -(33\pm 7)^\circ$ $ -(35\pm 6)^\circ$ $ -(34\pm 4)^\circ$
$B(\phi\to\pi^+\pi^-)$ $ (6.2\pm1.4)\cdot 10^{-5}$ $ (8.2\pm1.8)\cdot 10^{-5}$ $ (7.1\pm1.1)\cdot 10^{-5}$
The representation $Z_\pi=\mathbf{Re} Z_\pi+i\mathbf{Im} Z_\pi$ is suitable to present the different contributions into the systematic error of the interference parameters:
1. the calculation of the radiative corrections: $\mathbf{Re} Z_\pi$ — 5%, $\mathbf{Im} Z_\pi$ — 3%;
2. the subtraction of the non-resonant background: $\mathbf{Re} Z_\pi$ — 0.8%, $\mathbf{Im} Z_\pi$ — 0.6%;
3. the contribution of the process (\[f0g\]): $\mathbf{Re} Z_\pi$ — 1.5%, $\mathbf{Im} Z_\pi$ — 1.5%;
4. the model dependence on the choice of the function approximating the pion form factor: $\mathbf{Re} Z_\pi$ — 1%, $\mathbf{Im} Z_\pi$ — 8%;
5. the subtraction of the resonant background: $\mathbf{Im} Z_\pi$ — 3%.
The systematic error of $\mid F^\phi_\pi\mid^2$ is determined by the error of the detection efficiency $\varepsilon_\pi$ and the accuracy of luminosity determination (2%).
The final results are the following: $$\begin{aligned}
\label{result}
\mid F^\phi_\pi \mid^2 & = & 2.98\pm0.02\pm0.16, \\
Q_\pi & = & 0.073\pm0.005\pm0.004, \nonumber \\
\psi_\pi & = & -(34\pm4\pm3)^\circ, \nonumber \\
B(\phi\to\pi^+\pi^-) & = & (7.1\pm 1.1 \pm 0.9)\cdot
10^{-5}. \nonumber \end{aligned}$$ For another representation of $Z_\pi$ we obtained: $$\begin{aligned}
\mathbf{Re} Z_\pi & = &\quad 0.061\pm 0.005 \pm 0.003, \nonumber \\
\mathbf{Im} Z_\pi & = &- 0.041\pm 0.006 \pm 0.004. \nonumber \end{aligned}$$
Discussion
==========
The obtained value of the branching ratio $$B(\phi\to\pi^+\pi^-)=(7.1\pm 1.1\pm 0.9)\cdot 10^{-5}$$ agrees well with the world average value $B(\phi\to\pi^+\pi^-)=(8^{+5}_{-4})\cdot
10^{-5}$ [@PDG] and has a 3 times higher accuracy. However there is a discrepancy between our result and the preliminary result of CMD-2 experiment [@CMD2PI]: $B(\phi\to\pi^+\pi^-) = (18.1\pm 2.5 \pm 1.9)\cdot 10^{-5}.$
The measured value $\mathbf{Im} Z_\pi=- 0.041\pm 0.006 \pm 0.004$ agrees with the theoretical predictions [@Karn] while the value $\mathbf{Re} Z_\pi = 0.061\pm 0.005 \pm 0.003$ is 2.5 times lower than the expected value. The different models of the $\phi-\omega$ mixing were examined in the work [@Ach1]. The lowest value $\mathbf{Re} Z^{th}_\pi = 0.12$ from this work also contradicts our results. This disagreement could be understood if the direct decay $\phi\to\pi^+\pi^-$ exists or/and in case of nonstandard $\rho-\omega-\phi$ mixing. One can notice that the measured branching ratio of another rare decay $\phi\to\omega\pi^0$ [@omegapi], which violates OZI rule and G-parity, disagrees with theoretical predictions.
Acknowledgement
===============
This work is supported in part by Russian Fund for basic researches (grant 99-02-16815) and STP “Integration” (No.274).
[99]{} I.B.Vasserman et al., Phys. Let. B 99 (1981) 62. V.B.Golubev et al., Yad. Fiz., V.44 (1986) 633. Review of Particles Physics,\
Europ. Phys. Jour. C, V.3 (1998). V.A.Karnakov, Yad. Fiz., V.42 (1985) 1001. N.N.Achasov, A.A.Kozhevnikov,\
Inter. Jour. Mod. Phys. A, V.7, No.20 (1992) 4825. M.N.Achasov et al., hep-ex/9909015, submitted to NIM, Section A. M.N.Achasov et al., hep-ex/9907038. M.N.Achasov et al., Preprint Budker INP 98-65 (1998). M.N.Achasov et al., Phys. Let. B456 (1999) 303. A.B.Arbuzov et al., Radiative corrections for pion and kaon production at e+ e- colliders of energies below 2-GeV, hep-ph/9703456, JHEP 9710 (1997) 006. L.M.Barkov et al., Nucl. Phys. B 256 (1985) 365. R.R. Akhmetshin et al., Phys.Lett.B 462 (1999) 371. R.R.Akhmetshin et al., Preprint Budker INP 99-11 (1999). M.N.Achasov et al., Phys. Let. B449 (1999) 122.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Physics Education Research frequently investigates what students studying physics do on small time scales (e.g. single courses, observations within single courses), or post-education time scales (e.g., what jobs do physics majors get?) but there is little research into how students get from the beginning to the end of a physics degree. Our work attempts to visualize students paths through the physics major, and quantitatively describe the students who take physics courses, receive physics degrees, and change degree paths into and out of the physics program at Michigan State University.'
author:
- 'John M. Aiken'
- 'Marcos D. Caballero'
title: Methods for Analyzing Pathways through a Physics Major
---
Introduction {#sec:intro}
============
Recruiting and retaining students in the physics major is an important challenge that departments across the country are facing [@chen2013; @hodapbackpage]. Understanding the kinds of programs and practices that can support and sustain students intending to major or those currently majoring in physics is critical to grow a diverse population of physics graduates. The research that looks at specific student experiences to develop rich descriptions of how those experiences influence students’ perceptions and choices provides some understanding [@irving2013; @irving2015]. As does the work that uses prior student experiences to model eventual outcomes [@hazari2010]. Equally important is working to understand what might be learned using data from institutions themselves. For this project, we have collected student registration data at Michigan State University (MSU) in order to develop analytic methods that help unpack the pathways into and out of the major.
MSU has collected a wide body of data on students for the last 10+ years. This data set contains information on over 100,000 students who have taken math and physics courses at MSU. Two percent of these students have declared a physics major at some point in their academic career and 0.5% of students have earned a bachelor’s degree in physics. This data includes timestamped course and degree major choices, grades, and demographics such as gender, ethnicity, and family educational history.
In this methods paper, we are interested in (1) understanding the means of analysis that provide information on students’ paths into and out of the physics major, (2) developing visual representations of these analyses that communicate what paths students take through the major, and (3) describing a possible mechanism (inferred from the available data) that can explain what differentiates students who receive a degree in physics and those that do not. In doing this work, our aim is not to dismiss the rich work around retention and recruitment, e.g., Refs. [@irving2013; @irving2015; @hazari2010], but rather to provide additional context on that this (and other work) might draw. In this paper, we have not conducted an analysis using demographics.
Michigan State Physics {#sec:msu}
======================
MSU is a large, land grant university with approximately 39,000 undergraduate students currently enrolled. MSU has both a college of arts and sciences and a college of engineering and enrolls $>2000$ students in introductory physics courses annually. The student population is predominately white (65.7%) with a sizable minority population (34.3%). MSU has slightly more women enrolled than men (48% men, 52% women). The physics major enrolls a greater proportion white students (73.8%) in comparison to the general MSU population and graduates a greater proportion as well (83.1%). MSU physics graduate gender contrasts to the general population (83% men, 17% women) – a proportion that is typical of physics departments across the country [@womeninphys].
Student Pathways {#sec:paths}
================
We have begun to describe student pathways at two levels. One level looks at the starting major that students declare and the final major for which the student receives a degree. We visually represent the movement from start to finish using an alluvial diagram (FIG. \[fig:alluvial\]) [@alluvial]. This diagram helps visualize how student initial conditions affect graduation outcomes (e.g., what proportion of students graduate with their initially intended degree). A second level describes the order in which students complete each course required for the major. We represent this visually using a bubble diagram (FIG. \[fig:bubble\]). This level highlights the track that students take through the physics program and how completion of those courses relates to recommended, “on-track” schedule.
Approximately half (44.1%) of students who declare a physics major at MSU do so when they first arrive at MSU. The remaining students switch into the physics major from a different degree programs or have not declared a major. Graduating students who declare a physics major are likely to remain in a STEM degree program even if they move away from physics (FIG. \[fig:alluvial\]). Approximately one-third (33.7%) of students who attempt to get a degree in physics at MSU do so. An additional one-third (33.7%) complete a degree in an engineering program. The remaining students are likely to pursue other STEM offerings (e.g., chemistry or mathematics).
Students frequently complete physics courses outside of the recommended schedule by the physics department (FIG. \[fig:bubble\]). For example, students who declare a physics major but ultimately receive a different degree are most likely to take their introductory mechanics course (PHY 183) in their third semester and introductory electricity and magnetism course (PHY 184) in their fourth semester. Students who receive degrees in physics are more likely to take this introductory sequence prior to their third semester. Additionally, many students who eventually earn degrees in physics take senior-level E&M (PHY 481) up to 1 year before the recommended schedule. While we acknowledge there could be many reasons for students taking courses at different times, we are (currently) interested in finding useful representations that describe for a single institution what pathways students take through the major.
![\[Color online\] Approximately one third of students declaring a physics major go on to receive a degree in physics. Most students (87%) who declare a physics major eventually receive a degree in STEM if they graduate. Groups on the left are the initial major declared by the student. Groups on the right are the graduating major the students receive a degree in.[]{data-label="fig:alluvial"}](alluvial.pdf){width="0.95\linewidth"}
Earned Grades differentiate physics graduates from others {#sec:grades}
=========================================================
In this initial study, we found that grades earned in math and physics courses differentiate students that eventually earn physics degrees from other graduates. Because course grades are not normalized measurements, we cannot compare raw grades between different courses, different course instructors, and different semesters. Thus, we have used the standard score or “Z-score” [@zscore] to normalize students’ grades in a single course offering.
$$Z = \frac{x - \mu}{\sigma}$$
The Z-score provides a measure of what fraction of a standard deviation ($\sigma$) that a particular value ($x$) deviates from the mean ($\mu$) of a distribution of scores. By using the Z-score, we do assume that students’ scores within a given course offering are drawn from a normal distribution.
To compare groups of students, we first calculated the Z-score for each student in a particular course offering for every course from a restricted list of courses (described below). We grouped students by the degree they eventually earned (physics, engineering, STEM, and non-STEM) as well as by the condition of declaring physics as major any time in their academic career (yes, no) – leading to 7 total groups. We then calculated the mean Z-score and standard error for physics courses and math courses separately for the population of students in each of the seven groups of students. These mean Z-scores along with their standard errors are shown in both in TAB. \[tab:zscore\] and FIG. \[fig:zscore\].
We have restricted the courses from which we draw our data to only introductory courses (100 & 200 level) required to earn a bachelor’s degree in physics. Many different degree programs require these math and physics courses, thus they provide a large basis to compare students (TAB. \[tab:zscore\]). These courses include introductory mechanics (PHY 183), introductory electricity and magnetism (PHY 184), introductory lab courses (PHY 191, PHY 192), a third semester course covering thermodynamics and modern physics (PHY 215), and the calculus sequence from Calculus I to a first course in ordinary differential equations (MTH 132, MTH 133, MTH 234, MTH 235).
We find that students who receive a degree in physics perform above average in introductory math and physics (FIG. \[fig:zscore\]a; TAB. \[tab:zscore\]). We refer to these plots as “normalized comparisons.” Based on these normalized comparisons, students who declare a physics major but then move to other STEM programs/Engineering programs perform below average. Further, students who never declare a physics major and receive a degree in STEM/Engineering programs perform above average. We also find that students whose first declared major is an engineering program but ultimate degree is physics perform below average in physics and mathematics introductory courses.
Category (physics) N $Z_{math} \pm SE$ $Z_{phys} \pm SE$
-------------------- ------ ------------------- -------------------
Engineering (No) 4047 0.26$\pm$0.01 0.29$\pm$0.01
Non STEM (No) 4913 -0.07$\pm$0.01 -0.49$\pm$0.03
Other STEM (No) 5833 0.13$\pm$0.01 0.08$\pm$0.01
Engineering (Yes) 374 -0.20$\pm$0.02 -0.19$\pm$0.03
Non STEM (Yes) 134 -0.56$\pm$0.05 -0.67$\pm$0.07
Other STEM (Yes) 202 -0.14$\pm$0.04 -0.05$\pm$0.05
Physics (Yes) 369 0.11$\pm$0.03 0.24$\pm$0.023
: Numbers of students and their corresponding normalized scores for the groups represented in FIG. \[fig:zscore\]a. Students are labeled by their graduating major and whether or not they ever declared a physics major.[]{data-label="tab:zscore"}
Discussion & Conclusions {#sec:discussion}
========================
In this paper, we have analyzed data collected by the registrar at MSU over the last 10+ years. This data can begin to provide information about the pathways that students take into and out of the physics major (for a given institutional context). In this methods paper, we have presented 3 representations (FIGS. \[fig:alluvial\]-\[fig:zscore\]) that offer some shape to the story at MSU.
In particular, we have found that students earning physics degrees who have initially declared physics come from all areas of the university in roughly equal measure (FIG. \[fig:alluvial\]); contrary to departmental anecdotes. Students who leave the major also earn degrees in different areas in roughly equal measures, which is also is counter to the prevailing narrative in the department. Second, we find that students who earn physics degrees tend to follow the departmentally-recommended path up to the last year of their studies. We also find that students who eventually earn engineering degrees leave physics during or after the first E&M course (PHY 184; Green circles in FIG. \[fig:bubble\]) while students who eventually earn other STEM degrees leave much later (Pink circles in FIG. \[fig:bubble\]). Finally, we find that students who are physics degree earners perform better in math and physics than students who declare physics and eventually earn some other degree, but perform not as well on course work than their engineering colleagues who never declared physics as a major (FIG. \[fig:zscore\]).
Through this work, we are not claiming that we have uncovered the full story from our current analysis or that all possible representations have been generated to explain our claims. Rather, we are suggesting that we have developed some methods and representations (SEC. \[sec:paths\]) that provide some context for the paths that students take through the physics major at MSU as well as a possible mechanism for the observations of related to student attrition (SEC. \[sec:grades\]). While our results that show that students earning lower scores in math and physics courses are more likely to not earn degrees in physics (FIG. \[fig:zscore\]) are fairly obvious, we have also provided data that demonstrates that pathways of those degree earners are different from students earning degrees in other areas (FIG. \[fig:bubble\]).
The insight gained into the pathways that students take as gleaned from this data and our representations suggests there is a deeper and more interesting story that might exist in our data. For example, how do these pathways differ for different populations of students (e.g., based on incoming GPA, race, and ethnicity)? Furthermore, there are some analyses to be done that might provide additional context (e.g., how math and physics course enrollment and performance interact). While our analyses and representations provide some context and detail about student pathways through the major, we recognize that by assuming a particular pathway for students to earn a degree (FIG. \[fig:bubble\]) that we are demphasizing alternative pathways and, likely, marginalizing non-traditional students. Moreover, that we assume a particular course trajectory for students to earn a physics degree might paint an unreasonably narrow picture of how students earn physics degrees. We are in the process of developing additional analyses that are not predicated on the student taking courses in a particular order. What we suspect is that a more comprehensive diagram that demonstrates the relationship between math and physics courses taken (i.e., in what order) will support our analysis and provide new and interesting information on student progression through the course work.
Finally, our present analysis neglects demographic information that might be important for understanding how different groups of students might be affected differently. As we construct new analyses and produce different representations of our data, we might find that asking similar questions of the data from this perspective will offer new insights into the pathways that women and under-represented students take through the major. Such an analysis is necessary if we are meant to foster and grow a diverse population of physics graduates.
**ACKNOWLEDGMENTS**
This work was support by the College of Natural Sciences STEM Gateway Fellowship and Association of American Universities.
[2]{}
APS News, <https://goo.gl/9rWIcJ>, Retrieved June 13, 2016.
Chen, X. (2013) “STEM Attrition: College Students’ Paths into and out of STEM Fields. Statistical Analysis Report.” National Center for Education Statistics.
Irving, P. W., & Sayre, E. C. (2013). Journ. of the Schol. of Teach. and Learn., 13(4), 68-84.
Irving, P. W., & Sayre, E. C. (2015). Phys. Rev. Spec. Top.-Phys. Ed. Res., 11(2), 020120.
Hazari, Z., Sonnert, G., Sadler, P. M., & Shanahan, M. C. (2010). Journal of Research in Science Teaching, 47(8), 978-1003.
APS, <https://goo.gl/wuBN9E>, Retrieved June 13, 2016.
Rosvall M, Bergstrom CT (2010). PLoS ONE 5(1): e8694.
MSU Physics, <https://goo.gl/QLy1YD>, Retrieved June 12, 2016.
Kreyszig, E. (1979). Advanced Engineering Mathematics (Fourth ed.). Wiley. ISBN 0-471-02140-7.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present some applications of the Kudla-Millson and the Millson theta lift. The two lifts map weakly holomorphic modular functions to vector valued harmonic Maass forms of weight $3/2$ and $1/2$, respectively. We give finite algebraic formulas for the coefficients of Ramanujan’s mock theta functions $f(q)$ and $\omega(q)$ in terms of traces of CM-values of a weakly holomorphic modular function. Further, we construct vector valued harmonic Maass forms whose shadows are unary theta functions, and whose holomorphic parts have rational coefficients. This yields a rationality result for the coefficients of mock theta functions, i.e., harmonic Maass forms whose shadows lie in the space of unary theta functions. Moreover, the harmonic Maass forms we construct can be used to evaluate the Petersson inner products of unary theta functions with harmonic Maass forms, giving formulas and rationality results for the Weyl vectors of Borcherds products.'
address: 'Fachbereich Mathematik, Technische Universität Darmstadt, Schlossgartenstra[ß]{}e 7, D–64289 Darmstadt, Germany'
author:
- Jan Hendrik Bruinier and Markus Schwagenscheidt
bibliography:
- 'references.bib'
title: Algebraic formulas for the coefficients of mock theta functions and Weyl vectors of Borcherds products
---
Introduction
============
Over the last two decades, starting with the fundamental work of Borcherds [@Borcherds], theta lifts between spaces of integral and half-integral weight weakly holomorphic modular forms have become a powerful tool in number theory. For example, Funke and the first author [@BruinierFunke06] used the Kudla-Millson theta lift from weight $0$ to weight $3/2$ harmonic Maass forms to give a new proof and generalizations of Zagier’s [@ZagierTraces] famous result on the modularity of the generating series of traces of singular moduli. Further, in [@BruinierOnoAlgebraicFormulas], Ono and the first author used a variant of the Kudla-Millson theta lift to find a finite algebraic formula for the partition function $p(n)$ in terms of traces of CM-values of a certain non-holomorphic modular function. Recently, a similar theta lift was used in [@AGOR] to prove a refinement of a theorem of [@BruinierOnoHeegnerDivisors] connecting the vanishing of the central derivative of the twisted $L$-function of an even weight newform and the rationality of some coefficient of the holomorphic part of a half-integral weight harmonic Maass form. This so-called Millson theta lift, which maps weight $0$ to weight $1/2$ harmonic Maass forms, was studied in great detail by Alfes-Neumann in her thesis [@Alfesdiss], and by Alfes-Neumann and the second author in [@AlfesSchwagenscheidt]. The aim of this paper is to present some applications of the Kudla-Millson lift from [@BruinierFunke06] and the Millson lift studied in [@AlfesSchwagenscheidt]. We discuss the following three applications.
Algebraic formulas for Ramanujan’s mock theta functions {#algebraic-formulas-for-ramanujans-mock-theta-functions .unnumbered}
-------------------------------------------------------
We give finite algebraic formulas for the coefficients of Ramanujan’s order $3$ mock theta functions $f(q)$ and $\omega(q)$ in terms of traces of CM-values of a weakly holomorphic modular function (see Theorem \[MockThetaFormulas\]). For example, we show that the coefficients $a_{f}(n), n \geq 1$, of Ramanujan’s mock theta function $$\begin{aligned}
\label{Ramanujanf}
f(q) &= 1+\sum_{n=1}^{\infty}\frac{q^{n^{2}}}{(1+q)^{2}(1+q^{2})^{2}\cdots (1+q^{n})^{2}} = 1 + \sum_{n=1}^{\infty}a_{f}(n)q^{n}
\end{aligned}$$ are given by $$\begin{aligned}
a_{f}(n) = -\frac{1}{\sqrt{24n-1}}\Im\bigg(\sum_{Q \in \mathcal{Q}_{n}}\frac{F(z_{Q})}{\omega_{Q}}\bigg),
\end{aligned}$$ where $$\begin{aligned}
\label{InputF}
F(z) = -\frac{1}{40}\cdot\frac{E_{4}(z) + 4E_{4}(2z) - 9E_{4}(3z) - 36E_{4}(6z)}{(\eta(z)\eta(2z)\eta(3z)\eta(6z))^{2}} = q^{-1}-4-83q-296q^{2}+ \dots
\end{aligned}$$ is a $\Gamma_{0}(6)$-invariant weakly holomorphic modular function, $\mathcal{Q}_{n}$ is the (finite) set of $\Gamma_{0}(6)$-equivalence classes of positive definite integral binary quadratic forms $Q(x,y) = ax^{2}+ bxy+cy^{2}$ of discriminant $1-24n$ with $6 \mid a$ and $b \equiv 1(12)$, $z_{Q} \in {\mathbb{H}}$ is the CM-point characterized by $Q(z_{Q},1) = 0$, and $\omega_{Q}$ is half the order of the stabilizer of $Q$ in $\Gamma_{0}(6)$. Moreover, $E_{4}$ denotes the normalized Eisenstein series of weight $4$ for $\operatorname{SL}_{2}(\operatorname{\mathbb{Z}})$ and $\eta = q^{1/24}\prod_{n = 1}^{\infty}(1-q^{n})$ is the Dedekind eta function.
For the proof, we use Zwegers’ [@Zwegers] realization of Ramanujan’s mock theta functions as the holomorphic parts of vector valued harmonic Maass forms of weight $1/2$, then construct the corresponding harmonic Maass form as the Millson lift of $F$, and finally obtain the formula by comparing Fourier coefficients.
Rationality results for harmonic Maass forms {#rationality-results-for-harmonic-maass-forms .unnumbered}
--------------------------------------------
By applying the Kudla-Millson and the Millson theta lifts to a suitable weakly holomorphic input function, we construct harmonic Maass forms of weight $3/2$ and $1/2$ whose images under the differential operator $\xi_{k} = 2iv^{k}\overline{\frac{\partial}{\partial \bar{\tau}}}$ are vector valued unary theta functions of weight $1/2$ and $3/2$, and whose holomorphic parts (which are mock modular forms) are given by traces of CM-values of the input function (see Theorem \[ThetaLifts\]). This implies that these mock modular forms have rational coefficients (see Theorem \[RationalHolomorphicPart\]), which in turn yields a rationality result for the holomorphic parts of harmonic Maass forms that map to the space of unary theta functions under $\xi$ (see Theorem \[AlgebraicCoefficients\]).
More specifically, we show that if $f$ is a vector valued harmonic Maass form of weight $1/2$ whose principal part is defined over a number field $K$, and whose shadow lies in the space of unary theta functions, then all coefficients of the holomorphic part of $f$ lie in $K$. This contrasts a conjecture of Ono and the first author [@BruinierOnoHeegnerDivisors], stating that if $f$ is a harmonic Maass form of weight $1/2$ whose shadow is orthogonal to the space of unary theta functions, then all but a set of density $0$ of the non-vanishing coefficients of the holomorphic part of $f$ should be transcendental.
Inner product formulas and Weyl vectors of Borcherds products {#inner-product-formulas-and-weyl-vectors-of-borcherds-products .unnumbered}
-------------------------------------------------------------
We use our $\xi$-preimages to evaluate the regularized Petersson inner product of a harmonic Maass form $f$ and a unary theta function of weight $1/2$ (see Theorem \[InnerProductFormulas\]), and apply this to compute the Weyl vectors of the Borcherds lift of $f$ (see Corollary \[WeylVectors\]). For example, for $N = 1$ the Borcherds product associated with a weakly holomorphic modular form $f = \sum_{n \gg -\infty}c_{f}(n)q^{n}$ of weight $1/2$ for $\Gamma_{0}(4)$ in the Kohnen plus space with rational coefficients and integral principal part is given by $$\Psi(z,f) = q^{\rho_{f}}\prod_{n=1}^{\infty}(1-q^{n})^{c_{f}(n^{2})}, \qquad (q = e^{2\pi i z}),$$ where $\rho_{f}$ is the so-called Weyl vector of $f$. The product $\Psi(z,f)$ is a modular form of weight $c_{f}(0)$ for $\operatorname{SL}_{2}(\operatorname{\mathbb{Z}})$, whose divisor on ${\mathbb{H}}$ is a Heegner divisor. For higher level $N$, the orders of $\Psi(z,f)$ at the cusps of $\Gamma_{0}(N)$ are determined by Weyl vectors associated to the cusps. These vectors are essentially given by regularized inner products of $f$ with a unary theta function of weight $1/2$, and can be explicitly evaluated in terms of the coefficients of the holomorphic part of $f$ and the coefficients of the holomorphic part of a $\xi$-primage of the unary theta function. In particular, we show that all Weyl vectors associated to a harmonic Maass form with rational holomorphic coefficients are rational (see Corollary \[WeylVectorsRationality\]).\
We start with a section on the necessary definitions and notations, and then discuss each of the three applications in a separate section.
The starting point of this work was a talk given by Yingkun Li in march 2016 about his construction of real-analytic theta series of weight $1/2$. We thank him cordially for his inspiration and for helpful discussions.
Preliminaries
=============
We first need to set up some background material about unary theta functions, harmonic Maass forms, and traces of CM-values of modular functions. Throughout this work, we let $N$ be a positive integer.
Unary theta function for the Weil representation
------------------------------------------------
We consider the positive definite lattice $\operatorname{\mathbb{Z}}$ with the quadratic form $n\mapsto Nn^{2}$. Its dual lattice is $\frac{1}{2N}\operatorname{\mathbb{Z}}$, so its discriminant group can be identified with $\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$, equipped with the finite quadratic form ${r \mapsto r^{2}/4N \text{ mod } \operatorname{\mathbb{Z}}}$. For $r \in \operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$ we let $\operatorname{\mathfrak{e}}_{r}$ be the standard basis vectors of the group ring $\operatorname{\mathbb{C}}[\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}]$, and we let $\langle \cdot,\cdot \rangle$ be the standard inner product on $\operatorname{\mathbb{C}}[\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}]$ which satisfies $\langle \operatorname{\mathfrak{e}}_{r},\operatorname{\mathfrak{e}}_{r'}\rangle = \delta_{r,r'}$ and is antilinear in the second variable. The associated Weil representation is defined on the generators $T= \left(\left(\begin{smallmatrix}1 & 1 \\0 & 1 \end{smallmatrix}\right),1\right)$ and $S = \left(\left(\begin{smallmatrix}0 & -1 \\1 & 0 \end{smallmatrix}\right),\sqrt{\tau}\right)$ of the metaplectic group $\operatorname{Mp}_{2}(\operatorname{\mathbb{Z}})$ by $$\begin{aligned}
\rho(T)\operatorname{\mathfrak{e}}_{r} &= e(Q(r))\operatorname{\mathfrak{e}}_{r}, \qquad \rho(S)\operatorname{\mathfrak{e}}_{r} = \frac{e(-1/8)}{\sqrt{2N}}\sum_{r' (2N)}e(-(r,r'))\operatorname{\mathfrak{e}}_{r'},
\end{aligned}$$ where $e(z) = e^{2\pi i z}$ for $z\in \operatorname{\mathbb{C}}$. We consider the unary theta functions $$\begin{aligned}
\label{UnaryTheta12}
\theta_{1/2}(\tau) = \sum_{r(2N)}\sum_{\substack{b \in \operatorname{\mathbb{Z}}\\ b \equiv r(2N)}}q^{b^{2}/4N}\operatorname{\mathfrak{e}}_{r} \qquad \text{and} \qquad \theta_{3/2}(\tau) = \sum_{r(2N)}\sum_{\substack{b \in \operatorname{\mathbb{Z}}\\ b \equiv r(2N)}}bq^{b^{2}/4N}\operatorname{\mathfrak{e}}_{r},
\end{aligned}$$ where $q = e^{2\pi i \tau}$. They are holomorphic vector valued modular forms of weight $1/2$ and $3/2$ for $\rho$, as can be seen using [@Borcherds], Theorem 4.1, for example. From a more conceptual point of view, these are the theta functions associated to the archimedian Schwartz functions $e^{-\pi x^{2}}$ and $x e^{-\pi x^{2}}$ in the Schrödinger model of the Weil representation associated with a one-dimensional quadratic space, compare [@BruinierFunke04], Section 2. We sometimes write $\rho_{N}, \theta_{1/2,N}$ and $\theta_{3/2,N}$ if we want to emphasize the dependence on $N$.
Note that for $k \in 1/2+\operatorname{\mathbb{Z}}$ the space $M_{k,\rho}$ of holomorphic modular forms of weight $k$ for $\rho$ is isomorphic to the space $J_{k+1/2,N}^{*}$ of skew-holomorphic Jacobi forms of weight $k+1/2$ and index $N$, and $M_{k,\bar{\rho}}$ is isomorphic to the space $J_{k+1/2,N}$ of holomorphic Jacobi forms of weight $k+1/2$ and index $N$ (see [@EichlerZagier], Section 5). In particular, one can view $\theta_{1/2}$ and $\theta_{3/2}$ as skew-holomorphic Jacobi forms of weight $1$ and $2$. Further, we will make frequent use of a result by Skoruppa which asserts that $M_{1/2,\bar{\rho}} \cong J_{1,N} = \{0\}$ for all $N$ (see [@EichlerZagier], Theorem 5.7).
Besides the unary theta functions $\theta_{1/2}$ and $\theta_{3/2}$ themselves, we also want to regard modular forms which arise from these series by applications of certain simple operators as theta functions. There are two classes of such operators, which we describe now.
The automorphism group $\operatorname{Aut}(\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}})$ acts on vector valued modular forms $f = \sum_{r}f_{r}\operatorname{\mathfrak{e}}_{r}$ for $\rho$ or $\bar{\rho}$ by $f^{\sigma} = \sum_{r}f_{r}\operatorname{\mathfrak{e}}_{\sigma(r)}$. The elements of $\operatorname{Aut}(\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}})$ are all involutions, also called *Atkin-Lehner involutions*, and correspond to the exact divisors $c\mid \mid N$ (i.e., $c\mid N$ and $(c,N/c) =1$). The automorphism $\sigma_{c}$ corresponding to $c$ is defined by the equations $$\begin{aligned}
\label{AtkinLehnerInvolution}
\sigma_{c}(r) \equiv -r \ (2c) \quad \text{and} \quad \sigma_{c}(r) \equiv r \ (2N/c)
\end{aligned}$$ for $r \in \operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$, compare [@EichlerZagier], Theorem 5.2. Note that the Atkin-Lehner involutions only permute the components of a vector valued modular form.
For each positive integer $d$ there is a natural operator which maps modular forms for $\rho_{N}$ to forms for $\rho_{Nd^{2}}$ (see [@ScheithauerModularforms], Section 4). Under the isomorphism $J_{k+1/2,N}^{*}\cong M_{k,\rho_{N}}$ it corresponds to the index raising operator $U_{d}$ defined in [@EichlerZagier], Section 4. Its action on the Fourier expansion of a holomorphic modular form $f = \sum_{r(2N) }\sum_{\substack{D \equiv r^{2}(4N)} }c_{f}(D,r)e\left( \frac{D}{4N}\tau\right)\operatorname{\mathfrak{e}}_{r}$ for $\rho_{N}$ is given by $$\begin{aligned}
f|U_{d} &= \sum_{\substack{r(2Nd^{2}) \\ r \equiv 0(d)}}\sum_{\substack{D \geq 0 \\ D \equiv r^{2} (4Nd^{2})} }c_{f}(D/d^{2},r/d)e\left( \frac{D}{4Nd^{2}}\tau\right)\operatorname{\mathfrak{e}}_{r}. \label{UdOperator}
\end{aligned}$$ In particular, $U_{d}$ only distributes the components of $f$ in a certain way, but does not change the set of Fourier coefficients of $f$.
With our application to Weyl vectors of Borcherds products already in mind (compare ), we now define the *space of unary theta functions of weight $1/2$ for* $\rho_{N}$ as $$\sum_{\substack{d^{2} \mid N}}\sum_{c \mid \mid \frac{N}{d^{2}}}\operatorname{\mathbb{C}}\theta_{1/2,N/d^{2}}^{\sigma_{c}}|U_{d}\, .$$ The space of unary theta functions of weight $3/2$ for $\rho_{N}$ is defined analogously. We now show that the space of unary theta functions of weight $1/2$ agrees with the whole space $M_{1/2,\rho}$.
Let $\mathcal{D}(N)$ be the set of all positive divisors of $N$ modulo the equivalence relation $c \sim N/c$. Then the theta functions $$\theta_{1/2,N/(c,N/c)^{2}}^{\sigma_{c/(c,N/c)}}|U_{(c,N/c)}, \quad c \in \mathcal{D}(N),$$ form a basis of $M_{1/2,\rho}$.
Using a dimension formula for $J_{1,N}^{*} \cong M_{1/2,\rho}$ (see [@SkoruppaZagier], p. 130), we see that the number of given theta functions agrees with the dimension of $M_{1/2,\rho}$. On the other hand, by looking at the constant terms and the coefficients at $q^{d^{2}/4N}\operatorname{\mathfrak{e}}_{d}$ for $d \in \mathcal{D}(N)$, it is easy to see that the given functions are linearly independent.
\[SerreStarkRemark\] This result resembles the Serre-Stark theorem (see [@SerreStark]) which states that every scalar valued modular form of weight $1/2$ is a linear combination of elementary theta functions $\theta_{\psi,m}(\tau) = \sum_{n \in \operatorname{\mathbb{Z}}}\psi(n)n^{\nu}q^{mn^{2}}$, where $\nu = 0$ and $\psi$ is a primitive Dirichlet character with $\psi(-1) = 1$ (for $\nu = 1$ and $\psi(-1)= -1$ the theta function $\theta_{\psi,m}$ is a cusp form of weight $3/2$). If we split $\theta_{\psi,m}(\tau) = \sum_{r (2d)}\psi(r)\sum_{n \equiv r(2d)}n^{\nu}q^{mn^{2}}$, where $d$ is the conductor of $\psi$, we see that the inner sums are just the components of a vector valued unary theta function. Hence, many of the results of the present work, which are formulated for vector valued unary theta functions, immediatly imply the corresponding statements for the classical scalar valued theta functions.
Harmonic Maass forms
--------------------
Recall from [@BruinierFunke04] that a smooth function $f: {\mathbb{H}}\to \operatorname{\mathbb{C}}[\operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}]$ is called a *harmonic Maass form* of weight $k \in 1/2+\operatorname{\mathbb{Z}}$ with respect to $\bar{\rho}$ if it is annihilated by the weight $k$ hyperbolic Laplace operator $\Delta_k$, transforms like a modular form of weight $k$ for $\bar{\rho}$, and grows at most linearly exponentially at $\infty$. We denote the space of such functions by $H_{k,\bar{\rho}}$. Further, we let $M^{!}_{k,\bar{\rho}}$ be the subspace of weakly holomorphic modular forms, consisting of the forms in $H_{k,\bar{\rho}}$ which are holomorphic on ${\mathbb{H}}$. An important tool in the theory of harmonic Maass forms is the antilinear differential operator $\xi_k f=2iv^{k}\overline{\frac{\partial}{\partial \bar{\tau}}f(\tau)}$. It defines a surjective map $\xi_k: H_{k,\bar{\rho}}\rightarrow M^{!}_{2-k,\rho}$.
In this paper we only consider harmonic Maass forms that map to the space $M_{2-k,\rho}$ of holomorphic modular forms under $\xi_{k}$. Such a form for $\bar{\rho}$ uniquely decomposes into a *holomorphic* and a *non-holomorphic part* $f=f^{+}+f^{-}$ with Fourier expansions $$\begin{aligned}
f^{+}(\tau)&=\sum\limits_{r(2N)}\sum\limits_{\substack{D\gg \infty \\ D \equiv -r^{2}(4N)}}c_{f}^{+}(D,r)q^{D/4N} \mathfrak{e}_r,
\\
f^{-}(\tau)&=\sum\limits_{r(2N)}\bigg(c_{f}^{-}(0,r)v^{1-k} + \sum\limits_{\substack{D < 0 \\ D \equiv -r^{2}(4N)}}c_{f}^{-}(D,r)\Gamma(1-k,\pi |D|v/N)q^{D/4N}\bigg)\mathfrak{e}_r,\end{aligned}$$ where $\Gamma(s,x) = \int_{x}^{\infty}t^{s-1}e^{-t}dt$ denotes the incomplete gamma function. The finite Fourier polynomial $\sum_{r(2N)}\sum_{\substack{D \leq 0}}c_{f}^{+}(D,r)q^{D/4N} \mathfrak{e}_r$ is called the *principal part* of $f$. We briefly call the coefficients of the holomorphic part the *holomorphic coefficients* of $f$. The holomorphic part $f^{+}$ is sometimes called a *mock modular form*, and $\xi_{k}f = \xi_{k}f^{-}$ its *shadow*.
We remark that the operators $U_{d}$ and $\sigma_{c}$ defined in the last section also act on harmonic Maass forms, the action on the Fourier expansion being the same, and that they commute with the $\xi$-operator.
The regularized inner product of $f \in M^{!}_{k,\rho}$ and $g \in S_{k,\rho}$ for $k \in 1/2+\operatorname{\mathbb{Z}}$ is defined by $$\begin{aligned}
\label{PeterssonInnerProduct}
(f,g)^{\operatorname{reg}} = \lim_{T \to \infty} \int_{\mathcal{F}_{T}}\langle f(\tau),g(\tau) \rangle v^{k}\frac{du\, dv}{v^{2}}, \qquad (\tau = u+iv),
\end{aligned}$$ where $\mathcal{F}_{T} = \{\tau \in {\mathbb{H}}: |u|\leq \tfrac{1}{2}, |\tau| \geq 1 ,v \leq T\}$ is a truncated fundamental domain for the action of $\operatorname{SL}_{2}(\operatorname{\mathbb{Z}})$ on ${\mathbb{H}}$. For $k=1/2$ the regularized inner product also converges for $g \in M_{1/2,\rho}$. An application of Stokes’ theorem (compare Proposition 3.5 in [@BruinierFunke06]) shows that for $f \in M^{!}_{k,\rho}$ with coefficients $c_{f}(D,r)$, and $G \in H_{2-k,\bar{\rho}}$ with holomorphic coefficients $c_{G}^{+}(D,r)$ and $\xi_{2-k}G=g \in S_{k,\rho}$ (or $g \in M_{1/2,\rho}$ if $k =1/2$), the regularized inner product can be evaluated as $$\begin{aligned}
\label{InnerProductEvaluation}
(f,g)^{\operatorname{reg}} = (f,\xi_{2-k}G)^{\operatorname{reg}} =\sum_{r(2N)}\sum_{\substack{D \in \operatorname{\mathbb{Z}}\\ D \equiv r^{2}(4N)}}c_{f}(D,r)c_{G}^{+}(-D,r).
\end{aligned}$$ This can be employed to show the following useful lemma.
\[CuspFormLemma\] Let $G$ be a harmonic Maass form of weight $2-k \in 1/2+\operatorname{\mathbb{Z}}$ for $\rho$ or $\bar{\rho}$ whose principal part vanishes and which maps to a cusp form under $\xi_{2-k}$ (or a holomorphic modular form if $k=1/2$). Then $G$ is a cusp form.
Suppose that $g = \xi_{2-k} G$ is a cusp form. By we see that $(g,g)^{\operatorname{reg}} = (g,\xi_{k} G)^{\operatorname{reg}} = 0$, so $g = 0$. This means that $G$ is holomorphic, hence a cusp form. For $k=1/2$ the inner product $(g,g)^{\operatorname{reg}}$ also converges if $g \in M_{1/2,\rho}$, so the same argument applies in this case.
Traces of CM-values and theta lifts
-----------------------------------
For $k \in \operatorname{\mathbb{Z}}$ we let $M_{k}^{!}(N)$ denote the space of weakly holomorphic modular forms of weight $k$ and level $N$, i.e., the space of meromorphic modular forms of weight $k$ for $\Gamma_{0}(N)$ whose poles are supported at the cusps. We denote the subspace of holomorphic forms by $M_{k}(N)$.
For a discriminant $D < 0$ and $r \in \operatorname{\mathbb{Z}}$ with $D \equiv r^{2}(4N)$ we let $\operatorname{\mathcal{Q}}_{N,D,r}$ be the set of integral binary quadratic forms $ax^{2}+bxy + cy^{2}$ of discriminant $D = b^{2}-4ac$ with $N \mid a$ and $b \equiv r (2N)$. It splits into the sets $\operatorname{\mathcal{Q}}_{N,D,r}^{+}$ and $\operatorname{\mathcal{Q}}_{N,D,r}^{-}$ of positive and negative definite forms. The group $\Gamma_{0}(N)$ acts on both sets with finitely many orbits, and the order $w_{Q}= \frac{1}{2}|\Gamma_{0}(N)_{Q}|$ is finite. For each $Q \in \operatorname{\mathcal{Q}}_{N,D,r}$ there is an associated Heegner (or CM) point $z_{Q} = (-b+i\sqrt{|D|})/2a$, which is characterized by $Q(z_{Q},1) = 0$. We define two trace functions of $F \in M_{0}^{!}(N)$ by $$\begin{aligned}
\label{traces}
\operatorname{tr}_{F}^{+}(D,r) = \sum_{Q \in \operatorname{\mathcal{Q}}_{N,D,r}^{+}/\Gamma_{0}(N)}\frac{F(z_{Q})}{w_{Q}} \qquad \text{and} \qquad \operatorname{tr}_{F}^{-}(D,r) = \sum_{Q \in \operatorname{\mathcal{Q}}_{N,D,r}^{-}/\Gamma_{0}(N)}\frac{F(z_{Q})}{w_{Q}}.
\end{aligned}$$
The Kudla-Millson theta lift $\operatorname{\textit{I}_{KM}}(F,\tau)$ and the Millson theta lift $\operatorname{\textit{I}_{M}}(F,\tau)$ of a weakly holomorphic modular function $F \in M_{0}^{!}(N)$ are both defined as integrals of the shape $$I(F,\tau) = \int_{\Gamma_{0}(N)\setminus {\mathbb{H}}}F(z)\Theta(\tau,z)\frac{dx \, dy}{y^{2}}, \qquad (z = x + iy),$$ where $\Theta(\tau,z)$ is a suitable theta function which is $\Gamma_{0}(N)$-invariant in $z$ and transforms like a modular form of weight $3/2$ (for the Kudla-Millson lift) or weight $1/2$ (for the the Millson lift) for the representation $\bar{\rho}$ in $\tau$. We remark that we renormalize the Millson theta lift from [@AlfesSchwagenscheidt] by multiplication with $i/\sqrt{N}$ to simplify the formulas. The most important properties of the two theta lifts are studied in [@BruinierFunke06] and [@AlfesSchwagenscheidt]. It turns out that both lifts map modular functions for $\Gamma_{0}(N)$ to harmonic Maass forms for $\bar{\rho}$ whose $\xi$-images are linear combinations of unary theta functions. Further, the Fourier expansions of the holomorphic parts of the lifts are given by traces of $F$. We state simplified Fourier expansions of the lifts for special inputs $F$ in Theorem \[ThetaLifts\].
Algebraic formulas for the coefficients of Ramanujan’s $f(q)$ and $\omega(q)$
=============================================================================
As a first application of the Millson theta lift, we find finite algebraic formulas for the coefficients of Ramanujan’s third order mock theta functions $f(q)$ defined in and $$\begin{aligned}
\omega(q) &= 1+ \sum_{n=1}^{\infty}\frac{q^{2n^{2}+2n}}{(1-q)^{2}(1-q^{3})^{2}\cdots (1-q^{2n+1})^{2}} \\
&= 1 + 2q + 3q^{2} + 4q^{3} + 6q^{4} + 8q^{5}+ 10q^{6}+14q^{7}+\dots
\end{aligned}$$ in terms of the traces of a single modular function. We obtain the following result.
\[MockThetaFormulas\] Let $F \in M_{0}^{!}(6)$ be the function defined in .
1. For $n \geq 1$ the coefficients $a_{f}(n)$ of $f(q)$ are given by $$\begin{aligned}
a_{f}(n) = \frac{i}{2\sqrt{24n-1}}\big(\operatorname{tr}_{F}^{+}(1-24n,1)-\operatorname{tr}_{F}^{-}(1-24n,1)\big).
\end{aligned}$$
2. For $n \geq 1$ the coefficients $a_{\omega}(n)$ of $\omega(q)$ are given by $$\begin{aligned}
a_{\omega}(n) =
\begin{cases}
\frac{-i}{8\sqrt{24\frac{n+1}{2}-4}}\left(\operatorname{tr}_{F}^{+}\left(4-24\frac{n+1}{2},2\right) - \operatorname{tr}_{F}^{-}\left(4-24\frac{n+1}{2},2\right)\right), & n \text{ odd}, \\
\frac{-i}{8\sqrt{24(\frac{n}{2}+1)-16}}\big(\operatorname{tr}_{F}^{+}(16-24\left(\frac{n}{2}+1\right),4) - \operatorname{tr}_{F}^{-}(16-24\left(\frac{n}{2}+1\right),4)\big), & n \text{ even}.
\end{cases}
\end{aligned}$$
<!-- -->
1. The theorem extends results of Alfes-Neumann (see the example after Theorem 1.3 in [@alfes]), who gave similar formulas for the coefficients $a_{f}(n)$ with $1-24n$ being a fundamental discriminant, by looking at the Kudla-Millson theta lift of $F$ and employing a duality result between the Millson and the Kudla-Millson lift.
2. Using the Kudla-Millson theta lift, Ahlgren and Andersen [@AhlgrenAndersen] gave a formula for the smallest parts function in terms of traces of a modular function. The coefficients of Ramanujan’s mock theta functions are related to partitions as well. For example, $a_{f}(n)$ is the number of partitions of $n$ with even rank minus the number with odd rank, where the rank of a partition is its largest part minus the number of parts.
3. On of the main ingredients in the proof is Zwegers’ [@ZwegersPaper] realization of the mock theta functions $f(q)$ and $\omega(q)$ as the holomorphic parts of the components of a vector valued harmonic Maass form. Thus the same idea works for other mock theta functions as well, for example for the order $5$ and order $7$ mock theta functions treated in Zwegers’ thesis [@Zwegers]. The details will be part of an ongoing master’s thesis.
4. We checked the above formulas numerically using Sage [@Sage].
We illustrate our formulas by computing $a_{\omega}(1) = 2$. Note that $\operatorname{tr}_{F}^{-}(D,r) = \overline{\operatorname{tr}_{F}^{+}(D,r)}$ since $F$ has real coefficients, so we obtain $$a_{\omega}(1) = \frac{-i}{8\sqrt{20}}\left(\operatorname{tr}_{F}^{+}\left(-20,2\right) - \operatorname{tr}_{F}^{-}\left(-20,2\right)\right) = \frac{2}{8\sqrt{20}}\Im\left(\operatorname{tr}_{F}^{+}\left(-20,2\right) \right).$$ A system of representatives of the $\Gamma_{0}(6)$-classes of forms $ax^{2} + bxy + cy^{2}$ with $6 \mid a , a > 0, b \equiv 2 (12)$ and $D = b^{2}-4ac = -20$ is given by the two forms $$\begin{aligned}
Q_{1} = 6x^{2} -10xy + 5y^{2}, \qquad Q_{2} = 42x^2 - 34xy + 7y^2,
\end{aligned}$$ and the corresponding CM-points are $$\begin{aligned}
z_{Q_{1}} = \frac{10+i\sqrt{20}}{12},\qquad z_{Q_{2}} = \frac{34+i\sqrt{20}}{84}.
\end{aligned}$$ Plugging these values into the Fourier expansion of $F$, we find $$\begin{aligned}
F(z_{Q_{1}}) = F(z_{Q_{2}}) \sim i\cdot 17.888543820000.
\end{aligned}$$ Thus we obtain $$a_{\omega}(1) \sim \frac{2}{8\sqrt{20}}\cdot 2\cdot 17.888543820000 = 2.000000000000.$$
Zwegers [@ZwegersPaper] showed that the function $$(q^{-\frac{1}{24}}f(q),2q^{\frac{1}{3}}\omega(q^{\frac{1}{2}}),2q^{\frac{1}{3}}\omega(-q^{\frac{1}{2}}))^{T}$$ is the holomorphic part of a vector valued harmonic Maass form $H = (h_{0},h_{1},h_{2})^{T}$, transforming as $$\begin{aligned}
H(\tau+1) = \begin{pmatrix}\zeta_{24}^{-1} & 0 & 0 \\ 0 & 0 & \zeta_{3} \\ 0 & \zeta_{3} & 0 \end{pmatrix}H(\tau), \quad H\left(-\frac{1}{\tau}\right) = \sqrt{-i\tau}\begin{pmatrix}0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & -1 \end{pmatrix}H(\tau).
\end{aligned}$$ Further, $\xi_{1/2}H$ is a vector consisting of cuspidal unary theta functions of weight $3/2$.
One can check that $$\tilde{H} = (0,h_{0},h_{2}-h_{1},0,-h_{1}-h_{2},-h_{0},0,h_{0},h_{1}+h_{2},0,h_{1}-h_{2},-h_{0})^{T}$$ is a vector valued harmonic Maass form of weight $1/2$ for the dual Weil representation $\bar{\rho}$. We see that its principal part is given by $q^{-1/24}(\operatorname{\mathfrak{e}}_{1}-\operatorname{\mathfrak{e}}_{5}+\operatorname{\mathfrak{e}}_{7}-\operatorname{\mathfrak{e}}_{11})$.
The function $(\eta(z)\eta(2z)\eta(3z)\eta(6z))^{2}$ in the denominator of $F$ is a cusp form of weight $4$ for $\Gamma_{0}(6)$ which is invariant under all Atkin-Lehner involutions $W_{d}^{6}$ for $d \mid 6$, and the numerator of $F$ equals $E_{4}|(W_{1}^{6}+W_{2}^{6}-W_{3}^{6}-W_{6}^{6})$. Thus $F$ is an eigenfunction of all Atkin-Lehner involutions, with eigenvalue $1$ for $W_{1}^{6}$ and $W_{2}^{6}$, and eigenvalue $-1$ for $W_{3}^{6}$ and $W_{6}^{6}$. In particular, the Fourier expansions of $F$ at the cusps of $\Gamma_{0}(6)$ are essentially the same, up to a possible minus sign. Using the formula for the Fourier expansion of the Millson lift given in Theorem 5.1 of [@AlfesSchwagenscheidt], it is now straightforward to check that $F$ lifts to a harmonic Maass form $\operatorname{\textit{I}_{M}}(F,\tau)$ of weight $1/2$ for $\bar{\rho}$, having twice the principal part as $\tilde{H}$ (recall that we multiplied the Millson lift by $i/\sqrt{N}$). In view of Lemma \[CuspFormLemma\], this implies that the difference $\tilde{H} - \frac{1}{2}\operatorname{\textit{I}_{M}}(F,\tau)$ is a cusp form. But $S_{1/2,\bar{\rho}} \cong J^{\text{cusp}}_{1,6} = \{0\}$, so we find $\tilde{H} = \frac{1}{2}\operatorname{\textit{I}_{M}}(F,\tau)$. The holomorphic coefficients of $\operatorname{\textit{I}_{M}}(F,\tau)$ at $q^{(24n-r^{2})/24}\operatorname{\mathfrak{e}}_{r}$ for $r^{2}-24n < 0$ are given by $$\frac{i}{\sqrt{24n-r^{2}}}\big(\operatorname{tr}_{F}^{+}(r^{2}-24n,r)-\operatorname{tr}_{F}^{-}(r^{2}-24n,r)\big).$$ Comparing the holomorphic parts of $\tilde{H}$ and $\frac{1}{2}\operatorname{\textit{I}_{M}}(F,\tau)$, we obtain the stated formulas for the coefficients $a_{f}(n)$ and $a_{\omega}(n)$.
$\xi$-preimages of unary theta functions and rationality results
================================================================
In [@BringmannFolsomOno] and [@BringmannOno], Bringmann, Folsom and Ono constructed scalar valued harmonic Maass forms of weight $3/2$ and $1/2$ whose shadows are the components of the unary theta functions $\theta_{1/2}$ and $\theta_{3/2}$ above. In both cases, the proof of the modularity of their $\xi$-preimages relies on transformation properties of various hypergeometric functions and $q$-series. Here we construct $\xi$-preimages for both $\theta_{1/2}$ and $\theta_{3/2}$ using the Kudla-Millson and the Millson theta lift of a single weakly holomorphic modular function $F$ for $\Gamma_{0}(N)$. A nice feature of this approach is that the modularity is clear from the construction. Further, the coefficients of the holomorphic parts of these harmonic Maass forms are given by modular traces of $F$, and thus have good arithmetic properties. Therefore, we obtain rationality results for the holomorphic parts of harmonic Maass forms which map to the space of unary theta functions under $\xi$.
Let $\operatorname{\mathbb{C}}((q))$ be the ring of formal Laurent series and let $\operatorname{\mathbb{C}}[[q]]$ be the ring of formal power series in $q$. If $f=\sum_n a(n)q^n \in \operatorname{\mathbb{C}}((q))$, we call the polynomial $$P_f= \sum_{n\leq 0} a(n)q^{n}\in \operatorname{\mathbb{C}}[q^{-1}]$$ the *principal part* of $f$. There is a bilinear pairing $$\operatorname{\mathbb{C}}((q))\times \operatorname{\mathbb{C}}[[q]]\to \operatorname{\mathbb{C}}, \quad (f,g)\mapsto \{f,g\} := \text{coefficient of $q^0$
of $f\cdot g$}.$$ It only depends on the principal part of $f$.
Let $k>0$ be an even integer. We denote by $M_{k}^{!,\infty}(N) \subset M_{k}^{!}(N)$ the subspace of those weakly holomorphic modular forms which vanish at all cusps different from $\infty$. We view the space $M_{2-k}^{!,\infty}(N)$ as a subspace of $\operatorname{\mathbb{C}}((q))$ and view $M_k(N)$ as a subspace of $\operatorname{\mathbb{C}}[[q]]$ by taking $q$-expansions at the cusp $\infty$.
\[lem:var-bor\] Let $P\in \operatorname{\mathbb{C}}[q^{-1}]$. There exists an $F\in M_{2-k}^{!,\infty}(N)$ with prescribed principal part $P_F=P$ at the cusp $\infty$, if and only if $\{P,g\} = 0$ for all $g\in M_k(N)$.
This can be proved by varying the argument of [@borcherdsDuke], Theorem 3.1. By Serre duality it can be shown that the subspace $M_{2-k}^{!,\infty}(N)\subset \operatorname{\mathbb{C}}((q))$ is the orthogonal complement of $M_k(N)\subset \operatorname{\mathbb{C}}[[q]]$ with respect to the pairing $\{\cdot,\cdot\}$.
We use this lemma to construct a suitable input $F$ for the two theta lifts.
\[ConstructionModularFunction\] Let $k\in \operatorname{\mathbb{Z}}_{>0}$ be even. There exists an $F(z) = \sum_{n \gg -\infty}a(n)q^{n}\in M_{2-k}^{!,\infty}(N)$ with the following properties:
1. The Fourier coefficients $a(n)$ of $F$ at $\infty$ lie in $\operatorname{\mathbb{Q}}$.
2. The constant term $a(0)$ of $F$ at $\infty$ is non-zero.
If $N=1$ and $k=2$ then $M_k(N)$ is trivial. In this case we can take $F=1$. Therefore we exclude this case from now on, so that $M_k(N)$ is non-trivial. We let $M_{k,0}(N)\subset M_{k}(N)$ be the codimension $1$ subspace of those modular forms which vanish at the cusp $\infty$.
Since the cusp at $\infty$ of $X_0(N)$ is defined over $\operatorname{\mathbb{Q}}$, there exists an $E=\sum_{n\geq 0} c(n)q^n\in M_k(N)$ with rational coefficients which has value $1$ at $\infty$, i.e., $c(0)=1$. (Such an $E$ can be obtained explicitly as a linear combination of the Eisenstein series at the cusps $\infty$ and $0$.) It is well known that $M_k(N)$ has a basis consisting of modular forms with rational coefficients. Using $E$, we see that the space $M_{k,0}(N)$ also has a basis $g_1,\dots ,g_d$ consisting of forms with rational coefficients. Moreover, we have $$M_k(N)=M_{k,0}(N)\oplus \operatorname{\mathbb{C}}E.$$
The linear map $M_k(N)\to \operatorname{\mathbb{C}}[[q]]/\operatorname{\mathbb{C}}[q]$ induced by mapping a modular form to its $q$-expansion is injective. Hence, the images of $g_1,\dots,g_d$ and $E$ are linearly independent. Consequently, there exists a polynomial $$P_0=\sum_{n<0} a(n)q^{n}\in q^{-1}\operatorname{\mathbb{Q}}[q^{-1}]$$ such that $$\begin{aligned}
\{ P_0, g_i\} &= 0,\quad \text{for $i=1,\dots,d$,} \qquad \text{and} \qquad \{ P_0, E\} = -1.\end{aligned}$$ Put $P=P_0+1\in \operatorname{\mathbb{Q}}[q^{-1}]$. Then we have $$\begin{aligned}
\{ P, g_i\} &= 0,\quad \text{for $i=1,\dots,d$,} \qquad \text{and} \qquad \{ P, E\} = 0,\end{aligned}$$ and therefore $\{P,g\}=0$ for all $g\in M_k(N)$.
According to Lemma \[lem:var-bor\] there exists an $F\in M_{2-k}^{!,\infty}(N)$ with principal part $P_F=P$. We denote the Fourier expansion of $F$ by $F=\sum_n a(n) q^n$. The group $\operatorname{Aut}(\operatorname{\mathbb{C}}/\operatorname{\mathbb{Q}})$ acts on $M_{2-k}^!(N)$ by conjugation of the Fourier coefficients. Under the action of $\operatorname{Aut}(\operatorname{\mathbb{C}}/\operatorname{\mathbb{Q}})$ on $X_0(N)$ the cusp at $\infty$ is fixed, and the other cusps are permuted among themselves. Hence $\operatorname{Aut}(\operatorname{\mathbb{C}}/\operatorname{\mathbb{Q}})$ also acts on $M_{2-k}^{!,\infty}(N)$ by conjugation of the Fourier coefficients. Consequently, for $\sigma\in \operatorname{Aut}(\operatorname{\mathbb{C}}/\operatorname{\mathbb{Q}})$, the form $F^\sigma$ also belongs to $M_{2-k}^{!,\infty}(N)$. Since $P_F\in \operatorname{\mathbb{Q}}[q^{-1}]$, the form $F-F^\sigma$ has vanishing principal part and therefore vanishes identically. We find that $a(n)=a(n)^\sigma$ for all $n\in \operatorname{\mathbb{Z}}$. Therefore, all Fourier coefficients of $F$ are rational.
We now use the modular function $F\in M_{0}^{!,\infty}(N)$ constructed in Lemma \[ConstructionModularFunction\] as an input for the Kudla-Millson and the Millson theta lift. The following theorem is just a straightforward simplification of Theorem 4.5 from [@BruinierFunke06] and Theorem 5.1 from [@AlfesSchwagenscheidt], and will thus not be proved here. In order to simplify the formulas, we multiply the expansion of the Millson lift given in [@AlfesSchwagenscheidt] by $i/\sqrt{N}$.
\[ThetaLifts\] Let $F(z) = \sum_{n \gg -\infty}a(n)q^{n}\in M_{0}^{!,\infty}(N)$ be as in Lemma \[ConstructionModularFunction\] and $r \in \operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$.
1. The function $$\begin{aligned}
\operatorname{\textit{I}_{KM}}(F,\tau)_{r}^{+} &= \sum_{\substack{D < 0 \\ D \equiv r^{2}(4N)}}\big(\operatorname{tr}_{F}^{+}(D,r)+\operatorname{tr}_{F}^{-}(D,r)\big)q^{-D/4N} \\
&\quad + 4\delta_{0,r}\sum_{n \geq 0}a(-n)\sigma_{1}(n) - \sum_{b > 0}b\big(\delta_{b,r}+\delta_{b,-r} \big)\sum_{n>0}a(-bn)q^{-b^{2}/4N}
\end{aligned}$$ is the $r$-component of the holomorphic part of a harmonic Maass form of weight $3/2$ for $\bar{\rho}$ with $$\xi_{3/2}\left(\operatorname{\textit{I}_{KM}}(F,\tau)\right) = -\frac{\sqrt{N}}{4\pi}a(0)\theta_{1/2}(\tau).$$ Here $\delta_{r,r'}$ equals $1$ if $r \equiv r' (2N)$ and $0$ otherwise, and $\sigma_{1}(0) = -\frac{1}{24}$.
2. The function $$\begin{aligned}
\operatorname{\textit{I}_{M}}(F,\tau)_{r}^{+} &= \sum_{\substack{D < 0 \\ D \equiv r^{2}(4N)}}\frac{i}{\sqrt{|D|}}\big(\operatorname{tr}_{F}^{+}(D,r)-\operatorname{tr}_{F}^{-}(D,r)\big)q^{-D/4N} \\
&\quad +\sum_{b > 0}\big(\delta_{b,r}-\delta_{b,-r} \big)\sum_{n>0}a(-bn)q^{-b^{2}/4N}
\end{aligned}$$ is the $r$-component of the holomorphic part of a harmonic Maass form of weight $1/2$ for $\bar{\rho}$ with $$\xi_{1/2}\left(\operatorname{\textit{I}_{M}}(F,\tau)\right) = -\frac{1}{2\sqrt{N}}a(0)\theta_{3/2}(\tau).$$
The Kudla-Millson lift of the constant $1$-function gives a generalization of Zagier’s non-holomorphic Eisenstein series of weight $3/2$ from [@ZagierEisensteinSeries] to arbitrary level $N$, see also [@BruinierFunke06], Remark 4.6. The $\xi$-image of the Eisenstein series is a linear combination of unary theta series associated to lattices $(\operatorname{\mathbb{Z}},n \mapsto dn^{2})$ with $d \mid N$, and is invariant under all Atkin-Lehner involutions. Since $\theta_{1/2}$ is only Atkin-Lehner invariant if $N = 1$ or $N = p$ is prime, we can *not* take the Eisenstein series as a $\xi$-preimage of $\theta_{1/2}$ in general. Also note that usually the principal parts of the harmonic Maass forms given above are non-zero.
By the theory of complex multiplication, the rationality properties of traces of weakly holomorphic modular functions are well understood. Therefore, we obtain the following result on the rationality of the holomorphic parts of the harmonic Maass forms given above.
\[RationalHolomorphicPart\] Let $F \in M_{0}^{!}(N)$ and suppose that the Fourier coefficients of $F$ at $\infty$ lie in $\operatorname{\mathbb{Z}}$ and the expansions at all other cusps have coefficients in $\operatorname{\mathbb{Z}}[\zeta_{N}]$. Then for $D \equiv r^{2}(4N), D < 0,$ the numbers $$\begin{aligned}
\label{IntegralTraces}
6 \big(\operatorname{tr}_{F}^{+}(D,r)+\operatorname{tr}_{F}^{-}(D,r)\big) \quad \text{and} \quad 6 t \frac{i}{\sqrt{|D|}}\big(\operatorname{tr}_{F}^{+}(D,r)-\operatorname{tr}_{F}^{-}(D,r)\big)
\end{aligned}$$ are rational integers, where $D = t^{2}D_{0}$ with a negative fundamental discriminant $D_{0}$.
The assumption on the integrality of $F$ at $\infty$ implies that $F \in \operatorname{\mathbb{Q}}(j,j_{N})$. By the theory of complex multiplication (see Theorem 4.1 in [@BruinierOnoAlgebraicFormulas]), the values $F(z_{Q})$ of $F$ at Heegner points $z_{Q}$ of discriminant $D$ lie in the ring class field of the order $\mathcal{O}_{D}$ over $\operatorname{\mathbb{Q}}(\sqrt{D})$. Further, Lemma 4.3 in [@BruinierOnoAlgebraicFormulas] asserts that the values $F(z_{Q})$ are algebraic integers. The Galois group of the ring class field of $\mathcal{O}_{D}$ over $\operatorname{\mathbb{Q}}(\sqrt{D})$ permutes the Heegner points occuring in $\operatorname{tr}_{F}^{+}(D,r)$ and $\operatorname{tr}_{F}^{-}(D,r)$, see [@Gross]. It follows that $6\operatorname{tr}_{F}^{+}(D,r)$ and $6\operatorname{tr}_{F}^{-}(D,r)$ are algebraic integers in $\operatorname{\mathbb{Q}}(\sqrt{D})$, where the factor $6$ was added to get rid of possible factors $w_{Q}$ in the denominator. Using that $F$ has rational coefficients at $\infty$, we see that $\overline{\operatorname{tr}_{F}^{-}(D,r)} = \operatorname{tr}_{F}^{+}(D,r)$, and thus $$\begin{aligned}
&\operatorname{tr}_{F}^{+}(D,r)+\operatorname{tr}_{F}^{-}(D,r) \in \operatorname{\mathbb{Q}}\quad \text{and} \quad \operatorname{tr}_{F}^{+}(D,r)-\operatorname{tr}_{F}^{-}(D,r) \in \sqrt{D}\operatorname{\mathbb{Q}}.
\end{aligned}$$ This implies that the quantities in are rational integers.
In all the numerical examples we looked at, the numbers in were already integers without the factors $6$ and $6 t$. Possibly, this is always the case.
Combining Theorem \[ThetaLifts\] and Theorem \[RationalHolomorphicPart\] we obtain that if $F \in M_{0}^{!,\infty}(N)$ is as in Lemma \[ConstructionModularFunction\] and has rational principal part at $\infty$, then the holomorphic parts of $\operatorname{\textit{I}_{KM}}(F,\tau)$ and $\operatorname{\textit{I}_{M}}(F,\tau)$ have rational Fourier coefficients. This rationality result is remarkable since the holomorphic coefficients of a harmonic Maass form of weight $1/2$ which does not map to the space of unary theta functions under $\xi_{1/2}$ are conjectured to be transcendental almost always, see the conjecture and Corollary 1.4 in the introduction of [@BruinierOnoHeegnerDivisors].
\[AlgebraicCoefficients\] Let $K$ be a number field and let $H_{k,\bar{\rho}}(K)$ be the subspace of $H_{k,\bar{\rho}}$ consisting of forms whose principal part is defined over $K$.
1. Let $f \in H_{1/2,\bar{\rho}}(K)$ and suppose that $f$ is mapped to the space of unary theta functions by $\xi_{1/2}$. Then the coefficients of the holomorphic part $f^{+}$ of $f$ lie in $K$.
2. Let $f \in H_{3/2,\bar{\rho}}(K)$ and suppose that $f$ is mapped to the space of unary theta functions by $\xi_{3/2}$. Then there is a cusp form $f' \in S_{3/2,\bar{\rho}}$ such that the coefficients of $f^{+}-f'$ lie in $K$.
<!-- -->
1. The corresponding statement for the spaces $H_{1/2,\rho}$ and $H_{3/2,\rho}$ is also true, but a little less interesting. Since there are no unary theta functions for $\bar{\rho}$, it just says that a weakly holomorphic modular form of weight $1/2$ or $3/2$ for $\rho$, whose principal part is defined over $K$, has coefficients in $K$ up to addition of a cusp form. This follows immediatly from the fact that the spaces $M^{!}_{1/2,\rho}$ and $M^{!}_{3/2,\rho}$ have bases with rational coefficients (see [@McGraw]).
2. We have formulated the theorem using vector valued modular forms for the Weil representation since it best suits our applications. In view of Remark \[SerreStarkRemark\] the theorem implies a rationality result for the $\xi$-preimages of the scalar valued theta series $\theta_{\psi,m}(\tau)$ as well.
3. The holomorphic projection method developed in [@IRR] yields a recursive formula for the coefficients of the holomorphic part of $f$. It should also be possible to prove the above theorem using this technique.
We only prove the first claim, since the second one is similar. Let $f \in H_{1/2,\bar{\rho}}(K)$ and suppose that $\xi_{1/2}f$ lies in the space of unary theta functions. By Theorem \[ThetaLifts\], there is some $h \in H_{1/2,\bar{\rho}}$, which is a linear combination of harmonic Maass forms $h_{j}$ with rational holomorphic parts, such that $\xi_{1/2}f = \xi_{1/2}h$, i.e., $f -h$ is weakly holomorphic. We can write $f - h$ as a linear combination of forms $g_{i}$ with rational coefficients (see [@McGraw]). Having $f$ written in terms of the $g_{i}$ and $h_{j}$, we consider the system of linear equations obtained from comparing the principal parts. It is defined over $K$ and has a solution in $\operatorname{\mathbb{C}}$, so we can also solve it over $K$. Thereby we obtain a harmonic Maass form $\tilde{f}$ that has the same principal part as $f$ and still maps to the space of unary theta functions under $\xi_{1/2}$, but is now a linear combination of the $g_{i}$ and $h_{j}$ over $K$. In particular, the coefficients of $\tilde{f}^{+}$ lie in $K$. Then $f -\tilde{f}$ is a harmonic Maass forms which has vanishing principal part and maps to the space of unary theta function under $\xi_{1/2}$. By Lemma \[CuspFormLemma\], this implies that $f-\tilde{f}$ is a cusp form. But $S_{1/2,\bar{\rho}} \cong J_{1,N}^{\text{cusp}}= \{0\}$, so $f = \tilde{f}$, and thus $f^{+}$ has coefficients in $K$.
Regularized inner products and Weyl vectors of Borcherds products
=================================================================
The harmonic Maass forms constructed in Theorem \[ThetaLifts\] can be used to evaluate the regularized Petersson inner product of the unary theta functions $\theta_{1/2}$ and $\theta_{3/2}$ with harmonic Maass forms whose shadows are cusp forms.
\[InnerProductFormulas\] Let $F(z) = \sum_{n \gg -\infty}a(n)q^{n}\in M_{0}^{!,\infty}(N)$ be as in Lemma \[ConstructionModularFunction\].
1. Let $f \in H_{1/2,\rho}$ with holomorphic coefficients $c_{f}^{+}(D,r)$, where $D \equiv r^{2}(4N)$, and suppose that $\xi_{1/2}f \in S_{3/2,\bar{\rho}}$. Then $$\begin{aligned}
-\frac{\sqrt{N}}{4\pi}a(0)(f,\theta_{1/2})^{\operatorname{reg}} &= \sum_{r(2N)}\sum_{\substack{D < 0\\ D \equiv r^{2}(4N)}}c_{f}^{+}(D,r)\left( \operatorname{tr}_{F}^{+}(D,r) + \operatorname{tr}_{F}^{-}(D,r)\right) \\
& \quad + 4c_{f}^{+}(0,0)\sum_{n\geq 0}a(-n)\sigma_{1}(n) -2\sum_{b > 0}c_{f}^{+}(b^{2},b)b\sum_{n > 0}a(-bn).
\end{aligned}$$
2. Let $f \in H_{3/2,\rho}$ with holomorphic coefficients $c_{f}^{+}(D,r)$, where $D \equiv r^{2}(4N)$, and suppose that $\xi_{3/2}f \in S_{1/2,\bar{\rho}}$. Then $$\begin{aligned}
-\frac{1}{2\sqrt{N}}a(0)(f,\theta_{3/2})^{\operatorname{reg}} &= \sum_{r(2N)}\sum_{\substack{D < 0 \\ D \equiv r^{2}(4N)}}c^{+}_{f}(D,r)\frac{i}{\sqrt{|D|}}\left( \operatorname{tr}_{F}^{+}(D,r) - \operatorname{tr}_{F}^{-}(D,r)\right) \\
& \quad + 2\sum_{b>0}c_{f}^{+}(b^{2},b)\sum_{n>0}a(-bn).
\end{aligned}$$
We show the formula for $\theta_{1/2}$. Using Stokes’ theorem, we see as in the proof of [@BruinierFunke04], Proposition 3.5, that $$\begin{aligned}
(f,\theta_{1/2})^{\operatorname{reg}} = -(\operatorname{\textit{I}_{KM}}(F,\tau),\xi_{1/2}f)^{\operatorname{reg}} + \lim_{T \to \infty}\int_{-1/2}^{1/2}\langle f(u+iT),\overline{\operatorname{\textit{I}_{KM}}(F,u + iT)}\rangle du.
\end{aligned}$$ One can show as in [@alfes], Theorem 5.1, that $\operatorname{\textit{I}_{KM}}(F,\tau)$ and $\operatorname{\textit{I}_{M}}(F,\tau)$ are orthogonal to cusp forms, i.e., $(\operatorname{\textit{I}_{KM}}(F,\tau),\xi_{1/2}f)^{\operatorname{reg}} = 0$. The integral on the right-hand side picks out the zero-coefficient, to which only the holomorphic part of $f$ contributes. Hence we obtain a formula for $(f,\theta_{1/2})^{\operatorname{reg}}$ of the shape , involving only the coefficients of $f^{+}$. Plugging in the coefficients of $\operatorname{\textit{I}_{KM}}(F,z)$ from Theorem \[ThetaLifts\] yields the result.
\[ThetaNorms\] As a simple application of the last result, we show that the Petersson norms of $\theta_{1/2}$ and $\theta_{3/2}$ are given by $$\begin{aligned}
(\theta_{1/2},\theta_{1/2}) = \frac{\pi(N+1)}{3\sqrt{N}} \qquad \text{and} \qquad ( \theta_{3/2},\theta_{3/2}) = \frac{\sqrt{N}(N-1)}{6}.
\end{aligned}$$ These can of course also be evaluated using more direct methods, for instance, the Rankin-Selberg $L$-function, but it is interesting to see how the dependency on $F$ in Theorem \[InnerProductFormulas\] disappears if we plug in $\theta_{1/2}$ or $\theta_{3/2}$ for $f$.
We only show the formula for $\theta_{1/2}$, since the proof for $\theta_{3/2}$ is very similar. We can assume $a(0) = 1$. Let $$\begin{aligned}
E_{2}^{*}(z) = 1-24\sum_{n=1}^{\infty}\sigma_{1}(n)e(nz)-\frac{\pi}{3y}, \qquad \bigg(\sigma_{1}(n) = \sum_{d \mid N}d\bigg),
\end{aligned}$$ be the non-holomorphic Eisenstein series of weight $2$ for $\operatorname{SL}_{2}(\operatorname{\mathbb{Z}})$. Then $E_{2}^{*}(z)-NE_{2}^{*}(Nz)$ is a holomorphic modular form of weight $2$ for $\Gamma_{0}(N)$, and by applying the residue theorem to $F(z)(E_{2}^{*}(z)-NE_{2}^{*}(Nz))dz$, we find that $F$ satisfies $$\begin{aligned}
\label{ResidueFormula}
(1-N)-24\sum_{n>0}a(-n)(\sigma_{1}(n)-N\sigma_{1}(n/N)) = 0.
\end{aligned}$$ If we denote by $c_{\theta}(D,r)$ the coefficients of $\theta_{1/2}$, we see that $c_{\theta}(0,0) = 1$, and $c_{\theta}(b^{2},b) $ equals $2$ or $1$ for $b > 0$ depending on whether $b \equiv -b (2N)$ or not. Applying Theorem \[InnerProductFormulas\] and using the relation we obtain $$\begin{aligned}
-\frac{\sqrt{N}}{4\pi}(\theta_{1/2},\theta_{1/2}) &= 4\sum_{n\geq 0}a(-n)\sigma_{1}(n)-4\sum_{\substack{b > 0 \\ b \equiv 0(N)}}b\sum_{n>0}a(-bn) - 2\sum_{\substack{b>0 \\ b \not \equiv 0 (N)}}b\sum_{n>0}a(-bn) \\
&= 4\sigma_{1}(0) + 2\sum_{n>0}a(-n)(\sigma_{1}(n)-N\sigma_{1}(n/N)) = -\frac{1+N}{12}.
\end{aligned}$$ This yields the stated formula.
The formula given in the first item of Theorem \[InnerProductFormulas\] has applications in the theory of Borcherds products, see [@Borcherds]. We follow the exposition of [@BruinierOnoHeegnerDivisors]. Let $f \in H_{1/2,\rho}$ be a harmonic Maass form of weight $1/2$ for $\rho$ whose shadow is a cusp form, and assume that $c_{f}^{+}(D,r) \in \operatorname{\mathbb{R}}$ for all $D$ and $c_{f}^{+}(D,r) \in \operatorname{\mathbb{Z}}$ for $D \leq 0$. Then the infinite product $$\begin{aligned}
\Psi(z,f) = e(\rho_{f,\infty}z)\prod_{n =1}^{\infty}(1-e(nz))^{c_{f}^{+}(n^{2},n)}
\end{aligned}$$ is a meromorphic modular form of weight $c_{f}^{+}(0,0)$ for $\Gamma_{0}(N)$ and a unitary character, possibly of infinite order (see Theorems 6.1 and 6.2 in [@BruinierOnoHeegnerDivisors]). Here $\rho_{f,\infty}$ is the so-called *Weyl vector* at $\infty$, which is defined by $$\rho_{f,\infty} = \frac{\sqrt{N}}{8\pi}(f,\theta_{1/2})^{\operatorname{reg}}.$$ The Bocherds product $\Psi(z,f)$ has singularities at Heegner points in ${\mathbb{H}}$, which are prescribed by the principal part of $f$, and its orders at the cusps are determined by the corresponding Weyl vectors, which we describe now.
Each cusp of $\Gamma_{0}(N)$ can be represented by a reduced fraction $a/c$ with $c \mid N$, and the Weyl vector corresponding to $a/c$ is defined by $$\begin{aligned}
\label{WeylVectorAtCusp}
\rho_{f,a/c} = \frac{\sqrt{N}}{8\pi}(f,\theta_{1/2,N/(c,N/c)^{2}}^{\sigma_{c/(c,N/c)}}|U_{(c,N/c)})^{\operatorname{reg}},
\end{aligned}$$ where $\sigma_{c/(c,N/c)}$ is the Atkin-Lehner involution corresponding to the exact divisor $c/(c,N/c)$ of $N/(c,N/c)^{2}$ as in , and $U_{(c,N/c)}$ is the operator . Note that the Weyl vector at $a/c$ does not depend on $a$. Further, Theorem \[InnerProductFormulas\] yields a formula for the Weyl vector at each cusp $a/c$, involving only the principal part of $f$ and the coefficients $c_{f}^{+}(b^{2},r)$ for $b > 0$ and $r \in \operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$ with $r^{2}\equiv b^{2}(4N)$. Thus, we obtain the following rationality result.
\[WeylVectorsRationality\] Let $f \in H_{1/2,\rho}$ be a harmonic Maass form with $\xi_{1/2}f \in S_{3/2,\bar{\rho}}$. Suppose that $c_{f}^{+}(D,r) \in \operatorname{\mathbb{R}}$ for all $D$ and that $c_{f}^{+}(D,r) \in \operatorname{\mathbb{Z}}$ for $D \leq 0$. If $c_{f}^{+}(b^{2},r) \in \operatorname{\mathbb{Q}}$ for all $b > 0$ and all possible $r \in \operatorname{\mathbb{Z}}/2N\operatorname{\mathbb{Z}}$, then the Weyl vectors $\rho_{f,a/c}$ at all cusps are rational.
The formula for the Weyl vector $\rho_{f,a/c}$ obtained from Theorem \[InnerProductFormulas\] looks quite complicated in general. Thus, for simplicity, we only state it in the special case of a cusp $a/c$ with $c \mid \mid N$ and $(a,c) = 1$. Then $\theta_{1/2,N/(c,N/c)^{2}}^{\sigma_{c/(c,N/c)}}|U_{(c,N/c)} = \theta_{1/2,N}^{\sigma_{c}}$, and Theorem \[InnerProductFormulas\] gives the following formula.
\[WeylVectors\] Let $f \in H_{1/2,\rho}$ be a harmonic Maass form with $\xi_{1/2}f \in S_{3/2,\bar{\rho}}$. Suppose that $c_{f}^{+}(D,r) \in \operatorname{\mathbb{R}}$ for all $D$ and that $c_{f}^{+}(D,r) \in \operatorname{\mathbb{Z}}$ for $D \leq 0$. Let $c\mid \mid N$ and let $\sigma_{c}$ be the associated Atkin-Lehner involution as in . Let $F \in M_{0}^{!,\infty}(N)$ be as in Lemma \[ConstructionModularFunction\], normalized to $a(0) = 1$. Then the Weyl vector $\rho_{f,a/c}$ at the cusp $a/c$ is given by $$\begin{aligned}
\rho_{f,a/c}= \frac{\sqrt{N}}{8\pi}(f^{\sigma_{c}},\theta_{1/2})^{\operatorname{reg}} &= -\frac{1}{2}\sum_{r(2N)}\sum_{\substack{D < 0\\ D \equiv r^{2}(4N)}}c_{f}^{+}(D,\sigma_{c}(r))\left( \operatorname{tr}_{F}^{+}(D,r) + \operatorname{tr}_{F}^{-}(D,r)\right) \\
& \quad - 2c_{f}^{+}(0,0)\sum_{n\geq 0}a(-n)\sigma_{1}(n) + \sum_{b > 0}c_{f}^{+}(b^{2},\sigma_{c}(b))b\sum_{n > 0}a(-bn).
\end{aligned}$$
1. If $N$ is square free, the cusps of $\Gamma_{0}(N)$ are represented by the fractions $1/c$, where $c$ runs through the divisors of $N$. In this case all Weyl vectors can be computed with the above formula.
2. In [@Borcherds], Section 9, the Weyl vectors are computed in a similar way, using non-holomorphic Eisenstein series of weight $3/2$ as $\xi$-preimages for $\theta_{1/2}$. However, this only works if $N = 1$ or if $N = p$ is a prime. Otherwise, the Eisenstein series, and thus also its $\xi$-image, is invariant under all Atkin-Lehner involutions, but $\theta_{1/2}$ is not.
We consider the Borcherds lift of $f = \theta_{1/2}$, for $N$ arbitrary. By Example \[ThetaNorms\], the Weyl vector of $\theta_{1/2}$ at $\infty$ equals $(1+N)/24$, so its Borcherds product is given by $\Psi(z,\theta_{1/2}) = \eta(z)\eta(Nz)$. By a similar computation as in Example \[ThetaNorms\] we find $$(\theta_{1/2}^{\sigma_{c}},\theta_{1/2}) =\frac{\pi}{3\sqrt{N}}\left(\frac{N}{c} + c\right)$$ for $c \mid \mid N$. Hence the Weyl vector of $\Psi(z,f)$ at a cusp $a/c$ with $c \mid \mid N$ is given by $\frac{1}{24}\frac{N}{c}\big(1 + \frac{c^{2}}{N} \big)$.
Finally, we would like to mention that the harmonic Maass form $\operatorname{\textit{I}_{M}}(F,\tau)$ given in Theorem \[ThetaLifts\] can be used to construct rational functions on $X_{0}(N)$ with special divisors. Let $\Delta \neq 1$ be a fundamental discriminant and let $r \in \operatorname{\mathbb{Z}}$ with $\Delta \equiv r^{2}(4N)$. Further, let $\tilde{\rho} = \rho$ if $\Delta > 1$ and $\tilde{\rho} = \bar{\rho}$ if $\Delta < 0$. The twisted Borcherds product of a harmonic Maass form $f \in H_{1/2,\tilde{\rho}}$ with real holomorphic part and integral principal part is defined by $$\Psi_{\Delta,r}(z,f) = \prod_{n =1}^{\infty}\prod_{b(\Delta)}[1-e(b/\Delta)e(nz)]^{\left(\frac{\Delta}{b}\right)c_{f}^{+}(|\Delta|n^{2},rn)},$$ see [@BruinierOnoHeegnerDivisors], Theorem 6.1. Note that the Weyl vectors vanish for $\Delta \neq 1$. The function $\Psi_{\Delta,r}(z,f)$ is a meromorphic modular form of weight $0$ for $\Gamma_{0}(N)$ and a unitary character, which is of finite order if and only if the coefficients $c_{f}^{+}(|\Delta|n^{2},rn)$ are rational (see Theorem 6.2 in [@BruinierOnoHeegnerDivisors]).
If $F \in M_{0}^{!,\infty}(N)$ is as in Lemma \[ConstructionModularFunction\] and has integral principal part, then the Millson lift $\operatorname{\textit{I}_{M}}(F,\tau)$ given in Theorem \[ThetaLifts\] is a harmonic Maass form in $H_{1/2,\bar{\rho}}$ with rational holomorphic part and integral principal part. In particular, for $\Delta < 0$, some power of the twisted Borcherds product $\Psi_{\Delta,r}(z,\operatorname{\textit{I}_{M}}(F,\tau))$ defines a rational function on $X_{0}(N)$ whose zeros and poles lie on a twisted Heegner divisor.
| {
"pile_set_name": "ArXiv"
} |
[**KIAS P11035\
HIP-2011-18/TH**]{}
[**Production of two Higgses at the Large Hadron Collider in CP-violating supersymmetry**]{}\
\
[$^a$ Korea Institute for Advanced Study,\
Hoegiro 87, Dongdaemun-gu, Seoul 130-722, Korea]{}\
Introduction
============
CP violation is among the phenomena which are not fully understood in the context of the Standard Model (SM). Although CP violation exists in the SM, and agrees well with the laboratory experiments, there is an inconsistency between the amount of violation and matter content of the Universe, and it is argued that new sources of CP violation are needed.
Many of the proposals for beyond the SM (BSM) physics contain new sources for CP violation. In this work we will consider a supersymmetric model. Many parameters of the model are complex in general, and thus the common choice of real parameters can be thought of as very special. However, general complex parameters lead to the so-called SUSY CP problem, [*i.e.*]{} too large violation of CP [@cpphase]. In a more complete theory, a suitable amount of CP violation could be explained dynamically. We assume here that such a theory exists.
The physical observables in a general MSSM depend on two types of linearly independent phases [@cpphase], $$Arg[M_i\mu (m_{12}^2)^*],\,\, Arg[A_f\mu (m_{12}^2)^*].$$ Here $M_i$, $i=1,2,3$ are the soft gaugino mass parameters, $\mu$ is the higgsino mixing parameter, $m_{12}$ is a soft scalar parameter, and $A_f$ are the soft trilinear supersymmetry breaking parameters corresponding to each flavour $f$. The experimental upper bounds on the electric dipole moments (EDMs) of electrons and neutrons [@edm1; @edm2] as well as of mercury atoms [@edm3] constrain these phases, which appear also in the Higgs sector of the model via radiative corrections, and they can largely affect both the masses and couplings of the Higgs bosons. It is assumed that the mechanism of the electroweak symmetry breaking manifests itself in any case in the energy range of the LHC. Thus the related particle, the Higgs boson, may be the first one to inform on the BSM physics. If supersymmetry exists, it is indeed very possible that the CP violating effects are shining the Higgs sector.
Search for the Higgs boson(s) is one of the first goals of the Large Hadron Collider (LHC). For the SM Higgs boson, mass bound from the LEP collider is $m_h>114.4$ GeV [@Barate:2003sz; @Schael:2006cr], and LHC already excludes the SM Higgs boson with masses 145-466 GeV (except 288-296 GeV) at 95% CL \cite{}. In the MSSM with real and CP-conserving parameters, the lower limit on the lightest Higgs boson is $\sim$ 90 GeV [@susylim] for any $\tan{\beta}$. The lower bound on the mass of the lightest Higgs boson of the CP-conserving MSSM from LEP [@Schael:2006cr] can be drastically reduced or may even entirely vanish if non-zero CP-violating phases are allowed [@Bechtle:2006iw; @Carena:2000ks]. This can happen through radiative corrections to the Higgs potential, whereby the above mentioned phases of the $\mu$ parameter and the $A$ parameters enter into the picture [@Pilaftsis:1998pe; @Pilaftsis:1998dd]. We will concentrate on such a CP-violating scenario.
The three physical neutral states, which in the CP conserving case have definite CP properties, turn to three neutral states $h_i$ ($i$=1,2,3) with mixed CP properties. The collider search limits for all of them are modified since the physical Higgs bosons are mixtures of even and odd degrees of freedom. Due to this mixing through the loop effects, the lightest Higgs boson is almost CP-odd with strongly suppressed coupling to $ZZ$ pair, thus resulting in reduced production rates and consequent weakening of mass limits at collider experiments. The number of parameters in the general MSSM is very large, and thus it is customary to use benchmark points in order to map the relevant phenomenology. In this work we are particularly interested in the possibility that the Higgs boson is light and has not been seen because its couplings are suppressed at LEP. With this in mind we use the so-called CPX model [@Pilaftsis:1998pe; @Pilaftsis:1998dd] as our benchmark model.
In the CPX scenario the $ZZh_1$ coupling can be strongly reduced because of the CP violating phases, and the LEP mass limit for the lightest Higgs boson can be lowered to 50 GeV or even less, depending on $\tan\beta$. Thus the LEP searches leave a hole in $(m_{h_1},\tan\beta)$ parameter space [@Schael:2006cr]. Complementary channels such as $e^+e^-\to h_1h_2$ suffer also phase space suppression within the hole region. At Tevatron, the CP violation with Higgs production in the SM search channels [@Carena:2002bb], as well as CP violation specifically through $Wh_2$ production[@Das:2010ds] have been studied.
Within the hole region in addition to $ZZh_1$ coupling, $WWh_1$ and $tth_1$ are suppressed and thus the lightest Higgs boson $h_1$ is difficult to discover. On the other hand, the relatively heavy neutral Higgs bosons $h_{2,3}$ couple to $W$, $Z$ and $t$ favourably, but they can decay in non-standard channels, thus requiring a modification in search strategies. So far studies on possible signals of the CPX scenario at the LHC have been restricted to the production of $h_i$ ($i$=1,2,3) bosons in SM-like channels [@Accomando:2006ga], although all the decay channels have been considered. It has been concluded that parts of the holes in the $m_{H^\pm}-\tan{\beta}$ or the $m_{h_1}-\tan{\beta}$ parameter space can be plugged, although considerable portions of the hole, especially for low $\tan\beta$, may escape detection at the LHC even after accumulating 300 fb$^{-1}$ of integrated luminosity.
In [@Bandyopadhyay:2010tv] the cascade decay of the third generation scalar quarks, mainly $\tilde{t}_{1}\tilde{t}^{*}_{1}$ and $\tilde{b}_{1}\tilde{b}^{*}_{1}$ was explored in the context of Higgs production via charged Higgs decaying to $W$ and Higgs ($h_1$) bosons and the multi-channel analysis shows that the ‘hole’ can be probed at an integrated luminosity of 5-10 fb$^{-1}$ assuming $\sqrt{s}=14$ TeV. It was noted that in a general $CP$-violating MSSM, the cross section of ${\tilde{t}_1}{\tilde{t}_1^*}h_1$ production could be very much larger than that obtained by switching off the $CP$-violating phases [@Li:2006hq]. In [@Bandyopadhyay:2007cp], it was shown that the associated production of two Higgses could probe the ‘LEP hole’. These could be discovery channels, in cases where the $t$-$\bar{t}$-$h_1$ and $W$-$W$-$h_1$, $Z$-$Z$-$h_1$ couplings are highly suppressed. In [@Bandyopadhyay:2010tv; @Bandyopadhyay:2007cp] it has been found that although it is possible to probe the ‘LEP-hole’ in the CPX scenario, in most cases it is difficult to reconstruct all the Higgses ($h_1$, $h_2$ and $h_3$) due to the combinatorial backgrounds from the supersymmetric cascade and production processes. In the context of CP-conserving MSSM, cascade Higgs production has been studied in [@Bandyopadhyay:2008sd]. The process studied here, namely the Higgs pair production, is a clean channel as the final state has jets from Higgses only. This gives an extra motivation to look at the pair production channels along with the previously studied ones. We study the production of two Higgs bosons $h_ih_j$ at the LHC, and subsequent decays of the heavy Higgs bosons to $Z$ and the lightest Higgs boson, $h_{2,3}\rightarrow Z h_1$, where $Z$ furthermore decays leptonically. It turns out that with an integrated luminosity of 20 fb$^{-1}$ the hole can be covered, except in special parameter regions, where $h_2h_1$ production gets large enhancement and the required luminosity to cover the relevant parameter region is $\leq 1$ fb$^{-1}$. We study the signal and the Standard Model backgrounds from these production and decay processes, in order to probe the hole region.
We will also point out that in certain benchmark points the two Higgs production through coupling of three Higgs bosons is important, and thus we have a possibility to probe the Higgs potential at those points. Obviously construction of Higgs potential would be of fundamental importance.
The paper is organised as follows. In Section 2 we shortly review the relevant parts of the CPX model, including the resulting Higgs mass spectrum. All of our subsequent numerical analysis are in this framework. In Section 3 we discuss the proposed parton level signal. We do the collider simulation and devise the event selection criteria to reduce the SM backgrounds and present the final numerical results. We also discuss enhancements in the production cross-section for some parameter points. We summarise in Section 4.
CPX scenario
============
As indicated in the introduction, we adopt here the so-called CPX scenario in which the LEP analyses have been performed. In the CPX scenario, it is assumed that certain parameters are related: &&m\_ = m\_ = m\_ = M\_[SUSY]{}, |A\_t| = |A\_b| = |A\_| = 2 M\_[SUSY]{},\
&&arg(A\_[t]{})=arg(A\_[b]{})=arg(A\_)=90\^0. It has been observed [@Pilaftsis:1998pe; @Pilaftsis:1998dd] that the $CP$-violating quantum effects on the Higgs potential are proportional to $Im(\mu A_t)/M^2_{SUSY}$. The consequences of the CPX scenario have been studied in [@Ghosh:2004wr].
The corresponding inputs that we adopt here are compatible with the “hole” left out in the analysis. We take &&M\_[SUSY]{}=500[GeV]{},||= 1[TeV]{}, M\_2 = 2 M\_1 = 200[GeV]{},\
&&arg(A\_[b,]{}) = 90\^,arg() = 90\^, = 5 - 10 where the only difference to the reference [@Carena:2002bb] lies in a small tweaking in the mass ratio of the $U(1)$ and $SU(2)$ gaugino masses $M_1$ and $M_2$, aimed at ensuring gaugino mass unification at high scale. It has been checked that this difference does not affect the Higgs production or the decay rates. We vary $\phi_{A_t}$ to higher values ($\geq 90^\circ$), charged Higgs mass ($m_{H^\pm}$) and $\tan{\beta}$ to investigate the production of two Higgses. The value of the top quark mass has been taken to be 175 GeV[^1]
The first two generation sfermion masses are assumed to be heavy so that the experimental bounds (for example, the electric dipole moment of the neutron) are satisfied. Here we have not considered possible ways of bypassing such bounds, and will set the masses of the first two sfermion families at 10 TeV.
For the collider simulation and analysis to cover the regions of the ‘hole’, we select benchmark points in the ‘hole’. The benchmark points are given in Table \[tabbps\] with the corresponding neutral Higgs mass spectrum shown with the radiative corrections.
Parameters BP1 BP2 BP3 BP4
--------------- ------- ------- ------- -------
$\tan{\beta}$ 5 5 5 7
$\phi_{A_t}$ 112 122 124 163
$m_{H^\pm}$ 146 155 154 149
$m_{h_1}$ 31.0 30.8 12.6 4.0
$m_{h_2}$ 117.3 124.1 124.2 129.0
$m_{h_3}$ 146.1 152.8 151.5 139.8
: Benchmark points within the LEP-hole in $m_{h_1}$-$\tan\beta$ plane and the corresponding neutral Higgs masses.[]{data-label="tabbps"}
Numerical results for the benchmark points
==========================================
Production and decay of a $h_ih_j$ pair
---------------------------------------
As is clear from the Table \[tabbps\], in the CPX scenario the lightest Higgs mass can be very light. For some parameter points the heavy Higgs bosons ($h_{2,3}$) can decay to the lightest Higgs boson, $h_1$. We start by analysing the benchmark point BP1 in detail. For BP1 only $h_3h_i,\,i=$1,2,3 contributes to the signal we study ([*i.e.* ]{} jets and opposite sign lepton pair) and thus the signal cross section is the smallest of the chosen benchmark points. For the other benchmark points the analysis will be very similar except for benchmark point 4, where the light Higgs ($h_1$) is very light ($m_{h_1}\sim 4$ GeV). We give the relevant results for signal and background in the Tables in the next Section.
For the BP1 we list the most important Higgs decay branching ratios in Table \[brHiggses\] . It is seen that $h_3\to h_1 Z$ branching ratio is large. Since we want to include the opposite sign di-lepton (OSD) in our signal of the event, we consider the process $$\begin{aligned}
pp&\to & h_3h_i, \quad i=1,2,3\nonumber\\
&\to & h_1Zh_1h_1\,\, {\rm or}\,\, h_1Z h_1 \,\, {\rm or}\,\, h_1Zb\bar b\nonumber\\
&\to &(4b \,\, {\rm or}\,\, 6b)+OSD
\label{process}\end{aligned}$$ The cross-sections of the Higgs pair production for the benchmark points have been given in Table \[crossH\] for $E_{cm}=$14 TeV. We do not consider lower $E_{cm}$, since the cross section would be too small to be detected. The cross-sections are computed with [CalcHEP]{} [@Pukhov:2004ca] (interfaced with the program [CPSuperH]{} [@Ellis:2006eh; @Lee:2003nta]). [CTEQ6L]{} [@Lai:1999wy; @Pumplin:2002vw] parton distribution functions are used and the renormalization/factorization scale is set to $\sqrt{\hat{s}}$. The largest contribution for the signal in Eq. (\[process\]) for benchmark point BP1 comes from $h_3h_1$ for the $Z$ production. In BP1 $h_2$ decay to the lightest Higgs boson and $Z$ is not kinematically allowed but, as we will see later, this mode also gets opened up in the case of other benchmark points.
Br($h_3 \to h_1 Z$) Br($h_3 \to h_1 h_1$) Br($h_2 \to h_1 h_1 $) Br($h_2 \to b \bar{b}$) Br($h_1 \to b \bar{b}$)
--------------------- ----------------------- ------------------------ ------------------------- -------------------------
0.55 0.11 0.52 0.43 0.93
: Branching fractions for Higgs bosons in the CPX scenario BP1.[]{data-label="brHiggses"}
Benchmark point $\sigma_{h_3 h_i=1,2,3}$ \[fb\] $\sigma_{h_1 h_2}$ \[fb\]
----------------- --------------------------------- ---------------------------
BP1 226 285
BP2 206 323
BP3 248 7929
BP4 464 3119
: Production cross sections (in fb) for the signal processes at the LHC with $E_{cm}=14$ TeV in the CPX scenario for the benchmark points. []{data-label="crossH"}
The dominant background from the Standard Model for the signal in Eq. \[process\] are $t\bar{t}$, $t\bar{t}Z$ and $t\bar{t}b\bar{b}$. In the parton level $t\bar{t}$ seems no to be a background having $2b+ OSD +\ptmiss$ in the final state, but with ISR/FSR the number of jets can increase to a larger value and also due to the large cross-section contributes as one of the dominant backgrounds. Unlike the signal, the invariant mass of leptons will not peak around $Z$ mass. This will help to kill the $t\bar{t}$ backgrounds. Similarly we can also get rid of $t\bar{t}b\bar{b}$. On the other hand $t\bar{t}Z$ could be a true background if both the top quarks decay hadronically. In the next subsection we simulate both the signal and background through [PYTHIA]{} [@Sjostrand:2001yu] and define the cuts and signal topologies explicitly.
Signal analysis
---------------
In this study, [CalcHEP]{} (interfaced to the program [CPSuperH]{}) has been used for generating parton-level events for the relevant processes. The standard [CalcHEP-PYTHIA]{} interface [@alex], which uses the SLHA interface [@Skands:2003cj] was then used to pass the [CalcHEP]{}-generated events to [PYTHIA]{} [@Sjostrand:2001yu]. Furthermore, all relevant decay information is generated with [CalcHEP]{} and is passed to [PYTHIA]{} through the same interface. All these are required since there is no public implementation of CP violating MSSM in [PYTHIA]{}. Subsequent decays of the produced particles, hadronization and the collider analyses are done with [PYTHIA (version 6.4.22)]{}.
We use [CTEQ6L]{} parton distribution function (PDF) [@Lai:1999wy; @Pumplin:2002vw]. In [CalcHEP]{} we opted for the lowest order $\alpha_s$ evaluation, which is appropriate for a lowest order PDF like [CTEQ6L]{}. The renormalization/factorization scale in [CalcHEP]{} is set at $\sqrt{\hat{s}}$. This choice of scale results in a somewhat conservative estimate for the event rates.
In the CPX scenario, although $h_1$ decays dominantly into $b\bar{b}$, our simulation reveals that in a fairly large fraction of events neither of the $b$-quarks leads to sufficiently hard jets with reasonable $b$-tagging efficiency because of the lightness of $h_1$ in this scenario. To illustrate this, we present in Figure \[bptdis\] the ordered $p_T$ distributions for the four parton-level $b$-quarks in the signal from $h_3h_{i=1,2,3}$ and corresponding $p^{jet}_T$ distribution including ISR/FSR. It is clear from this figure that the $b$-quark which is coming from $h_1$ has the smallest $p_T$ in a given event, and is often below 40 GeV or thereabout. The other $b$ coming from $h_3$ or $h_2$ has larger $p_T$. For this analysis we have required a $b$-jet tagging efficiency ($\geq$ 50%) [@Baer:2007ya].
-15pt -12pt
For hadronic level simulation we have used [PYCELL]{}, the toy calorimeter simulation provided in [PYTHIA]{}, with the following criteria:
- the calorimeter coverage is $\rm |\eta| < 4.5$ and the segmentation is given by $\rm\Delta\eta\times\Delta\phi= 0.09 \times 0.09 $ which resembles a generic LHC detector
- a cone algorithm with $\Delta R = \sqrt{\Delta\eta^{2}+\Delta\phi^{2}} = 0.5$ has been used for jet finding
- $ p_{T,min}^{jet} = 20$ GeV and jets are ordered in $p_{T}$
- leptons ($\rm \ell=e,~\mu$) are selected with $p_T \ge 20$ GeV and $\rm |\eta| \le 2.5$
- no jet should match with a hard lepton in the event
The jet-multiplicity distribution is shown in Figure \[jetm\] where the jets are generated using PYCELL. It is evident that the Higgs pair signal has lower jet-multiplicity compared to that of $t\bar{t}$. Thus, jet-multiplicity ($n_{jet}\le 4$) can reduce the $t\bar{t}$ background considerably.
From Figure \[leppt\] we can see the $p_T$ distributions of the leptons where the leptons come from the $Z$ boson which originates from the decay of either $h_2$ or $h_3$ via $h_{2,3}\to h_1 Z$. The invariant distribution of these leptons would also peak at the $Z$ mass. This is the best handle to tackle the background like $t\bar{t}$ or $t\bar{t}b\bar{b}$ where the leptons come from two different $W^\pm$s from the decays of tops. We also put the isolation criteria on the leptons from jets or leptons as mentioned above. In Figure \[lepmt\] we plot an isolated lepton multiplicity distribution which shows that there are a few isolated lepton pairs at the end of all the cuts. Figure \[lpminv\] shows the invariant mass distribution of the isolated leptons for the signal and $t\bar{t}$. For the signal we can reconstruct the $Z$ peak.
No. Signal topology BP1 BP2 BP3
----- ------------------------------------------------------------------------------ ------ ------ ------
1 $n_{jet}\le 4( b-{\rm jet}\ge 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{\l_1}_T \le 90+ p^{\l_2}_T \le 90 $ 0.27 0.33 9.9
$p^{j_3}_T \le 40+ M_{eff}\le 200 +|M_{\l\l}-90| \le 3 $
2 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_i(i\ne 1)}_T \le 70 $ 0.25 0.30 6.9
$M_{eff}\le 200+|M_{\l\l}-90| \le 3 +\phi_{j_2,\l_1}\le 1.6 $
3 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 70 + p^{j_2)}_T \le 70 $ 0.16 0.20 3.9
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 +0.5 \le\phi_{j_2,\l_1}\le 1.8 $
4 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_2)}_T \le 70 $ 0.31 0.38 10.1
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 $
5 $n_{jet}\le 3( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{j_3}_T \le 40 $ 0.06 0.08 2.5
$+|M_{\l\l}-90| \le 2.5 $
: Event rates for the CPX benchmark points with an integrated luminosity of 1 fb$^{-1}$[]{data-label="tabsig"}
No. Signal topology $t\bar{t}$ $t\bar{t}Z$ $t\bar{t}b\bar{b} $
----- ------------------------------------------------------------------------------ ------------ ------------- ---------------------
1 $n_{jet}\le 4( b-{\rm jet}\ge 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{\l_1}_T \le 90+ p^{\l_2}_T \le 90 $ 0.10 0.005 0.0
$p^{j_3}_T \le 40+ M_{eff}\le 200 +|M_{\l\l}-90| \le 3 $
2 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_i(i\ne 1)}_T \le 70 $ 0.07 0.004 0.0
$M_{eff}\le 200+|M_{\l\l}-90| \le 3 +\phi_{j_2,\l_1}\le 1.6 $
3 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 70 + p^{j_2)}_T \le 70 $ 0.07 0.003 0.0
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 +0.5 \le\phi_{j_2,\l_1}\le 1.8 $
4 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_2)}_T \le 70 $ 0.10 0.005 0.0
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 $
5 $n_{jet}\le 3( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{j_3}_T \le 40 $ 0.04 0.001 0.0
$+|M_{\l\l}-90| \le 2.5 $
: Event rates for the backgrounds with an integrated luminosity of 1 fb$^{-1}$[]{data-label="backg"}
Considering all the distributions we have defined five signal topologies with different cuts. Table \[tabsig\] presents these five different signal topologies and the event numbers at 1 fb$^{-1}$. The corresponding background events are given in Table \[backg\]. Table \[sigbps\] gives the corresponding significance and required integrated luminosity for $5\sigma$ significance. Comparing the signal and backgrounds we can see that for 5$\sigma$ signal significance we have to achieve at least an integrated luminosity of 108 fb$^{-1}$ for signal 4 of BP1.
Similar analyses for the other benchmark points (BP2 and BP3) for signal and backgrounds are summarised in Tables \[tabsig\] and \[backg\]. In order to achieve 5$\sigma$ significance for signal 4 of BP2 (Table \[sigbps\]) an integrated luminosity of 84 fb$^{-1}$ is needed. For BP3 the corresponding required luminosity is only 2.5 fb$^{-1}$. This happens because the cross-sections are enhanced for this benchmark point as shown in Table \[crossH\]. The huge enhancement is due to the increase in the couplings affecting $h_1h_2$ cross section, as discussed later.
----------- -------- --------------- -----------------------------------
Benchmark Significance Required luminosity \[fb$^{-1}$\]
Points Signal at 1fb$^{-1}$ 5$\sigma$ significance
sig1 0.44 129
sig2 0.44 130
BP1 sig3 0.33 225
sig4 0.48 108
sig5 0.19 701
sig1 0.5 100
sig2 0.5 103
BP2 sig3 0.4 169
sig4 0.5 84
sig5 0.2 473
sig1 3.1 2.6
sig2 2.6 3.7
BP3 sig3 2.0 6.5
sig4 3.2 2.5
sig5 1.6 10.2
----------- -------- --------------- -----------------------------------
: Significance at 1 fb$^{-1}$ and the required luminosity for 5$\sigma$ significance.[]{data-label="sigbps"}
-12pt
After calculating the significance we look at the possibility to construct the invariant mass peak for $h_1$. The left part of Figure \[invmass\] represents the parton level $b\bar{b}$ invariant mass distribution. From that figure we can see that the peaks corresponding to $h_1$, $Z$, $h_2$ and $h_3$ are clearly visible. To mimic a more realistic situation, we use the PYCELL jets and require $\geq 4\,\, {\rm jets} + 2 \,\, \rm{isolated} \,\, \rm{leptons}$ in the final state. The jet level invariant mass distribution with such a final state topology is shown in the right part of Figure \[invmass\]. We can clearly see the light Higgs mass peak, i.e., $m_{h_1}$ peak, near 30 GeV. This will remove all the model backgrounds on top of the SM backgrounds. Also we can reconstruct the light Higgs mass ($\leq 50$ GeV), which is an artifact of the ‘CPX’ scenario. It has already been noted in the literature that the size and the exact location of the hole in the parameter space depend on the method of calculating the loop correction [@Frank:2006yh; @Hahn:2006np]. However, the calculations agree qualitatively and confirm the presence of the hole.
Since the cross section of the signal for the benchmark point BP3 is much larger than for the other benchmark points, we consider PB3 at LHC with $E_{CM}=7$ TeV as well. The production cross sections for Higgs pairs are $\sigma(h_1h_2)=2$ pb and $\sigma(h_1h_3)=87$ fb. Thus clearly the main contribution for the signals comes from the production of $h_1h_2$. In Table \[bp37tev\] we present the amount of the signal events for BP3 and the SM background events at $E_{CM}=7$ TeV and integrated luminosity = 1 fb$^{-1}$. From the Table one sees that with an integrated luminosity of 6-13 fb$^{-1}$ the significance is 5$\sigma$ over the SM backgrounds for all the final states.
----- ------------------------------------------------------------------------------ -------- ------------ -------------- --------------------- ---------------
Required
Signal luminosity
in fb$^{-1}$
No. Final State BP3 $t\bar{t}$ $t\bar{t}Z$ $t\bar{t}b\bar{b} $ for $5\sigma$
Signi
-ficance
1 $n_{jet}\le 4( b-{\rm jet}\ge 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{\l_1}_T \le 90+ p^{\l_2}_T \le 90 $ 3.67 0.026 0.0 0.0 6.9
$p^{j_3}_T \le 40+ M_{eff}\le 200 +|M_{\l\l}-90| \le 3 $
2 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_i(i\ne 1)}_T \le 70 $ 2.46 0.026 0.0 0.0 10.3
$M_{eff}\le 200+|M_{\l\l}-90| \le 3 +\phi_{j_2,\l_1}\le 1.6 $
3 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 70 + p^{j_2)}_T \le 70 $ 1.99 0.0 0.0 0.0 12.6
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 +0.5 \le\phi_{j_2,\l_1}\le 1.8 $
4 $n_{jet}\le 4( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 90 + p^{j_2)}_T \le 70 $ 3.67 0.052 0.0 0.0 6.9
$M_{eff}\le 200+|M_{\l\l}-90| \le 2.5 $
5 $n_{jet}\le 3( b-{\rm jet}= 3) + \l\ge 2(\rm{OSD}\ge1) + \ptmiss \le 20$
$p^{j_1}_T \le 75 + p^{j_2}_T \le 50 +p^{j_3}_T \le 40 $ 0.80 0.013 0.0 0.0 5.6
$+|M_{\l\l}-90| \le 2.5 $
----- ------------------------------------------------------------------------------ -------- ------------ -------------- --------------------- ---------------
: Event rates for the backgrounds with an integrated luminosity of 1 fb$^{-1}$ at the LHC for ECM=7 TeV.[]{data-label="bp37tev"}
The scenario for benchmark point BP4 is a little different, as the lightest neutral Higgs ($h_1$) mostly decays to tau pairs (BR$(h_1\rightarrow \tau\tau ) \sim 90\%$). Thus the search mode for Higgs as well as the final state consist of taus. As in the previous case, the heavier Higgses ($h_{2,3}$) also decay to $Z$ and $h_1$. The final state then consists of four $\tau$s and two opposite sign leptons, invariant mass for which peaks around $Z$ mass. The $\tau$s coming from $h_1$ can be very soft as $h_1$ in this case is very light ($\sim 4$ GeV, see Table \[tabbps\]). Boost of the light Higgs ($h_1$) of course increases the $p_T$ of $\tau$s from decays. Figure \[pttau\] shows the $p_T$ distribution of the partonic $\tau$ coming from $h_1$ which establishes the fact that indeed there are some boosted $\tau$s, tagging of which might be possible.
Taus coming from Higgs then decay to pions through one prong or three prong decay. In the present study, we would be using the one prong (one charged track) hadronic decays of the $\tau$-leptons which have a collective branching fraction of about 50% of which almost 90% is comprised of final states with $\pi^\pm$, $\rho$ and $a_1$ mesons. To establish a jet as a $\tau$-jet we take the following approach. We first check, for each jet coming out of PYCELL within $|\eta| \le 2.5$, if there is a partonic $\tau$ within a cone of $\Delta{R}\le 0.4$ about the jet-axis. If there is one, then we further ensure that there is a single charged track within a cone of $\Delta{R} \le 0.1$ of the same jet axis. This marks a narrow jet character of a $\tau$-jet. Of course there is an efficiency associated to such kind of a geometric requirement which is a function of $p_T$ of the concerned jet and has been demonstrated in the literature [@tau1; @tau2]. We closely reproduce the values of the $\tau$-tagging efficiencies for the $p_T$ range we adopted for the concerned jets as indicated in references.
Tau tagging efficiency thus decreases when one tries to tag more $\tau$-jets. We optimize our signal with 4-jets among which three are reconstructed $\tau$-jets with one charged track and two opposite sign leptons peak around $Z$ mass. From the event numbers given below, we can see that $\ptmiss \le 20$ GeV cut is not necessary as the signal is background free without that as well. Unlike the previous three benchmark points, here only $t\bar{t}Z$ is the possible dominant background that can give two potential $\tau$s from the two $W^\pm$s coming from the top and another from some jet faking as $\tau$-jet. One can see from Table \[sigbp4\] that the background is zero and the signal has roughly two events at 1 fb$^{-1}$ of integrated luminosity without the missing $p_T$ cut which further reduces the number of signal events. Both of the final states are background free and have a reach with relatively small integrated luminosity.
----- --------------------------------------- -------- ------------- ---------------- ---------------------------
No. Final state Signal Background Significance Required
topology BP4 $t\bar{t}Z$ at 1 fb$^{-1}$ luminosity \[fb$^{-1}$\],
5$\sigma$ significance
1 $n_{jet}\le 4( \tau-{\rm jet}\ge 3)$
$+ \l\ge 2({\rm{OSD}}\ge1)$
$ + |M_{ll}|\le 2.5 \rm{GeV}$ 1.8 0.0 1.3 14
2 $n_{jet}\le 4(\tau-{\rm jet}\ge 3)$
$ + \l\ge 2(\rm{OSD}\ge1)$
$+ |M_{ll}|\le 2.5 + \ptmiss \le 20$ 0.6 0.0 0.77 42
----- --------------------------------------- -------- ------------- ---------------- ---------------------------
: Event rates for the CPX benchmark point BP4 and background $t\bar{t}Z$ with an integrated luminosity of 1 fb$^{-1}$[]{data-label="sigbp4"}
-8pt\
(a)-12pt (b)\
(c)
For benchmark points BP3 and BP4 the total cross-section gets an enhancement due to the large value of $\phi_{A_t}$. Figures \[ggh3\](a) and \[ggh3\](b) show how the odd parts of $g-g-h_3$ and $b-b-h_3$ increase as $\phi_{A_t}$ increases. For large $\phi_{A_t}$, also $h_3\sim h^{odd}_3 $ ($h_3=h^{r}_3+ih^{odd}_3$). On top of that $g_{h_3h_2h_1}$ also increases at large $\phi_{At}$ (see Figure \[ggh3\](c)). These two effects increase $\sigma(h_1h_2)$ for large $\phi_{A_t}$. Thus, for benchmark point BP3, the contribution of $\sigma(h_1h_2)=7.93$ pb (Table \[crossH\]) whereas $\sigma(h_ih_3)=0.248$ pb, where i=1,2,3 have been included. Similarly for BP4, the contributions are $\sigma(h_1h_2)=3.12$ pb and $\sigma(h_ih_3)=0.464$ pb. In the case of a low $\phi_{A_t}$, corresponding to BP1 and BP2; the contribution of $\sigma(h_1h_2)=0.285, 0.323$ pb and $\sigma(h_ih_3)=0.226, 0.206$ pb for BP1 and BP2, respectively.
Summary and discussion
======================
From our analysis it is clear that the Higgs pair production is interesting in spite of being electroweak production process. We have seen that for some signal topology an integrated luminosity of 20 fb$^{-1}$ is enough for $5\sigma$ significance. There are some points where we have a very large cross-section (BP3 and BP4) due to the enhancement of couplings. Thus, it is very easy to get discovery significance for these points. We have seen that it is also possible to reconstruct the Higgs mass peak, and we can get rid of the model backgrounds. Thus the signal topologies coming from Higgs pair productions are very different from CP-conserving case.
Production of two Higgs bosons has contribution from the subprocess including coupling of three Higgs bosons. Especially this contribution is important for our benchmark points BP3 and BP4. Since the three Higgs boson vertex comes from the Higgs potential, the production process can give information on the potential, and is thus of great interest and importance. Lastly, BP4 analysis in $\tau$ mode is relatively clean as it is background free.
[**Acknowledgments:**]{}
KH acknowledges support from the Academy of Finland (Project No 137960). PB wants to thank Helsinki Institute of Physics for the visit during the project and KIAS overseas travel grant. PB also thanks Prof. Jae Sik Lee for useful discussions.
[99]{}
M. Dugan, B. Grinstein and L. J. Hall, Nucl. Phys. B [**255**]{} (1985) 413. S. Dimopoulos and S. D. Thomas, Nucl. Phys. B [**465**]{} (1996) 23 \[arXiv:hep-ph/9510220\]. Y. Kizukuri, N. Oshimo, Phys. Rev. [**D46**]{}, 3025-3033 (1992).
P. Nath, Phys. Rev. Lett. [**66**]{}, 2565 (1991); Y. Kizukuri and N. Oshimo, Phys. Rev. D [**46**]{}, 3025 (1992); T. Ibrahim and P. Nath Phys. Lett. B [**418**]{}, 98 (1998); Phys. Rev. D [**57**]{}, 478 (1998); [*ibid*]{} D [**58**]{}, 019901(E) (1998); [*ibid*]{} D [**60**]{}, 079903 (1999); [*ibid*]{} D [**60**]{}, 119901 (1999); M. Brhlik, G.J. Good and G.L. Kane, Phys. Rev. D [**59**]{}, 115004 (1999); A. Bartl, T. Gajdosik, W. Porod, P. Stockinger and H. Stremnitzer, Phys. Rev. D [**60**]{}, 073003 (1999); D. Chang, W.-Y. Keung and A. Pilaftsis, Phys. Rev. Lett. [**82**]{}, 900 (1999); S. Pokorski, J. Rosiek and C.A. Savoy, Nucl. Phys. B [**570**]{}, 81 (2000); E. Accomando, R. Arnowitt and B. Dutta, Phys. Rev. D [**61**]{}, 115003 (2000); S. Abel, S. Khalil and O. Lebedev, Nucl. Phys. B [**606**]{}, 151 (2001); U. Chattopadhyay, T. Ibrahim and D.P. Roy, Phys. Rev. D [**64**]{}, 013004 (2001); D.A. Demir, M. Pospelov and A. Ritz, hep-ph/0208257.
A. Pilaftsis, Nucl. Phys. B [**644**]{}, 263 (2002).
T. Falk, K.A. Olive, M. Pospelov and R. Roiban, Nucl. Phys. B [**60**]{}, 3 (1999).
R. Barate [*et al.*]{} \[LEP Working Group for Higgs boson searches\], Phys. Lett. B [**565**]{}, 61 (2003) \[arXiv:hep-ex/0306033\];\
see also $http://lephiggs.web.cern.ch/LEPHIGGS/www/Welcome.html$ S. Schael [*et al.*]{} \[ALEPH Collaboration\], Eur. Phys. J. C [**47**]{}, 547 (2006) \[arXiv:hep-ex/0602042\];\
see also $http://lephiggs.web.cern.ch/LEPHIGGS/www/Welcome.html$
T. Aaltonen [*et al.*]{} \[CDF and D0 Collaborations\], Phys. Rev. Lett. [**104**]{} (2010) 061802 \[arXiv:1001.4162 \[hep-ex\]\].
See: LEP SUSY Working Group, [http://lepsusy.web.cern.ch/lepsusy]{}, and LEP Higgs Working Group,[LHWG-Note 2004-01]{}.
P. Bechtle \[LEP Collaboration\], PoS [**HEP2005**]{}, 325 (2006) \[arXiv:hep-ex/0602046\]. M. S. Carena, J. R. Ellis, A. Pilaftsis and C. E. M. Wagner, Phys. Lett. B [**495**]{}, 155 (2000) \[arXiv:hep-ph/0009212\]. A. Pilaftsis, Phys. Rev. D [**58**]{}, 096010 (1998) \[arXiv:hep-ph/9803297\]. A. Pilaftsis, Phys. Lett. B [**435**]{}, 88 (1998) \[arXiv:hep-ph/9805373\].
M. S. Carena, J. R. Ellis, S. Mrenna, A. Pilaftsis and C. E. M. Wagner, Nucl. Phys. B [**659**]{}, 145 (2003) \[arXiv:hep-ph/0211467\]. S. P. Das, M. Drees, \[arXiv:1010.3701 \[hep-ph\]\]. S. P. Das, M. Drees, \[arXiv:1010.2129 \[hep-ph\]\].
E. Accomando [*et al.*]{}, \[arXiv:hep-ph/0608079\]. P. Bandyopadhyay, \[arXiv:1008.3339 \[hep-ph\]\]. Z. Li, C. S. Li and Q. Li, Phys. Rev. D [**73**]{}, 077701 (2006) \[arXiv:hep-ph/0601148\]; P. Bandyopadhyay, A. Datta, A. Datta [*et al.*]{}, Phys. Rev. [**D78** ]{} (2008) 015017. \[arXiv:0710.3016 \[hep-ph\]\].
P. Bandyopadhyay, JHEP [**0907** ]{} (2009) 102. \[arXiv:0811.2537 \[hep-ph\]\]; P. Bandyopadhyay, A. Datta, B. Mukhopadhyaya, Phys. Lett. [**B670** ]{} (2008) 5-11. \[arXiv:0806.2367 \[hep-ph\]\]; K. Huitu, R. Kinnunen, J. Laamanen [*et al.*]{}, Eur. Phys. J. [**C58** ]{} (2008) 591-608. \[arXiv:0808.3094 \[hep-ph\]\]; A. Datta, A. Djouadi, M. Guchait [*et al.*]{}, Nucl. Phys. [**B681** ]{} (2004) 31-64. \[hep-ph/0303095\]; A. Datta, A. Djouadi, M. Guchait [*et al.*]{}, Phys. Rev. [**D65** ]{} (2002) 015007. \[hep-ph/0107271\]; G. D. Kribs, A. Martin, T. S. Roy and M. Spannowsky, arXiv:1006.1656 \[hep-ph\]; G. D. Kribs, A. Martin, T. S. Roy and M. Spannowsky, Phys. Rev. D [**81**]{} (2010) 111501 \[arXiv:0912.4731 \[hep-ph\]\].
D. K. Ghosh, S. Moretti, Eur. Phys. J. [**C42** ]{} (2005) 341-347 \[hep-ph/0412365\]; D. K. Ghosh, R. M. Godbole, D. P. Roy, Phys. Lett. [**B628** ]{} (2005) 131-140 \[hep-ph/0412193\]; A. Pilaftsis and C. E. M. Wagner, Nucl. Phys. B [**553**]{}, 3 (1999) \[arXiv:hep-ph/9902371\]; D. A. Demir, Phys. Rev. D [**60**]{}, 055006 (1999) \[arXiv:hep-ph/9901389\]; S. Y. Choi, M. Drees and J. S. Lee, Phys. Lett. B [**481**]{}, 57 (2000) \[arXiv:hep-ph/0002287\]; G. L. Kane and L. T. Wang, Phys. Lett. B [**488**]{}, 383 (2000) \[arXiv:hep-ph/0003198\]; S. Y. Choi, K. Hagiwara and J. S. Lee, Phys. Rev. D [**64**]{}, 032004 (2001) \[arXiv:hep-ph/0103294\]; S. Y. Choi, K. Hagiwara and J. S. Lee, Phys. Lett. B [**529**]{}, 212 (2002) \[arXiv:hep-ph/0110138\]; S. Heinemeyer, Eur. Phys. J. C [**22**]{}, 521 (2001) \[arXiv:hep-ph/0108059\]; T. Ibrahim and P. Nath, Phys. Rev. D [**66**]{}, 015005 (2002) \[arXiv:hep-ph/0204092\]; S. W. Ham, S. K. Oh, E. J. Yoo, C. M. Kim and D. Son, Phys. Rev. D [**68**]{}, 055003 (2003) \[arXiv:hep-ph/0205244\]. J. S. Lee, AIP Conf. Proc. [**1078**]{} (2009) 36 \[arXiv:0808.2014 \[hep-ph\]\].
A. Pukhov, “CalcHEP 3.2: MSSM, structure functions, event generation, batchs, and generation of matrix elements for other packages”, \[arXiv:hep-ph/0412191\]. J. R. Ellis, J. S. Lee and A. Pilaftsis, Mod. Phys. Lett. A [**21**]{}, 1405 (2006) \[arXiv:hep-ph/0605288\]. J. S. Lee, A. Pilaftsis, M. S. Carena, S. Y. Choi, M. Drees, J. R. Ellis and C. E. M. Wagner, Comput. Phys. Commun. [**156**]{}, 283 (2004) \[arXiv:hep-ph/0307377\]. J. S. Lee, M. Carena, J. Ellis, A. Pilaftsis, C. E. M. Wagner,
Comput. Phys. Commun. [**180**]{}, 312-331 (2009). \[arXiv:0712.2360 \[hep-ph\]\].
H. L. Lai [*et al.*]{} \[CTEQ Collaboration\], Eur. Phys. J. C [**12**]{}, 375 (2000) \[arXiv:hep-ph/9903282\]. J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. Nadolsky and W. K. Tung, JHEP [**0207**]{}, 012 (2002) \[arXiv:hep-ph/0201195\]. T. Sjostrand, L. Lonnblad and S. Mrenna, \[arXiv:hep-ph/0108264\]. See “http://hep.pa.msu.edu/people/belyaev/public/calchep/index.html”
P. Skands [*et al.*]{}, JHEP [**0407**]{}, 036 (2004) \[arXiv:hep-ph/0311123\];\
see also $http://home.fnal.gov/\tilde{}skands/slha/$
H. Baer, V. Barger, G. Shaughnessy, H. Summy and L. t. Wang, Phys. Rev. D [**75**]{}, 095010 (2007) \[arXiv:hep-ph/0703289\]. M. Frank, T. Hahn, S. Heinemeyer, W. Hollik, H. Rzehak and G. Weiglein, JHEP [**0702**]{}, 047 (2007) \[arXiv:hep-ph/0611326\]. T. Hahn, S. Heinemeyer, W. Hollik, H. Rzehak, G. Weiglein and K. Williams, \[arXiv:hep-ph/0611373\].
G. L. Bayatian [*et al.*]{} \[ CMS Collaboration \], J. Phys. G [**G34** ]{} (2007) 995-1579.
G. Bagliesi, \[arXiv:0707.0928 \[hep-ex\]\].
[^1]: The central value of $m_t$ has shifted frequently during the years. These shifts change the size of the hole, although the location remains the same. The resulting uncertainty is no bigger than the theoretical uncertainties resulting from the quantum corrections causing the CP violating effects.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $X$ be a separable Banach function space on the unit circle ${\mathbb{T}}$ and $H[X]$ be the abstract Hardy space built upon $X$. We show that the set of analytic polynomials is dense in $H[X]$ if the Hardy-Littlewood maximal operator is bounded on the associate space $X''$. This result is specified to the case of variable Lebesgue spaces.'
address: |
Centro de Matemática e Aplicações,\
Departamento de Matemática,\
Faculdade de Ciências e Tecnologia,\
Universidade Nova de Lisboa,\
Quinta da Torre,\
2829–516 Caparica, Portugal
author:
- 'Alexei Yu. Karlovich'
title: |
Density of Analytic Polynomials\
in Abstract Hardy Spaces
---
[^1]
Introduction
============
For $1\le p\le\infty$, let $L^p:=L^p({\mathbb{T}})$ be the Lebesgue space on the unit circle ${\mathbb{T}}:=\{z\in{\mathbb{C}}:|z|=1\}$ in the complex plane ${\mathbb{C}}$. For $f\in L^1$, let $$\widehat{f}(n):=\frac{1}{2\pi}
\int_{-\pi}^\pi f(e^{i\varphi})e^{-in\varphi}\,d\varphi,
\quad n\in{\mathbb{Z}},$$ be the sequence of the Fourier coefficients of $f$. The classical Hardy spaces $H^p$ are given by $$H^p:=\big\{f\in L^p\ :\ \widehat{f}(n)=0\quad\mbox{for all}\quad n<0\big\}.$$ A function of the form $$q(t)=\sum_{k=0}^n\alpha_k t^k,
\quad
t\in{\mathbb{T}},
\quad
\alpha_0,\dots,\alpha_n\in{\mathbb{C}},$$ is said to be an analytic polynomial on ${\mathbb{T}}$. The set of all analytic polynomials is denoted by ${\mathcal{P}}_A$. It is well known that that the set ${\mathcal{P}}_A$ is dense in $H^p$ whenever $1\le p<\infty$ (see, e.g., [@C91 Chap. III, Corollary 1.7(a)]).
Let $X$ be a Banach space continuously embedded in $L^1$. Following [@Xu92 p. 877], we will consider the abstract Hardy space $H[X]$ built upon the space $X$, which is defined by $$H[X]:=\big\{f\in X:\ \widehat{f}(n)=0\quad\mbox{for all}\quad n<0\big\}.$$ It is clear that if $1\le p\le\infty$, then $H[L^p]$ is the classical Hardy space $H^p$. The aim of this note is to find sufficient conditions for the density of the set ${\mathcal{P}}_A$ in the space $H[X]$ when $X$ falls into the class of so-called Banach function spaces.
We equip ${\mathbb{T}}$ with the normalized Lebesgue measure $dm(t)=|dt|/(2\pi)$. Let $L^0$ be the space of all measurable complex-valued functions on ${\mathbb{T}}$. As usual, we do not distinguish functions, which are equal almost everywhere (for the latter we use the standard abbreviation a.e.). Let $L^0_+$ be the subset of functions in $L^0$ whose values lie in $[0,\infty]$. The characteristic function of a measurable set $E\subset{\mathbb{T}}$ is denoted by $\chi_E$.
Following [@BS88 Chap. 1, Definition 1.1], a mapping $\rho: L_+^0\to [0,\infty]$ is called a Banach function norm if, for all functions $f,g, f_n\in L_+^0$ with $n\in{\mathbb{N}}$, for all constants $a\ge 0$, and for all measurable subsets $E$ of ${\mathbb{T}}$, the following properties hold: $$\begin{aligned}
{\rm (A1)} & &
\rho(f)=0 \Leftrightarrow f=0\ \mbox{a.e.},
\
\rho(af)=a\rho(f),
\
\rho(f+g) \le \rho(f)+\rho(g),\\
{\rm (A2)} & &0\le g \le f \ \mu-\mbox{a.e.} \ \Rightarrow \
\rho(g) \le \rho(f)
\quad\mbox{(the lattice property)},\\
{\rm (A3)} & &0\le f_n \uparrow f \ \mbox{a.e.} \ \Rightarrow \
\rho(f_n) \uparrow \rho(f)\quad\mbox{(the Fatou property)},\\
{\rm (A4)} & & m(E)<\infty\ \Rightarrow\ \rho(\chi_E) <\infty,\\
{\rm (A5)} & &\int_E f(t)\,dm(t) \le C_E\rho(f)\end{aligned}$$ with the constant $C_E \in (0,\infty)$ that may depend on $E$ and $\rho$, but is independent of $f$. When functions differing only on a set of measure zero are identified, the set $X$ of all functions $f\in L^0$ for which $\rho(|f|)<\infty$ is called a Banach function space. For each $f\in X$, the norm of $f$ is defined by $\|f\|_X :=\rho(|f|)$. The set $X$ under the natural linear space operations and under this norm becomes a Banach space (see [@BS88 Chap. 1, Theorems 1.4 and 1.6]). If $\rho$ is a Banach function norm, its associate norm $\rho'$ is defined on $L_+^0$ by $$\rho'(g):=\sup\left\{
\int_{\mathbb{T}}f(t)g(t)\,d\mu(t) \ : \
f\in L_+^0, \ \rho(f) \le 1
\right\}, \ g\in L_+^0.$$ It is a Banach function norm itself [@BS88 Chap. 1, Theorem 2.2]. The Banach function space $X'$ determined by the Banach function norm $\rho'$ is called the associate space (Köthe dual) of $X$. The associate space $X'$ can be viewed a subspace of the (Banach) dual space $X^*$.
The distribution function $m_f$ of an a.e. finite function $f\in L^0$ is defined by $$m_f(\lambda):=m\{t\in {\mathbb{T}}:|f(t)|>\lambda\},\quad\lambda\ge 0.$$ Two a.e. finite functions $f,g\in L^0$ are said to be equimeasurable if $$m_f(\lambda)=m_g(\lambda)
\quad\mbox{for all}\quad \lambda\ge 0.$$ The non-increasing rearrangement of an a.e. finite function $f\in L^0$ is defined by $$f^*(x):=\inf\{\lambda: m_f(\lambda)\le x\},\quad x\ge 0.$$ We refer to [@BS88 Chap. 2, Section 1] and [@KPS82 Chap. II, Section 2] for properties of distribution functions and non-increasing rearrangements. A Banach function space $X$ is called rearrangement-invariant if for every pair of a.e. finite equimeasurable functions $f,g \in L^0$, one has the following property: if $f\in X$, then $g\in X$ and the equality $\|f\|_{X}=\|g\|_{X}$ holds. Lebesgue spaces $L^p$, $1\le p\le\infty$, as well as, more general Orlicz spaces, Lorentz spaces, and Marcinkiewicz spaces are classical examples of rearrangement-invariant Banach function spaces (see [@BS88; @KPS82]). For more recent examples of rearrangement-invariant spaces, like Cesàro, Copson, and Tandori spaces, we refer to the paper of Maligranda and Leśnik [@LM16].
One of our motivations for this work is the recent progress in the study of Harmonic Analysis in the setting of variable Lebesgue spaces [@CF13; @DHHR11; @KMRS16]. Let $\mathfrak{P}({\mathbb{T}})$ be the set of all measurable functions $p: {\mathbb{T}}\to[1,\infty]$. For $p\in\mathfrak{P}({\mathbb{T}})$, put $${\mathbb{T}}_\infty^{p(\cdot)} :=\{t\in {\mathbb{T}}\ :\ p(t)=\infty\}.$$ For a measurable function $f: {\mathbb{T}}\to{\mathbb{C}}$, consider $$\varrho_{p(\cdot)}(f)
:=
\int_{{\mathbb{T}}\setminus {\mathbb{T}}_\infty^{p(\cdot)}}|f(t)|^{p(t)}dm(t)
+\|f\|_{L^\infty({\mathbb{T}}_\infty^{p(\cdot)})}.$$ According to [@CF13 Definition 2.9], the variable Lebesgue space $L^{p(\cdot)}$ is defined as the set of all measurable functions $f:{\mathbb{T}}\to{\mathbb{C}}$ such that $\varrho_{p(\cdot)}(f/\lambda)<\infty$ for some $\lambda>0$. This space is a Banach function space with respect to the Luxemburg-Nakano norm given by $$\|f\|_{L^{p(\cdot)}}:=\inf\{\lambda>0:
\varrho_{p(\cdot)}(f/\lambda)\le 1\}$$ (see, e.g., [@CF13 Theorems 2.17, 2.71 and Section 2.10.3]). If $p\in\mathfrak{P}({\mathbb{T}})$ is constant, then $L^{p(\cdot)}$ is nothing but the standard Lebesgue space $L^p$. If $\in\mathfrak{P}({\mathbb{T}})$ is not constant, then $L^{p(\cdot)}$ is not rearrangement-invariant [@CF13 Example 3.14]. Variable Lebesgue spaces are often called Nakano spaces. We refer to Maligranda’s paper [@M11] for the role of Hidegoro Nakano in the study of variable Lebesgue spaces. The associate space of $L^{p(\cdot)}$ is isomorphic to the space $L^{p'(\cdot)}$, where $p'\in\mathfrak{P}({\mathbb{T}})$ is defined so that $1/p(t)+1/p'(t)=1$ for a.e. $t\in{\mathbb{T}}$ with the usual convention $1/\infty:=0$ [@DHHR11 Theorem 3.2.13]. For $p\in\mathfrak{P}({\mathbb{T}})$, put $$p_-:=\operatornamewithlimits{ess\,inf}_{t\in{\mathbb{T}}}p(t),
\quad
p_+:=\operatornamewithlimits{ess\,sup}_{t\in{\mathbb{T}}}p(t).$$ The space variable Lebesgue space $L^{p(\cdot)}$ is separable if and only if $p_+<\infty$ (see, e.g., [@CF13 Theorem 2.78]).
The following result is a kind of folklore.
\[th:density-analytic-polynomials-RI\] Let $X$ be a separable rearrangement-invariant Banach function space on ${\mathbb{T}}$. Then the set of analytic polynomials ${\mathcal{P}}_A$ is dense in the abstract Hardy space $H[X]$. Moreover, for every $f\in H[X]$, there is a sequence of analytic polynomials $\{p_n\}$ such that $\|p_n\|_X\le\|f\|_X$ for all $n\in{\mathbb{N}}$ and $p_n\to f$ in the norm of $X$ as $n\to\infty$.
Surprisingly enough, we could not find in the literature neither Theorem \[th:density-analytic-polynomials-RI\] explicitly stated nor any result on the density of ${\mathcal{P}}_A$ in abstract Hardy spaces $H[X]$ in the case when $X$ is an arbitrary Banach function space beyond the class of rearrangement-invariant spaces. The aim of this note is to fill in this gap.
Given $f\in L^1$, the Hardy-Littlewood maximal function is defined by $$(Mf)(t):=\sup_{I\ni t}\frac{1}{m(I)}\int_I|f(\tau)|\,dm(\tau),
\quad t\in{\mathbb{T}},$$ where the supremum is taken over all arcs $I\subset{\mathbb{T}}$ containing $t\in{\mathbb{T}}$. The operator $f\mapsto Mf$ is called the Hardy-Littlewood maximal operator.
\[th:density-analytic-polynomials-BFS\] Suppose $X$ is a separable Banach function space on ${\mathbb{T}}$. If the Hardy-Littlewood maximal operator $M$ is bounded on the associate space $X'$, then the set of analytic polynomials ${\mathcal{P}}_A$ is dense in the abstract Hardy space $H[X]$.
To illustrate this result in the case of variable Lebesgue spaces, we will need the following classes of variable exponents. Following [@CF13 Definition 2.2], one says that $r:{\mathbb{T}}\to{\mathbb{R}}$ is locally log-Hölder continuous if there exists a constant $C_0>0$ such that $$|r(x)-r(y)|=C_0/(-\log|x-y|))\quad\mbox{for all}\quad x,y\in{\mathbb{T}},\quad |x-y|<1/2.$$ The class of all locally log-Hölder continuous functions is denoted by $LH_0({\mathbb{T}})$. If $p_+<\infty$, then $p\in LH_0({\mathbb{T}})$ if and only if $1/p\in LH_0({\mathbb{T}})$. By [@CF13 Theorem 4.7], if $p\in\mathfrak{P}({\mathbb{T}})$ is such that $1<p_-$ and $1/p\in LH_0({\mathbb{T}})$, then the Hardy-Littlewood maximal operator $M$ is bounded on $L^{p(\cdot)}$. This condition was initially referred to as “almost necessary" (see [@CF13 Section 4.6.1] for further details). However, Lerner [@L05] constructed an example of discontinuous variable exponent such that the Hardy-Littlewood maximal operator is bounded on $L^{p(\cdot)}$.
Kapanadze and Kopaliani [@KK08] developed further Lerner’s ideas. They considered the following class of variable exponents. Recall that a function $f\in L^1$ belongs to the space $BMO$ if $$\|f\|_*:=\sup_{I\subset{\mathbb{T}}}\frac{1}{m(I)}\int_I|f(t)-f_I|\,dm(t)<\infty,$$ where $f_I$ is the integral average of $f$ on the arc $I$ and the supremum is taken over all arcs $I\subset{\mathbb{T}}$. For $f\in BMO$, put $$\gamma(f,r):=\sup_{m(I)\le r}\frac{1}{m(I)}\int_I|f(t)-f_I|\,dm(t).$$ Let $VMO^{1/|\log|}$ be the set of functions $f\in BMO$ such that $$\gamma(f,r)=o(1/|\log r|)\quad\mbox{ as }\quad r\to 0.$$ Note that $VMO^{1/|\log|}$ contains discontinuous functions. We will say that $p\in\mathfrak{P}({\mathbb{T}})$ belongs to the Kapanadze-Kopaliani class $\mathfrak{K}({\mathbb{T}})$ if $1<p_-\le p_+<\infty$ and $p\in VMO^{1/|\log|}$. It is shown in [@KK08 Theorem 2.1] that if $p\in\mathfrak{K}({\mathbb{T}})$, then the Hardy-Littelwood maximal operator $M$ is bounded on the variable Lebesgue space $L^{p(\cdot)}$.
Suppose $p\in\mathfrak{P}({\mathbb{T}})$. If $p_+<\infty$ and $p\in LH_0({\mathbb{T}})$ or if $p'\in\mathfrak{K}({\mathbb{T}})$, then the set of analytic polynomials ${\mathcal{P}}_A$ is dense in the abstract Hardy space $H[L^{p(\cdot)}]$ built upon the variable Lebesgue space $L^{p(\cdot)}$.
The paper is organized as follows. In Section \[sec:preliminaries\], we prove that the separability of a Banach function space $X$ is equivalent to the density of the set of trigonometric polynomials ${\mathcal{P}}$ in $X$ and to the density of the set of all continuous functions $C$ in $X$. Further, we recall a pointwise estimate of the Fejér means $f*K_n$, where $K_n$ is the $n$-th Fejér kernel, by the Hardy-Littlewood maximal function $Mf$. In Section \[sec:proofs\] we show that the norms of the operators $F_nf=f*K_n$ are uniformly bounded on a Banach function space $X$ if $X$ is rearrangement-invariant or if the Hardy-Littlewood maximal operator is bounded on $X'$. Moreover, if $X$ is rearrangement-invariant, then $\|F_n\|_{{\mathcal{B}}(X)}\le 1$ for all $n\in{\mathbb{N}}$. Further, we prove that under the assumptions of Theorem \[th:density-analytic-polynomials-RI\] or \[th:density-analytic-polynomials-BFS\], $\|f*K_n-f\|_X\to 0$ as $n\to\infty$. It remains to observe that $f*K_n\in{\mathcal{P}}_A$ if $f\in H[X]$, which will complete the proof of Theorems \[th:density-analytic-polynomials-RI\] and \[th:density-analytic-polynomials-BFS\].
Preliminaries {#sec:preliminaries}
=============
Elementary lemma
----------------
We start with the following elementary lemma, whose proof can be found, e.g., in [@C91 Chap. III, Proposition 1.6(a)]. Here and in what follows, the space of all bounded linear operators on a Banach space $E$ will be denoted by ${\mathcal{B}}(E)$.
\[le:elementary\] Le $E$ be a Banach space and $\{T_n\}$ be a sequence of bounded operators on $E$ such that $$\sup_{n\in{\mathbb{N}}}\|T_n\|_{{\mathcal{B}}(E)}<\infty.$$ If $D$ is a dense subset of $E$ and for all $x\in D$, $$\label{eq:elementary}
\|T_n x-x\|_E\to 0
\quad\mbox{as}\quad n\to\infty,$$ then holds for all $x\in E$.
Density of continuous function and trigonometric polynomials in Banach function spaces
--------------------------------------------------------------------------------------
A function of the form $$q(t)=\sum_{k=-n}^n\alpha_k t^k,
\quad
t\in{\mathbb{T}},
\quad
\alpha_{-n},\dots,\alpha_n\in{\mathbb{C}},$$ is said to be a trigonometric (or Laurent) polynomial on ${\mathbb{T}}$. The set of all trigonometric polynomials is denoted by ${\mathcal{P}}$.
\[le:density-polynomials\] Let $X$ be a Banach function space on ${\mathbb{T}}$. The following statements are equivalent:
1. the set ${\mathcal{P}}$ of all trigonometric polynomials is dense in $X$;
2. the space $C$ of all continuous functions on ${\mathbb{T}}$ is dense in $X$;
3. the Banach function space $X$ is separable.
The proof is developed by analogy with [@K00 Lemma 1.3].
\(a) $\Rightarrow$ (b) is trivial because ${\mathcal{P}}\subset C\subset X$.
\(b) $\Rightarrow$ (c). Since $C$ is separable and $C\subset X$ is dense in $X$, we conclude that $X$ is separable.
\(c) $\Rightarrow$ (a). Assume that $X$ is separable and ${\mathcal{P}}$ is not dense in $X$. Then by the corollary of the Hahn-Banach theorem (see, e.g., [@BSU96 Chap. 7, Theorem 4.2]), there exists a nonzero functional $\Lambda\in X^*$ such that $\Lambda(p)=0$ for all $p\in{\mathcal{P}}$. Since $X$ is separable, from [@BS88 Chap. 1, Corollaries 4.3 and 5.6] it follows that the Banach dual $X^*$ of $X$ is canonically isometrically isomorphic to the associate space $X'$. Hence there exists a nonzero function $h\in X'\subset L^1$ such that $$\int_{\mathbb{T}}p(t)h(t)\,dm(t)=0
\quad\mbox{for all}\quad p\in{\mathcal{P}}.$$ Taking $p(t)=t^n$ for $n\in{\mathbb{Z}}$, we obtain that all Fourier coefficients of $h\in L^1$ vanish, which implies that $h=0$ a.e. on ${\mathbb{T}}$ by the uniqueness theorem of the Fourier series (see, e.g., [@Kat76 Chap. I, Theorem 2.7]). This contradiction proves that ${\mathcal{P}}$ is dense in $X$.
Pointwise estimate for the Fejér means
--------------------------------------
Recall that $L^1$ is a commutative Banach algebra under the convolution multiplication defined for $f,g\in L^1$ by $$(f* g)(e^{i\theta})=\frac{1}{2\pi}\int_{-\pi}^\pi
f(e^{i\theta-i\varphi})g(e^{i\varphi})\,d\varphi,
\quad
e^{i\theta}\in{\mathbb{T}}.$$ For $n\in{\mathbb{N}}$, let $$K_n(e^{i\theta}):=\sum_{k=-n}^n\left(1-\frac{|k|}{n+1}\right)e^{i\theta k}
=
\frac{1}{n+1}\left(
\frac{\sin\frac{n+1}{2}\theta}{\sin\frac{\theta}{2}}
\right)^2,
\quad e^{i\theta}\in{\mathbb{T}},$$ be the $n$-th Fejér kernel. It is well-known that $\|K_n\|_{L^1}\le 1$. For $f\in L^1$, the $n$-th Fejér mean of $f$ is defined as the convolution $f*K_n$. Then $$\label{eq:Fejer-mean}
(f*K_n)(e^{i\theta})=
\sum_{k=-n}^n\widehat{f}(k)\left(1-\frac{|k|}{n+1}\right)e^{i\theta k},
\quad e^{i\theta}\in{\mathbb{T}}$$ (see, e.g., [@Kat76 Chap. I]). This means that if $f\in L^1$, then $f* K_n\in\mathcal{P}$. Moreover, if $f\in H^1=H[L^1]$, then $f* K_n\in\mathcal{P}_A$.
\[le:Fejer-pointwise\] For every $f\in L^1$ and $t\in{\mathbb{T}}$, $$\label{eq:Fejer-pointwise-1}
\sup_{n\in{\mathbb{N}}}|(f*K_n)(t)|\le\frac{\pi^2}{2}(Mf)(t).$$
Since $|\sin\varphi|\ge 2|\varphi|/\pi$ for $|\varphi|\le\pi/2$, we have for $\theta\in[-\pi,\pi]$, $$\begin{aligned}
K_n(e^{i\theta})
&\le
\frac{\pi^2}{n+1}\frac{\sin^2\left(\frac{n+1}{2}\theta\right)}{\theta^2}
\nonumber\\
&=
\frac{\pi^2}{4}(n+1)
\frac{\sin^2\left(\frac{n+1}{2}\theta\right)}
{\left(\frac{n+1}{2}\theta\right)^2}
\nonumber\\
&\le
\frac{\pi^2}{4}(n+1)\min\left\{1,\left(\frac{n+1}{2}\theta\right)^{-2}\right\}
\nonumber\\
&\le
\frac{\pi^2}{2}\frac{n+1}{1+\left(\frac{n+1}{2}\theta\right)^2}
=:\Psi_n(\theta).
\label{eq:Fejer-pointwise-2}\end{aligned}$$ It is easy to see that $$\label{eq:Fejer-pointwise-3}
\frac{1}{2\pi}\int_{-\pi}^\pi\Psi_n(\theta)\,d\theta\le\frac{\pi^2}{2}
\quad\mbox{for all}\quad n\in{\mathbb{N}}.$$ From [@MS13 Lemma 2.11] and estimates – we immediately get estimate .
Proofs of the main results {#sec:proofs}
==========================
Norm estimates for the Fejér means
----------------------------------
First we consider the case of rearrangement-invariant Banach function spaces.
\[le:uniform-boundednes-Fejer-means-RI\] Let $X$ be a rearrangement-invariant Banach function space on ${\mathbb{T}}$. Then for each $n\in{\mathbb{N}}$, the operator $F_nf=f*K_n$ is bounded on $X$ and $$\sup_{n\in{\mathbb{N}}}\|F_n\|_{{\mathcal{B}}(X)}\le 1.$$
By [@BS88 Chap. 3, Lemma 6.1], for every $f\in X$ and every $n\in{\mathbb{N}}$, $$\|f*K_n\|_X\le \|K_n\|_{L^1}\|f\|_X.$$ It remains to recall that $\|K_n\|_{L_1}\le 1$ for all $n\in{\mathbb{N}}$.
Now we will show the corresponding results for Banach function spaces such that the Hardy-Littlewood maximal operator is bounded on $X'$.
\[th:uniform-boundednes-Fejer-means-BFS\] Let $X$ be a Banach function space on ${\mathbb{T}}$ such that the Hardy-Littlewood maximal operator $M$ is bounded on its associate space $X'$. Then for each $n\in{\mathbb{N}}$, the operator $F_nf=f*K_n$ is bounded on $X$ and $$\sup_{n\in{\mathbb{N}}}\|F_n\|_{{\mathcal{B}}(X)}\le\pi^2\|M\|_{X'\to X'}.$$
The idea of the proof is borrowed from the proof of [@CF13 Theorem 5.1]. Fix $f\in X$ and $n\in{\mathbb{N}}$. Since $K_n\ge 0$, we have $|f*K_n|\le|f|*K_n$. Then from the Lorentz-Luxemburg theorem (see, e.g., [@BS88 Chap. 1, Theorem 2.7]) we deduce that $$\begin{aligned}
\|f*K_n\|_X
&\le
\|\,|f|*K_n\|_X=\|\,|f|*K_n\|_{X''}
\\
&=
\sup\left\{
\int_{\mathbb{T}}(|f|*K_n)(t)|g(t)|\,dm(t)\ : \ g\in X',\ \|g\|_{X'}\le 1
\right\}.\end{aligned}$$ Hence there exists a function $h\in X'$ such that $h\ge 0$, $\|h\|_{X'}\le 1$, and $$\label{eq:uniform-boundednes-Fejer-means-BFS-1}
\|f*K_n\|_X\le 2\int_{\mathbb{T}}(|f|*K_n)(t)h(t)\,dm(t).$$ Taking into account that $K_n(e^{i\theta})=K_n(e^{-i\theta})$ for all $\theta\in{\mathbb{R}}$, by Fubini’s theorem, we get $$\int_{\mathbb{T}}(|f|*K_n)(t)h(t)\,dm(t)=\int_{\mathbb{T}}(h*K_n)(t)|f(t)|\,dm(t).$$ From this identity and Hölder’s inequality for $X$ (see, e.g., [@BS88 Chap. 1, Theorem 2.4]), we obtain $$\label{eq:uniform-boundednes-Fejer-means-BFS-2}
\int_{\mathbb{T}}(|f|*K_n)(t)h(t)\,dm(t)\le\|f\|_X\|h*K_n\|_{X'}.$$ Applying Lemma \[le:Fejer-pointwise\] to $h\in X'\subset L^1$, by the lattice property, we see that $$\label{eq:uniform-boundednes-Fejer-means-BFS-3}
\|h*K_n\|_{X'}\le\frac{\pi^2}{2}\|Mh\|_{X'}.$$ Combining estimates – and taking into account that $M$ is bounded on $X'$ and that $\|h\|_{X'}\le 1$, we arrive at $$\|f*K_n\|_{X}\le \pi^2\|M\|_{X'\to X'}\|f\|_X.$$ Hence $$\sup_{n\in{\mathbb{N}}}\|F_n\|_{{\mathcal{B}}(X)}
=
\sup_{n\in{\mathbb{N}}}\sup_{f\in X\setminus\{0\}}
\frac{\|f*K_n\|_X}{\|f\|_X}\le \pi^2\|M\|_{X'\to X'}<\infty,$$ which completes the proof.
Convergence of the Fejér means in the norm
------------------------------------------
The following statement is the heart of the proof of the main results.
\[th:Fejer-norm-convergence\] Suppose $X$ is a separable Banach function space on ${\mathbb{T}}$. If $X$ is rearrangement-invariant or the Hardy-Littlewood maximal operator is bounded on the associate space $X'$, then for every $f\in X$, $$\label{eq:Fejer-norm-convergence}
\lim_{n\to\infty}\|f*K_n-f\|_X=0.$$
It is well-known that for every $f\in C$, $$\lim_{n\to\infty}\|f*K_n-f\|_C=0$$ (see, e.g., [@C91 Chap. III, Theorem 1.1(a)] or [@Kat76 Theorem 2.11]). From the definition of the Banach function space $X$ it follows that $C\subset X\subset L^1$, where both embeddings are continuous. Then, for every $f\in C$, is fulfilled. From Lemma \[le:density-polynomials\] we know that the set $C$ is dense in the space $X$. By Lemma \[le:uniform-boundednes-Fejer-means-RI\] and Theorem \[th:uniform-boundednes-Fejer-means-BFS\], $$\sup_{n\in{\mathbb{N}}}\|F_n\|_{{\mathcal{B}}(X)}<\infty,$$ where $F_nf= f*K_n$. It remains to apply Lemma \[le:elementary\].
This statement for rearrangement-invariant Banach function spaces is contained, e.g., in [@DVL93 p. 268]. Notice that the assumption of the separability of $X$ is hidden there.
Now we formulate the corollary of the above theorem in the case of variable Lebesgue spaces.
Suppose $p\in\mathfrak{P}({\mathbb{T}})$. If $p_+<\infty$ and $p\in LH_0({\mathbb{T}})$ or if $p'\in\mathfrak{K}({\mathbb{T}})$, then for every $f\in L^{p(\cdot)}$, $$\lim_{n\to\infty}\|f*K_n-f\|_{L^{p(\cdot)}}=0.$$
For variable exponents $p\in\mathfrak{P}({\mathbb{T}})$ satisfying $p_+<\infty$ and $p\in LH_0({\mathbb{T}})$, this result was obtained by Sharapudinov [@Sh96 Section 3.1]. For $p\in\mathfrak{K}({\mathbb{T}})$, the above corollary is new.
Proofs of Theorems \[th:density-analytic-polynomials-RI\] and \[th:density-analytic-polynomials-BFS\]
------------------------------------------------------------------------------------------------------
If $f\in H[X]$, then $p_n=f*K_n\in{\mathcal{P}}_A$ for all $n\in{\mathbb{N}}$ in view of . By Theorem \[th:Fejer-norm-convergence\], $\|p_n-f\|_X\to 0$ as $n\to\infty$. Thus the set ${\mathcal{P}}_A$ is dense in in the abstract Hardy space $H[X]$ built upon $X$.
Moreover, if $X$ is a rearrangement-invariant Banach function space, then from Lemma \[le:uniform-boundednes-Fejer-means-RI\] it follows that $\|p_n\|_X\le\|f\|_X$ for all $n\in{\mathbb{N}}$.
[XX]{} C. Bennett and R. Sharpley, *Interpolation of Operators*, Academic Press, Boston, 1988.
Yu. M. Berezansky, Z. G. Sheftel, and G. F. Us, *Functional Analysis, Vol. 1*, Birkhäuser, Basel, 1996.
J. B. Conway, *The Theory of Subnormal Operators*, American Mathematical Society, Providence, RI, 1991.
D. Cruz-Uribe and A. Fiorenza, *Variable Lebesgue Spaces*, Birkhäuser, Basel, 2013.
R. A. DeVore and G. G. Lorentz, *Constructive Approximation*, Springer, Berlin, 1993.
L. Diening, P. Harjulehto, P. Hästö, and M. Ržička, *Lebesgue and Sobolev Spaces with Variable Exponents*, Springer, Berlin, 2011.
E. Kapanadze and T. Kopaliani, *A note on maximal operator on $L^{p(t)}(\Omega)$ spaces*, Georgian Math. J. **15** (2008), 307–316.
A. Yu. Karlovich, *On the essential norm of the Cauchy singular integral operator in weighted rearrangement-invariant spaces*, Integr. Equ. Oper. Theory **38** (2000), 28–50.
Y. Katznelson, *An Introduction to Harmonic Analysis*, Dower Publications, Inc., New York, 1976.
V. Kokilashvili, A. Meskhi, H. Rafeiro, and S. Samko, *Integral Operators in Non-Standard Function Spaces. Volume 1: Variable Exponent Lebesgue and Amalgam Spaces*. Birkhäuser Verlag, Basel, 2016.
S. G. Krein, Yu. I. Petunin, and E. M. Semenov, *Interpolation of Linear Operators*, Amer. Math. Soc., Providence, 1982.
A. Lerner, *Some remarks on the Hardy-Littlewood maximal function on variable $L^p$ spaces*, Math. Z. **251** (2005), 509–521.
K. Leśnik and L. Maligranda, *Interpolation of abstract Cesàro, Copson and Tandori spaces*, Indag. Math., New Ser. **27** (2016), 764–785.
L. Maligranda, *Hidegoro Nakano (1909-1974) – on the centenary of his birth*, M. Kato, (ed.) et al., Proceedings of the 3rd international symposium on Banach and function spaces (ISBFS 2009), Kitakyushu, Japan, September 14–17, 2009. Yokohama, Yokohama Publishers, pp. 99–171, 2011.
C. Muscalu and W. Schlag, *Classical and Multilinear Harmonic Analysis. Vol. I.* Cambridge University Press, Cambridge, 2013.
I. I. Sharapudinov, *Uniform boundedness in $L^p$ ($p=p(x)$) of some families of convolution operators*, Math. Notes **59** (1996), 205–212.
Q. Xu, *Notes on interpolation of Hardy spaces*, Ann. Inst. Fourier **42** (1992), 875–889.
[^1]: This work was supported by the Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the project UID/MAT/00297/2013 (Centro de Matemática e Aplicações).
| {
"pile_set_name": "ArXiv"
} |
by 0.14in by +0.35in
Ultracold atoms trapped in optical lattices is an active area of research in many-body physics [@jaksch; @stringari; @pedri; @pethick; @moritz; @shin]. At zero temperature and without disorder, bosons in a lattice with integer filling exhibit two distinct phases – a superfluid phase, where the phase of the wavefunctions are sharply defined on sites, and a Mott insulating phase, where the site occupation numbers are sharply defined [@fisher; @frey; @fisher2; @greiner; @stoferley]. This system allows novel many-body phenomena associated with unusual correlated states and quantum phase transitions [@recati; @buchler; @kollath; @demler; @lieb; @giamarchi; @kuklov; @zwerger].
In this work we will study a system of bosons in a one-dimensional (1D) optical lattice that is tunnel coupled to a three-dimensional (3D) Bose-condensed particle reservoir. Our goal is to understand how the reservoir impacts the states of the bosons in the optical lattice. We find that the system supports three phases: a 3D superfluid phase in which the 1D system becomes phase-locked with the reservoir; a decoupled phase (in which, under circumstances described below, the 1D system behaves like a metal); and a Mott insulating phase. These states are separated by deconfinement transitions [@fertig], either of vortices or of tunneling events, as we describe below. The states may be distinquished both by their low-energy excitation spectra, and by their conduction properties.
[*Model:*]{} Our analysis begins with a Hamiltonian for lattice bosons (in number-phase representation) [@fisher] $$\begin{aligned}
{\cal H} &=& -t\sum_{\langle ij \rangle, \tau} \cos(\phi_{i\tau} - \phi_{j\tau}) + \frac {U}{2}\sum_{i,\tau} n_{i\tau}^2 \cr
&-&t_R\sum_{\langle {\bf R}{\bf R'} \rangle, \tau} \cos(\Phi_{{\bf R}\tau} - \Phi_{{\bf R'} \tau}) + \frac
{U_R}{2}\sum_{{\bf R},\tau} N_{{\bf R}\tau}^2 \cr
&-&\sum_{i,\tau} J_i\cos (\phi_{i\tau} -\Phi_{{\bf R}(i) \tau}),
\label{ham}\end{aligned}$$ where $(n_{i\tau},\phi_{i\tau})$ and $(N_{{\bf R}\tau}, \Phi_{{\bf R}\tau})$ are the canonically conjugate occupation number fluctuations and phases of the bosons in the chain and in the 3D reservoir, respectively, $t$ and $U>0$ ($t_R$ and $U_R>0$) are the nearest neighbor interchain hopping and on-site repulsion terms for the 1D (3D) system, and $J_i$ is a tunneling amplitude between a site on the chain $[i]$ and a site in the reservoir $[{\bf R}(i)]$. While the particle numbers in the real optical lattices are typically small, we nevertheless allow the fluctuations to vary from $-\infty$ to $+\infty$. This simplifies the calculation, and should not affect the allowed phases because the symmetries of the resulting Hamiltonian are unchanged [@cardy]. Our geometry could be realized as a linear array of sites using red detuned light focused near the edge of a BEC cloud [@recati; @dumke; @schlosser; @kolomeisky]. Note that the tight-binding form for the reservoir is adopted purely as a matter of convenience.
Following standard procedure [@negele], we construct a path-integral representation of the partition function and then use the Villain model for the three cosine terms in Eq. \[ham\] [@herbut]. Assuming that vortex rings in the 3D reservoir are unimportant (which will always be valid if the reservoir is sufficiently dilute), we integrate out the internal degrees of freedom of the reservoir to arrive at a partition function of the form $$\begin{aligned}
{\cal Z}_{VM}&=&
\sum_{m_{x\tau},n_{x\tau}}\exp\Bigl\{-\frac 1{2} \sum_{x,\tau}
\Bigl[\varepsilon U |n_{x\tau}|^2
+ \frac 1{\varepsilon t} |m_{x\tau}|^2 \Bigr] \cr
&-& \frac 1{2L_x\beta}\sum_{q_x,\omega_n} \frac 1{h({\bf q})}\Bigl|iq_xm({\bf q})+i\omega_n n({\bf q}) \Bigr|^2\Bigr \}.
\label{ZVM}\end{aligned}$$ Here $\beta = (kT)^{-1} \rightarrow \infty$, $L_x$ is the number of sites in the 1D chain, and $\varepsilon$ is the time slice interval. Physically, the integer variables $m_{x\tau}$ may be understood as bond currents and $n_{x\tau}$ as the fluctuations in the site occupation number. Eq. \[ZVM\] is a form of the Bose-Hubbard model. The coefficient $1/h({\bf q})$ contains information about the reservoir degrees of freedom, in particular the gapless collective mode it supports due to its own superfluidity, $$\frac 1{h({\bf q})} = \frac 1{\varepsilon J} + \gamma \ln \Bigl(1 + \frac{\Lambda^2 c_R^2}{c_R^2 q_x^2 +
\omega_n^2}\Bigr),
\label{heqn}$$ where $c_R = \sqrt{\varepsilon^2 t_RU_R}$, $\Lambda $ is the momentum cut-off, and $\gamma =(1/4\pi \varepsilon t_R)$ [@log].
The partition function ${\cal Z}_{VM} \equiv \sum_{m,n} \exp(-{\cal H}_{VM})$ may be reexpressed in terms of another pair of integer fields $\phi(x,\tau)$ and $A(x,\tau)$ with $m(x,\tau) = -\partial_\tau \phi (x,\tau)$ and $n(x,\tau) = \partial_x \phi (x,\tau) + A(x,t)$, so that $$\begin{aligned}
&&{\cal H}_{VM} = \frac 1{2K}\sum_{x,\tau} \Bigl\{c^2\Bigl|\partial_x \phi(x,\tau)+A(x,\tau)\Bigr|^2 \cr
&+& \Bigl|\partial_\tau \phi(x,\tau)\Bigr|^2 \Bigr\}
+ \frac 1{2L_x\beta}\sum_{q_x,\omega_n} \frac {1}{h({\bf q})}\Bigl|i\omega_n A({\bf q})\Bigr|^2.
\label{HamVM2}\end{aligned}$$ Here, $K = \varepsilon t$ and $c=\sqrt{\varepsilon^2 Ut}$. Configurations for which $\vec{\nabla}\phi \ne 0$ and $A=0$ contain closed loops which may be understood as worldlines of particle-hole pairs that separate and recombine. The $A$ field, which should be regarded as residing on the time interval links, can cancel the gradient energy $(\partial_{x}\phi)^2$ on part of a closed loop configuration to form individual particle or hole trajectories; the endpoints (occuring where $\partial_{\tau}A(x,\tau) \ne 0$) represent tunneling events between the 3D and 1D systems. Tunneling events may be shown [@fertig] to be dual to the vortices. Moreover, the model defined by Eq. \[HamVM2\] may be generalized to include a core energy $E_c$ for the vortices.
As we discuss below, this generalized Hamiltonian supports three phases. When tunneling events proliferate through the system, the 1D and 3D systems exchange particles freely (equivalently, vortices of the 1+1 dimensional system are linearly confined), and their phases become locked together to form a single superfluid. As fluctuations in $\phi$ and $A$ are decreased (by decreasing $E_c$ and/or $h$, or decreasing $K$), tunneling events bind into pairs which conserve the overall particle number in the 1D system, although closed particle-hole worldlines remain proliferated. This state may be understood as one in which Josephson vortices between the 3D and 1D system proliferate due to quantum fluctuations, effectively decoupling the two systems. For still smaller fluctuations, particle-hole worldline loops of arbitrarily large size become irrelevant, indicating that the fluctuations in the particle number on each site have been suppressed, and the system is a Mott insulator.
[*RG analysis:*]{} A method for constructing a momentum shell RG for Hamiltonians such as Eq. \[HamVM2\] was developed in Ref. . We replace the integer fields $\phi$ and $A$ with the continuous fields $\varphi(x,\tau)$ and $a(x,\tau)$ in Eq. \[HamVM2\] and add terms of the form $-y\int dxd\tau \cos[2\pi\varphi(x, \tau)]$ and $- y_a \int dxd\tau \cos[2\pi a(x,\tau)]$ so that the resulting effective Hamiltonian has the same symmetries as the original one [@cardy]. In this replacement, $y=\exp(-E_c)$ is the usual vortex fugacity, and choosing $y_a^2 \sim \int d^2q 1/h(q)$ approximately reproduces the action associated with a worldline endpoint (i.e., a tunneling event). We then integrate out short wavelength degrees of freedom \[$\Lambda/b <|q_x|,|w_n|/c < \Lambda$ with $b=\exp(l)$\] to lowest order in $y$ and $y_a$, and rescale lengths, times, and fields according to $x=bx', \tau = b\tau', \varphi'(x',\tau') = \varphi(x,\tau)$, and $a(x,\tau)=a'(x', \tau')/b$. This choice preserves the terms that are lowest order in gradients in the quadratic part of the Hamiltonian. Because the $a$ field shrinks upon rescaling, it is natural to expand $\cos[2\pi a(x,\tau)]$ in its argument, producing a quadratic term of the form $\frac 1{2}\rho |a(x,\tau)|^2$ that contributes to the fixed point. The higher order vertices generated by this expansion contribute to the renormalization of $\rho$ but are themselves irrelevant. Note the initial value of $\rho$ is $4\pi^2 y_a$.
The fixed points that emerge from this procedure have the form $$\begin{aligned}
{\cal H}_{*} = \frac 1{2K}\int\!dx d\tau &\Bigl[& c^2\Bigl|\partial_x \varphi(x,\tau)+a(x,\tau)\Bigr|^2 \cr
&+&
\Bigl|\partial_\tau \varphi(x,\tau)\Bigr|^2
+ \rho K |a(x,\tau)|^2 \Bigr].
\label{Hfixed}\end{aligned}$$ The last term in ${\cal H}_{*}$ is very important: since tunneling events are specified by $\partial_{\tau}a \ne 0$, when $\rho \ne 0$ they are bound into equal and opposite pairs. Unbinding occurs if $\rho(\ell)$ scales to zero. To lowest order in $y_a$, its scaling relation is $${{d \ln \rho} \over {d\ell}} =
-\Bigl(\frac{2\pi \Lambda^2 {K}}{\sqrt{\rho K}}\Bigr)e^{-2\ell} .
\label{rhoflow}$$ For small $\rho$ one can easily show that the term $-y\int dxd\tau \cos[2\pi\varphi(x, \tau)]$ is strongly irrelevant. The resulting scaling flows are shown in Fig. \[phase\]. As may be seen, in general $\rho$ scales to a point along a fixed line; if its initial value is small enough (as occurs for small $\gamma$ and large $J$), that point is at $\rho=0$. A remarkable feature of Eq. \[rhoflow\] for such flows is that $\rho(\ell)=0$ at a [*finite*]{} value of $\ell=\ell^*$, because of the singularity as $\rho \rightarrow 0$. If the initial value of $\rho$ is increased, $\ell^*$ eventually diverges, defining a transition point above which $\rho$ scales to a non-zero value. [*This represents a confinement transition for tunneling events*]{} [@fertig; @com]. The dual representation of this system is obtained by performing a Poisson resummation on $A$ and $\phi$ in Eq. \[HamVM2\]. The resulting Hamiltonian is very similar in form to Eq. \[HamVM2\] [@nelsonbook], and may be understood as a model with vortex degrees of freedom rather than tunneling events. Because of this, there is a second transition, in which vortices go from a bound to an unbound state, in the same continuous fashion as seen above for tunneling events. In terms of $\varphi$ and $a$ degrees of freedom, this corresponds to a phase in which arbitrarily large particle-hole loops are negligible in the partition function. Pairs of tunneling events are then linearly confined [@fertig].
The phases described above have physically different characteristics. This can be seen most directly in the collective mode spectra, which in principle may be measured in light scattering experiments [@unpub; @scattering]. We find these by examining the density-density correlation function $\langle n(-q_x,-\omega)n(q_x,\omega) \rangle $ near the fixed point Hamiltonians representing the various phases. Generally, this contains a broad response for $\omega>c_R q_x$ due to modes in the reservoir. Beyond this, in the superfuid phase \[Fig. \[mode\](a)\], we find [@unpub] a sharp resonance (collective mode) at $\omega_1 = c_R q_x - \delta \omega$, where $\delta \omega = c_R \Lambda \exp\{ -(c^2/\gamma K)/(q_x^2[c_R^2-c^2]) \}$. This rapidly approaches the reservoir continuum as $q_x \rightarrow 0$, indicating that number fluctuations in the 1D system will strongly mix with those of the reservoir in the long wavelength limit. This is consistent with our interpretation of this phase as a single 3D superfluid. (Note that although the term containing the logarithmic singularity is irrelevant in the RG sense, it nevertheless has the physical effect of “pulling” the collective mode very close to the edge of the reservoir modes at small enough wavevectors.) In the intermediate phase we find two collective modes, the superfluid mode of the reservoir at $\omega_1$ and another linear mode at $\omega_2 =
cq_x \sqrt{\rho K/(\rho K + c^2)}$ \[Fig. \[mode\](b)\]. For an appropriate geometry we will show this leads to [*metallic*]{} behavior. In the Mott insulator phase, the reservoir becomes decoupled from the 1D chain in the long-wavelength limit, and supports a gapped mode at $\omega_3 = \sqrt{c^2q_x^2 + 2\pi^2 K/E_c}$ \[Fig. \[mode\](c)\].
[*Conductance:*]{} The different characters of the phases may also be seen in the conductance [@schultz] of the system. To measure this, one has to attach a particle source and lead as shown in Fig. \[JJgeometry\]. The conductance quantifies the current injected by this source with chemical potential $\mu_s>0$, draining into the 3D reservoir which is held at zero chemical potential. (Note that a link has been removed to ensure that the current flows in the 1D chain before tunneling into the reservoir.) This system could be fabricated in a Josephson-junction array [@fisher].
The source and lead introduce two extra degrees of freedom $[\phi_L(\tau), N_L(\tau)]$ to our model. We treat the particle source as an ideal reservoir with Hamiltonian $H_L=\mu_s N_L$; current conservation at the site $(x=0)$ where current is injected specifies $N_L$ in terms of the other variables through a constraint in the partition function, $ \Pi_{\tau}\lbrace \partial_x m(x=0,\tau)+\partial_\tau n(x=0,\tau)
- dN_L(\tau)/d\tau = 0 \rbrace$. The conductance is given by $G(\omega) =
\omega \langle N_L(-\omega) N_L(\omega) \rangle$.
The analysis proceeds in a fashion similar to what is described above for the uniform chain. However, because of the lead there is an additional term in our effective Hamiltonian of the form $-\delta y \int d\tau \cos [2\pi a(x=0,\tau)]$, reflecting the fact that the rate of tunneling events at the $x=0$ site is different than at other sites. The scaling relation for $\delta y(\ell)$ takes the form [@unpub] $$\frac {d \delta y (\ell)}{d\ell} \approx \delta y (\ell)
\Bigr[1 - \frac {\pi^2 K \Lambda^2 [\alpha (\ell)]^2}{\sqrt{\rho K}}\Bigl].
\label{yeqn1}$$ for small $y$, $y_a$, and $\delta y$, with $\alpha(\ell)=\exp(-\ell)$. It is important to recognize that in the 3D superfluid state, $\rho(\ell) \rightarrow 0$ for [*finite*]{} $\ell \equiv \ell^*$, so that $\delta y(\ell)$ will be driven to zero at $\ell=\ell^*$. Thus the non-uniformity of tunneling events along the chain is irrelevant in this state. In this situation, we can compute the conductance using the fixed point Hamiltonian (Eq. \[Hfixed\]) for $\rho=0$ to find a true superconducting response, $G(\omega) \propto -i/(\omega+i\delta)$.
By contrast, if $\rho$ remains finite as $\ell \rightarrow \infty$, $\delta y$ will necessarily grow, and we need to look for a new fixed point. To do this, we integrate the RG flows to a scale $\ell_0$ for which the irrelevant operators may be ignored. We then integrate out $\varphi$ and $a(x \ne 0,\tau)$ and recollect some irrelevant terms to restore the ${1 \over 2} \rho a(x=0,\tau)^2$ term to its cosine form $\cos [2\pi \alpha(\ell_0) a(x=0,\tau)]$, and arrive at an effective Hamiltonian for the lead site in the chain, $$\begin{aligned}
{\cal H}_{eff}^L &=&
\frac {c}{2K}\sqrt{\frac{\rho K}{\rho K + c^2}}\frac 1{\beta}\sum_{\omega_n}|\omega_n| |a(x=0,\omega_n)|^2 \cr
&+& y_0(\ell_0)\sum_{\tau}\cos \bigl[2\pi \alpha(\ell_0)a(x=0,\tau)\bigr].
\label{RGH}\end{aligned}$$ We then modify the RG so that $a^{\prime}(x=0,\tau^{\prime})=a(x=0,\tau)$ to preserve the form of the quadratic term in ${\cal H}_{eff}^L$; the scaling relation obeyed by $y_0$ is then $$\frac {dy_0(\ell)}{d\ell} \approx y_0 (\ell)
\Bigr[1 - 2\pi^2 K \sqrt{\frac{\rho K + c^2}{\rho K c^2}}[\alpha (\ell_0)]^2\Bigl].
\label{yeqn2}$$ It is important to recognize that $\alpha(\ell_0)$ does [*not*]{} shrink as it did in Eq. \[yeqn1\]. Thus since $\rho$ is small, $y_0$ is irrelevant. Our fixed point is then ${\cal H}_{eff}^L $ with $y_0=0$. In this case the conductance is finite as $\omega \rightarrow 0$, with $G(\omega) \propto 1/\sqrt{\rho}$. This metallic behavior is surprising in light of the linear mode supported by this phase. It is a result of the very limited phase space available for fluctuations in one dimension [@schultz]. For this reason the transport properties of this intermediate phase is distinct from that of the 3D superfluid phase.
The conductance in the deconfined vortex state can be computed straightforwardly using the dual representation of the model. The result unsurprisingly is an insulating response $G(\omega) \sim \omega$. (Details will be presented elsewhere [@unpub].)
In summary, we have shown that bosons in a one-dimensional optical lattice that exchanges particles with a bulk superfluid supports three distinct states, with different collective mode spectra and conductances.
The authors thank K. MacAdam for helpful discussions. Support was provided by the NSF under Grant Nos. PHY9907949, DMR0108451, and [DMR0454699]{}. HAF thanks the KITP for its hospitality.
[19]{} D. Jaksch [*et. al*]{}, Phys. Rev. Lett. [**81**]{}, 3108 (1998). F. Dalfovo, S. Giorgini, L. P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. [**71**]{}, 463 (1999). P. Pedri [*et. al*]{}, Phys. Rev. Lett. [**87**]{}, 220401 (2001). C. J. Pethik and H. Smith, [*Bose-Einstein Condensation in Dilute Gases*]{}, Cambridge University Press, New York, 2002. H. Moritz, T. Stöferle, M. Köhl, and T. Esslinger, Phys. Rev. Lett. [**91**]{}, 250402 (2003). Y. Shin [*et. al*]{} Phys. Rev. Lett. [**92**]{}, 050405 (2004). M. P. A. Fisher, P. B. Weichman, G. Grinstein, and D. S. Fisher, Phys. Rev. B [**40**]{}, 546 (1989). E. Frey and L. Balents, Phys. Rev. B [**55**]{}, 1050 (1997). M. P. A. Fisher and D. H. Lee, Phys. Rev. B [**39**]{}, 2756 (1989). M. Greiner [*et. al*]{}, Phys. Rev. Lett. [**87**]{}, 160405 (2001); Nature [**415**]{}, 39 (2002). T. Stöferle [*et. al*]{}, Phys. Rev. Lett. [**92**]{}, 130403 (2004). A. Recati [*et. al*]{}, Phys. Rev. Lett. [**90**]{}, 020401 (2003). H. P. Büchler, G. Blatter, and W. Zwerger, Phys. Rev. Lett. [**90**]{}, 130401 (2003). C. Kollath, U. Schollöck, J. von Delft, and W. Zwerger, Phys. Rev. A [**69**]{}, 031601 (2004). E. Altman, W. Hofstetter, E. Demler, and M. D. Lukin, New J. Phys. [**5**]{}, 113 (2003). M. Aizenman [*et. al*]{}, Phys. Rev. A [**70**]{}, 023612 (2004). A. F. Ho, M. A. Cazalilla, and T. Giamarchi, Phys. Rev. Lett. [**92**]{}, 130405 (2004). A. Kuklov, N. Prokof’ev, and B. Svistunov, Phys. Rev. Lett. [**92**]{}, 050402 (2004). A. Recati [*et. al*]{}, Phys. Rev. Lett. [**94**]{}, 040404 (2005). H. A. Fertig, Phys. Rev. Lett. [**89**]{}, 035703 (2002); H. A. Fertig and K. Majumdar, Ann. Phys. [**305**]{}, 190 (2003). J. Cardy, [*Scaling and Renormalization in Statistical Physics*]{}, Cambridge University Press, New York, 1996. R. Dumke [*et. al*]{}, Phys. Rev. Lett. [**89**]{}, 097903 (2002). N. Schlosser, G. Reymond, I.Protsenko, P. Grangier, Nature [**411**]{}, 1024 (2001). E.B. Kolomeisky, J. P. Straley, and R. M. Kalas, Phys. Rev. A [**69**]{}, 063401 (2004). J. Negele and H. Orland, [*Quantum Many Particle Systems*]{}, Addison-Wesley, Reading, MA, 1988. I. F. Herbut, Phys. Rev. B [**57**]{}, 13 729 (1998). The logarithmic singularity in Eq. \[heqn\] is specific to a 3D reservoir. The last term in Eq. \[ZVM\] for which $1/h({\bf q})$ is the coefficient is irrelevant in the RG sense; it does not contribute directly to the fixed point Hamiltonian, but sets the initial conditions for the Hamiltonian flows. A 2D reservoir has qualitatively similar behavior but a 1D does not, and instead leads to coupled Luttinger liquid behavior with different fixed points. It is interesting to speculate that, for low boson (fractional) densities and weak coupling to the reservoir, the decoupled phase will be in the Tonks-Girardeau regime, for which the bosons display behavior similar to fermions. See B. Paredes [*et.al*]{}, Nature [**429**]{}, 277 (2004). Our model is specific to integer filling and cannot describe this regime. D. R. Nelson, [*Defects and Geometry in Condensed Matter Physics*]{}, Cambridge University Press, New York, 2002. K. Majumdar and H. A. Fertig, Unpublished. M. Köhl [*et. al*]{}, J. Low. Temp. Phys. [**138**]{}, 635 (2005). Note that the [*conductivity*]{} does not distinquish between a superconducting and a metallic state in one dimension. See H. J. Schultz, G. Cunibeti, and P. Pieri, cond-mat/9807366.
=2.2in
0.3cm
\[phase\]
=2.4in
0.3cm
\[mode\]
=3.0in
.2cm
\[JJgeometry\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work, we present a novel robust distributed beamforming (RDB) approach to mitigate the effects of channel errors on wireless networks equipped with relays based on the exploitation of the cross-correlation between the received data from the relays at the destination and the system output. The proposed RDB method, denoted cross-correlation and subspace projection (CCSP) RDB, considers a total relay transmit power constraint in the system and the objective of maximizing the output signal-to-interference-plus-noise ratio (SINR). The relay nodes are equipped with an amplify-and-forward (AF) protocol and we assume that the channel state information (CSI) is imperfectly known at the relays and there is no direct link between the sources and the destination. The CCSP does not require any costly optimization procedure and simulations show an excellent performance as compared to previously reported algorithms.'
author:
- |
Hang Ruan \* and Rodrigo C. de Lamare\*$^\#$\
$*$ Department of Electronics, The University of York, England, YO10 5BB\
[$^\#$CETUC, Pontifical Catholic University of Rio de Janeiro, Brazil]{}\
Emails: [email protected], [email protected]
title: 'Study of Robust Distributed Beamforming Based on Cross-Correlation and Subspace Projection Techniques'
---
Introduction
============
Distributed beamforming has been widely investigated in wireless communications and sensor array signal processing in recent years [@r1; @r2; @r3] [@TDS_clarke; @TDS_2; @switch_int; @switch_mc; @smce; @TongW; @jpais_iet; @TARMO; @did; @badstc; @baplnc; @rdb]. Such algorithms are key for situations in which the channels between the sources and the destination have poor quality so that devices cannot communicate directly and employ relays that receive and forward the signals. In [@r2], relay network problems are described as optimization problems and related transformations and implications are provided and discussed. The work in [@r7] formulates an optimization problem that maximizes the output signal-to-interference-plus-noise ratio (SINR) under total relay transmit power constraints, by computing the beamforming weight vector with only local information. The work in [@r4] focuses on multiple scenarios with different optimization problem formulations, in order to optimize the beamforming weight vector and increase the system signal-to-noise ratio (SNR), with the assumption that the global channel state information (CSI) is perfectly known. Other works like in [@r5; @r6] analyze power control methods based on channel magnitude, whereas the powers of each relay are adaptively adjusted according to the qualities of their associated channels.
However, in most scenarios encountered, the channels observed by the relays may lead to performance degradation because of inevitable measurement, estimation and quantization errors in CSI [@r12] as well as propagation effects. These impairments result in imperfect CSI that can affect most distributed beamforming methods , which either fail or cannot provide satisfactory performance. In this context, robust distributed beamforming (RDB) techniques are hence in demand to mitigate the channel errors or uncertainties and preserve the relay system performance. The studies in [@r9; @r11; @r12; @r25] minimize the total relay transmit power under an overall quality of service (QoS) constraint, using either a convex semi-definite programme (SDP) relaxation method or a convex second-order cone programme (SOCP). The works in [@r9; @r12] consider the channel errors as Gaussian random vectors with known statistical distributions between the source to the relay nodes and the relay nodes to the destination, whereas [@r11] models the channel errors with their covariance matrices as a type of matrix perturbation. The work in [@r11; @r14; @r21] presents a robust design, which ensures that the SNR constraint is satisfied for imperfect CSI by adopting a worst-case design and formulates the problem as a convex optimization problem that can be solved efficiently.
In this work, we propose an RDB technique that achieves very high estimation accuracy in terms of channel mismatch with reduced computational complexity, in scenarios where the global CSI is imperfect and local communication is unavailable. Unlike existing RDB approaches, we aim to maximize the system output SINR subject to a total relay transmit power constraint using an approach that exploits the cross-correlation between the beamforming weight vector and the system output and then projects the obtained cross-correlation vector onto subspaces computed from the statistics of second-order imperfect channels, namely, the cross-correlation and subspace projection (CCSP) RDB technique. Unlike our previous work on centralized beamforming [@r17], the CCSP RDB technique is distributed and has marked differences in the way the subspace processing is carried out. In the CCSP RDB method, the covariance matrices of the channel errors are modeled by a certain type of additive matrix perturbation methods [@r8], which ensures that the covariance matrices are always positive-definite. We consider multiple source signals and assume that there is no direct link between them and the destination. The proposed CCSP RDB technique shows outstanding SINR performance as compared to the existing distributed beamforming techniques, which focus on transmit power minimization over input SNR values.
The rest of this work is organized as follows: Section II presents the system model. Section III devises the proposed CCSP RDB method. Section IV illustrates and discusses the simulation results. Section V states the conclusion.
System Model
============
We consider a wireless communication network consisting of $K$ signal sources (one desired signal source with the others as interferers), $M$ distributed single-antenna relays and a destination. We assume that that direct links are not reliable and therefore not considered. The $M$ relays receive data transmitted by the signal sources and then retransmit to the destination by employing beamforming, in which a two-step amplify-and-forward (AF) protocol is considered for cooperative communications. In the first step, the $k$ sources transmit the signals to the $M$ single-antenna relays according to the model given by $${\bf x}={\bf F}{\bf s}+{\boldsymbol \nu}, \label{eq1}$$ where the vector ${\bf s}=[s_1, s_2, \dotsb, s_K]^T \in {\mathbb
C}^{K \times 1}$ contains signals with zero mean denoted by $s_k=\sqrt{P_{s,k}}b_k$ for $k=1,2,\dotsb,K$, where $E[|b_k|^2]=\sigma^2_{b_k}$, $P_{s,k}$ and $b_k$ are the transmit power and the information symbol of the $k$th signal source, respectively. We assume that $s_1$ is the desired signal while the remaining source signals are treated as interferers. The matrix ${\bf F}=[{\bf f}_1, {\bf f}_2, \dotsb, {\bf f}_K] \in {\mathbb
C}^{M \times K}$ is the channel matrix between the signal sources and the relays, ${\bf f}_k=[f_{1,k}, f_{2,k}, \dotsb, f_{M,k}]^T \in
{\mathbb C}^{M \times 1}$, $f_{m,k}$ denotes the channel between the $m$th relay and the $k$th source ($m=1,2, \dotsb, M$, $k=1,2,\dotsb,
K$). ${\boldsymbol \nu}=[\nu_1, \nu_2, \dotsb, \nu_M]^T \in {\mathbb
C}^{M \times 1}$ is the complex Gaussian noise vector at the relays and $\sigma_{\nu}^2$ is the noise variance at each relay ($\nu_m$ \~ ${\mathcal CN}(0,\sigma_{\nu}^2)$, which refers to the complex Gaussian distribution with zero mean and variance $\sigma_{\nu}^2$). The vector ${\bf x} \in {\mathbb C}^{M \times 1}$ represents the received data at the relays. In the second step, the relays transmit ${\bf y} \in {\mathbb C}^{M \times 1}$, which is an amplified and phase-steered version of ${\bf x}$ that can be written as $${\bf y}={\bf W}{\bf x}, \label{eq2}$$ where ${\bf W}={\rm diag}([w_1, w_2, \dotsb, w_M]) \in {\mathbb
C}^{M \times M}$ is a diagonal matrix whose entries denote the beamforming weights, where ${\rm diag}(.)$ denote the diagonal entry of a matrix. Then the signal received at the destination is given by $$z={\bf g}^T{\bf y}+n, \label{eq3}$$ where $z$ is a scalar, ${\bf g}=[g_1, g_2, \dotsb, g_M]^T \in
{\mathbb C}^{M \times 1}$ is the complex Gaussian channel vector between the relays and the destination, $n$ ($n$ \~ ${\mathcal
CN}(0,\sigma_n^2)$) is the noise at the destination and $z$ is the received signal at the destination. Here we assume that the noise samples at each relay and the destination have the same power, which means we have $P_n=\sigma_n^2=\sigma_{\nu}^2$.
Both channel matrices ${\bf F}$ and ${\bf g}$ are modeled as Rayleigh distributed random variables, i.e., distance based large-scale channel propagation effects that include distance based fading and shadowing are considered. An exponential based path loss model is described by [@r16] $$\gamma=\frac{\sqrt{L}}{\sqrt{d^{\rho}}}, \label{eq4}$$ where $\gamma$ is the distance-based path loss, $L$ is the known path loss at the destination, $d$ is the distance of interest relative to the destination and $\rho$ is the path loss exponent, which can vary due to different environments and is typically set within $2$ to $5$, with a lower value representing a clear and uncluttered environment, which has a slow attenuation and a higher value describing a cluttered and highly attenuating environment. Shadow fading can be described as a random variable with a probability distribution for the case of large scale fading given by $$\beta=10^{(\frac{\sigma_s{\mathcal N}(0,1)}{10})}, \label{eq5}$$ where $\beta$ is the shadowing parameter, ${\mathcal N}(0,1)$ means the Gaussian distribution with zero mean and unit variance, $\sigma_s$ is the shadowing spread in dB. The shadowing spread reflects the severity of the attenuation caused by shadowing, and is given between $0$dB to $9$dB [@r16]. The channels modeled with both path-loss and shadowing can be represented as: $${\bf F}=\gamma\beta{\bf F}_0, \label{eq6}$$ $${\bf g}=\gamma\beta{\bf g}_0, \label{eq7}$$ where ${\bf F}_0$ and ${\bf g}_0$ denote the Rayleigh distributed channels without large-scale propagation effects [@r16].
The received signal at the $m$th relay can be expressed as: $$x_m=\sum_{k=1}^K \underbrace{\sqrt{P_{s,k}}{b_k}}_{s_k}
f_{m,k}+{\nu}_m, \label{eq8}$$ then the transmitted signal at the $m$th relay is given by $$y_m={w_m}x_m. \label{eq9}$$ The transmit power at the $m$th relay is equivalent to $E[|y_m|^2]$ so that can be written as $\sum_{m=1}^ME[|y_m|^2]=\sum_{m=1}^ME[|{w_m}x_m|^2]$ or in matrix form as ${\bf w}^H{\bf D}{\bf w}$ where ${\bf D}={\rm
diag}\big(\sum_{k=1}^KP_{s,k}\sigma^2_{b_k}\big[E[|f_{1,k}|^2],
E[|f_{2,k}|^2], \dotsb, E[|f_{M,k}|^2]\big]+P_n\big)$ is a full-rank matrix, where $(.)^H$ denotes the Hermitian transpose operator. The signal received at the destination can be expanded by substituting and in , which yields $$\begin{gathered}
z=\underbrace{\sum_{m=1}^M{w_m}g_m\sqrt{P_{s,1}}f_{m,1}b_1}_{\text{desired signal}}
+\underbrace{\sum_{m=1}^M{w_m}g_m\sum_{k=2}^K\sqrt{P_{s,k}}f_{m,k}b_k}_{\text{interferers}}\\
+\underbrace{\sum_{m=1}^M{w_m}g_m\nu_m+n}_{\text{noise}}.
\label{eq10}\end{gathered}$$ By taking expectation of the components of , we can compute the desired signal power $P_{z,1}$, the interference power $P_{z,i}$ and the noise power $P_{z,n}$ at the destination as follows: $$\begin{split}
\hspace{-0.2em}P_{z,1}&=E\Big[\sum_{m=1}^M({w_m}g_m\sqrt{P_{s,1}}f_{m,1}b_1)^2\Big] \\
&= P_{s,1}\sigma^2_{b_1}\underbrace{\sum_{m=1}^M
E\Big[w_m^*(f_{m,1}g_m)(f_{m,1}g_m)^*w_m\Big]}_{{\bf w}^H E[({\bf
f}_1 \odot {\bf g})({\bf f}_1 \odot {\bf g})^H]{\bf w}},
\label{eq11}
\end{split}$$ $$\begin{split}
\hspace{-0.2em}P_{z,i}& =E\Big[\sum_{m=1}^M({w_m}g_m\sum_{k=2}^K\sqrt{P_{s,k}}f_{m,k}b_k)^2\Big] \\
&=\sum_{k=2}^KP_{s,k}\sigma^2_{b_k}\underbrace{\sum_{m=1}^M
E\Big[w_m^*(f_{m,k}g_m)(f_{m,k}g_m)^*w_m\Big]}_{{\bf w}^HE[({\bf
f}_k \odot {\bf g})({\bf f}_k \odot {\bf g})^H]{\bf w}} \label{eq12}
\end{split}$$ $$\hspace{-0.2em}P_{z,n}=E\Big[\sum_{m=1}^M({w_m}g_m\nu_m+n)^2\Big]
=P_n(1+\underbrace{\sum_{m=1}^ME\Big[w_m^*g_mg_m^*w_m\Big]}_{{\bf
w}^HE[{\bf g}{\bf g}^H]{\bf w}}), \label{eq13}$$ where $*$ denotes complex conjugation. By defining $${\bf R}_k
\triangleq P_{s,k}\sigma^2_{b_k}E[({\bf f}_k \odot {\bf g})({\bf
f}_k \odot {\bf g})^H],$$ where $\odot$ is the Schur-Hadamard product for $k=1,2,\dotsb,K$, $${\bf Q} \triangleq P_nE[{\bf g}{\bf
g}^H],$$ and the SINR is computed as: $$\begin{gathered}
SINR=\frac{P_{z,1}}{P_{z,i}+P_{z,n}} =\frac{{\bf w}^H{\bf R}_1{\bf
w}}{P_n +{\bf w}^H({\bf Q}+\sum_{k=2}^K{\bf R}_k){\bf w}}.
\label{eq14}\end{gathered}$$ It should be noted that in , the quantities ${\bf R}_k$, $k=1,\dotsb,K$ and ${\bf Q}$ only consist of the second-order statistics of the channels, which means that if the channels have no mismatches, those quantities describe the perfect knowledge of CSI. At this point, in order to introduce errors described by ${\bf
E}=[{\bf e}_1,\dotsb,{\bf e}_K] \in {\mathbb C}^{M \times K}$ and ${\bf e} \in {\mathbb C}^{M \times 1}$ to the channels $\hat{\bf F}$ and $\hat{\bf g}$, we have $$\hat{\bf f}_k={\bf f}_k+{\bf e}_k, k=1,2,\dotsb,K, \label{eq15}$$ $$\hat{\bf g}={\bf g}+{\bf e}, k=1,2,\dotsb,K, \label{eq15+}$$ where $\hat{\bf f}_k$ is the $k$th mismatched channel component of ${\bf F}$. The elements of ${\bf e}_k$ for any $k=1,\dotsb,K$ and ${\bf e}$ are assumed to be for simplicity independent and identically distributed (i.i.d) Gaussian variables so that the covariance matrices ${\bf R}_{{\bf e}_k}=E[{\bf e}_k{\bf e}_k^H]$ and ${\bf R}_{\bf e}=E[{\bf e}{\bf e}^H]$ are diagonal, in which case we can directly impose the effects of the uncertainties to all the matrices associated with ${\bf f}_k$ and ${\bf g}$ in . By assuming that the channel errors are uncorrelated with the channels so that $E[{\bf e}_k \odot {\bf g}]={\bf 0}$, $E[{\bf e} \odot {\bf f}_k]={\bf 0}$, $E[{\bf e} \odot {\bf g}]={\bf
0}$ and $E[{\bf e}_k \odot {\bf f}_k]={\bf 0}$, then we can use an additive Frobenius norm matrix perturbation method as introduced in [@r8], thus we can have the following: $$\hat{\bf R}_k={\bf R}_k+{\bf R}_{{\bf e}_k}={\bf R}_k+\epsilon||{\bf R}_k||_F{\bf I}_M, k=1,\dotsb,K, \label{eq16}$$ $$\hat{\bf Q}={\bf Q}+{\bf R}_{\bf e}={\bf R}_k+\epsilon||{\bf Q}||_F{\bf I}_M, k=1,\dotsb,K, \label{eq16+}$$ $$\hat{\bf D}={\bf D}+\epsilon||{\bf D}||_F{\bf I}_M, \label{eq17}$$ where $\hat{\bf R}_k$, $\hat{\bf Q}$ and $\hat{\bf D}$ are the matrices perturbed after the channel mismatches are taken into account, $\epsilon$ is the perturbation parameter uniformly distributed within $(0,{\epsilon}_{max}]$ where ${\epsilon}_{max}$ is a predefined constant which describes the mismatch level. The matrix ${\bf I}_M$ represents the identity matrix of dimension $M$ and it is clear that $\hat{\bf R}_k$, $\hat{\bf Q}$ and $\hat{\bf
D}$ are positive definite, i.e. $\hat{\bf R}_k \succ {\bf 0}
(k=1,\dotsb,K)$, $\hat{\bf Q} \succ {\bf 0}$ and $\hat{\bf D} \succ
{\bf 0}$. According to , the robust optimization problem that maximizes the output SINR with a total relay transmit power constraint is written as $$\begin{aligned}
\underset{\bf w}{\rm max}~~ \frac{{\bf w}^H\hat{\bf R}_1{\bf w}}{P_n+{\bf w}^H(\hat{\bf Q}+\sum_{k=2}^K\hat{\bf R}_k){\bf w}} \\
{\rm subject} ~~ {\rm to} ~~~~ {\bf w}^H\hat{\bf D}{\bf w} \leq P_T. \label{eq18}
\end{aligned}$$ The optimization problem has a similar form to the optimization problem in [@r4] and hence can be solved in a closed form using an eigen-decomposition method that only requires quantities or parameters with known second-order statistics.
Proposed CCSP RDB Algorithm
===========================
In this section, the proposed CCSP RDB algorithm is introduced. The algorithm is considered for a system with imperfect CSI, works iteratively to estimate and obtain the channel statistics over snapshots. The algorithm is based on the exploitation of cross-correction vector between the relay received data and the system output, as well as the construction of eigen-subspaces. By projecting the so obtained cross-correlation vector onto the subspaces at the relays, the channel errors can be mitigated and the result leads to a precise estimate of the beamformers. To this end, the sample cross-correlation vector (SCV) $\hat{\bf q}(i)$ in the $i$th iteration can be estimated by $$\hat{\bf q}(i)=\frac{1}{i}\sum\limits_{j=1}^i{\bf x}(j){z^*}(j), \label{eq19}$$ which uses sample averages that take into account all the data observations from snapshot one to the current snapshot, where ${\bf
x}(i)$ and ${z^*}(i)$ refer to the data observation vector in the $i$th snapshot at the relays and the system output in the $i$th snapshot at the destination, respectively, in the presence of channel errors. Then, we decompose the mismatched channel matrix $\hat{\bf F}(i)$ into $K$ components as $\hat{\bf F}(i)=[\hat{\bf
f}_1(i), \hat{\bf f}_2(i), \dotsb,\hat{\bf f}_K(i)]$ and for each of them we construct a separate projection matrix. For the $k$th ($1\leq{k}\leq{K}$) component, we compute the covariance matrix for $\hat{\bf f}_k(i)$ and use it as an estimate of the true channel covariance matrix instead of the mismatched channel covariance matrices: $$\hat{\bf R}_{\hat{\bf f}_k}(i)=\frac{1}{i}\sum\limits_{j=1}^i\hat{\bf f}_k(j)\hat{\bf f}_k^H(j). \label{eq20}$$ $$\hat{\bf R}_{\hat{\bf g}}(i)=\frac{1}{i}\sum\limits_{j=1}^i\hat{\bf g}(j)\hat{\bf g}^H(j). \label{eq20+}$$ Here we take an approximation for the time-averaged estimate of the covariance matrices so that we have ${\bf R}_{{\bf
f}_k}(i)=\frac{1}{i}\sum\limits_{j=1}^i{\bf f}_k(j){\bf
f}_k^H(j)\approx\frac{1}{i}\sum\limits_{j=1}^i\hat{\bf
f}_k(j)\hat{\bf f}_k^H(j)$ and ${\bf R}_{\bf
g}(i)=\frac{1}{i}\sum\limits_{j=1}^i{\bf g}(j){\bf
g}^H(j)\approx\frac{1}{i}\sum\limits_{j=1}^i\hat{\bf g}(j)\hat{\bf
g}^H(j)$. Then the error covariance matrices ${\bf R}_{{\bf
e}_k}(i)$ and ${\bf R}_{\bf e}(i)$ can be computed as $${\bf R}_{{\bf e}_k}(i)=\epsilon||{\bf R}_{{\bf f}_k}(i)||_F{\bf I}_M. \label{eq21}$$ $${\bf R}_{{\bf e}}(i)=\epsilon||{\bf R}_{\bf g}(i)||_F{\bf I}_M. \label{eq21+}$$ In order to eliminate or reduce the errors ${\bf e}_k(i)$ from $\hat{\bf f}_k(i)$ and ${\bf e}$ from $\hat{\bf g}(i)$, the SCV obtained in can be projected onto the subspaces described by $${\bf P}_k(i)=[{\bf c}_{1,k}(i),\dotsb,{\bf c}_{N,k}(i)][{\bf
c}_{1,k}(i),\dotsb,{\bf c}_{N,k}(i)]^H, \label{eq22}$$ and $${\bf P}(i)=[{\bf c}_{1}(i),\dotsb,{\bf c}_{N}(i)][{\bf
c}_{1}(i),\dotsb,{\bf c}_{N}(i)]^H, \label{eq22+}$$ where ${\bf c}_{1,k}(i),{\bf c}_{2,k}(i),\dotsb,{\bf c}_{N,k}(i)$ and ${\bf c}_{1}(i),{\bf c}_{2}(i),\dotsb,{\bf c}_{N}(i)$ are the $N$ principal eigenvectors of the error spectrum matrix ${\bf
C}_k(i)$ and ${\bf C}(i)$, respectively, defined by $$\begin{split}
{\bf C}_k(i) & \triangleq
\int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}E[\hat{\bf
f}_k(i)\hat{\bf f}_k^H(i)]d\epsilon
%& = \int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}E[({\bf f}_k(i)+{\bf e}_k(i))({\bf f}_k(i)+{\bf e}_k(i))^H]d\epsilon,
\vspace{-0.5em}\label{eq23}
\end{split}$$ and $$\begin{split}
\vspace{-0.5em}{\bf C}(i) & \triangleq \int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}E[\hat{\bf g}(i)\hat{\bf g}^H(i)]d \epsilon. %\\
%& =\int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}E[({\bf
%g}(i)+{\bf e}(i))({\bf g}(i)+{\bf e}(i))^H]d\epsilon.
\label{eq23+}
\end{split}$$ Since we have already assumed that ${\bf e}_k(i)$ and ${\bf e}(i)$ are uncorrelated with ${\bf f}_k(i)$ and ${\bf g}(i)$, if $\epsilon$ follows a uniform distribution over the sector $(0,\epsilon_{max}]$, by approximating $E[{\bf f}_k(i){\bf f}_k^H(i)]\approx{\bf R}_{{\bf
f}_k}(i)$, $E[{\bf e}_k(i){\bf e}_k^H(i)]\approx{\bf R}_{{\bf
e}_k}(i)$, $E[{\bf g}(i){\bf g}^H(i)]\approx{\bf R}_{\bf g}(i)$ and $E[{\bf e}(i){\bf e}^H(i)]\approx{\bf R}_{\bf e}(i)$, and can be simplified as $$\begin{split}
{\bf C}_k(i) %&
%= \int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}({\bf
%R}_{{\bf f}_k}(i) + {\bf R}_{{\bf e}_k}(i))d\epsilon \\
& = \epsilon_{max}{\bf R}_{{\bf f}_k}(i) +
\frac{\epsilon_{max}^2}{2}||{\bf R}_{{\bf f}_k}(i)||_F{\bf I}_M,
\vspace{-0.5em} \label{eq24}
\end{split}$$ and $$\begin{split}
\vspace{-0.5em}{\bf C}(i) %& = \int\limits_{\epsilon\rightarrow{0}^{+}}^{\epsilon_{max}}({\bf R}_{\bf g}(i) +{\bf R}_{\bf e}(i))d\epsilon \\
& =\epsilon_{max}{\bf R}_{\bf g}(i) +
\frac{\epsilon_{max}^2}{2}||{\bf R}_{\bf g}(i)||_F{\bf I}_M,
\label{eq24+}
\end{split}$$ Then the mismatched channel components are then estimated by $$\hat{\bf f}_k(i)=\frac{{\bf P}_k(i)\hat{\bf q}(i)}{{\lVert{{\bf P}_k(i)\hat{\bf q}(i)}\rVert}_2}, \label{eq25}$$ $$\hat{\bf g}(i)=\frac{{\bf P}(i)\hat{\bf q}(i)}{{\lVert{{\bf P}(i)\hat{\bf q}(i)}\rVert}_2}. \label{eq25+}$$ To this point, all the $K$ channel components of $\hat{\bf f}_k(i)$ are obtained so that we have $\hat{\bf F}_k(i)=[\hat{\bf f}_1(i),
\hat{\bf f}_2(i), \dotsb, \hat{\bf f}_K(i)]$. In the next step, we use the so obtained channel components to provide estimates for the matrix quantities $\hat{\bf R}_k(i)$ ($k=1,\dotsb,K$), $\hat{\bf Q}(i)$ and $\hat{\bf D}(i)$ in as follows: $$\hat{\bf R}_k(i)=P_{s,k}E[(\hat{\bf f}_k(i) \odot \hat{\bf g}(i))(\hat{\bf f}_k(i) \odot \hat{\bf g}(i))^H], \label{eq26}$$ $$\hat{\bf Q}(i)=P_nE[\hat{\bf g}(i)\hat{\bf g}^H(i)], \label{eq26+}$$ $$\hat{\bf D}(i)={\rm diag}\Big(\sum_{k=1}^KP_{s,k}[E[|\hat{f}_{1,k}(i)|^2], \dotsb, E[\hat{f}_{M,k}(i)|^2]]+P_n\Big). \label{eq27}$$ To proceed further, we define $\hat{\bf U}(i)=\hat{\bf Q}(i)+\sum_{k=2}^K\hat{\bf R}_k(i)$ so that can be written as $$\begin{aligned}
\underset{{\bf w}(i)}{\rm max}~~ \frac{{\bf w}^H(i)\hat{\bf R}_1(i){\bf w}(i)}{P_n+{\bf w}^H(i)\hat{\bf U}(i){\bf w}(i)} \\
{\rm subject} ~~ {\rm to} ~~~~ {\bf w}^H(i)\hat{\bf D}(i){\bf w}(i)
\leq P_T. \label{eq28}
\end{aligned}$$ To solve the optimization problem in , the weight vector is rewritten as $${\bf w}(i)=\sqrt{p}{\bf D}^{-1/2}(i)\tilde{\bf w}(i), \label{eq29}$$ where $\tilde{\bf w}(i)$ satisfies $\tilde{\bf w}^H(i)\tilde{\bf w}(i)=1$. Then can be rewritten as $$\begin{aligned}
\underset{p,\tilde{\bf w}(i)}{\rm max}~~\frac{p\tilde{\bf w}^H(i)\tilde{\bf R}_1(i)\tilde{\bf w}(i)}{p\tilde{\bf w}^H(i)\tilde{\bf U}(i)\tilde{\bf w}(i)+P_n} \\
{\rm subject} ~~ {\rm to} ~~~~ ||\tilde{\bf w}(i)||^2=1, p \leq P_T, \label{eq30}
\end{aligned}$$ where $\tilde{\bf R}_1(i)=\hat{\bf D}^{-1/2}(i)\hat{\bf R}_1(i){\bf
D}^{-1/2}(i)$ and $\tilde{\bf U}(i)=\hat{\bf D}^{-1/2}(i)\hat{\bf U}(i)\hat{\bf
D}^{-1/2}(i)$. As the objective function in increases monotonically with $p$ regardless of $\tilde{\bf w}(i)$, which means the objective function is maximized when $p=P_T$, hence can be simplified to $$\begin{aligned}
\underset{\tilde{\bf w}(i)}{\rm max}~~\frac{P_T\tilde{\bf w}^H(i)\tilde{\bf R}_1(i)\tilde{\bf w}(i)}{P_T\tilde{\bf w}^H(i)\tilde{\bf U}(i)\tilde{\bf w}(i)+P_n} \\
{\rm subject} ~~ {\rm to} ~~~~ ||\tilde{\bf w}(i)||^2=1, \label{eq31}
\end{aligned}$$ or equivalently as $$\begin{aligned}
\underset{\tilde{\bf w}(i)}{\rm max}~~\frac{P_T\tilde{\bf w}^H(i)\tilde{\bf R}_1(i)\tilde{\bf w}(i)}{\tilde{\bf w}^H(i)(P_n{\bf I}_M+P_T\tilde{\bf U}(i))\tilde{\bf w}(i)} \\
{\rm subject} ~~ {\rm to} ~~~~ ||\tilde{\bf w}(i)||^2=1, \label{eq32}
\end{aligned}$$ in which the objective function is maximized when $\tilde{\bf w}(i)$ is chosen as the principal eigenvector of $(P_n{\bf I}_M+P_T{\tilde{\bf
U(i)}})^{-1}\tilde{\bf R}_1(i)$ [@r4], which leads to the solution for the weight vector of the distributed beamformer with channel errors given by $$\begin{gathered}
{\bf w}(i)=\sqrt{P_T}\hat{\bf D}^{-1/2}(i){\mathcal P}\{(P_n{\bf I}_M \\ +\hat{\bf D}^{-1/2}(i)\hat{\bf U}(i)\hat{\bf D}^{-1/2}(i))^{-1} \hat{\bf D}^{-1/2}(i)\hat{\bf R}_1(i)\hat{\bf D}^{-1/2}(i)\}, \label{eq33}\end{gathered}$$ where ${\mathcal P}\{.\}$ denotes the principal eigenvector corresponding to the largest eigenvalue. Then the maximum achievable SINR of the system in the presence of channel errors is given by $$\begin{gathered}
{\rm SINR}_{max}=P_T{\lambda}_{max}\{(P_n{\bf I}_M+\hat{\bf D}^{-1/2}(i)\hat{\bf U}(i)\hat{\bf D}^{-1/2}(i))^{-1} \\
\hat{\bf D}^{-1/2}(i)\hat{\bf R}_1(i)\hat{\bf D}^{-1/2}(i)\}, \label{eq34}\end{gathered}$$ where $\lambda_{max}$ is the maximum eigenvalue. In order to reproduce the proposed CCSP RDB algorithm, we use -, -, and for each iteration.
Simulations
===========
In the simulations, we compare the proposed CCSP RDB algorithm with several existing robust approaches [@r11; @r12; @r14; @r20; @r21; @r25] (i.e. worst-case SDP online programming) in the presence of imperfect CSI. The simulation metrics considered include the system output SINR versus input SNR, snapshots as well as the maximum allowable total transmit power $P_T$. In some scenarios, we consider that the interferers are strong enough as compared to the desired signal and the noise. In all simulations, the system input SNR is known and can be controlled by adjusting only the noise power. Both channels ${\bf F}$ and ${\bf g}$ follow the Rayleigh distribution, whereas the mismatch is only considered for ${\bf F}$. The shadowing and path loss parameters employ $\rho=2$, the source-to-destination power path loss is $L=10$dB and the shadowing spread is $\sigma_s=3$dB. The distances of the source-to-relay links $d_{s,r_m}$ ($m=1,\dotsb,M$) are modeled as pseudo-random in an area defined by a range of relative distances based on the source-to-destination distance $d_{s,d}$ which is set to $1$, so as the source-to-relay link distances $d_{s,r_m}$ are decided by a set of uniform random variables distributed between $0.5$ to $0.9$, with corresponding relay-source-destination angles $\theta_{r_m,s,d}$ randomly chosen from an angular range of $-\pi/2$ to $\pi/2$. Therefore, the relay-to-destination distances $d_{r_m,d}$ can be calculated using the trigonometric identity given by $$d_{r_m,d}=\sqrt{d_{s,r_m}^2+1-2d_{s,r_m}\cos\theta_{r_m,s,d}}.$$ The total number of relays and signal sources are set to $M=8$ and $K=3$, respectively. The system interference-to-noise ratio (INR) is specified in each scenario and $100$ snapshots are considered. The number of principal components is manually selected to optimize the performance for the CCSP RDB algorithm.
We first examine the SINR performance versus a variation of maximum allowable total transmit power $P_T$ (i.e. $1$dBW to $5$dBW) by setting both SNR and INR to $10$dB. We consider that all interferers have the same power. We also set the matrix perturbation parameter to $\epsilon_{max}=0.5$ for all algorithms. Fig. \[figure4\] shows that the output SINR increases as we lift up the limit for the maximum allowable transmit power. The proposed CCSP RDB method outperforms the worst-case SDP algorithm and perform close to the case with perfect CSI.
\#1\#2[0.99]{}
In the second example, we increase the system INR from $10$dB to $20$dB, consider $K=3$ users but rearrange the powers of the interferers so that one of them is much stronger than the other. We then examine the algorithms in an incoherent scenario and set the power ratio of the stronger interferer over the weaker to $10$. The maximum allowable total transmit power $P_T$ and the perturbation parameter $\epsilon_{max}$ are fixed to $1$dBW and $0.2$, respectively. We observe the SINR performance versus SNR for these algorithms and illustrate the results in Fig. \[figure5\]. Then we set the system SNR to $10$dB and observe the output SINR performance versus snapshots as in Fig. \[figure6\]. It can be seen that all algorithms have performance degradation due to strong interferers as well as their power distribution. However, the CCSP RDB algorithm has excellent robustness in terms of SINR against the presences of strong interferers with unbalanced power distribution.
\#1\#2[0.99]{}
\#1\#2[0.99]{}
Conclusion
==========
We have devised the CCSP RDB approach based on the exploitation of the cross-correlation between the received data from the relays at the destination and the system output. The proposed CCSP RDB method does not require any costly online optimization procedure and the results show an excellent performance as compared to prior art.
[100]{}
R. Mudumbai, D.R. Brown III, U. Madhow, and H.V. Poor, “Distributed Transmit Beamforming Challenges and Recent Progress," *IEEE Communications Magazine*, Vol. 47, Issue. 2, pp. 102-110, 2009.
A. B. Gershman, N. D. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson, and B. Ottersten, “Convex Optimization-Based Beamforming," *IEEE Signal Processing Magazine*, Vol. 27, Issue. 3, pp. 62-75, May 2010.
J. Uher, T. A. Wysocki, and B. J. Wysocki, “Review of Distributed Beamforming," *Journal of Telecommunications and Inf. Technology*, 2011.
V. H. Nassab, S. Shahbazpanahi, A. Grami, and Z. Luo, “Distributed Beamforming for Relay Networks Based on Second-Order Statistics of the Channel State Information," *IEEE Trans. Signal Process.*, Vol. 56, No. 9, pp. 4306-4316, Sep 2008.
Y. Jing and H. Jafarkhani, “Network Beamforming Using Relays With Perfect Channel Information," *IEEE Trans. Information Theory*, Vol. 55, No. 6, pp. 2499-2517, June 2009.
V. Havary-Nassab, S. Shahbazpanahi, and A. Grami, “Optimal Distributed Beamforming for Two-Way Relay Networks," *IEEE Trans. Signal Process.*, Vol. 58, No. 3, pp. 1238-1250, March 2010.
K. Zarifi, S. Zaidi, S. Affes, and A. Ghrayeb, “A Distributed Amplify-and-Forward Beamforming Technique in Wireless Sensor Networks," *IEEE Trans. Signal Process.*, Vol. 59, No. 8, pp. 3657-3674, Aug 2011.
M. Dahleh, M. A. Dahleh, G. Verghese, “Dynamic Systems and Control," *Department of Electrical Engineering and Computer Science Massachuasetts Institute of Technology*, 2011.
M. A. Maleki, S. Mehrizi, M. Ahmadian, “Distributed Robust Beamforming in Multi-User Relay Network," *IEEE Wireless Communications and Networking Conference (WCNC)*, pp. 904-907, 2014.
B. Mahboobi, M. Ardebilipour, A. Kalantari and E. Soleimani-Nasab, “Robust Cooperative Relay Beamforming," *IEEE Wireless Communications Letters*, Vol. 2, Issue. 4, pp. 399-402, May 2013.
D. Ponukumati, F. Gao, and C. Xing, “Robust Peer-to-Peer Relay Beamforming A Probabilistic Approach," *IEEE Communications Letters*, Vol. 17, Issue. 2, pp. 305-308, Jan 2013.
P. Ubaidulla; A. Chockalingam, “Robust distributed beamforming for wireless relay networks," *IEEE 20th International Symposium on Personal, Indoor and Mobile Radio Communications*, Sep 2009.
T. Hesketh, R, C. de Lamare and S. Wales, “Joint maximum likelihood detection and link selection for cooperative MIMO relay systems," *IET Communications*, Vol. 8, Issue 14, pp. 2489-2499, Sep 2014.
H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming Based on Low-Rank and Cross-Correlation Techniques," *IEEE Trans. Signal Process.*, Vol. PP, Issue. 99, pp. 1, April 2016.
J. V. Stone, “Principal Component Analysis and Factor Analysis," *MIT Press*, Edition. 1, pp. 129-135, 2004.
A. G. Fabregas, A. Martinez and G. Caire, “Bit-Interleaved Coded Modulation," *Foundations and Trends in Communications and Information Theory*, Vol. 5, No. 1–2, pp 1-153, 2008. http://dx.doi.org/10.1561/0100000019
H. Chen and L. Zhang, “Worst-Case Based Robust Distributed Beamforming for Relay Networks," *Proc. IEEE Int. Conf. Acoustic, Speech and Signal Processing*, pp. 4963-4967, May 2013.
S. Salari, M. Z. Amirani, I. Kim, D. Kim and J. Yang, “Worst-Case Based Robust Distributed Beamforming for Relay Networks," *IEEE Trans. Wireless Commun.*, Vol. 15, Issue. 6, pp. 4455-4469, March 2016.
Y. Zhang, E. Dall’Anese and G. B. Giannakis, “Distributed Optimal Beamformers for Cognitive Radios Robust to Channel Uncertainties," *IEEE Trans. Signal Process.*, Vol. 60, Issue. 12, pp. 6495-6508, Sep 2012.
P. Hsieh, Y. Lin, and S. Chen, “Robust distributed beamforming design in amplify-and-forward relay systems with multiple user pairs," *23rd International Conference on Software, Telecommunications and Computer Networks (SoftCOM)*, pp. 371-375, Sep 2015.
H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming Using a Low-Complexity Shrinkage-Based Mismatch Estimation Algorithm," *IEEE Sig. Proc. Letters.*, Vol. 21, No. 1, pp. 60-64, Nov 2013.
P. Clarke and R. C. de Lamare, “Joint Transmit Diversity Optimization and Relay Selection for Multi-Relay Cooperative MIMO Systems Using Discrete Stochastic Algorithms,” *IEEE Communications Letters*, vol.15, no.10, pp.1035-1037, October 2011.
P. Clarke and R. C. de Lamare, “Transmit Diversity and Relay Selection Algorithms for Multirelay Cooperative MIMO Systems” *IEEE Transactions on Vehicular Technology*, vol.61, no. 3, pp. 1084-1098, October 2011.
Y. Cai, R. C. de Lamare, and R. Fa, “Switched Interleaving Techniques with Limited Feedback for Interference Mitigation in DS-CDMA Systems," IEEE Transactions on Communications, vol.59, no.7, pp.1946-1956, July 2011.
Y. Cai, R. C. de Lamare, D. Le Ruyet, “Transmit Processing Techniques Based on Switched Interleaving and Limited Feedback for Interference Mitigation in Multiantenna MC-CDMA Systems," IEEE Transactions on Vehicular Technology, vol.60, no.4, pp.1559-1570, May 2011.
T. Wang, R. C. de Lamare, and P. D. Mitchell, “Low-Complexity Set-Membership Channel Estimation for Cooperative Wireless Sensor Networks," IEEE Transactions on Vehicular Technology, vol.60, no.6, pp.2594-2607, July 2011.
T. Wang, R. C. de Lamare and A. Schmeink, “Joint linear receiver design and power allocation using alternating optimization algorithms for wireless sensor networks,” *IEEE Trans. on Vehi. Tech.*, vol. 61, pp. 4129-4141, 2012.
R. C. de Lamare, “Joint iterative power allocation and linear interference suppression algorithms for cooperative DS-CDMA networks", IET Communications, vol. 6, no. 13 , 2012, pp. 1930-1942.
T. Peng, R. C. de Lamare and A. Schmeink, “Adaptive Distributed Space-Time Coding Based on Adjustable Code Matrices for Cooperative MIMO Relaying Systems”, *IEEE Transactions on Communications*, vol. 61, no. 7, July 2013.
P. Li and R. C. de Lamare, “Distributed Iterative Detection With Reduced Message Passing for Networked MIMO Cellular Systems,” in IEEE Transactions on Vehicular Technology, vol. 63, no. 6, pp. 2947-2954, July 2014.
T. Peng and R. C. de Lamare, “Adaptive Buffer-Aided Distributed Space-Time Coding for Cooperative Wireless Networks," *IEEE Transactions on Communications*, vol. 64, no. 5, pp. 1888-1900, May 2016.
J. Gu, R. C. de Lamare and M. Huemer, “Buffer-Aided Physical-Layer Network Coding with Optimal Linear Code Designs for Cooperative Networks," *IEEE Transactions on Communications*, 2017.
H. Ruan and R. C. de Lamare, “Robust distributed beamforming based on cross-correlation and subspace projection techniques", *Digital Signal Processing (DSP)*, 2017.
Z. Xu and M.K. Tsatsanis, “Blind adaptive algorithms for minimum variance CDMA receivers," *IEEE Trans. Communications*, vol. 49, No. 1, January 2001.
R. C. de Lamare and R. Sampaio-Neto, “Low-Complexity Variable Step-Size Mechanisms for Stochastic Gradient Algorithms in Minimum Variance CDMA Receivers", *IEEE Trans. Signal Processing*, vol. 54, pp. 2302 - 2317, June 2006.
C. Xu, G. Feng and K. S. Kwak, “A Modified Constrained Constant Modulus Approach to Blind Adaptive Multiuser Detection," *IEEE Trans. Communications*, vol. 49, No. 9, 2001.
Z. Xu and P. Liu, “Code-Constrained Blind Detection of CDMA Signals in Multipath Channels," *IEEE Sig. Proc. Letters*, vol. 9, No. 12, December 2002.
R. C. de Lamare and R. Sampaio Neto, “Blind Adaptive Code-Constrained Constant Modulus Algorithms for CDMA Interference Suppression in Multipath Channels”, *IEEE Communications Letters*, vol 9. no. 4, April, 2005.
L. Landau, R. C. de Lamare and M. Haardt, “Robust adaptive beamforming algorithms using the constrained constant modulus criterion," IET Signal Processing, vol.8, no.5, pp.447-457, July 2014.
R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters", *Electronics Letters*, vol. 44, no. 9, 2008.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Reduced-Rank Processing Based on Joint and Iterative Interpolation, Decimation and Filtering", *IEEE Transactions on Signal Processing*, vol. 57, no. 7, July 2009, pp. 2503 - 2514.
R. C. de Lamare and Raimundo Sampaio-Neto, “Reduced-rank Interference Suppression for DS-CDMA based on Interpolated FIR Filters", *IEEE Communications Letters*, vol. 9, no. 3, March 2005.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Reduced-Rank MMSE Filtering with Interpolated FIR Filters and Adaptive Interpolators", *IEEE Signal Processing Letters*, vol. 12, no. 3, March, 2005.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive Interference Suppression for DS-CDMA Systems based on Interpolated FIR Filters with Adaptive Interpolators in Multipath Channels", *IEEE Trans. Vehicular Technology*, Vol. 56, no. 6, September 2007.
R. C. de Lamare, L. Wang and R. Fa “Adaptive reduced-rank LCMV beamforming algorithms based on joint iterative optimization of filters: Design and analysis," Signal Processing, Feb. 2010.
Adaptive reduced-rank LCMV beamforming algorithms based on joint iterative optimization of filters: Design and analysis
R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank adaptive filtering based on joint iterative optimization of adaptive filters", *IEEE Signal Process. Lett.*, vol. 14, no. 12, pp. 980-983, Dec. 2007.
R. C. de Lamare, M. Haardt, and R. Sampaio-Neto, “Blind Adaptive Constrained Reduced-Rank Parameter Estimation based on Constant Modulus Design for CDMA Interference Suppression", *IEEE Transactions on Signal Processing*, June 2008.
M. Yukawa, R. C. de Lamare and R. Sampaio-Neto, “Efficient Acoustic Echo Cancellation With Reduced-Rank Adaptive Filtering Based on Selective Decimation and Adaptive Interpolation," IEEE Transactions on Audio, Speech, and Language Processing, vol.16, no. 4, pp. 696-710, May 2008.
R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank space-time adaptive interference suppression with joint iterative least squares algorithms for spread-spectrum systems," *IEEE Trans. Vehi. Technol.*, vol. 59, no. 3, pp. 1217-1228, Mar. 2010.
R. C. de Lamare and R. Sampaio-Neto, “Adaptive reduced-rank equalization algorithms based on alternating optimization design techniques for MIMO systems," *IEEE Trans. Vehi. Technol.*, vol. 60, no. 6, pp. 2482-2494, Jul. 2011.
R. C. de Lamare, L. Wang, and R. Fa, “Adaptive reduced-rank LCMV beamforming algorithms based on joint iterative optimization of filters: Design and analysis," Signal Processing, vol. 90, no. 2, pp. 640-652, Feb. 2010.
R. Fa, R. C. de Lamare, and L. Wang, “Reduced-Rank STAP Schemes for Airborne Radar Based on Switched Joint Interpolation, Decimation and Filtering Algorithm," *IEEE Transactions on Signal Processing*, vol.58, no.8, Aug. 2010, pp.4182-4194.
L. Wang and R. C. de Lamare, “Low-Complexity Adaptive Step Size Constrained Constant Modulus SG Algorithms for Blind Adaptive Beamforming”, *Signal Processing*, vol. 89, no. 12, December 2009, pp. 2503-2513.
L. Wang and R. C. de Lamare, “Adaptive Constrained Constant Modulus Algorithm Based on Auxiliary Vector Filtering for Beamforming," IEEE Transactions on Signal Processing, vol. 58, no. 10, pp. 5408-5413, Oct. 2010.
L. Wang, R. C. de Lamare, M. Yukawa, “Adaptive Reduced-Rank Constrained Constant Modulus Algorithms Based on Joint Iterative Optimization of Filters for Beamforming,” *IEEE Transactions on Signal Processing*, vol.58, no.6, June 2010, pp.2983-2997.
L. Qiu, Y. Cai, R. C. de Lamare and M. Zhao, “Reduced-Rank DOA Estimation Algorithms Based on Alternating Low-Rank Decomposition," *IEEE Signal Processing Letters*, vol. 23, no. 5, pp. 565-569, May 2016.
L. Wang, R. C. de Lamare and M. Yukawa, “Adaptive reduced-rank constrained constant modulus algorithms based on joint iterative optimization of filters for beamforming", IEEE Transactions on Signal Processing, vol.58, no. 6, pp. 2983-2997, June 2010.
L. Wang and R. C. de Lamare, “Adaptive constrained constant modulus algorithm based on auxiliary vector filtering for beamforming", IEEE Transactions on Signal Processing, vol. 58, no. 10, pp. 5408-5413, October 2010.
R. Fa and R. C. de Lamare, “Reduced-Rank STAP Algorithms using Joint Iterative Optimization of Filters," IEEE Transactions on Aerospace and Electronic Systems, vol.47, no.3, pp.1668-1684, July 2011.
Z. Yang, R. C. de Lamare and X. Li, “L1-Regularized STAP Algorithms With a Generalized Sidelobe Canceler Architecture for Airborne Radar," IEEE Transactions on Signal Processing, vol.60, no.2, pp.674-686, Feb. 2012.
Z. Yang, R. C. de Lamare and X. Li, “Sparsity-aware space–time adaptive processing algorithms with L1-norm regularisation for airborne radar", IET signal processing, vol. 6, no. 5, pp. 413-423, 2012.
Neto, F.G.A.; Nascimento, V.H.; Zakharov, Y.V.; de Lamare, R.C., “Adaptive re-weighting homotopy for sparse beamforming,” in Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European , vol., no., pp.1287-1291, 1-5 Sept. 2014
Almeida Neto, F.G.; de Lamare, R.C.; Nascimento, V.H.; Zakharov, Y.V.,“Adaptive reweighting homotopy algorithms applied to beamforming," IEEE Transactions on Aerospace and Electronic Systems, vol.51, no.3, pp.1902-1915, July 2015.
L. Wang, R. C. de Lamare and M. Haardt, “Direction finding algorithms based on joint iterative subspace optimization," IEEE Transactions on Aerospace and Electronic Systems, vol.50, no.4, pp.2541-2553, October 2014.
S. D. Somasundaram, N. H. Parsons, P. Li and R. C. de Lamare, “Reduced-dimension robust capon beamforming using Krylov-subspace techniques," IEEE Transactions on Aerospace and Electronic Systems, vol.51, no.1, pp.270-289, January 2015.
H. Ruan and R. C. de Lamare, “Robust adaptive beamforming using a low-complexity shrinkage-based mismatch estimation algorithm," *IEEE Signal Process. Lett.*, vol. 21 no. 1 pp. 60-64, Nov. 2013.
H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming Based on Low-Rank and Cross-Correlation Techniques," *IEEE Transactions on Signal Processing*, vol. 64, no. 15, pp. 3919-3932, Aug. 2016.
S. Xu and R.C de Lamare, , *Distributed conjugate gradient strategies for distributed estimation over sensor networks*, Sensor Signal Processing for Defense SSPD, September 2012.
S. Xu, R. C. de Lamare, H. V. Poor, “Distributed Estimation Over Sensor Networks Based on Distributed Conjugate Gradient Strategies", IET Signal Processing, 2016 (to appear).
S. Xu, R. C. de Lamare and H. V. Poor, *Distributed Compressed Estimation Based on Compressive Sensing*, IEEE Signal Processing letters, vol. 22, no. 9, September 2014.
S. Xu, R. C. de Lamare and H. V. Poor, “Distributed reduced-rank estimation based on joint iterative optimization in sensor networks," in Proceedings of the 22nd European Signal Processing Conference (EUSIPCO), pp.2360-2364, 1-5, Sept. 2014
S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive link selection strategies for distributed estimation in diffusion wireless networks," in Proc. IEEE International Conference onAcoustics, Speech and Signal Processing (ICASSP), , vol., no., pp.5402-5405, 26-31 May 2013.
S. Xu, R. C. de Lamare and H. V. Poor, “Dynamic topology adaptation for distributed estimation in smart grids," in Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2013 IEEE 5th International Workshop on , vol., no., pp.420-423, 15-18 Dec. 2013.
S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive Link Selection Algorithms for Distributed Estimation", EURASIP Journal on Advances in Signal Processing, 2015.
L. Zhang, Y. Cai, C. Li, R. C. de Lamare, “Variable forgetting factor mechanisms for diffusion recursive least squares algorithm in sensor networks", *EURASIP Journal on Advances in Signal Processing*, 2017.
S. Xu, R. C. de Lamare, H. V. Poor, “Distributed Low-Rank Adaptive Estimation Algorithms Based on Alternating Optimization", *Signal Processing*, 2017
T. G. Miller, S. Xu, R. C. de Lamare and H. V. Poor, “Distributed Spectrum Estimation Based on Alternating Mixed Discrete-Continuous Adaptation," *IEEE Signal Processing Letters*, vol. 23, no. 4, pp. 551-555, April 2016.
N. Song, R. C. de Lamare, M. Haardt, and M. Wolf, “Adaptive Widely Linear Reduced-Rank Interference Suppression based on the Multi-Stage Wiener Filter," IEEE Transactions on Signal Processing, vol. 60, no. 8, 2012.
N. Song, W. U. Alokozai, R. C. de Lamare and M. Haardt, “Adaptive Widely Linear Reduced-Rank Beamforming Based on Joint Iterative Optimization," IEEE Signal Processing Letters, vol.21, no.3, pp. 265-269, March 2014.
R.C. de Lamare, R. Sampaio-Neto and M. Haardt, “Blind Adaptive Constrained Constant-Modulus Reduced-Rank Interference Suppression Algorithms Based on Interpolation and Switched Decimation,” *IEEE Trans. on Signal Processing*, vol.59, no.2, pp.681-695, Feb. 2011.
Y. Cai, R. C. de Lamare, “Adaptive Linear Minimum BER Reduced-Rank Interference Suppression Algorithms Based on Joint and Iterative Optimization of Filters," IEEE Communications Letters, vol.17, no.4, pp.633-636, April 2013.
R. C. de Lamare and R. Sampaio-Neto, “Sparsity-Aware Adaptive Algorithms Based on Alternating Optimization and Shrinkage," IEEE Signal Processing Letters, vol.21, no.2, pp.225,229, Feb. 2014.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider bilinear pseudo-differential operators whose symbols posses Gevrey type regularity and may have a sub-exponential growth at infinity, together with all their derivatives. It is proved that those symbol classes can be described by the means of the short-time Fourier transform and modulation spaces. Our first main result is the invariance property of the corresponding bilinear operators. Furthermore we prove the continuity of such operators when acting on modulation spaces. As a consequence, we derive their continuity on anisotropic Gelfand-Shilov type spaces. We consider both Beurling and Roumieu type symbol classes and Gelfand-Shilov spaces.'
address:
- 'Department of Mathematics, University of Turin, Italy'
- 'Department of Mathematics, University of Turin, Italy'
- 'Department of Mathematics and Informatics, University of Novi Sad, Serbia'
author:
- Ahmed Abdeljawad
- Sandro Coriasco
- Nenad Teofanov
title: '**Bilinear pseudo-differential operators with Gevrey-H[ö]{}rmander symbols**'
---
Introduction {#sec0 .unnumbered}
============
The study of multilinear operators has been influenced by the Calderón-Zygmund theory. Indeed, one of the achievements of Coifman-Meyer’s pioneer work [@CoMe] is the realization of pseudo-differential operators in terms of singular integrals of Calderón-Zygmund type. Their approach is based on a multilinear point of view and have had a far reaching impact in operator theory and partial differential equations. For example, boundedness of a class of translation invariant bilinear operators on Lebesgue spaces is proved in [@CoMe]. Furthermore, the bilinear Calderón-Zygmund theory developed by Grafakos and Torres [@GT] paved the way to the extension of those results to bilinear pseudo-differential operators which are non-translation invariant, i.e. whose symbols may depend on the space variable as well. We refer to [@BO] for a brief survey and discussion of applications to partial differential equations, and to [@BeMaNaTo] for a systematic study of bilinear pseudo-differential operators with symbols in bilinear Hörmander classes. See also [@KT] for a recent contribution in the context of Triebel-Lizorkin and local Hardy spaces.
Another type of results concerns bilinear (and multilinear) operators whose symbols are not necessarily smooth. Their continuity properties on modulation spaces were first observed in [@BKasso2004]. In contrast to classical bilinear pseudo-differential operators considered in e.g. [@CoMe], these operators are treated by the techniques of time-frequency analysis, see also [@BKasso2006; @BeGrHeOk; @BeMaNaTo; @CKasso; @MOPf].
In this paper, we employ the techniques of time-frequency analysis and modulation spaces, and consider bilinear pseudo-differential operators of Gevrey-H[ö]{}rmander type whose symbols are of infinite order and may have a (super-)exponential growth at infinity, together with all their derivatives. The linear counterpart of such operators is considered in [@CaTo], within the environment of isotropic Gelfand-Shilov spaces of functions and distributions, see also [@AbCoTo; @To25], and extended to the anisotropic setting in [@AbCaTo; @AbTo]. The main purpose of this paper is to extend some boundedness results given there to the bilinear case.
More precisely, we consider Gevrey-H[ö]{}rmander type symbols $a \in
{\Gamma _{(\omega)}^{\sigma,\mathbf{s}}} ({\mathbf R^{3d}})$ (or $ a\in {\Gamma _{(\omega)}^{\sigma,\mathbf{s},0}} ({\mathbf R^{3d}})$), see Definition \[Def:GammaSymb2\], and the corresponding pseudo-differential operators, denoted by ${\operatorname{Op}}_{r, t}(a)$, see below. When $r=t=0$ we recover the Kohn-Nirenberg correspondence considered, e.g., in [@BKasso2004; @BeGrHeOk], while for $r=t=1/2$ we obtain the Weyl correspondence, considered in [@Te5; @Te7]. We remark that, in view of this choice of symbol classes, we cannot rely on arguments based on standard (e.g., Littlewood-Paley) localization techniques. The substitute for this is, from the very beginning, a “global approach”, aimed at obtaining and employing appropriate characterizations of the involved objects, in terms of suitable estimates which hold true on the whole Euclidean spaces.
The paper is organized as follows. In Section \[sec1\] we collect necessary definitions, background material and basic facts on Gelfand-Shilov spaces, weight functions, modulation spaces, symbol classes and the corresponding bilinear pseudo-differential operators. In Section \[sec2\], we first study exponential-type operators on Gelfand-Shilov space and prove the corresponding invariance properties. We proceed with a characterization of the symbol spaces in terms of their regularity and decay properties, and suitable estimates related to modulation spaces. Finally, we prove our main results about the continuity of bilinear operators in Section \[sec3\]. The proofs of a few technical results are collected in the Appendix.
Preliminaries {#sec1}
=============
In this section we provide notation and background material which will be used throughout the paper. Proofs and details are in general omitted, since they can be found, e.g., in [@CoPiRoTe05; @Fe2; @Fe3; @Fe4; @FG1; @FG2; @FG4; @Gro; @Te5; @To2; @To7; @To8].
We use the standard notation for Euclidean spaces and multiindeces, cf. [@Ho1]. For example, if $x=(x_{1},...,x_{d})$ $\in{\mathbf R^{d}}$ and $\alpha=(\alpha_{1},...,\alpha _{d})\in{\mathbf N^{d}}$, then $x^{\alpha}=x_{1}^{\alpha_{1}}\dots x_{d}^{\alpha_{d}}$, $\partial_{x}^{\alpha}=\partial_{x_{1}}^{\alpha_{1}} \dots\partial_{x_{d}}^{\alpha_{d}}$, $|\alpha|=\alpha_{1}+\dots+\alpha_{d}$ and $\alpha!=\alpha_{1}!\dots\alpha_{d}!$. Here $ \mathbf{N} $ denotes the set of non-negative integers. If $\alpha\in{\mathbf N^{d}}$, then $\alpha>0$ means that $\alpha_{j} > 0$ for every $j=1,\dots, d,$ and similarly for $\alpha \geq 0$. We write $A(\theta )\lesssim B(\theta )$, $\theta \in \Omega$, if there is a constant $c>0$ such that $|A(\theta )|\le c|B(\theta )|$ for all $\theta \in \Omega$. We write $A(\theta)\asymp B(\theta)$ if $A(\theta )\lesssim B(\theta )$ and $B(\theta )\lesssim A(\theta )$ for all $\theta \in \Omega$. Here $\Omega$ is an open subset of ${\mathbf R^{N}}$. If $\mathcal{B}_1$ and $\mathcal{B}_2$ are topological vector spaces, then $ \mathcal{B}_1 \hookrightarrow \mathcal{B}_2$ means that $\mathcal{B}_1$ is continuously embedded into $\mathcal{B}_2$. By $\mathscr S ({\mathbf R^{d}})$ we denote the Schwartz space of rapidly decreasing functions, and $\mathscr S'({\mathbf R^{d}})$ denotes its dual space of tempered distributions.
Gelfand-Shilov spaces {#subsec1.2}
---------------------
Let $h,s,\sigma > 0$ be fixed. Then ${\mathcal S}_{s;h}^{\sigma}({\mathbf R^{d}})$ is the Banach space of all $f\in C^\infty ({\mathbf R^{d}})$ such that $$\label{gfseminorm}
{\Vert f\Vert _{{\mathcal S}_{s;h}^{\sigma}}}\equiv \sup_{\alpha ,\beta \in
\mathbf N^d} \sup_{x \in {\mathbf R^{d}}} \frac {|x^\alpha \partial ^\beta
f(x)|}{h^{|\alpha + \beta |} \alpha !^s\, \beta !^\sigma}<\infty,$$ endowed with the norm . Obviously, ${\mathcal S}_{s;h}^{\sigma}({\mathbf R^{d}})$ increases as $h$, $s$ and $\sigma$ increase, and it is contained in $\mathscr S ({\mathbf R^{d}})$ for every $h,s,\sigma > 0$.
The [Gelfand-Shilov spaces]{} ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ and $\Sigma _{s}^{\sigma}({\mathbf R^{d}})$ are defined as the inductive and projective limits respectively of ${\mathcal S}_{s;h}^{\sigma}({\mathbf R^{d}})$, i.e. $$\label{GSspacecond1}
{\mathcal S}_{s}^{\sigma}({\mathbf R^{d}}) = \bigcup _{h>0}{\mathcal S}_{s;h}^{\sigma}({\mathbf R^{d}})
\quad \text{and}\quad \Sigma _{s}^{\sigma}({\mathbf R^{d}}) =\bigcap _{h>0}
{\mathcal S}_{s;h}^{\sigma}({\mathbf R^{d}}),$$ with the usual inductive and projective limit topologies. Note that $\Sigma _{s}^{\sigma}({\mathbf R^{d}})\neq \{ 0\}$, if and only if $s+\sigma \ge 1$ and $(s,\sigma )\neq
(\frac 12,\frac 12)$, and ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})\neq \{ 0\}$, if and only if $s+\sigma \ge 1$, see [@GS; @Pi88]. For every $s,\sigma >0$ we have $$\label{Eq:GSEmbeddings}
\Sigma _s^\sigma ({\mathbf R^{d}})
\hookrightarrow
{\mathcal S}_s^\sigma ({\mathbf R^{d}})
\hookrightarrow
\Sigma _{s+{\varepsilon}}^{\sigma +{\varepsilon}}({\mathbf R^{d}})
\hookrightarrow
{\mathscr S}({\mathbf R^{d}})$$ for every ${\varepsilon}>0$. If $s+\sigma \ge 1$, then the last two inclusions in are dense, and if in addition $(s,\sigma )\neq (\frac 12,\frac 12)$ then the first inclusion in is dense. Moreover, for $\sigma<1$ the elements of ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ admit entire extensions to $\mathbf{C}^d$ satisfying suitable exponential bounds, [@GS].
Note that the original definition in [@GS] is given with $s,\sigma \ge 0$. Then ${\mathcal S}_{0}^{\sigma}({\mathbf R^{d}})\neq \{ 0\}$ ($\Sigma _{0}^{\sigma}({\mathbf R^{d}})\neq \{ 0\}$) and ${\mathcal S}_{s}^{0}({\mathbf R^{d}})\neq \{ 0\}$ ($\Sigma _{s}^{0}({\mathbf R^{d}})\neq \{ 0\}$), if and only if $\sigma > 1$ and $s>1$ respectively, see [@GS Chapter IV 8.1]. In fact, ${\mathcal S}_{0}^{\sigma}({\mathbf R^{d}})\neq \{ 0\}$ consists of compactly supported Gevrey functions, while ${\mathcal S}_{s}^{0}({\mathbf R^{d}})\neq \{ 0\}$ contains functions whose Fourier transforms are compactly supported Gevrey functions.
The spaces ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ and $\Sigma _{s}^{\sigma}({\mathbf R^{d}})$ combine global regularity with suitable decay properties at infinity, thus offering an abstract functional analytic framework for some problems in mathematical physics, [@Gram; @NR]. The following result is a well-known characterization of ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ and $\Sigma _{s}^{\sigma}({\mathbf R^{d}})$ in terms of the exponential decay of derivatives of their elements. Although the proof is standard, it contains some tools relevant for our investigations, and is therefore included in Appendix \[appendix\].
\[GSlemma\] Let $f $ be a smooth function on ${\mathbf R^{d}}$, $f\in C^{\infty}({\mathbf R^{d}}).$ Then $f \in {\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ (respectively $f \in \Sigma _{s}^{\sigma}({\mathbf R^{d}})$) if and only if for every $\alpha \in {\mathbf N^{d}}$ $$\label{GSspaces}
|\partial^\alpha f(x)| \lesssim l^{|\alpha|} \alpha!^\sigma e^{-h|x|^{\frac 1s}}, \;\;\; x \in {\mathbf R^{d}},$$ for some $l, h>0$ (respectively for every $l,h>0$).
The *Gelfand-Shilov distribution spaces* $({\mathcal S}_{s}^{\sigma})'({\mathbf R^{d}})$ and $(\Sigma _{s}^{\sigma})'({\mathbf R^{d}})$ are the projective and inductive limit respectively of $({\mathcal S}_{s;h}^{\sigma})'({\mathbf R^{d}})$: $$\tag*{(\ref{GSspacecond1})$'$}
({\mathcal S}_{s}^{\sigma})'({\mathbf R^{d}}) = \bigcap _{h>0}({\mathcal S}_{s;h}^{\sigma})'({\mathbf R^{d}})\quad
\text{and}\quad (\Sigma _{s}^{\sigma})'({\mathbf R^{d}}) =\bigcup _{h>0}({\mathcal S}_{s;h}^{\sigma})'({\mathbf R^{d}}).$$ It follows that ${\mathscr S}'({\mathbf R^{d}})\hookrightarrow
({\mathcal S}_s^\sigma)'({\mathbf R^{d}})$ when $s+\sigma \ge 1$, and if in addition $(s,\sigma )\neq (\frac 12,\frac 12)$, then $({\mathcal S}_s^\sigma)'({\mathbf R^{d}})
\hookrightarrow (\Sigma _s^\sigma)'({\mathbf R^{d}})$.
The Fourier transform $\mathscr F$ is the linear and continuous map on ${\mathscr S}({\mathbf R^{d}})$, given by the formula $$(\mathscr Ff)(\xi )= \widehat f(\xi ) \equiv (2\pi )^{-\frac d2}\int _{{\mathbf R^{d}}} f(x)e^{-i{\langle x,\xi\rangle} }\, dx, \;\;\; \xi \in {\mathbf R^{d}},$$ when $f\in {\mathscr S}({\mathbf R^{d}})$. Here ${\langle {\, \cdot \, },{\, \cdot \, }\rangle}$ denotes the usual scalar product on ${\mathbf R^{d}}$. The Fourier transform extends uniquely to homeomorphisms from $({\mathcal S}_{s}^{\sigma})'({\mathbf R^{d}})$ to $({\mathcal S}_{\sigma}^{s})'({\mathbf R^{d}})$, and from $(\Sigma _{s}^{\sigma})'({\mathbf R^{d}})$ to $(\Sigma _{\sigma}^{s})'({\mathbf R^{d}})$. Furthermore, it restricts to homeomorphisms from ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ to ${\mathcal S}_{\sigma}^{s}({\mathbf R^{d}})$, and from $\Sigma _{s}^{\sigma}({\mathbf R^{d}})$ to $\Sigma _{\sigma}^{s}({\mathbf R^{d}})$.
Next we rewrite the definition of Gelfand-Shilov spaces in the notation which is convenient for our analysis, see also [@CaTo; @GS; @GZ]. We put $${\mathbf R^{d_0+\dots+d_k}} = {\mathbf R^{d_0}} \times {\mathbf R^{d_1}} \times \dots \times {\mathbf R^{d_k}} = {\mathbf R^{(d_0,\dots,d_k)}} = {\mathbf R^{\mathbf{d}}}.$$
Let $k \in \mathbf N$, $ \mathbf{\sigma} = (\sigma_0,\dots,\sigma_k) >0$, $ \mathbf{s} = (s_0,\dots,s_k)>0$, and $ \mathbf{d} = d_0 + \dots + d_k $. The Gelfand-Shilov spaces $$\begin{aligned}
{3}
&{\mathcal S}_\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}}) = {\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}}) &
&\quad \text{and} \quad &
&\Sigma _\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}}) = \Sigma _{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}}),\end{aligned}$$ consist of all $F\in C^\infty ({\mathbf R^{d_0+\dots+d_k}})$ such that $$|x_0^{\alpha_0 }\dots x_k^{\alpha_k}\partial _{x_0}^{\beta _0}\dots
\partial _{x_k}^{\beta _k}F(x_0,\dots,x_k)|
\lesssim
h^{|\alpha _0+\beta _0+\dots+\alpha _k+\beta _k|}
\prod_{j=0}^{k}\alpha_j!^{s_j}\beta_j!^{\sigma_j}$$for some $h>0$ and for every $h>0$ respectively, where $x_j\in {\mathbf R^{d_j}}$, $\alpha_j,\beta_j\in \mathbf N^{d_j}$, $j=0,\dots,k$. The dual spaces of $ {\mathcal S}_\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}}) $ and $\Sigma _\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}})$ are denoted by $$({\mathcal S}_\mathbf{s} ^\mathbf{\sigma})' ({\mathbf R^{\mathbf{d}}})
=
({\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}})$$ and $$(\Sigma _{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})' ({\mathbf R^{\mathbf{d}}}) = (\Sigma _{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}})$$ respectively.
The space $ {\mathcal S}_\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}})$ is nontrivial if and only if $s_j + \sigma_j \geq 1$, for each $ j = 0,\dots, k$ and $ \Sigma _\mathbf{s} ^\mathbf{\sigma} ({\mathbf R^{\mathbf{d}}})$ is nontrivial if and only if $s_j + \sigma_j \geq 1$, and $(s_j,\sigma_j) \neq ( \frac{1}{2}, \frac{1}{2}) $ for each $ j = 0,\dots, k$.
Obviously, if $\sigma_j=\sigma$, $s_j=s$ and $d_j=d$, $j=0,\dots, k$, then $${\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}})
\equiv
{\mathcal S}_s^\sigma({\mathbf R^{(k+1)d}})
,\quad
\Sigma_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}})
\equiv
\Sigma _s ^\sigma({\mathbf R^{(k+1)d}}),$$ $$({\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}})
\equiv
({\mathcal S}_s^\sigma)'({\mathbf R^{(k+1)d}}),$$ and $$( \Sigma _{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}})
\equiv
(\Sigma _s ^\sigma)'({\mathbf R^{(k+1)d}}).$$
The Fourier transform is a homeomorphism between $ {\mathcal S}_s ^\sigma({\mathbf R^{d}})$ and $ {\mathcal S}^s _\sigma({\mathbf R^{d}})$ (and between $ \Sigma _s ^\sigma({\mathbf R^{d}})$ and $ \Sigma ^s _\sigma({\mathbf R^{d}})$), cf. [@GS]. This, together with the kernel theorem for Gelfand-Shilov spaces (see [@LCPT; @Prang; @Te3] implies the following mapping properties of partial Fourier transforms on Gelfand-Shilov spaces. The proof is therefore omitted. Here, ${\mathscr F}_jF$ is the partial Fourier transforms of $F(x_0,x_1,\dots,x_{k})$ with respect to $x_j\in {\mathbf R^{d_j}}$, $j=0,\dots,k$.
\[propBroadGSSpaceChar\] Let $k\in \mathbf N$, $s_j,\sigma _j>0$, $j=0,\dots,k$. Then the following is true:
1. the mapping ${\mathscr F}_j$ on ${\mathscr S}({\mathbf R^{d_0+\dots+d_k}})$ restrict to homeomorphism $$\begin{aligned}
{\mathscr F}_j \, &: \, {\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}})
\to
{\mathcal S}_{s _0,\dots,s_{j-1},\sigma_j,s_{j+1},\dots,s_k}
^{\sigma _0,\dots,\sigma_{j-1},s_j,\sigma_{j+1},\dots,\sigma_k}
({\mathbf R^{d_0+\dots+d_k}});\end{aligned}$$
2. the mapping ${\mathscr F}_j$ on ${\mathscr S}({\mathbf R^{d_0+\dots+d_k}})$ is uniquely extendable to homeomorphism $$\begin{aligned}
{\mathscr F}_j \, &: \, ({\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}}) \to
({\mathcal S}_{s _0,\dots,s_{j-1},\sigma_j,s_{j+1},\dots,s_k}
^{\sigma _0,\dots,\sigma_{j-1},s_j,\sigma_{j+1},\dots,\sigma_k})'
({\mathbf R^{d_0+\dots+d_k}}).\end{aligned}$$
The same holds true if the ${\mathcal S}_{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}$-spaces and their duals are replaced by corresponding $\Sigma _{s_0,\dots,s_k}^{\sigma_0,\dots,\sigma_k}$ -spaces and their duals in each occurrence.
The result analogous to Proposition \[propBroadGSSpaceChar\] holds for partial Fourier transforms with respect to some choice of variables. In particular the (full) Fourier transform on ${\mathscr S}({\mathbf R^{\mathbf{d}}})$ restricts to to homeomorphism $$\begin{aligned}
{\mathscr F}\, &: \, {\mathcal S}_{\mathbf{s}}^{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}})
\to
{\mathcal S}^{\mathbf{s}} _{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}});\end{aligned}$$ and is uniquely extendable to homeomorphism $$\begin{aligned}
{\mathscr F}\, &: \, ({\mathcal S}_{\mathbf{s}}^{\mathbf{\sigma}}
)'
({\mathbf R^{\mathbf{d}}}) \to
({\mathcal S}^{\mathbf{s}} _{\mathbf{\sigma}}
)'
({\mathbf R^{\mathbf{d}}}),\end{aligned}$$ and same holds true if ${\mathcal S}_{\mathbf{s}}^{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}})$ spaces and their duals are replaced by corresponding $\Sigma _{\mathbf{s}}^{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}})$ spaces and their duals.
Alternatively, this result is contained in the following Proposition.
\[prop:GScharac\] Let $k\in \mathbf N$, $ \mathbf{\sigma} = (\sigma_0,\dots,\sigma_k) >0$, $ \mathbf{s} = (s_0,\dots,s_k)>0$, and $ \mathbf{d} = d_1 + \dots + d_k $. Then the following conditions are equivalent.
1. $F\in {\mathcal S}_{\mathbf{s}}^{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}}) $($F\in\Sigma _{\mathbf{s}}^{\mathbf{\sigma}}
({\mathbf R^{\mathbf{d}}})$);
2. for some $r>0$ (for every $r>0$) it holds $$\begin{aligned}
\displaystyle{|F(x_0,\dots,x_k)|\lesssim
e^{-r\left(|x_0|^{\frac 1{s_0}}+\dots+|x_k|^{\frac 1{s_k}}
\right)}}
\intertext{and}
\displaystyle{|\widehat F(\xi _0,\dots,\xi _k)|
\lesssim
e^{-r\left(|\xi _0|^{\frac 1{\sigma _0}}+\dots
+ |\xi _k|^{\frac 1{\sigma _k}} \right)}}.\end{aligned}$$
3. for every $\alpha = (\alpha_0, \dots, \alpha_k) \in {\mathbf N^{\mathbf{d}}}$ and for some $h,r>0$ (for every $h,r>0$) it holds $$|\partial ^{\alpha} F(x_0,\dots,x_k)|\lesssim
h ^{|\alpha|} \prod_{j=0} ^k \alpha_j ^{\sigma_j}
e^{-r\left(|x_0|^{\frac 1{s_0}}+\dots+|x_k|^{\frac 1{s_k}}
\right)}.$$
The equivalence between (1) and (2) follows from [@ChChKi96], and (1) $ \Leftrightarrow $ (2) can be proved by a slight modification of the proof of Lemma \[GSlemma\] (cf. Appendix \[appendix\]) and we therefore leave it for the reader.
Weight functions {#subsec1.1}
----------------
A function $\omega$ is called a *weight* or *weight function* on ${\mathbf R^{d}}$, if $\omega,1/\omega\in L^\infty _{loc}({\mathbf R^{d}})$ are positive everywhere. Without loss of generality we may assume that the weight functions are continuous on ${\mathbf R^{d}}$. Let $\omega$ and $v$ be weights on ${\mathbf R^{d}}$. Then $\omega$ is called *$v$-moderate* or *moderate*, if $$\label{e1.1}
\omega (x_1+x_2)\lesssim \omega (x_1) v(x_2),\quad x_1,x_2\in {\mathbf R^{d}} .$$
If $v$ can be chosen as polynomial, then $\omega$ is called a weight of polynomial type. A weight function $v$ is *submultiplicative*, if it is symmetric in each coordinate and $$v (x_1+x_2)\lesssim v (x_1) v(x_2),\quad x_1,x_2\in {\mathbf R^{d}} .$$ From now on, $v$ always denote a submultiplicative weight if nothing else is stated. In particular, if holds and $v$ is submultiplicative, then $$\label{eq:2Next}
\frac {\omega (x_1)}{v(x_2)}
\lesssim \omega(x_1 + x_2) \lesssim \omega(x_1)v(x_2),
\quad x_1,x_2\in{\mathbf R^{d}}.$$
If $\omega$ is a moderate weight on ${\mathbf R^{d}}$, then there exists a submultiplicative weight $v$ on ${\mathbf R^{d}}$ such that and hold, cf. [@To8; @To11; @To18]. Moreover if $v$ is submultiplicative on ${\mathbf R^{d}}$, then $$\label{Eq:CondSubWeights}
1\lesssim v(x) \lesssim e^{c|x|}$$ for some constant $c>0$ (cf. [@Gc2.5 Lemma 4.2]).
In particular, if $\omega$ is moderate, then $$\omega (x+y)\lesssim \omega (x)e^{c|y|}
\quad \text{and}\quad
e^{-c|x|}\lesssim \omega (x)\lesssim e^{c|x|},\quad
x,y\in {\mathbf R^{d}}$$for some $c>0$.
For a given $k\in \mathbf N$, we let ${\mathscr P}_E({\mathbf R^{d_0+\dots+d_k}})$ be the set of all moderate weights on ${\mathbf R^{d_0+\dots+d_k}}$, and ${\mathscr P}({\mathbf R^{d_0+\dots+d_k}})$ be the subset of ${\mathscr P}_E({\mathbf R^{d_0+\dots+d_k}})$ which consists of weights of polynomial type.
If $\omega \in {\mathscr P}_E({\mathbf R^{d_0+\dots+d_k}})$ then there exists a submultiplicative weight $v$ on ${\mathbf R^{d_0+\dots+d_k}}$, such that $$\label{eq:2Nextbis}
\frac {\omega (x_0, \dots,x_k)}{v(y_0, \dots,y_k)}
\lesssim \omega(x_0+y_0, \dots,x_k+y_k)
\lesssim \omega(x_0, \dots, x_k)v(y_0, \dots,y_k),$$ where $x_j, y_j \in{\mathbf R^{d_j}}$, $j=0,\dots,k$. Note that for a submultiplicative weight $v$ on ${\mathbf R^{d_0+\dots+d_k}}$ becomes $$\label{eq:expEstonsubmultipicative}
1\lesssim v(x_0,\dots,x _k) \lesssim e^{r(|x_0|+\dots+|x _k|)},
\quad x_j\in {\mathbf R^{d_j}},\ j=0,\dots,k,$$ for some $r>0$.
Next we extend the weight functions considered in [@AbCaTo; @AbCoTo] to the case when ${\mathbf R^{d}} = {\mathbf R^{d_0+\dots+d_k}}$.
\[classesofweights\] Let $k\in \mathbf N$ and $s_j>0$, $j=0,\dots,k$. Then, the set ${\mathscr P}_{s_0,\dots,s_k}
({\mathbf R^{d_0+\dots+d_k}})$ (${\mathscr P}_{s_0,\dots,s_k}^0({\mathbf R^{d_0+\dots+d_k}})$) consists of all weights $\omega\in {\mathscr P}_E({\mathbf R^{d_0+\dots+d_k}})$ such that $$\label{eq:omegaEst}
\omega(x_0+y_0, \dots,x_k+y_k)
\lesssim \omega (x_0, \dots,x_k )
e^{r(|y_0|^{\frac 1{s_0}}+\dots+|y_k |^{\frac 1{s_k}})},\;
x_j,y_j\in {\mathbf R^{d_j}}$$ holds for some (for every) $r>0$.
In particular, if $\omega \in {\mathscr P}_{s_0,\dots,s_k}
({\mathbf R^{d_0+\dots+d_k}})$ (${\mathscr P}_{s_0,\dots,s_k}^0({\mathbf R^{d_0+\dots+d_k}})$), then $$e^{-r(|x_0|^{\frac 1{s_0}}+\dots+|x_k |^{\frac 1{s_k}})}
\lesssim
\omega (x_0, \dots,x_k )
\lesssim
e^{r(|x_0|^{\frac 1{s_0}} +\dots+|x_k |^{\frac 1{s_k}})}$$ for some $r>0$ (for every $r>0$).
By and it follows that $${\mathscr P}_{s_0,\dots,s_k}^0({\mathbf R^{d_0+\dots+d_k}}) =
{\mathscr P}_{\tilde s_0,\dots, \tilde s_k} ({\mathbf R^{d_0+\dots+d_k}}) =
{\mathscr P}_E ({\mathbf R^{d_0+\dots+d_k}})$$ when $s_j<1$ and $\tilde s_j\le 1$, $j=0,\dots,k$. For convenience we set $${\mathscr P}^0_E({\mathbf R^{d_0+\dots+d_k}})={\mathscr P}^0_{E,1}({\mathbf R^{d_0+\dots+d_k}}).$$
The following extension of [@AbCoTo Proposition 1.6], shows that for any weight in ${\mathscr P}_E ({\mathbf R^{d_0+\dots+d_k}})$, there are equivalent weights that satisfy the anisotropic Gevrey regularity.
\[Prop:EquivWeights\] Let there be given $\omega \in {\mathscr P}_E({\mathbf R^{d_0+d_1+\dots+d_k}})$. Then there exists a weight $\omega _0\in {\mathscr P}_E({\mathbf R^{d_0+d_1+\dots+d_k}})
\cap C^\infty ({\mathbf R^{d_0+d_1+\dots+d_k}})$ such that the following is true:
1. $\omega _0\asymp \omega $;
2. for every (multiindex) $\alpha_j\geq 0$, $ j =0,\dots,k,$ we have $$\begin{gathered}
|\partial _{x} ^{\alpha_0}\partial _{\xi_1} ^{\alpha_1}
\dots \partial _{\xi_k} ^{\alpha_k} \omega_0 (x, \xi_1,\dots,\xi_k)|
\lesssim
h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j}
\omega (x, \xi_1,\dots,\xi_k)
\\[1ex]
\asymp
h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j} \omega_0 (x, \xi_1,\dots,\xi_k),
\;\;\;
x\in {\mathbf R^{d}}_0, \; \xi_j \in {\mathbf R^{d_j}},\; j=1,\dots, k,\end{gathered}$$ for every $h>0$ and $ s_j>0$, $ j =0,\dots,k$.
The proof is given in Appendix \[appendix\].
Modulation spaces {#subsec1.3}
-----------------
Modulation spaces, originally introduced by Feichtinger in [@Fe4], are recognized as appropriate family of spaces when dealing with problems of time-frequency analysis, see [@Fe4; @FG1; @FG2; @FG4; @GaSa; @Gro; @RSTT; @Te2], to mention just a few references. A broader family of modulation spaces is recently studied in [@AbCoTo; @PfeuToft; @To25].
Let $s,\sigma>0$, such that $s+\sigma\geq 1$, and let $\phi \in {\mathcal S}_s ^\sigma ({\mathbf R^{d}})$ be fixed. Then the *short-time Fourier transform* $V_\phi f$ of $f\in ({\mathcal S}_s^\sigma )'
({\mathbf R^{d}})$ with respect to the *window function* $\phi$ is defined by $$V_\phi f(x,\xi ) \equiv
{\mathscr F}(f \, \cdot \, \overline {\phi ({\, \cdot \, }-x)})(\xi),\;\;\; x,\xi \in {\mathbf R^{d}}.$$ This definition makes sense as a Gelfand-Shilov distribution [@AbCaTo Remark 1.5].
If $f ,\phi \in {\mathcal S}_s^\sigma ({\mathbf R^{d}})$, then $$V_\phi f(x,\xi ) = (2\pi )^{-\frac d2}\int f(y)\overline {\phi
(y-x)}e^{-i{\langle y,\xi\rangle}}\, dy .$$
Let $s,\sigma>0$, such that $s+\sigma\geq 1$. Let there be given $\phi \in {\mathcal S}_s^\sigma({\mathbf R^{d}})\setminus 0$, $p,q\in [1,\infty ]$ and $\omega \in{\mathscr P}_E({\mathbf R^{2d}})$. Then the *modulation space* $M^{p,q}_{(\omega )}({\mathbf R^{d}})$ consists of all Gelfand-Shilov distributions $f$ on ${\mathbf R^{d}}$ such that $$\label{modnorm}
{\Vert f\Vert _{M^{p,q}_{(\omega )}}} \equiv \Big ( \int \Big ( \int |V_\phi f(x,\xi
)\omega (x,\xi )|^p\, dx\Big )^{q/p} \, d\xi \Big )^{1/q} <\infty$$ (with the obvious changes if $p=\infty$ and/or $q=\infty$). If $p=q$ we simply write $M^p_{(\omega )}$ instead of $M^{p,p}_{(\omega )}$, and if $\omega =1$, then we set $M^{p,q}=M^{p,q}_{(\omega )}$ and $M^{p}=M^{p}_{(\omega )}$.
The spaces $M_{(\omega)}^{p,q}$ are Banach spaces and every $\phi \in M^r_{(v)} \setminus 0$ yields an equivalent norm in and so $M_{(\omega)}^{p,q}$ is independent on the choice of $\phi \in M^r_{(v)}$ [@To25 Proposition 1.1].
Gelfand-Shilov spaces and their dual spaces can be described as projective or inductive limits of modulation spaces [@To18 Theorem 3.9]. In particular, we have the following characterization of Gelfand-Shilov spaces by the means of the short-time Fourier transform. We refer to [@GZ] for the proof, see also [@Te1; @Te6; @To18].
\[Prop:STFTGelfand2\] Let $k \in \mathbf N$, $ \mathbf{\sigma} = (\sigma_0,\dots,\sigma_k) >0$, $ \mathbf{s} = (s_0,\dots,s_k)>0$, and $ \mathbf{d} = d_1 + \dots + d_k $. Also let $\phi \in \mathcal S_{\mathbf{s}}^{\mathbf{\sigma} }({\mathbf R^{\mathbf{d}}})
\setminus 0$. Then the following is true:
1. $F \in {\mathcal S}_{\mathbf{s}}^{\mathbf{\sigma} }({\mathbf R^{\mathbf{d}}})$ if and only if $$\label{stftexpest2}
|V_\phi F(x_0,\dots,x_k,\xi _0,\dots,\xi _k )|
\lesssim
e^{-r \left(|x_0|^{\frac 1{s_0}}+\dots+|x_k|^{\frac 1{s_k}} + |\xi _0|^{\frac 1{\sigma _0}}+\dots
+ |\xi _k|^{\frac 1{\sigma _k}} \right)}$$ holds for some $r > 0$;
2. if, in addition, $\phi \in \Sigma _{\mathbf{s}}^{\mathbf{\sigma} }({\mathbf R^{\mathbf{d}}}) \setminus 0$, then $F\in \Sigma _{\mathbf{s}}^{\mathbf{\sigma} }({\mathbf R^{\mathbf{d}}}) $ if and only if holds for every $r > 0$.
Symbol classes and Pseudo-differential operators {#sec:PrePdo}
------------------------------------------------
First we introduce function spaces related to symbol classes of the multilinear pseudo-differential operators. We consider $a \in C^{\infty} ({\mathbf R^{d_0+\dots+d_k}})$ which obey various conditions of the form $$\begin{gathered}
|\partial _x^{\alpha} \partial _{\xi_1}^{\beta_1}
\dots, \partial _{\xi_k}^{\beta_k}
a(x,\xi_1,\dots,\xi_k )|
\\[1ex]
\lesssim
h ^{|\alpha +\beta_1+\dots+\beta_k|}
\alpha !^\sigma
\prod_{j=1} ^k \beta_j !^{s_j}
\cdot\omega (x,\xi_1,\dots,\xi_k),
$$ $w \in {\mathscr P}_E ({\mathbf R^{d_0+d_1+\dots+d_k}})$, $\alpha\in{\mathbf N^{d_0}}$, $\beta_j\in {\mathbf R^{d_j}}$, $s_j,\sigma,h>0$, $j = 1,\dots, k$.
When $k=1$ we recover the condition (1.14) from [@AbTo]. Similarly to [@AbTo], for a given $w \in {\mathscr P}_E ({\mathbf R^{d_0+d_1+\dots+d_k}})$ and $s_j,\sigma,h>0$, $j = 1,\dots, k$, we consider norms of the form $$\begin{gathered}
\label{norm}
{\Vert a\Vert _{\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}}}
\equiv
\\
\sup _{\substack{\alpha \in {\mathbf N^{d_0}}\\\beta_j \in {\mathbf N^{d_j}}}}
\left ( \sup _{\substack{x\in {\mathbf R^{d_0}}\\\xi_j\in{\mathbf R^{d_j}}}}
\left (
\frac {|\partial _x^{\alpha}
\partial _{\xi_1}^{\beta_1}\dots, \partial _{\xi_k}^{\beta_k}
a(x,\xi_1,\dots,\xi_k )|}
{h ^{|\alpha +\beta_1 +\dots+\beta_k|}
\alpha !^\sigma
\prod_{j=1} ^k \beta_j !^{s_j} \cdot \omega (x,\xi_1,\dots,\xi_k)}
\right )
\right ).\end{gathered}$$
More precisely, we are interested in invariance and continuity for bilinear pseudo-differential operators when symbols belong to the following symbol classes.
\[Def:GammaSymb2\] Let there be given $ \sigma, s_j, h>0$, $j=1,\dots,k$ and $\omega \in {\mathscr P}_E ({\mathbf R^{d_0+d_1+\dots+d_k}})$, and set $\mathbf{s} = (s_1, \dots, s_k)$.
1. The set ${\Gamma _{(\omega)}^{\sigma, \mathbf{s};h}}
({\mathbf R^{d_0+\dots+d_k}})$ consists of all $a \in C^\infty({\mathbf R^{d_0+\dots+d_k}})$ such that $${\Vert a\Vert _{\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}}} < \infty,$$ where the norm ${\Vert \cdot\Vert _{\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}}} $ is given by .
2. The sets ${\Gamma _{(\omega)}^{\sigma,\mathbf{s}}}
({\mathbf R^{d_0+\dots+d_k}})$ and ${\Gamma _{(\omega)}^{\sigma,\mathbf{s};0}}
({\mathbf R^{d_0+\dots+d_k}})$ are given by $$\begin{aligned}
{\Gamma _{(\omega)}^{\sigma,\mathbf{s} }}
({\mathbf R^{d_0+\dots+d_k}})
\equiv
\bigcup _{h>0}
{\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}}
({\mathbf R^{d_0+\dots+d_k}})
\intertext{and}
{\Gamma _{(\omega)}^{\sigma,\mathbf{s} ;0}}
({\mathbf R^{d_0+\dots+d_k}})
\equiv
\bigcap _{h>0}
{\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}}
({\mathbf R^{d_0+\dots+d_k}}),\end{aligned}$$ and their topologies are, respectively, the inductive and the projective limit topologies of ${\Gamma _{(\omega)}^{\sigma,\mathbf{s};h}} ({\mathbf R^{d_0+\dots+d_k}})$ with respect to $h>0$.
As it is common in the theory of ultradifferentiable functions, we say that (the inductive limit) ${\Gamma _{(\omega)}^{\sigma,\mathbf{s}}} ({\mathbf R^{d_0+\dots+d_k}})$ is a Roumieu class, and (the projective limit) ${\Gamma _{(\omega)}^{\sigma,\mathbf{s};0}}
({\mathbf R^{d_0+\dots+d_k}})$ is a Beurling class of test functions.
Notice that ${\Gamma _{(\omega)}^{\sigma,\mathbf{s}}}
({\mathbf R^{d_0+\dots+d_k}})$ and ${\Gamma _{(\omega)}^{\sigma,\mathbf{s};0}}
({\mathbf R^{d_0+\dots+d_k}})$ are nontrivial for any $ \sigma, s_j, h>0$, $j=1,\dots,k$. For instance by Proposition \[Prop:EquivWeights\] for any $\omega \in {\mathscr P}_E({\mathbf R^{d_0+d_1+\dots+d_k}})$ there exist a smooth function $\omega_0 \in {\mathscr P}_E({\mathbf R^{d_0+d_1+\dots+d_k}}) $ such that $\omega_0 \in {\Gamma _{(\omega)}^{\sigma,\mathbf{s};0}} ({\mathbf R^{d_0+\dots+d_k}})$.
When $k=1$ we put $\sigma_1 =\sigma $ and recover the symbol classes $ {\Gamma _{(\omega)}^{\sigma, s }} ({\mathbf R^{d_0+d_1}}) $ and $ {\Gamma _{(\omega)}^{\sigma, s;0 }} ({\mathbf R^{d_0+d_1}}) $ considered in [@AbTo].
Next we recall some facts on pseudo-differential operators. The pseudo-differential operator ${\operatorname{Op}}_t (a)$ is the linear and continuous operator on ${\mathcal S}({\mathbf R^{d}})$, defined by the formula $${\operatorname{Op}}_t (a)f(x) = \frac{1}{(2\pi ) ^{d}}\iint a(x- t(x-y),\xi )f(y)e^{i{\langle x-y,\xi\rangle} }\,
dyd\xi, \;\;\; x \in {\mathbf R^{d}} .$$More generally, the definition of ${\operatorname{Op}}_t (a)$ extends uniquely to $a\in {\mathcal S}'({\mathbf R^{2d}})$, and then ${\operatorname{Op}}_t (a)$ is continuous from ${\mathcal S}({\mathbf R^{d}})$ to ${\mathcal S}'({\mathbf R^{d}})$.
Let $\mathbf{t} = (t_1, t_2, \dots, t_m) \in [0,1]^m,$ be such that $ \sum_{j=1} ^m t_j \leq 1,$ and put $ \vec{f} = (f_1, f_2, \dots, f_m) \in {\mathscr S}({\mathbf R^{md}})$. The multilinear pseudo-differential operator ${\operatorname{Op}}_{\mathbf{t}}(a)$ from ${\mathscr S}({\mathbf R^{md}})$ to ${\mathscr S}' ({\mathbf R^{d}})$ is defined by the formula $${\operatorname{Op}}_{\mathbf{t}}(a) \vec{f} (x)=
\frac{1}{(2\pi)^{2md}}\iint e^{-i\psi(x,\mathbf{y},\mathbf{\xi})}
a_{\mathbf{t}}(x,\mathbf{y},\mathbf{\xi})\prod_{j=1} ^m f_j(y_j)
d\mathbf{y} d\mathbf{\xi},$$where $$a_{\mathbf{t}}(x,\mathbf{y},\mathbf{\xi}) =a(x+\sum_{j=1} ^m ( t_j y_j -x), \xi, \eta), \;\;\;
x, y_j, \xi_j \in {\mathbf R^{d}},$$ and the phase function $\psi$ is defined by $$\psi(x,\mathbf{y},\mathbf{\xi})= \sum_{j=1} ^m {\langle y_j -x,\xi_j\rangle},
\;\;\; x,y_j,\xi_j \in {\mathbf R^{d}}.$$
When $m=2$ we obtain bilinear pseudo-differential operators ${\operatorname{Op}}_{r, t}(a)$. That is, ${\operatorname{Op}}_{r, t}(a)$ is the bilinear and continuous operator from ${\mathscr S}({\mathbf R^{d}})\otimes{\mathscr S}({\mathbf R^{d}})$ to ${\mathscr S}' ({\mathbf R^{d}})$, defined by the formula $$\begin{gathered}
\label{eq:BilPdo}
\left({\operatorname{Op}}_{r, t}(a)(f,g)\right)(x)=
\\[1ex]
(2\pi)^{-2d}\iiiint e^{-i\psi(x,y,z,\xi,\eta)}
a_{r,t}(x,y,z, \xi, \eta) f(y)g(z) \,dydzd\xi d\eta, \; x \in {\mathbf R^{d}},\end{gathered}$$ where $(r,t)\in[0,1]\times[0,1]$, $ r+t\leq 1$, $$a_{r,t}(x,y,z, \xi, \eta)=a(x+r(y-x)+t(z-x), \xi, \eta), \;\;\;
x,y,z, \xi, \eta \in {\mathbf R^{d}},$$ and the phase function $\psi$ is defined by $$\psi(x,y,z,\xi,\eta)={\langle y-x,\xi\rangle}+{\langle z-x,\eta\rangle},
\;\;\; x,y,z, \xi, \eta \in {\mathbf R^{d}}.$$
If $r=t=0$, then the definition of ${\operatorname{Op}}_{0}(a)$ coincides with the definition of bilinear pseudo-differential operators $$T_a(f,g)(x)=(2\pi)^{-d}\iint e^{i{\langle x,\xi+\eta\rangle}}a(x,\xi,\eta)
\widehat{f}(\xi)\widehat{g}(\eta)\, d\xi d\eta, \;\;\; x \in {\mathbf R^{d}},$$considered in e.g [@BeGrHeOk], and the corresponding multilinear extension is studied in [@MOPf].
In fact, in Sections \[sec2\] and \[sec3\] we will consider the action of ${\operatorname{Op}}_{r, t}(a)$ when restricted to different Gelfand-Shilov spaces, and related unique extension of such operators to Gelfand-Shilov distributions.
We will use the following results about the continuity of linear pseudo-differential operators with symbols in Gevrey-Hörmander classes, and we refer to [@AbTo Theorem 2.1] and [@AbCaTo Theorem 3.7] respectively, for the proofs.
\[thm:anisCntp\] Let $s,\sigma \ge 1$, $p,q\in [1,\infty]$, $\omega ,\omega _0\in{\mathscr P}_{s,\sigma}^0({\mathbf R^{2d}})$, and $a\in \Gamma _{(\omega_0 )}^{\sigma,s}({\mathbf R^{2d}})$. Then ${\operatorname{Op}}_t (a)$ is a continuous operators from $M^{p,q}_{(\omega _0\omega)} ({\mathbf R^{d}})$ to $M^{p,q}_{(\omega)} ({\mathbf R^{d}})$ for any $t \in [0,1]$.
Note that, in the notation of Definition \[classesofweights\] we have ${\mathscr P}_{s,\sigma}^0({\mathbf R^{2d}}) = {\mathscr P}_{s_0,s_1}^0({\mathbf R^{d_0 + d_1}})$ when $s_0 =s, $ $s_1 = \sigma$ and $d_1 = d_2 =d.$
\[Thm:theorem2\] Let $A \in \mathbf M (d,R)$, $s,\sigma >0$ be such that $s+\sigma \ge 1$, $\omega\in{\mathscr P}_{s,\sigma}^0({\mathbf R^{2d}})$ and let $a\in \Gamma _{0}^{\sigma ,s;h}({\mathbf R^{2d}})$ for some $h>0$. Then ${\operatorname{Op}}_A (a)$ is continuous on ${\mathcal S}_s^\sigma ({\mathbf R^{d}})$ and on $({\mathcal S}_s^\sigma )'({\mathbf R^{d}})$.
Characterization and invariance property for bilinear pseudo-differential operators {#sec2}
===================================================================================
Our aim in this section is to show that $\Gamma ^{ \sigma,s_1,s_2}_{(\omega )} ({\mathbf R^{3d}})$ and $\Gamma ^{ \sigma, s_1,s_2;0}_{(\omega )} ({\mathbf R^{3d}})$ can be characterized in terms of estimates of short-time Fourier transforms and modulation spaces. This is done in subsection \[subsec2.2\]. We refer to [@AbCaTo; @To24] for similar results related to “standard” pseudo-differential operators. As a preparation, we show that ${\operatorname{Op}}_{r, t}(a)$ is independent of the choice of $r$ and $t$, which gives the invariance property for bilinear operators, Theorem \[Thm:CalculiTransf\]. The counterpart of Theorem \[Thm:CalculiTransf\] for “standard” pseudo-differential is proved in e.g. [@AbCaTo; @AbTo; @To24]. The key tools we employ to achieve the desired characterizations and invariance properties in this section are mapping results for exponentials of certain linear operators, similar to the typical ones often appearing in the “usual” Weyl-Hörmander calculus, whose description is given here below.
Mapping properties of exponential-type operators on Gelfand-Shilov spaces
-------------------------------------------------------------------------
For the study of mapping properties of the operator $e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}$ we need the following auxiliary result. By $\mathbf M(d,{\mathbf R})$ we denote the set of all $d\times d$-matrices with entries in ${\mathbf R}$.
\[lem:FTInvOpr\] Let $A,B\in \mathbf M(d,{\mathbf R})$ and $a \in {\mathscr S}({\mathbf R^{3d}})$. Then $$\label{eq:linid}
\big ({\mathscr F}_{2,3}^{-1}(e^{i {\langle AD_\xi+BD_\eta,D_x\rangle}} a)\big )(x+Ay+Bz,y,z) =
({\mathscr F}_{2,3}^{-1}a)(x,y,z),$$ $ x,y,z \in {\mathbf R^{d}}.$
Throughout the proof, the integrals are observed as either Fourier transforms or inverse Fourier transforms of appropriate distributions. The left-hand side of is given by $$\begin{gathered}
\big ({\mathscr F}_{2,3}^{-1}(e^{i {\langle AD_\xi+BD_\eta,D_x\rangle}} a)\big )(x+Ay+Bz,y,z)
\\[1ex]
=
\iint e^{i({\langle y,\xi\rangle}+{\langle z,\eta\rangle})}\left(e^{i {\langle AD_\xi+BD_\eta,D_x\rangle}} a)\right )
(x+Ay+Bz,\xi,\eta) \,d\xi d\eta,
$$ and $$\begin{gathered}
\left(e^{i {\langle AD_\xi+BD_\eta,D_x\rangle}} a)\right )
(x+Ay+Bz,\xi,\eta)
\\[1ex]
=
\iiint e^{i({\langle x+Ay+Bz,\zeta\rangle}+{\langle y_1,\xi\rangle}+{\langle z_1,\eta\rangle})}
e^{i ({\langle Ay_1+Bz_1,\zeta\rangle})} \widehat{a}(\zeta,y_1,z_1)\, d\zeta dy_1dz_1,\end{gathered}$$ where $$\widehat{a}(\zeta,y_1,z_1)
\\[1ex]
=
\iiint e^{-i({\langle x_1,\zeta\rangle}+{\langle y_1,\xi _1\rangle}+{\langle z_1,\eta _1\rangle})}
a(x_1,\xi_1,\eta_1) \, dx_1 d\xi _1 d\eta _1,$$ $ x_1,\xi_1,\eta_1 \in {\mathbf R^{d}}$. Let $\Psi\equiv \Psi(x,x_1,y,y_1,z,z_1,\zeta,\xi,\eta,\xi _1, \eta _1)$ be given by $$\begin{gathered}
\Psi={\langle y+y_1,\xi\rangle}+{\langle z+z_1,\eta\rangle}
+ {\langle x+Ay+Bz,\zeta\rangle}
\\
+{ ({\langle Ay_1+Bz_1,\zeta\rangle})}
{-({\langle x_1,\zeta\rangle}+{\langle y_1,\xi _1\rangle}+{\langle z_1,\eta _1\rangle})},\end{gathered}$$ $x,x_1,y,y_1,z,z_1,\zeta,\xi,\eta,\xi _1, \eta _1 \in {\mathbf R^{d}}$. It follows that $$\begin{gathered}
\label{eq:FTInvOpr2}
\big ({\mathscr F}_{2,3}^{-1}(e^{i {\langle AD_\xi+BD_\eta,D_x\rangle}} a)\big )(x+Ay+Bz,y,z)
\\[1ex]
=
\iiiint\!\!\!\iiiint
e^{i\Psi(x,x_1,y,y_1,z,z_1,\zeta,\xi,\eta,\xi _1, \eta _1)}
\\
\times a(x_1,\xi_1,\eta_1) \, dx_1 d\xi _1 d\eta _1
\, d\zeta dy_1dz_1
\,d\xi d\eta, \;\;\; x,y,z \in {\mathbf R^{d}}.\end{gathered}$$ Since $$\int e^{i{\langle y+y_1,\xi\rangle}} \, d\xi=\delta({y+y_1}),
\quad \text { and }
\int e^{i{\langle z+z_1,\eta\rangle}} \, d\eta=\delta ({z+z_1}),$$ where $\delta$ is the Dirac delta distribution, it follows that reduces to $$\begin{aligned}
\iiiint
&e^{i(
{\langle x,\zeta\rangle}
{-{\langle x_1,\zeta\rangle}+{\langle y,\xi _1\rangle}+{\langle z,\eta _1\rangle}})
}
a(x_1,\xi_1,\eta_1) \, d\zeta dx_1 d\xi _1 d\eta _1
\\[1ex]
&=\iint
e^{i({\langle y,\xi _1\rangle}+{\langle z,\eta _1\rangle})}
a(x,\xi_1,\eta_1) \, d\xi _1 d\eta _1
=\left({\mathscr F}^{-1}_{2,3}a\right)(x,y,z),\end{aligned}$$ and the claim follows.
Next we show some mapping properties of the operator $e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}$ which are an important ingredient in our analysis.
\[Thm:CalculiTransf\] Let $s_j ,\sigma _j$, $j=1,2,3$, be such that $$\label{eq:conditions}
s_j+\sigma _j\ge 1,\;
\quad 0< s_2,\, s _3\le s_1,
\quad \text{and} \quad 0< \sigma _1\le \sigma _2,\, \sigma _3$$ and let $ r,t\in[0,1]$ be such that $r+t\leq 1$. Then the following is true:
1. $e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}$ on ${\mathscr S}({\mathbf R^{3d}})$ restricts to a homeomorphism on ${\mathcal S}_{s_1,\sigma _2,\sigma _3}^{\sigma _1,s_2,s_3}
({\mathbf R^{3d}})$, and extends uniquely to a homeomorphism on $({\mathcal S}_{s_1,\sigma _2,\sigma _3}^{\sigma _1,s_2,s_3})'({\mathbf R^{3d}})$;
2. if in addition $(s_j,\sigma _j)\neq (\frac 12 ,\frac 12)$, $j=1,2,3$, then $e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}$ on ${\mathscr S}({\mathbf R^{3d}})$ restricts to a homeomorphism on $\Sigma _{s_1,\sigma _2,\sigma _3}^{\sigma _1,s_2,s_3}({\mathbf R^{3d}})$, and extends uniquely to a homeomorphism on $(\Sigma _{s_1,\sigma _2,\sigma _3}^{\sigma _1,s_2,s_3})'({\mathbf R^{3d}})$.
We only prove (1) and leave (2) for the reader.
Let $a\in {\mathscr S}({\mathbf R^{3d}})$ and let $U_{r,t}$ be the map given by $$(U_{r,t} F)(x,y)=F(x-ry-tz,y,z), \;\;\; x,y \in {\mathbf R^{d}}.$$ By Lemma \[lem:FTInvOpr\], we have $$\big ({\mathscr F}_{2,3}^{-1}(e^{i {\langle rD_\xi+tD_\eta,D_x\rangle}} a)\big )(x+ry+tz,y,z) =
({\mathscr F}_{2,3}^{-1}a)(x,y,z), \;\;\; x,y,z \in {\mathbf R^{d}},$$ wherefrom $e^{i {\langle rD_\xi+tD_\eta,D_x\rangle}} = {\mathscr F}_{2,3}\circ U_{r,t} \circ {\mathscr F}_{2,3}^{-1}$. Therefore it only remains to show that the mapping $U_{r,t}$ is continuous on ${\mathcal S}_{s_1,s_2,s_3} ^{\sigma_1,\sigma _2,\sigma _3}$.
Since the Fourier transform with respect to the 2$^{nd}$ and 3$^{rd}$ variables switches between the corresponding decay and regularity properties on Gelfand-Shilov spaces we consider $G=U_{r,t}F$, where $F\in {\mathcal S}_{s_1,\sigma_2,\sigma_3} ^{\sigma_1,s_2,s_3}$. Then $$G(x,y,z)= F(x-ry-tz,y,z)
\quad \text{and}\quad
\widehat G(\zeta,\xi ,\eta )= \widehat F(\zeta ,\zeta+r\xi,\zeta +t\eta ),$$ $x,y,z, \zeta,\xi ,\eta \in {\mathbf R^{d}}$. In view of Proposition \[propBroadGSSpaceChar\] and from the assumptions on $s_j$ and $\sigma _j$, it follows that there exist constants $c,r_0>0$, where $c$ depends on $r,t$, $s_j$ and $\sigma _j$ only, such that $$\begin{gathered}
|G(x,y,z)|= |F(x-ry-tz,y,z)|
\\
\lesssim e^{-r_0(|x-ry-tz|^{\frac 1{s_1}}
+|y|^{\frac 1{s_2} }+|z|^{\frac 1{s_3} } ) }
\lesssim
e^{-cr_0(|x|^{\frac 1{s_1}}+|y|^{\frac 1{s_2}}
+|z|^{\frac 1{s_3}})},\end{gathered}$$ $x,y,z \in {\mathbf R^{d}},$ and $$\begin{gathered}
|\widehat G(\zeta,\xi ,\eta )|= |\widehat F(\zeta ,\zeta+r\xi,\zeta +t\eta )|
\\
\lesssim e^{-r(|\zeta |^{\frac 1{\sigma _1}}+|\zeta+r\xi | ^{\frac 1{\sigma_2}}
|\zeta+t\eta| ^{\frac 1{\sigma_3}})}
\lesssim
e^{-cr(|\zeta |^{\frac 1{\sigma_1}}+|\xi |^{\frac 1{\sigma_2}}
+|\eta|^{\frac 1{\sigma _3}})},\end{gathered}$$ $ \zeta,\xi ,\eta \in {\mathbf R^{d}}$. The result follows since by Proposition \[prop:GScharac\] the topology in ${\mathcal S}_{s_1,s_2,s_3} ^{\sigma_1,\sigma _2,\sigma _3} ({\mathbf R^{3d}})$ can be defined by the above estimates.
Let $s,\sigma >0$ be such that $s+\sigma \ge 1$ and $\sigma \le s$. Then $e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}$ is a homeomorphism on ${\mathcal S}_{s}^{\sigma}({\mathbf R^{3d}})$, $\Sigma _{s}^{\sigma}({\mathbf R^{3d}})$, $({\mathcal S}_{s}^{\sigma})'({\mathbf R^{3d}})$ and on $(\Sigma _{s}^{\sigma})'({\mathbf R^{3d}})$.
Next we study an invariance property of bilinear pseudo-differential operators ${\operatorname{Op}}_{r, t}(a)$ given by . More precisely, it can be shown that for every Gelfand-Shilov distribution $a$ there is a unique distribution $b$ in the same Gelfand-Shilov class such that ${\operatorname{Op}}_{r_1,t_1}(a) = {\operatorname{Op}}_{r_2,t_2} (b)$, when $r_j,t_j\in[0,1]$ and $r_j+ t_j\leq 1$. The following result, which explains the relation between such $a$ and $b$, follows from Theorem \[Thm:CalculiTransf\] when the conditions in are fulfilled. We give an independent proof in Appendix \[appendix\].
\[Prop:CalculiTransfer\] Let $r_j,t_j\in[0,1]$ be such that $r_j+ t_j\leq 1$, and let $a,b \in ({\mathcal S}_{s_1,\sigma_2,\sigma_3} ^{\sigma_1, s_2,s_3} )'({\mathbf R^{3d}})$, where $s_j ,\sigma _j > 0$, and $s_j +\sigma _j \geq 1$, $j=1,2,3.$ Then $$\begin{aligned}
\label{calculitransform}
{\operatorname{Op}}_{r_1,t_1}(a) &= {\operatorname{Op}}_{r_2,t_2}(b)\nonumber
\\
\quad &\Leftrightarrow \quad
\\
e^{-i{\langle r_1D_\xi+t_1D_\eta,D_x\rangle}}a(x,\xi,\eta )
&= e^{-i{\langle r_2D_\xi+t_2D_\eta,D_x\rangle}} b(x,\xi,\eta ), \;\;\;
x,\xi,\eta \in {\mathbf R^{d}}.\nonumber\end{aligned}$$
Note that the latter equality in makes sense since it is equivalent to $$e^{-i{\langle r_1y+t_1z,\zeta\rangle}}\widehat a(\zeta ,y,z)
=e^{-i{\langle r_2y+t_2z,\zeta\rangle}}\widehat b(\zeta ,y,z), \;\;\; \zeta ,y,z \in {\mathbf R^{d}}.$$ Moreover, by using the similar arguments as in e.g. [@AbCaTo; @Tr; @CaTo], it can be shown that the map $a\mapsto e^{-i{\langle ry+tz,\zeta\rangle} }a$ is continuous on $({\mathcal S}_{s_1,\sigma_2,\sigma_3}^{\sigma_1, s_2,s_3} )'({\mathbf R^{3d}})$.
The following corollary is a consequence of [@CaTo Theorem 4.6] and Proposition \[Prop:CalculiTransfer\].
\[Somega\] Let $s_j,\sigma _j>0$ be such that $s_j +\sigma_j \ge 1$, $(s_j,\sigma_j ) \neq (\frac 12,\frac 12)$, $j=1,2,3$ and $r_j,t_j\in[0,1]$ be such that $r_j+ t_j\leq 1$, $j=1,2$. If $a,b\in (\Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2,s_3})'({\mathbf R^{3d}})$, are such that ${\operatorname{Op}}_{r_1,t_1}(a)={\operatorname{Op}}_{r_2,t_2}(b)$ Then $$\begin{aligned}
{3}
a &\in \Gamma ^{\sigma_1,s_2,s_3;0}_{(\omega)}({\mathbf R^{3d}})&\qquad
&\Leftrightarrow &
\qquad b &\in \Gamma ^{\sigma_1,s_2,s_3;0}_{(\omega)}({\mathbf R^{3d}})
\intertext{and}
a &\in \Gamma^{\sigma_1,s_2,s_3}_{(\omega)}({\mathbf R^{3d}})&\qquad
&\Leftrightarrow &
\qquad b &\in \Gamma^{\sigma_1,s_2,s_3}_{(\omega)}({\mathbf R^{3d}}),\end{aligned}$$ for any given $\omega \in {\mathscr P}_{E}({\mathbf R^{3d}})$.
Passages between different kinds of pseudo-differential calculi have been considered before [@Ho1; @Tr]. On the other hand, for the bilinear pseudo-differential calculi, it seems that the representation $a\mapsto {\operatorname{Op}}_{r,t}(a)$ for $(r,t)\in [0,1]\times[0,1]$ such that $r+t\leq 1$, has not been considered so far.
Gevrey-type symbol classes characterizations {#subsec2.2}
--------------------------------------------
. Our first result concerns the Roumieu case of symbols in $\Gamma ^{ \sigma,s,s}_{(\omega )} ({\mathbf R^{3d}})$. It can be deduced from [@CaTo Proposition 4.3], see also [@AbCaTo Proposition 2.4]. For the sake of completeness, we give the proof.
\[Prop:CharGammaSTFT\] Let $s_j ,\sigma _j >0$, $j=1,2,3$, be such that the conditions in hold, let $\omega \in {\mathscr P}_{s_1,\sigma_2,\sigma_3} ^0({\mathbf R^{3d}})$ and let $a$ be a Gelfand-Shilov distribution on ${\mathbf R^{3d}}$.
Then the following conditions are equivalent:
1. $ a \in \Gamma ^{\sigma_1 ,s_2, s_3}_{(\omega )} ({\mathbf R^{3d}})$, that is, $a\in C^\infty ({\mathbf R^{3d}})$ and $$|\partial_x ^\alpha \partial _\xi ^\beta \partial _\eta^\gamma a(x, \xi, \eta)|
\lesssim
h ^{|\alpha +\beta+ \gamma|} \alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3} \omega (x, \xi, \eta), \;\;\;
x, \xi, \eta \in {\mathbf R^{d}},$$for every $ \alpha,\beta,\gamma\in {\mathbf N^{d}}$ and some $h >0$;
2. For every $\phi \in {\mathcal S}_{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3}({\mathbf R^{3d}})\setminus 0$, there exist constants $h,R>0$ such that for every $\alpha,\beta,\gamma\in {\mathbf N^{d}}$, $$\begin{gathered}
\label{stftconda}
|\partial _x^\alpha \partial _\xi ^\beta \partial _\eta ^\gamma
{\left(e^{i({\langle x,\zeta\rangle} +{\langle y,\xi\rangle} +{\langle z,\eta\rangle} )}
V_\phi a(x,\xi, \eta, \zeta, y, z)\right)}|
\\[1ex]
\lesssim
h^{|\alpha+\beta+\gamma|} \alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3}
\omega (x, \xi, \eta)e^{-R (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})},\end{gathered}$$ $x,\xi, \eta, \zeta, y, z\in {\mathbf R^{d}}$.
3. For every $\phi \in {\mathcal S}_{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3}({\mathbf R^{3d}})\setminus 0$, there exist a constant $R >0$ such that $$\label{stftcondaA}
| V_\phi a(x,\xi, \eta, \zeta, y, z)|
\lesssim
\omega (x,\xi,\eta)e^{-R (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})}, \;\;\;
x,\xi, \eta, \zeta, y, z\in {\mathbf R^{d}}.$$
That (2) implies (3) is immediate, since is equal to when $\alpha =\beta =\gamma = 0.$
Let $X=(x,\xi,\eta),\ Y=(x_1,\xi_1,\eta_1 ), Z=(\zeta,y,z)\in {\mathbf R^{3d}}$, and set $$F_a(X,Y) = a (X+Y)\phi(Y) =
a(x+x_1,\xi +\xi_1, \eta+ \eta_1 )\phi(x_1, \xi_1, \eta_1).$$ Assume that (1) holds true. By the Leibniz rule, and Proposition \[prop:GScharac\] we obtain $$\begin{gathered}
|\partial _x^\alpha \partial _\xi ^\beta \partial _\eta^\gamma
F_a(x,\xi, \eta, x_1, \xi_1, \eta_1)|
\\[1ex]
\lesssim
h^{|\alpha + \beta +\gamma|}\alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3}
\omega(x,\xi,\eta )e^{-R(|x_1|^{\frac 1{s_1}}+
|\xi_1 |^{\frac 1{\sigma_2}}+|\eta_1 |^{\frac 1{\sigma_3}})},\end{gathered}$$ $x,\xi, \eta, x_1, \xi_1, \eta_1 \in {\mathbf R^{d}} $, for some constants $h,R >0$.
It follows that the set $$\Big \{ G_{a, h, X}(Y) \;\; | \;\;
G_{a, h,x,\xi, \eta }( x_1, \xi_1, \eta_1 ) =
\frac{\partial _x^\alpha \partial _\xi ^\beta
\partial _\eta^\gamma F_a(x,\xi, \eta, x_1, \xi_1, \eta_1 )}
{h^{|\alpha +\beta +\gamma|} \alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3} \omega(x,\xi, \eta )} \Big \}$$ is bounded in $ {\mathcal S}_{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3}(\mathbf{R}^{3d})$. If ${\mathscr F}_2 F_a$ denotes the partial Fourier transform of $F_a(X, Y)$ with respect to the $Y$-variable, then we get $$\begin{gathered}
|\partial _x^\alpha \partial _\xi ^\beta \partial_\eta^\gamma
({\mathscr F}_2 F_a)(x,\xi, \eta, \zeta, y,z)|
\\[1ex]
\lesssim
h^{|\alpha +\beta +\gamma |} \alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3}
\omega(x,\xi, \eta )e^{-R(|y|^{\frac 1{\sigma_1}}+
|z|^{\frac 1{s_2}} + |\zeta|^{\frac 1{s_3}})},\end{gathered}$$ $x,\xi, \eta, x_1, \xi_1, \eta_1 \in {\mathbf R^{d}} $, for some constants $h,R>0$. This, together with the Leibnitz rule applied to $ \partial _x^\alpha \partial _\xi ^\beta \partial _\eta ^\gamma
{\left(e^{i({\langle x,\zeta\rangle} +{\langle y,\xi\rangle} +{\langle z,\eta\rangle} )}
V_\phi a(x,\xi, \eta, \zeta, y, z)\right)}$ gives (2).
Assume now that (3) holds. By the inversion formula we get $$\label{Eq:STFTInversionFormula}
a(X) = \frac{ (2\pi)^{-\frac {3d}2}}{
{\Vert \phi\Vert _{L^2}}^{2} } \iint V_\phi a(Y,Z)
\phi (X-Y)e^{i{\langle X,Z\rangle}}\,dYdZ, \;\;\; X \in {\mathbf R^{3d}},$$ in the weak sense. Since $\phi\in {\mathcal S}_{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3}({\mathbf R^{3d}})$ we notice that $$\begin{aligned}
(X,Y,Z ) &\mapsto V_\phi a(Y,Z)\phi (X-Y)e^{i{\langle X,Z\rangle}}
\intertext{is a smooth map, and}
(Y,Z) &\mapsto Z^{{\boldsymbol{\alpha}}}
V_\phi a(Y,Z )\partial ^{{\boldsymbol{\beta}}} \phi (X-Y)e^{i{\langle X,Z\rangle}}\end{aligned}$$ is an integrable function for every $X\in {\mathbf R^{3d}}$, ${\boldsymbol{\alpha}}=(\alpha_1,\alpha_2,\alpha_3)\in {\mathbf N^{3d}}$, and ${\boldsymbol{\beta}}=(\beta_1,\beta_2,\beta_3)\in {\mathbf N^{3d}}$ in view of (3). Hence the derivatives of $a$ in satisfy the following estimates: $$\begin{gathered}
|\partial ^{\boldsymbol{\alpha}} a(X)|
\le
\sum _{{\boldsymbol{\beta}} \le {\boldsymbol{\alpha}}} {{\boldsymbol{\alpha}} \choose {\boldsymbol{\beta}}}
\iint |Z ^{{\boldsymbol{\beta}}} V_\phi a(Y,Z )
(\partial ^{{\boldsymbol{\alpha}} -{\boldsymbol{\beta}} }\phi )(X-Y)|\, dYdZ
\\[1ex]
\lesssim
\sum _{{\boldsymbol{\beta}} \le {\boldsymbol{\alpha}}} {{\boldsymbol{\alpha}} \choose {\boldsymbol{\beta}}}
\iint |Z^{{\boldsymbol{\beta}}}
\omega (Y)e^{-R (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})}
(\partial ^{{\boldsymbol{\alpha}} -{\boldsymbol{\beta}} }\phi )(X-Y)|
\, dYdZ
\\[1ex]
\lesssim
\sum _{{\boldsymbol{\beta}} \le {\boldsymbol{\alpha}}} {{\boldsymbol{\alpha}} \choose {\boldsymbol{\beta}}}
h_2 ^{|{\boldsymbol{\alpha}}- {\boldsymbol{\beta}}|} (\alpha _1- \beta _1 )!^{\sigma_1}
(\alpha _2- \beta _2 )!^{s_2} (\alpha _3- \beta _3 )!^{s_3}
\\[1ex]
\times \iint |Z^{{\boldsymbol{\beta}}}|
e^{-R (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})}
\omega (Y)
e^{-h_1(|x-x_1|^{\frac 1{s_1}}+|\xi-\xi _1|^{\frac 1{\sigma_2}}+
|\eta-\eta _1|^{\frac 1{\sigma_3}})}\, dYdZ,\end{gathered}$$ $X\in {\mathbf R^{3d}}$, for some constants $h_1, h_2>0$, and we used Lemma \[GSlemma\]. For any ${\boldsymbol{\beta}} \in {\mathbf N^{3d}}$, $\sigma, s_j >0$ such that $s_j + \sigma\ge 1$, $j=1,2$, and $h_2, R >0$, it holds $$|\zeta ^{\beta_1 }y^{\beta_2}z^{\beta_3}
e^{-R (|\zeta|^{\frac 1\sigma}+|y |^{\frac 1s}+|z |^{\frac 1s})}|
\lesssim h_2^{|{\boldsymbol{\beta}}|}\beta_1 !^{\sigma_1}
\beta _2 !^{s_2}\beta _3 !^{s_3}
e^{-\frac R2 (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})},$$$ \zeta,y,z \in {\mathbf R^{d}} $, so that $$\begin{gathered}
\label{Eq:GSTypeEst1}
|\partial ^{{\boldsymbol{\alpha}}} a(X)|
\\[1ex]
\lesssim h_2 ^{|{\boldsymbol{\alpha}} |}
\sum _{{\boldsymbol{\beta}} \le {\boldsymbol{\alpha}}}
{{\boldsymbol{\alpha}} \choose {\boldsymbol{\beta}}}
(\beta _1 !(\alpha _1-\beta _1)!)^{\sigma_1}
(\beta _2 !(\alpha _2-\beta _2)!)^{s_2}
(\beta _3 !(\alpha _3-\beta _3)!)^{s_3}
\\[1ex]
\times \iint e^{-\frac R2 (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})}
\omega (Y)
e^{-h_1(|x-x_1|^{\frac 1{s_1}}+|\xi-\xi _1|^{\frac 1{\sigma_2}}+
|\eta-\eta _1|^{\frac 1{\sigma_3}})} \, dYdZ
\\[1ex]
\lesssim (2h_2)^{|{\boldsymbol{\alpha}} |}
\alpha _1!^{\sigma_1} \alpha _2!^{s_2} \alpha _3!^{s_3}
\\[1ex]
\times
\int \omega(X+(Y-X))
e^{-h_1(|x-x_1|^{\frac 1{s_1}}+|\xi-\xi _1|^{\frac 1{\sigma_2}}+
|\eta-\eta _1|^{\frac 1{\sigma_3}})} \, dY, \;\;\; X \in {\mathbf R^{3d}}.\end{gathered}$$
Since $\omega \in {\mathscr P}_{s_1,\sigma_2, \sigma_3 } ^0$, that is holds for every $r>0$, by choosing $r\in(0,{h_1}/{2})$, from it follows that $$|\partial ^{{\boldsymbol{\alpha}}} a(X)|
\lesssim (2h_2)^{|{\boldsymbol{\alpha}}|}
\alpha _1!^{\sigma_1} \alpha _2!^{s_2} \alpha _3!^{s_3}
\omega (X), \;\;\; X \in {\mathbf R^{3d}},$$ for some constant $h_2>0$ (and we conclude that holds also in the pointwise sense). Therefore (3) implies (1) and the result follows.
The Beurling case follows by similar arguments as in the proof of Proposition \[Prop:CharGammaSTFT\].
\[prop:GammaOmegaChar\] Let $s_j ,\sigma _j >0$, $(s_j,\sigma_j )\neq (\frac 12,\frac 12)$, $j=1,2,3$, and let the conditions in hold. Also, let $\omega \in {\mathscr P}_{s_1,\sigma_2, \sigma_3 } ({\mathbf R^{3d}})$ and let $a$ be a Gelfand-Shilov distribution on ${\mathbf R^{3d}}$. Then the following conditions are equivalent:
1. $ a\ in \in \Gamma ^{\sigma_1 ,s_2, s_3; 0}_{(\omega )} ({\mathbf R^{3d}})$, i.e. $a\in C^\infty ({\mathbf R^{3d}})$ and $$|\partial_x ^\alpha \partial _\xi ^\beta \partial _\eta^\gamma a(x, \xi, \eta)|
\lesssim
h ^{|\alpha +\beta+ \delta|}\alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3} \omega (x, \xi, \eta), \;\;\;
x, \xi, \eta \in {\mathbf R^{d}},$$for every $ \alpha,\beta,\gamma\in {\mathbf N^{d}}$ and for every $h >0$;
2. For every $\phi \in \Sigma _{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3} ({\mathbf R^{3d}}) \setminus 0$, $h,R>0$ and $\alpha,\beta,\gamma\in {\mathbf N^{d}}$, it holds $$\begin{gathered}
|\partial _x ^\alpha \partial _\xi ^\beta \partial _\eta ^\gamma
{\left(e^{i({\langle x,\zeta\rangle} +{\langle y,\xi\rangle} +{\langle z,\eta\rangle} )}
V_\phi a(x,\xi, \eta, \zeta, y, z)\right)}|
\\[1ex]
\lesssim
h^{|\alpha+\beta+\gamma|} \alpha !^{\sigma_1} \beta!^{s_2} \gamma!^{s_3}
\omega (x, \xi, \eta)e^{-R (|\zeta|^{\frac 1{\sigma_1}}+ |y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})},\end{gathered}$$ $x,\xi, \eta, \zeta, y, z\in {\mathbf R^{d}}$.
3. For every $\phi \in \Sigma _{s_1,\sigma_2,\sigma_3}^{\sigma_1 ,s_2, s_3} ({\mathbf R^{3d}}) \setminus 0$, and $R >0$ it holds $$|V_\phi a(x,\xi, \eta, \zeta, y, z)|
\lesssim
\omega (x,\xi,\eta)e^{-R (|\zeta|^{\frac 1{\sigma_1}}+
|y |^{\frac 1{s_2}}+|z |^{\frac 1{s_3}})}, \;\;\;
x,\xi, \eta, \zeta, y, z\in {\mathbf R^{d}}.$$
In the next result we consider the Beurling case. It gives a description of the symbol class $\Gamma ^{\sigma_1,s_2,s_3;0}_{(\omega)} ({\mathbf R^{3d}})$ in terms of modulation spaces $M^{\infty,q}_{(1/\omega _R)} ({\mathbf R^{3d}})$ for $q \in[1,\infty]$ and $\omega_R$ defined in below. To prove Proposition \[Prop:GammaModIdent\*\] we follow arguments analogous to those used in the proofs of [@AbCaTo Proposition 3.5] and [@CaTo Proposition 4.4].
\[Prop:GammaModIdent\*\] Let $R>0$, $q\in [1,\infty ]$, $s_j,\sigma _j>0$, $(s_j,\sigma_j ) \neq (\frac 12,\frac 12)$, $j=1,2,3$, and let the conditions in hold. Also, let $\phi, \phi_0 \in \Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2 ,s_3} ({\mathbf R^{3d}})\setminus 0$, $\omega \in {\mathscr P}_{s_1,\sigma_2, \sigma_3}(\mathbf{R}^{3d})$, and let $$\label{eq:omegaR}
\omega _R(x,\xi, \eta, \zeta,y,z) =
\omega(x,\xi,\eta) e^{-R(|\zeta|^{\frac 1{\sigma_1}}+|y|^{\frac 1{s_2}}+|z|^{\frac 1{s_3}})}.$$ Then $$\label{iden*}
\Gamma ^{\sigma_1,s_2,s_3;0} _{(\omega)} ({\mathbf R^{3d}})=
\bigcap _{R>0}{\{ \, a\in
(\Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2,s_3})'
({\mathbf R^{3d}})\, ;\, {\Vert \omega _R^{-1}V_\phi a\Vert _{L^{\infty ,q}
({\mathbf R^{3d}}\times{\mathbf R^{3d}})}} <\infty \, \} }.$$
When $q =\infty$, becomes $$\Gamma ^{\sigma_1,s_2,s_3;0}_{(\omega)}=
\bigcap_{R>0} M^\infty_{(1/\omega _R)}
(\mathbf{R}^{3d}),$$ which is a straightforward consequence of Proposition \[prop:GammaOmegaChar\]. Therefore it is enough to prove that $$\begin{gathered}
\bigcap_{R>0} M^\infty_{(1/\omega _R)}
(\mathbf{R}^{3d})
\\[1ex]
=\bigcap _{R>0}{\{ \, a\in
(\Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2,s_3})'
({\mathbf R^{3d}})\, ;\, {\Vert \omega _R^{-1}V_\phi a\Vert _{L^{\infty ,q}
({\mathbf R^{3d}}\times{\mathbf R^{3d}})}} <\infty \, \} }.\end{gathered}$$
Put $$\begin{gathered}
V_{0,a}(X,Y)=|(V_{\phi _0}a)(x,\xi,\eta, \zeta, y, z)|,
\quad
V_{a}(X,Y)=|(V_{\phi}a)(x,\xi,\eta, \zeta, y, z)|
\\[1ex]
\text{and}\quad
G(x,\xi,\eta, \zeta, y, z)=|(V_\phi {\phi _0})(x,\xi,\eta, \zeta, y, z)|,\end{gathered}$$ where $X=(x,\xi, \eta) \in \mathbf{R}^{3d}$ and $Y=(\zeta, y, z) \in \mathbf{R}^{3d}$. By Proposition \[Prop:STFTGelfand2\] we have $$\label{GEst1}
0\le G(x,\xi,\eta, \zeta, y, z) \lesssim
e^{-R(|x|^{\frac 1{s_1}}+|\xi|^{\frac 1{\sigma_2}}+|\eta|^{\frac 1{\sigma_3}}
+|\zeta|^{\frac 1{\sigma_1}}+|y|^{\frac 1{s_2}}+|z|^{\frac 1{s_3}})},$$ $x,\xi,\eta, \zeta, y, z \in {\mathbf R^{d}}$, for every $R>0 .$
From [@Gro Lemma 11.3.3] (when extended to the duality between Gelfand-Shilov spaces and their dual spaces of distributions) it follows that $V_a \lesssim V_{0,a} \ast G$, so we obtain $$\label{pippo}
(\omega_R^{-1} \cdot V_a)(X,Y)
\lesssim
\left((\omega^{-1}_{cR} \cdot V_{0,a})*G_1\right)(X,Y),
, \;\;\;
X,Y \in \mathbf{R}^{3d},$$ for some $G_1$ which satisfies , and for a constant $c>0$ independent of $R$. By applying the $L^\infty$-norm on the both sides of we obtain $$\begin{gathered}
{\Vert \omega _R^{-1}V_a\Vert _{L^{\infty}({\mathbf R^{6d}})}} =
\sup _Y \sup _X {\left|\omega_R^{-1}V_a(X,Y)\right|}
\\[1ex]
\lesssim \sup _Y \sup _X {\left|\omega_{cR}^{-1}V_{0,a}*G_{1}(X,Y)\right|}
\\[1ex]
\lesssim \sup_{Y}\left(\iint
\big ( \sup _X(\omega^{-1}_{cR} \cdot V_{0,a})(X , Y-Y_1)\big )
G_1(X_1, Y_1)\, dX_1 dY_1\right)
\\[1ex]
\le
{\Vert \omega^{-1}_{cR} \cdot V_{0,a}\Vert _{L^{\infty ,q}}}
{\Vert G_1\Vert _{L^{1,q'}}}
\asymp {\Vert \omega^{-1}_{cR} \cdot V_{0,a}\Vert _{L^{\infty ,q}}},\end{gathered}$$ wherefrom $$\begin{gathered}
\bigcap _{R>0}{\{ \, a\in
(\Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2,s_3})'
({\mathbf R^{3d}})\, ;\, {\Vert \omega _R^{-1}V_\phi a\Vert _{L^{\infty ,q}
({\mathbf R^{3d}}\times{\mathbf R^{3d}})}} <\infty \, \} } \\[1ex]
\subset \bigcap_{R>0} M^\infty_{(1/\omega _R)} ({\mathbf R^{3d}}).\end{gathered}$$
For the opposite inclusion we put $K_j=\omega^{-1}_{jcR} \cdot V_{0,a}$, $j=1,2$. By and Minkowski’s inequality we have $$\begin{gathered}
\| \omega^{-1}_{R} \cdot V_{a} \| _{L^{\infty,q}}^q
\lesssim
\| \left((K_1*G\right)\| _{L^{\infty,q}}^q
\\[1ex]
\lesssim
\int \sup _X\left(\iint K _1 (X-X_1, Y-Y_1)G(X_1, Y_1)
\, dX_1 dY_1\right)^q\, dY
\\[1ex]
\lesssim \int
\left(\iint \sup \left(K _2({\, \cdot \, }, Y-Y_1)\right)G(X_1, Y_1) \right.
\\[1ex]
\left. \times e^{
-cR(|\zeta-\zeta_1|^{\frac 1{\sigma_1}}+|y-y_1|^{\frac 1{s_2}}+
|z-z_1|^{\frac 1{s_3}})}
\, dX_1 dY_1 \right)^q dY
\\[1ex]
\lesssim \| K _2 \|_{L^\infty}^q
\int \left(\iint G(X_1, Y_1)
e^{-cR(|\zeta-\zeta_1|^{\frac 1{\sigma_1}}+|y-y_1|^{\frac 1{s_2}}+
|z-z_1|^{\frac 1{s_3}})}\, dX_1 dY_1 \right)^q\, dY
\\[1ex]
\lesssim \|K _2\|_{L^\infty}^q\equiv\| \omega^{-1}_{2cR} \cdot V_{0,a} \|_{L^\infty}^q.\end{gathered}$$ Finally, by interchanging the roles of $\phi$ and $\phi _0$ we get $$\| \omega^{-1}_{R} \cdot V_{0,a} \| _{L^{\infty,q}} \lesssim
\| \omega^{-1}_{2cR} \cdot V_{a} \|_{L^\infty},$$ i.e. $$\begin{gathered}
\bigcap_{R>0} M^\infty_{(1/\omega _R)}
({\mathbf R^{3d}}) \\[1ex]
\subset
\bigcap _{R>0}{\{ \, a\in
(\Sigma _{s_1,\sigma_2, \sigma_3}^{\sigma_1,s_2,s_3})'
({\mathbf R^{3d}})\, ;\, {\Vert \omega _R^{-1}V_\phi a\Vert _{L^{\infty ,q}
({\mathbf R^{3d}}\times{\mathbf R^{3d}})}} <\infty \, \} },\end{gathered}$$ and the result follows.
We leave for the reader to write down Proposition \[Prop:GammaModIdent\*\] when the (Roumieu case) symbol class $\Gamma^{ \sigma_1,s_2, s_3}_{(\omega)}({\mathbf R^{3d}}) $ is considered instead.
In [@CaTo Theorem 4.1], it is shown that if $A$ is a $d\times d$-matrix with real entries, then the operator $e^{i{\langle AD_\xi,D_x\rangle}}$ is a homeomorphism between certain classes of symbols. We proceed with an analogous result in the context of the symbol class $\Gamma _{(\omega)} ^{\sigma_1,s_2,s_3;0}({\mathbf R^{3d}})$.
\[thm:CalculiTransfbis\] Let $s_j ,\sigma _j > 0$ be such that the conditions in hold, and let $r,t\in [0,1]$ be such that $r+t \leq 1.$
If $\omega \in {\mathscr P}_{s_1,\sigma_2,\sigma_3}^0({\mathbf R^{3d}})$, then $a\in \Gamma _{(\omega)} ^{\sigma_1,s_2,s_3}({\mathbf R^{3d}})$ if and only if $$e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}
a\in \Gamma _{(\omega)} ^{\sigma_1,s_2,s_3}({\mathbf R^{3d}}).$$
If $\omega \in {\mathscr P}_{s_1,\sigma_2,\sigma_3}({\mathbf R^{3d}})$ instead, and if, in addition to , $(s_j,\sigma_j ) \neq (\frac 12,\frac 12)$, $j = 1,2,3$, then $a\in \Gamma _{(\omega)} ^{\sigma_1,s_2,s_3;0}({\mathbf R^{3d}})$ if and only if $$e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}} a\in \Gamma _{(\omega)} ^{\sigma_1,s_2,s_3;0}({\mathbf R^{3d}}).$$
We give the proof for the Beurling case, and the Roumieu case is left for the reader.
We will use the result of Proposition \[Prop:GammaModIdent\*\]. Therefore we fix a window function $\phi \in \Sigma _{s_1,\sigma_2,\sigma_3}^{\sigma_1, s_2,s_3}({\mathbf R^{3d}})$ and let $\phi _{r,t} =e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}\phi$. Then, in view of Theorem \[Thm:CalculiTransf\] (2), $\phi _{r,t}$ belongs to $\Sigma _{s_1,\sigma_2,\sigma_3}^{\sigma_1, s_2,s_3} ({\mathbf R^{3d}})$.
By similar arguments as in the proof of Lemma \[lem:FTInvOpr\], we get $$\begin{gathered}
\label{eq:ModSTFTequa}
|(V_{\phi _{r,t}} (e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a))(x,\xi ,\eta,\zeta ,y,z)|
\\
=
|(V_\phi a)(x+ry+tz,\xi +r\zeta,\eta+t\zeta ,\zeta ,y,z)|,\end{gathered}$$ $ x,\xi ,\eta,\zeta ,y,z \in {\mathbf R^{d}}$. Then using and a change of variables argument, we get $${\Vert \omega _{0,0;R}^{-1}V_{\phi} a\Vert _{L^{p,q}}} =
{\Vert \omega _{r,t;R}^{-1}V_{\phi _{r,t}}
(e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a)\Vert _{L^{p,q}}},$$ where $$\omega _{r,t;R}(x,\xi ,\eta,\zeta ,y,z) = \omega (x+ry+tz,\xi +r\zeta,\eta+t\zeta )
e^{-R(|y|^{\frac 1{s_2}}+|z|^{\frac 1{s_3}}+|\zeta |^{\frac 1{\sigma_1}})},$$ and $p,q \in [1,\infty]$.
Hence Proposition \[Prop:GammaModIdent\*\], and the fact that there exists a constant $c>0$ such that $$\omega _{0,0;R+c}\lesssim \omega _{r,t;R}\lesssim \omega _{0,0;R-c},$$ give $$\begin{gathered}
a\in\Gamma _{(\omega)} ^{\sigma_1,s_2,s_3;0}({\mathbf R^{3d}})
\quad \Leftrightarrow \quad
{\Vert \omega _{0,0;R}^{-1}V_\phi a\Vert _{L^\infty}}<\infty \;
\text{for every}
\ R>0
\\[1ex]
\Leftrightarrow \quad
{\Vert \omega _{r,t;R}^{-1}V_{\phi _{r,t}}
(e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a)\Vert _{L^\infty}} <\infty \;
\text{for every} \ R>0
\\[1ex]
\Leftrightarrow \;
{\Vert \omega _{0,0;R}^{-1}V_{\phi _{r,t}}
(e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a)\Vert _{L^\infty}}
<\infty \;
\text{for every} \ R>0
\\[1ex]
\Leftrightarrow \quad
e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a \in
\Gamma _{(\omega)} ^{\sigma_1,s_2,s_3;0} ({\mathbf R^{3d}}),\end{gathered}$$ and the result follows.
Continuity of bilinear pseudo-differential operators with symbols of Gevrey-regularity and infinite order {#sec3}
=========================================================================================================
We first discuss the continuity of bilinear operators in ${\operatorname{Op}}(\Gamma ^{\sigma_1,s_2,s_3}_{(\omega)})$ and ${\operatorname{Op}}(\Gamma ^{\sigma_1,s_2,s_3;0}_{(\omega)})$ when acting on products modulation spaces. In particular, Theorem \[thm:p3.2\] can be considered as an extension of [@To14 Theorem 3.2] to bilinear operators and a more general class of weights.
\[thm:p3.2\] Let $s_j,\sigma _j>0$ be such that the conditions in hold. Also, let $v _1\in {\mathscr P}_{s_1} ^0({\mathbf R^{d}})$, $v _j\in{\mathscr P}_{\sigma_j} ^0({\mathbf R^{d}})$, $j=2,3$, $\omega _0,\omega \in{\mathscr P}_{s_1,\sigma_2,\sigma_3}^0({\mathbf R^{3d}})$, and let $\omega _0$ be $\otimes _{j=1}^3 v_j$-moderate. Furthermore, let $r,t\in [0,1]$ such that $r+t\leq 1$, and let $p,q \in [1,\infty]$. If $a\in \Gamma _{(\omega _0)}^{\sigma_1,s_2,s_3}({\mathbf R^{3d}})$, then there exists $R>0$ such that ${\operatorname{Op}}_{r,t}(a)$ is continuous from $M_{(\omega _0\omega)}^{p,q} ({\mathbf R^{d}}) \times M _{(1/\omega_R)}^{\infty,\infty} ({\mathbf R^{d}}) $ to $M _{(\omega)}^{p,q} ({\mathbf R^{d}}) $, where $$\omega _R(x,\xi,\eta)=e^{-R(|x|^{\frac 1{s_1}} +|\xi|^{\frac 1{\sigma_2}}+|\eta|^{\frac 1{\sigma_3}})},
\;\;\; x,\xi,\eta \in {\mathbf R^{d}}.$$
\[rem:BiltoLin\] We will use estimates similar to those obtained in the proof of [@BeMaNaTo Theorem 6.1]. We observe that out arguments are anyway different since, in view of the fact that we employ Gevrey tpye symbols, we cannot rely on standard localization techniques. The idea is that for a fixed function $g$ in appropriate space of test functions, $a\in \Gamma _{(\omega )}^{\sigma_1,s_2,s_3}({\mathbf R^{3d}})$, and $r=t=0$, the operator ${\operatorname{Op}}_{0,0}(a)({\, \cdot \, },g)\equiv T_a({\, \cdot \, }, g)$ can be regarded as a linear pseudo-differential operator (with symbol depending on $g$), that is, $${\operatorname{Op}}_{0,0}(a)(f,g)(x)=(2\pi)^{-\frac{d}{2}}
\int e^{i {\langle x,\xi\rangle}} a_g(x,\xi)\widehat{f}(\xi)\,d\xi,$$ where $$\label{eq:a_gDef}
a_g(x,\xi)=(2\pi)^{-\frac{d}{2}}
\int e^{i {\langle x,\eta\rangle}} a(x,\xi,\eta)\widehat{g}(\eta)\,d\eta.$$ If the symbol $a_g$ belongs to $\Gamma ^{\sigma_1,s_2}_{(\omega_0)} ( {\mathbf R^{2d}})$, then the continuity of ${\operatorname{Op}}_{0,0}(a)({\, \cdot \, },g)$ from $M^{p.q}_{(\omega_0\omega)}$ to $M^{p.q}_{(\omega)}$ follows by Proposition \[thm:anisCntp\].
\[lemma:a\_gClass\] Let $s_j,\sigma _j>0$ be such that the conditions in hold. Also, let $v _1\in {\mathscr P}_{s_1} ^0({\mathbf R^{d}})$, $v _j\in{\mathscr P}_{\sigma_j} ^0({\mathbf R^{d}})$, $j=2,3$, $\omega _0,\omega \in{\mathscr P}_{s_1,\sigma_2,\sigma_3}^0({\mathbf R^{3d}})$, and let $\omega _0$ be $\otimes _{j=1}^3 v_j$-moderate.
If $g\in {\mathcal S}_{s_1}^{\sigma_1}({\mathbf R^{d}})$ and $a\in \Gamma _{(\omega_0 )}^{\sigma_1,s_2,s_3}({\mathbf R^{3d}})$, then the symbol $a_g$ given by belongs to $\Gamma _{(\omega )}^{\sigma_1,s_2}({\mathbf R^{2d}})$, where $\omega(x,\xi)
\equiv \omega _0(x,\xi,0) \in{\mathscr P}_{s_1,\sigma_2}^0({\mathbf R^{2d}})$.
By it follows that $a_g$ is a smooth function. Indeed, $$(x,\xi,\eta)\mapsto
e^{i {\langle x,\eta\rangle}} a(x,\xi,\eta)\widehat{g}(\eta)$$ is a smooth mapping and $$\eta\mapsto
\eta^\gamma e^{i {\langle x,\eta\rangle}} \partial _x^\alpha
\partial _\xi^\beta a(x,\xi,\eta)\widehat{g}(\eta)$$ is an integrable function for every $x,\xi, \alpha,\beta$ and $\gamma$.
Since $\widehat{g} \in {\mathcal S}^{s_1} _{\sigma_1}({\mathbf R^{d}})$ (cf. Proposition \[propBroadGSSpaceChar\]) and since $\omega _0\in {\mathscr P}_{s_1,\sigma_2,\sigma_3}^0 ({\mathbf R^{3d}})$, it follows that $$\begin{gathered}
{\left|\partial _x^\alpha\partial _\xi^\beta a_g(x,\xi)\right|}
\lesssim
\sum _{\gamma \leq \alpha}\binom{\alpha}{\gamma}
\int {\left|\eta^{\gamma}\partial _x^{\alpha-\gamma}
\partial _\xi^\beta a(x,\xi,\eta)\widehat{g}(\eta)\right|}\, d\eta
\\[1ex]
\lesssim
\sum _{\gamma \leq \alpha}\binom{\alpha}{\gamma}
\int
h^{|\alpha+\beta-\gamma|}(\alpha-\gamma)!^{\sigma_1} \beta!^{s_2}
{\left|\omega _0(x,\xi,\eta)
\eta^\gamma e^{-r|\eta|^{\frac 1{\sigma_1}}}\right|}
\, d\eta
\\[1ex]
\lesssim
\sum _{\gamma \leq \alpha}\binom{\alpha}{\gamma}
\int
h^{|\alpha+\beta-\gamma|}(\alpha-\gamma)!^{\sigma_1} \beta!^{s_2}
\omega _0(x,\xi,0){\left|e^{r_0|\eta|^{\frac 1{\sigma_3}}}
\eta^\gamma e^{-r|\eta|^{\frac 1{\sigma_1}}}\right|}
\, d\eta,\end{gathered}$$ for every $r_0>0$, and some constants $r,h>0$. Since $r_0$ can be chosen such that $r_0<r$, and since $${\left|\eta^\gamma e^{-(r-r_0)|\eta|^{\frac 1{\sigma_1}}}\right|}
\lesssim h^{|\gamma|}\gamma!^{\sigma_1}
e^{-\frac{(r-r_0)}2|\eta|^{\frac 1{\sigma_1}}},$$we get $$\begin{gathered}
{\left|\partial _x^\alpha\partial _\xi^\beta a_g(x,\xi)\right|}
\\[1ex]
\lesssim
h^{|\alpha+\beta|} \beta!^{s_2}
\sum _{\gamma \leq \alpha}\binom{\alpha}{\gamma}
{\left((\alpha-\gamma)!\gamma!\right)}^{\sigma_1} \int
\omega _0(x,\xi,0)
e^{-(r-r_0)|\eta|^{\frac 1{\sigma_1}}}
\, d\eta,
\\[1ex]
\lesssim (4h)^{|\alpha+\beta|}\alpha!^{\sigma_1} \beta!^{s_2}
\omega(x,\xi), \;\;\; x,\xi \in {\mathbf R^{d}},\end{gathered}$$ for some constant $h>0$, where $\omega(x,\xi)
\equiv \omega _0(x,\xi,0)\in {\mathscr P}_{s_1,\sigma_2}^0({\mathbf R^{2d}})$. This gives the desired result.
In view of the invariance properties for the bilinear pseudo-differential operators given in Theorem \[thm:CalculiTransfbis\], we may assume $r=t=0$ without loss of generality.
By Proposition \[thm:anisCntp\] and Lemma \[lemma:a\_gClass\], it follows that ${\operatorname{Op}}(a)(f,g)$ is a continuous mapping from $M^{p.q}_{(\omega_0\omega)} ({\mathbf R^{d}}) \times{\mathcal S}_{s_1} ^{\sigma_1} ({\mathbf R^{d}})$ to $M^{p.q}_{(\omega)} ({\mathbf R^{d}})$. Now the result follows from Proposition \[Prop:STFTGelfand2\].
In a similar way, we get the result for the Beurling case. The details are left for the reader.
\[thm:p3.2bis\] Let $s_j,\sigma _j>0$, $(s_j,\sigma_j ) \neq (\frac 12,\frac 12)$, $j=1,2,3$, be such that the conditions in hold. Also, let $v _1\in {\mathscr P}_{s_1} ^0({\mathbf R^{d}})$, $v _j\in{\mathscr P}_{\sigma_j} ^0({\mathbf R^{d}})$, $j=2,3$, $\omega _0,\omega \in{\mathscr P}_{s_1,\sigma_2,\sigma_3}^0({\mathbf R^{3d}})$, and let $\omega _0$ be $\otimes _{j=1}^3 v_j$-moderate. Furthermore, let $r,t\in [0,1]$ such that $r+t\leq 1$, and let $p,q \in [1,\infty]$.
If $a\in \Gamma _{(\omega _0)}^{\sigma_1,s_2,s_3;0}({\mathbf R^{3d}})$, then for any $R>0$ the operator ${\operatorname{Op}}_{r,t}(a)$ is continuous from $M_{(\omega _0\omega)}^{p,q} ({\mathbf R^{d}}) \times M _{(1/\omega_R)}^{\infty,\infty} ({\mathbf R^{d}}) $ to $M _{(\omega)}^{p,q} ({\mathbf R^{d}}) $, where $$\omega _R(x,\xi,\eta)=e^{-R(|x|^{\frac 1{s_1}} +|\xi|^{\frac 1{\sigma_2}}+|\eta|^{\frac 1{\sigma_3}})},
\;\;\; x,\xi,\eta \in {\mathbf R^{d}}.$$
Finally, the characterization of Gelfand-Shilov spaces via modulation spaces gives the following result (cf. [@GZ; @Te6; @To18]), see also Proposition \[Prop:STFTGelfand2\].
\[thm:CntGSspaces\] Let there be given $s,\sigma >0$ such that $s+\sigma \ge 1$, $v _1\in {\mathscr P}_s^0({\mathbf R^{d}})$, $v _j\in{\mathscr P}_\sigma^0({\mathbf R^{d}})$, $j=2,3$, and $\omega_0 \in{\mathscr P}_{s,\sigma,\sigma}^0({\mathbf R^{3d}})$, such that $\omega_0 $ is $\otimes _{j=1}^3 v_j$-moderate. If $r,t\in [0,1]$, such that $r+t\leq 1$, and $a\in \Gamma _{(\omega )}^{\sigma,s,s}({\mathbf R^{3d}})$ then ${\operatorname{Op}}_{r,t}(a)$ is continuous from ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})\times
{\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ to ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$, and from $({\mathcal S}_s^\sigma )'({\mathbf R^{d}})\times({\mathcal S}_s^\sigma )'({\mathbf R^{d}})$ to $({\mathcal S}_s^\sigma )'({\mathbf R^{d}})$.
In view of Theorem \[thm:CalculiTransfbis\], it is enough to consider the case when $r=t=0$, i.e. ${\operatorname{Op}}_{0,0}(a)$.
By Proposition \[Thm:theorem2\] and Remark \[rem:BiltoLin\], it is enough to show that $a_g$ given by belongs to $ \Gamma _{(\omega)} ^{\sigma ,s}({\mathbf R^{2d}})$ for $\omega(x,\xi) \equiv \omega _0(x,\xi,0)\in {\mathscr P}_{s,\sigma}^0({\mathbf R^{2d}})$. This follows from Lemma \[lemma:a\_gClass\]. Now the the continuity of ${\operatorname{Op}}_{r,t}(a)$ from ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})\times {\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ to ${\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ follows from Theorem \[thm:p3.2\] and Proposition \[Prop:STFTGelfand2\].
The continuity of ${\operatorname{Op}}_{r,t}(a)$ from $({\mathcal S}_s^\sigma )'({\mathbf R^{d}})\times({\mathcal S}_s^\sigma )'({\mathbf R^{d}})$ to $({\mathcal S}_s^\sigma )'({\mathbf R^{d}})$ follows by duality.
The analogous result given below follows by similar arguments and is left for the reader.
\[thm:CntGSspacesbis\] Let there be given $s,\sigma >0$ such that $s+\sigma \ge 1$ and $ (s,\sigma) \neq (\frac{1}{2},\frac{1}{2}).$ Also, let $v _1\in {\mathscr P}_s^0({\mathbf R^{d}})$, $v _j\in{\mathscr P}_\sigma^0({\mathbf R^{d}})$, $j=2,3$, and $\omega_0 \in{\mathscr P}_{s,\sigma,\sigma}^0({\mathbf R^{3d}})$, such that $\omega_0 $ is $\otimes _{j=1}^3 v_j$-moderate. If $r,t\in [0,1]$, such that $r+t\leq 1$, and $a\in \Gamma _{(\omega )}^{\sigma,s,s;0}({\mathbf R^{3d}})$ then ${\operatorname{Op}}_{r,t}(a)$ is continuous from $\Sigma _{s}^{\sigma}({\mathbf R^{d}})\times
\Sigma _{s}^{\sigma}({\mathbf R^{d}})$ to $\Sigma _{s}^{\sigma}({\mathbf R^{d}})$, and from $(\Sigma _s^\sigma )'({\mathbf R^{d}})\times( \Sigma _s^\sigma )'({\mathbf R^{d}})$ to $(\Sigma _s^\sigma )'({\mathbf R^{d}})$.
Proofs of some auxiliary results {#appendix}
================================
In this concluding section we collect the proofs of some technical results employed above.
We give the proof for the case $ \Sigma _{s}^{\sigma}({\mathbf R^{d}})$ (the Beurling case) only, and the (simpler) case $ {\mathcal S}_{s}^{\sigma}({\mathbf R^{d}})$ (the Roumieu case) is left for the reader. For simplicity (to avoid the use of inequalities related to multi-indices) we here set $d= 1$.
Note that $$\sup_{x} |x^p f^{(q)} (x)| \leq C k^p l^{q} p!^s q!^{\sigma}, \;\;\; \forall k,l >0,$$ is equivalent to $$\sup_{x} |x^p f^{(q)} (x)| \leq C h^{p + q} p!^s q!^{\sigma}, \;\;\; \forall h >0.$$ In fact, if the former condition holds, then we may put $l=k=h$ for any given $h>0$, so the later condition holds. If the later condition holds instead, then for any given $ k,l >0$ we may choose $h = \min\{ k,l\}$ to obtain $$\sup_{x} |x^p f^{(q)} (x)| \leq C \min\{ k,l\}^{p + q} p!^s q!^{\sigma} \leq C k^p l^{q} p!^s q!^{\sigma},$$ and the equivalence follows.
Assume that $f \in \Sigma _{s}^{\sigma}({\mathbf R})$, that is, for every $p,q \in {\mathbf N^{}}$, $$\label{prviuslov}
\sup_{x \in {\mathbf R}} | x^p f^{(q)} (x)| \lesssim k^p l^q p!^s q!^{\sigma}, \;\;\; \forall k,l >0.$$ Put $F_q (x, l) = f^{(q)} (x) / (l^{q} q!^{\sigma}),$ $ x \in {\mathbf R}$. Then, by $$\sup_{x} k^{-p} p !^{-s} | x|^{p} | F_q (x,l) | < \infty \;\;
\Rightarrow \;\;
\sup_{x} k^{-\frac{p}{s}} p !^{-1} | x|^{\frac{p}{s}} | F_q (x,l) |^{\frac{1}{s}} \leq C,$$ for some constant $C>0,$ uniformly in $p,$ so that $$\sup_{x} \sum_{p \in \mathbb{N}} \left ( \frac{(| x|k^{-1})^{1/s}}{2} \right )^{p} p !^{-1}
| F_q (x,l) |^{1/s} \leq C \sum_{p \in \mathbb{N}} \frac{1}{2^{p}}.$$ For any given $h>0$ we may choose $ k = (s/ 2h )^s $ to obtain $$\begin{gathered}
\sup_{x} \sum_{p \in \mathbb{N}} ( \frac{\left ( \left(| x|\left (\frac{2h}{s}\right )^s \right )^{1/s} \right )^{p} }{2^{p} p !}
| F_q (x,l) |^{1/s} \\
=
\sup_{x} \sum_{p \in \mathbb{N}} \frac{\left ( | x|^{1/s} \frac{h}{s} \right )^p}{ p !}
| F_q (x,l) |^{1/s}
\leq \tilde C.\end{gathered}$$ Therefore $$e^{\frac{h}{s} |x|^{1/s}} \sup_{x} | F_q (x,l) |^{1/s} \leq \tilde C \;\;\; \Rightarrow \;\;\;
| F_q (x,l) |^{1/s} \leq \tilde C e^{-\frac{h}{s} |x|^{1/s}}, \;\;\; \forall h,l>0,$$ and we conclude that $$\frac{|f^{(q)} (x)|}{l^{q} q!^{\sigma}} \leq \tilde{C} ^s e^{-h |x|^{1/s}}
\;\;\; \Leftrightarrow \;\;\;
|f^{(q)} (x)| \leq C l^{q} q!^{\sigma} e^{-h |x|^{1/s}}, \;\;\; \forall h,l >0,$$ which proves the first part of Lemma.
Now, assume that holds, so that for $F_q (x,l) = f^{(q)} (x) / (l^{q} q!^{\sigma})$ we have $$|F_q (x, l)|^{1/s} e^{\frac{h}{s} |x|^{1/s}} \leq C, \;\;\; \forall h,l>0,$$ for some $C>0$ uniformly in $x\in {\mathbf R}$. Therefore $$\sup_{x} \sum_{p \in \mathbb{N}} \frac{1}{p !}
(\frac{h}{s})^p |x|^{p/s} \left | F_q (x, l) \right|^{1/s} < \infty,$$ which implies $$\frac{1}{p !} \left (\frac{h}{s} \right )^p |x|^{p/s} \left | F_q (x,l) \right|^{1/s} < \infty
\;\;\;
\Leftrightarrow
\;\;\;
\frac{1}{p !^s} \left (\frac{h}{s} \right )^{ps} |x|^{p} \left | F_q (x, l) \right| < \infty,$$ that is $$|x^{p} f^{(q)} (x) | \leq C \left (\frac{s^s}{h^s} \right )^{p} l^{q} p !^s q!^{\sigma},
\;\;\; \forall h,l >0.$$ Finally, for a given $k>0$ by choosing $h = s/ k^{1/s}$, we have $ k = (s/h)^s $ and $$|x^{p} f^{(q)} (x) | \leq C k^{p} l^{q} p !^s q!^{\sigma},$$ which completes the proof.
It is enough to prove (2) for every $0<s_j < 1$, $j=1,\dots, k$. In addition, we assume that $(s_0,\dots,s_k)\neq (1/2,\dots,1/2)$, and leave the necessary modifications when $(s_0,\dots,s_k)= (1/2,\dots,1/2)$ for the reader.
Let $\phi_0 \in \Sigma _{1-s_0,\dots,1-s_k}^{s_0,\dots, s_k}({\mathbf R^{d_0 +\dots + d_k}})\setminus 0$, and $\phi =|\phi _0|^2$. Then $\phi \in _{1-s_0,\dots,1-s_k}^{s_0,\dots, s_k}({\mathbf R^{d_0 +\dots + d_k}})\setminus 0$, i.e. $$\begin{gathered}
|\partial _{x} ^{\alpha_0}\partial _{\xi_1} ^{\alpha_1}
\dots \partial _{\xi_k} ^{\alpha_k} \phi (x, \xi_1,\dots,\xi_k)|
\lesssim
h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j}
\\[1ex]
\times
e^{-c(|x|^{\frac 1{1-s_0}}+|\xi_1|^{\frac 1{1-s_1}}+\dots +|\xi_k|^{\frac 1{1-s_k}})},\end{gathered}$$ for every $h>0$ and $c>0$.
Now let $\omega _0=\omega *\phi$. Then we have $$\begin{gathered}
|\partial _{x} ^{\alpha_0}\partial _{\xi_1} ^{\alpha_1}
\dots \partial _{\xi_k} ^{\alpha_k}
\omega _0 (x, \xi_1,\dots,\xi_k)|
\\[1ex]
= | \int_{{\mathbf R^{d_0 +\dots + d_k}}} \omega (x -y, \xi_1 - \eta_1, \dots, \xi_k - \eta_k)
\\[1ex]
\times
(\partial _{x} ^{\alpha_0}\partial _{\xi_1} ^{\alpha_1}
\dots \partial _{\xi_k} ^{\alpha_k} \phi )
(y, \eta_1, \dots, \eta_k)\, dy d\eta_1 \dots d\eta_k |\end{gathered}$$ $$\begin{gathered}
\lesssim
h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j}
\times
\iint\cdots\int \omega (x -y, \xi_1 - \eta_1, \dots, \xi_k - \eta_k)
\\[1ex]
\times
e^{-c(|y|^{\frac 1{1-s_0}}+|\eta_1|^{\frac 1{1-s_1}}+\dots +|\eta_k|^{\frac 1{1-s_k}})}
\, dy d\eta_1 \cdots d\eta_k\end{gathered}$$ $$\begin{gathered}
\lesssim
h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j} \times\omega (x, \xi_1,\dots,\xi_k)
\\[1ex]
\times
\iint\cdots\int
e^{-\frac{c}{2}(|y|^{\frac 1{1-s_0}}+|\eta_1|^{\frac 1{1-s_1}}+\dots +|\eta_k|^{\frac 1{1-s_k}})}
\, dy d\eta_1 \dots d\eta_k\end{gathered}$$ $$\asymp h^{|\alpha_0 +\alpha_1+\dots+\alpha_k|}\prod_{j=0} ^k \alpha_j !^{s_j} \times\omega (x, \xi_1,\dots,\xi_k),$$ where we used , , and the fact that for any $r,c>0$ and $s\in (0,1)$ it holds $ \displaystyle e^{r|\cdot|} \lesssim e^{\frac{c}{2}|\cdot|^{1/s}}.$ This gives the first part of (2) and $ \omega _0 \lesssim \omega$, by choosing $\alpha_j = 0$, $ j =0,\dots,k$.
Now, for $ \omega _0 \gtrsim \omega $ we note that and imply $$\begin{gathered}
\omega (x -y, \xi_1 - \eta_1, \dots, \xi_k - \eta_k) \gtrsim
\frac{\omega (x , \xi_1, \dots, \xi_k )}{v (-y, - \eta_1, \dots, - \eta_k)} \\[1ex]
\gtrsim \omega (x , \xi_1, \dots, \xi_k ) e^{-r(|y|+|\eta_1|+\dots +|\eta_k|)},\end{gathered}$$ for some $r>0.$ Therefore, $$\begin{gathered}
|\omega _0 (x, \xi_1,\dots,\xi_k)| = | \omega *\phi(x, \xi_1,\dots,\xi_k)|
\\[1ex]
\left | \iint\cdots\int
\omega (x -y, \xi_1 - \eta_1, \dots, \xi_k - \eta_k)
\phi (y, \eta_1, \dots, \eta_k)\, dy d\eta_1 \cdots d\eta_k \right |
\\[1ex]
\gtrsim \omega (x , \xi_1, \dots, \xi_k ) \\[1ex]
\times
\iint\cdots\int e^{-r(|y|^{\frac 1{1-s_0}}+|\eta_1|^{\frac 1{1-s_1}}+\dots +|\eta_k|^{\frac 1{1-s_k}})}
\phi (y, \eta_1, \dots, \eta_k)\, dy d\eta_1 \cdots d\eta_k
\\[1ex]
\asymp \omega (x, \xi_1,\dots,\xi_k),\end{gathered}$$ for some $r>0,$ so that $ \omega _0 \gtrsim \omega $, and (1) follows. This also gives the second part of (2).
It is no restriction to assume that $r_2=t_2=0$. Let $(r,t)=(r_1,t_1)$. The proof is a straightforward application of the Fourier inversion formula, see also [@To24 Proposition 1.1] and [@Ho1 Section 18.5] .
The equality ${\operatorname{Op}}_{r,t}(a)={\operatorname{Op}}(b)$ is the same as $${\mathscr F}_{2,3} (b(x,{\, \cdot \, },\cdot\cdot ))(y-x,z-x)
=
{\mathscr F}_{2,3} (a(x+r(y-x)+t(z-x),{\, \cdot \, },\cdot\cdot ))(y-x,z-x)$$$$\Leftrightarrow$$ $$\begin{gathered}
\iint b(x,\xi_1,\eta_1 )e^{i({\langle y,\xi_1\rangle} + {\langle z,\eta_1\rangle} )}\, d\xi_1 d\eta_1
\\[1ex]
= \iint a(x+ry+tz,\xi_1,\eta_1 )e^{i({\langle y,\xi_1\rangle} + {\langle z,\eta_1\rangle} )}\, d\xi_1 d\eta_1\end{gathered}$$$$\Leftrightarrow$$ $$b(x,\xi,\eta ) = \frac{1}{(2\pi )^{2d}} \iiiint a(x+ry+tz,\xi_1,\eta_1 )
e^{i({\langle y,\xi_1-\xi\rangle} + {\langle z,\eta_1-\eta\rangle} )}\, dydzd\xi_1 d\eta_1.$$Let $Y=(y,z)$, $Y_2=(y _2,z _2)$, $\Xi=(\xi,\eta)$ and $\Xi _1=(\xi _1,\eta _1)$. Then by the Fourier inversion formula we get $$\begin{gathered}
b(x,\xi,\eta )=
\\[1ex]
(2\pi )^{-\frac{7d}{2}}\iint\!\!\!\iiint\widehat a(\zeta ,y_2,z_2 )
e^{i( {\langle x+ry+tz,\zeta\rangle}+{\langle Y_2,\Xi_1\rangle} +{\langle Y,\Xi_1-\Xi\rangle} )}
\,dz_2dy_2d\zeta dYd\Xi_1
\\[1ex]
= (2\pi )^{-\frac{3d}{2}}
\iiint \widehat a(\zeta ,y_2,z_2 )
e^{i( {\langle x,\zeta\rangle}-{\langle ry_2+tz_2,\zeta\rangle}
+{\langle y_2,\xi\rangle}+{\langle z_2,\eta\rangle})}
\,dz_2dy_2d\zeta
\\[1ex]
= (2\pi )^{-\frac{3d}{2}}
\iiint e^{i({\langle y_2,\xi\rangle}+{\langle z_2,\eta\rangle}+{\langle x,\zeta\rangle})}
{\left(e^{-i{\langle ry_2+tz_2,\zeta\rangle}}\widehat{a}(\zeta ,y_2,z_2)\right)}
\,dz_2dy_2d\zeta
\\[1ex]
=
e^{-i{\langle rD_\xi+tD_\eta,D_x\rangle}}a(x,\xi,\eta ),\end{gathered}$$ which gives the result.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors are grateful to professor Joachim Toft for valuable remarks and suggestions. N. Teofanov is partially supported by MPNTR Project No 174024.
[2000]{} A. Abdeljawad, M. Cappiello, J. Toft, *Pseudo-differential calculus in an anisotropic Gelfand-Shilov setting*, Integr. Equ. Oper. Theory (2019) 91: 26.
A. Abdeljawad, S. Coriasco, J. Toft, *Liftings for ultra-modulation spaces, and one-parameter groups of Gevrey type pseudo-differential operators*, (preprint), arXiv:1712.04338.
A. Abdeljawad, J. Toft, *Anisotropic Gevrey-H[ö]{}rmander pseudo-differential operators on modulation spaces*, (preprint), arXiv:1806.10002.
Á. Bényi, T. Oh, *On a Class of Bilinear Pseudodifferential Operators*, Journal of Function Spaces and Applications, vol. 2013, Article ID 560976, 5 pages, 2013. https://doi.org/10.1155/2013/560976.
Á. Bényi, K. A. Okoudjou, *Bilinear pseudodifferential operators on modulation spaces*, J. Fourier Anal. Appl. **10** (2004), 301–313.
Á. Bényi, K. A. Okoudjou, *Modulation spaces estimates for multilinear pseudodifferential operators*, Studia Math., **172** (2006), 169–180.
Á. Bényi, K. Groechenig, C. Heil, K. Okoudjou, *Modulation spaces and a class of bounded multilinear pseudodifferential operators*, J. Operator Theory, **54** (2005), 389–401.
Á. Bényi, D. Maldonado, V. Naibo, R.H. Torres, *On the Hörmander classes of bilinear pseudodifferential operators*, Integral Equations and Operator Theory **67** (2010), 341-364.
M. Cappiello, J. Toft, *Pseudo-differential operators in a Gelfand-Shilov setting*, Math. Nachr. **290** (2017), 738–755.
, *Characterizations of the Gelfand-Shilov spaces via Fourier transforms*, Proc. Amer. Math. Soc. **124** (1996), 2101–2108.
R.R. Coifman, Y. Meyer, *Au delá des opérateurs pseudo-differentiels*, Astérisque **57**, Société Math. de France (1978)
E. Cordero, S. Pilipović, L. Rodino, N. Teofanov, *Localization operators and exponential weights for modulation spaces*. Mediterr. J. Math. **2** (2005), 381–394.
E. Cordero, K. A. Okoudjou, *Multilinear localization operators*, J. Math. Anal. Appl. **325** (2007), 1103–1116.
H. G. Feichtinger, *Banach spaces of distributions of Wiener’s type and interpolation,* in: P. Butzer, B. Sz. Nagy and E. G[ö]{}rlich (Eds), Proc. Conf. Oberwolfach, Functional Analysis and Approximation, August 1980, Int. Ser. Num. Math. **69** Birkh[ä]{}user Verlag, Basel, Boston, Stuttgart, 1981, pp. 153–165.
H. G. Feichtinger, *Banach convolution algebras of Wiener’s type, [in: Proc. Functions, Series, Operators in Budapest]{}*, [Colloquia Math. Soc. J. Bolyai, North Holland Publ. Co.]{}, Amsterdam Oxford NewYork, 1980.
H. G. Feichtinger, *Modulation spaces on locally compact abelian groups. Technical report*, [University of Vienna]{}, Vienna, 1983; also in: M. Krishna, R. Radha, S. Thangavelu (Eds) Wavelets and their applications, Allied Publishers Private Limited, NewDehli Mumbai Kolkata Chennai Hagpur Ahmedabad Bangalore Hyderbad Lucknow, 2003, pp.99–140.
, *Banach spaces related to integrable group representations and their atomic decompositions, I*, J. Funct. Anal. **86** (1989), 307–340.
, *Banach spaces related to integrable group representations and their atomic decompositions, II*, Monatsh. Math. **108** (1989), 129–148.
, *Gabor frames and time-frequency analysis of distributions*, [J. Functional Anal. (2)]{} **146** (1997), 464–495.
Y. V. Galperin, S. Samarah, *Time-frequency analysis on modulation spaces $M^{p,q}_m$, $0<p,q\le \infty$*, Appl. Comput. Harmon. Anal. **16** (2004), 1–18.
, *Generalized functions, I–III*, Academic Press, NewYork London, 1968.
, *Multilinear Calderón–Zygmund theory*, Adv. Math., **165** (2002), 124–164/
, *Gelfand-Shilov spaces: structural properties and applications to pseudodifferential operators in $\mathbb{R}^n$*, in Quantization, PDEs, and geometry, 1–68, Oper. Theory Adv. Appl. 251, Birkhäuser/Springer, 2016.
, *Foundations of Time-Frequency Analysis*, [Birkh[ä]{}user]{}, Boston, 2001.
K. Gröchenig, *Weight functions in time-frequency analysis [in: L. Rodino, M. W. Wong (Eds) Pseudodifferential Operators: Partial Differential Equations and Time-Frequency Analysis]{}*, Fields Institute Comm., **52** 2007, 343–366.
K. Gr[ö]{}chenig, G. Zimmermann, *Spaces of test functions via the STFT* J. Funct. Spaces Appl. **2** (2004), 25–53.
L. H[ö]{}rmander, *The Analysis of Linear Partial Differential Operators*, vol [I,II, III]{}, [Springer-Verlag]{}, Berlin Heidelberg, 1983, 1985.
K. Koezuka, N. Tomita, *Bilinear Pseudo-differential Operators with Symbols in $ BS ^m _{1,1} $ on Triebel–Lizorkin Spaces*, J. Fourier Anal. Appl, **24** 2018, 309–319.
Z. Lozanov–Crvenković, D. Perišić, M. Tasković, *Gelfand-Shilov spaces, Structural and Kernel theorems*, Preprint. arXiv:0706.2268v2
F. Nicola, L. Rodino, *Global Pseudo-differential calculus on Euclidean spaces, Pseudo-Differential Operators. Theory and Applications* **4**, [B]{}irkh[ä]{}user Verlag (2010)
S. Molahajloo, K. A. Okoudjou, G. E. Pfander, *Boundedness of Multilinear Pseudodifferential Operators on Modulation Spaces*, J. Fourier Anal. Appl., **22** (2016), 1381–1415.
, *Tempered Ultradistributions*, [Bollettino U.M.I.]{} (7) **2**-B (1988), 235–251.
C. Pfeuffer, J. Toft, Compactness Properties for Modulation Spaces, J. Complex Anal. Oper. Theory (2019), https://doi.org/10.1007/s11785-019-00903-4
B. Prangoski, Pseudodifferential operators of infinite order in spaces of tempered ultradistributions, Journal of Pseudo-Differential Operators and Applications, 4 (4), 495–549 (2013)
M. Ruzhansky, M. Sugimoto, J. Toft, N. Tomita, *Changes of variables in modulation and Wiener amalgam spaces*, Math. Nachr. **284**, (2011) 2078–2092.
, *Ultradistributions and time-frequency analysis* , in: P. Boggiatto, L. Rodino, J. Toft, M.W. Wong, (Eds) Oper. Theory Adv. Appl., **164**, 173–191, Birkhäuser, Verlag (2006)
, *Modulation spaces, Gelfand-Shilov spaces and pseudodifferential operators*, Sampl. Theory Signal Image Process, **5** (2006), 225–242.
N. Teofanov, *Gelfand-Shilov spaces and localization operators*, Funct. Anal. Approx. Comput. **2** (2015), 135–158.
N. Teofanov, *Bilinear localization operators on modulation spaces*, J. Funct. Spaces 2018, Art. ID 7560870, 10 pp.
N. Teofanov, *The Grossmann-Royer transform, Gelfand-Shilov spaces, and continuity properties of localization operators on modulation spaces,* in: L. Rodino, J. Toft (eds), Mathematical Analysis and Applications – Plenary Lectures, Springer Nature, 2018, pp. 161–207,
N. Teofanov, *Continuity properties of multilinear localization operators on modulation spaces*, in P. Boggiatto, E. Cordero, H. Feichtinger, M. de Gosson, F. Nicola, A. Oliaro, A. Tabacco (eds), Landscapes of Time-frequency Analysis, Springer Nature (2019)
, *Subalgebras to a Wiener type Algebra of Pseudo-Differential operators*, [Ann. Inst. Fourier (5)]{} **51** (2001), 1347–1383.
, *Continuity properties for modulation spaces with applications to pseudo-differential calculus, I*, [J. Funct. Anal. (2)]{}, **207** (2004), 399–429.
, *Continuity properties for modulation spaces with applications to pseudo-differential calculus, II*, [Ann. Global Anal. Geom.]{}, **26** (2004), 73–106.
, *Continuity and Schatten properties for pseudo-differential operators on modulation spaces [[in: J. Toft, M. W. Wong, H. Zhu (Eds) Modern Trends in Pseudo-Differential Operators,]{}]{}* Operator Theory: Advances and Applications **172**, Birkh[ä]{}user Verlag, Basel, 2007, pp. 173–206.
, *Pseudo-differential operators with smooth symbols on modulation spaces*, Cubo, **11** (2009), 87–107.
J. Toft, *The Bargmann transform on modulation and Gelfand-Shilov spaces, with applications to Toeplitz and pseudo-differential operators*, J. Pseudo-Differ. Oper. Appl. **3** (2012), 145–227.
J. Toft, *Matrix parameterized pseudo-differential calculi on modulation spaces [[in: M. Oberguggenberger, J. Toft, J. Vindas, P. Wahlberg (eds)]{}]{}, Generalized functions and Fourier analysis*, Operator Theory: Advances and Applications Vol 260, Birkh[ä]{}user/Springer, Basel, 2017, pp. 215–235.
J.Toft, *Continuity of Gevrey-H[ö]{}rmander pseudo-differential operators on modulation spaces*, J. Pseudo-Differ. Oper. Appl. (2019) https://doi.org/10.1007/s11868-018-0273-9
G. Tranquilli, *Global normal forms and global properties in function spaces for second order Shubin type operators* PhD Thesis, 2013.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend some classical theorems in the theory of orthogonal polynomials on the unit circle to the matrix case. In particular, we prove a matrix analogue of Szegő’s theorem. As a by-product, we also obtain an elementary proof of the distance formula by Helson and Lowdenslager.'
author:
- 'Maxim Derevyagin, Olga Holtz , Sergey Khrushchev, and Mikhail Tyaglov'
date: 'December 9, 2011'
title: 'Szegő’s theorem for matrix orthogonal polynomials'
---
Matrix orthogonal polynomials, operator functions, Szegő’s theory.
42C05, 30C10, 62M10, 60G25, 46G10, 46G25, 15A54, 15A16, 15A45.
Historical background and motivation
====================================
The classical work [@Szego] of Szegő was the first to address the asymptotics of orthogonal polynomials on the unit circle $\mathbb{T}$ under the assumption that the entropy of the underlying measure $\sigma$ is finite, i.e., $$\int_{\mathbb T} \log \sigma' {d\theta \over 2\pi} > - \infty.$$ Further aspects of Szegő’s theory were developed by Geronimus, Verblunsky and others, which led to a number of other formulas, in various setups, involving the entropy such as the formula of Helson-Lowdenslager [@HelsonLowdenslager] for multivariate random processes (for a historical account, see [@SimonI §1.1]).
Verblunsky [@Verblunsky formulas (v) and (vi)] showed that, for any probability measure $\sigma$ on the unit circle ${\mathbb T}$, $$\label{GeronimSzego}
\lim_n\prod_{k=0}^n(1-|a_k|^2)=\exp\int_{\mathbb{T}}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}.$$ Here $\{a_k\}_{k\geq 0}$ is a sequence of points in the unit disc $\mathbb{D}$ called the *parameters* of $\sigma$ [@Khrushchev §8.3] and $\sigma^{\,\prime}=2\pi d\sigma/d\theta$ is the Lebesgue derivative of $\sigma$. The numbers $\{a_k\}_{k\geq 0}$ have different names depending on the area where they are considered. In the theory of orthogonal polynomials they are known as the [*Szegő recurrence coefficients, Verblunsky parameters, Geronimus parameters,*]{} in Schur’s theory they are [*Schur’s parameters,*]{} in inverse scattering problems they are [*reflection coefficients,*]{} see [@SimonI §1.1].
In the matrix setting, $\mathbf{\sigma}$ is a Borel measure on $\mathbb{T}$ with values in the set ${\mathcal M}_\ell^{+}$ of all nonnegative definite matrices in ${\mathcal M}_\ell$, the set of all $\ell\times \ell$ matrices with complex entries. We denote by $\textsf{P}_{\ell}(\mathbb{T})$ the set of all matrix-valued nonnegative measures $\sigma$ on $\mathbb{T}$ that are normalized, i.e., $$\sigma(\mathbb{T})=\mathbf{1}$$ to the unit matrix $\mathbf{1}$ in ${\mathcal M}_\ell^{+}$. We refer to $\textsf{P}_{\ell}(\mathbb{T})$ as the class of [*matrix probability measures*]{}.
The matrix case is important in multivariate Time Series and Prediction Theory [@Hannan; @HelsonLowdenslager; @Masani; @Rozanov; @Wiener]. As far as we know, the first Szegő-type results on matrix-valued orthogonal polynomials were obtained by Delsarte, Genin and Kamp [@DGK:78]; this line of research was continued by Aptekarev and Nikishin [@AN:83].
Our method is a combination of the recent theory of matrix orthogonal polynomials presented in [@DPS:08] and the approach to Szegő’s theory developed in [@Khrushchev2001; @Khrushchev]. This combination allows us to avoid using factorization theory in matrix Hardy classes. Instead, we use only methods of Real Analysis and Matrix/Operator Theory.
Our main goal is to respond to the following challenge of Damanik, Pushnitski, and Simon [@DPS:08]:
[*Among the deepest and most elegant methods in OPUC are those of Khrushchev \[125, 126, 101\]. We have not been able to extend them to MOPUC! We regard their extension as an important open question...*]{}
Below we provide a full matrix-valued version of Szegő’s theorem, yielding the previously known trace versions as corollaries of our matrix formula.
Throughout the paper, we mostly follow the notation and terminology of [@DPS:08].
Main results
============
In the matrix case, the parameters $\alpha_k$ are matrices in ${\mathcal M}_\ell$ with norms $\Vert{\alpha}_k\Vert$ not exceeding $1$. Here $\Vert{\alpha}\Vert$ is the norm of the linear operator defined by the matrix $\alpha$ subordinate to the usual Euclidean vector norm ($2$-norm) on ${\mathbb C}^\ell$. This operator norm is also known as the [*spectral*]{} or the [*Euclidean*]{} norm. This norm is well known to equal the largest singular value of the matrix $\alpha$; in particular, if $\alpha$ is self-adjoint, the norm $\Vert{\alpha}\Vert$ equals the spectral radius of $\alpha$.
We denote by $\alpha^\dagger$ the Hermitian conjugate of $\alpha\in{\mathcal M}_\ell$. The symbol $\,^*\,$ is reserved for the Szegő dual, so we do not use it for the adjoint (see ).
We assume that the matrix $$\label{matrixpoly}
\int_{\mathbb{T}}p(e^{i\theta})^\dagger d\sigma(\theta)p(e^{i\theta})$$ is positive (definite) for any polynomial $p$ with coefficients in ${\mathcal M}_\ell$. Condition is equivalent to the requirement that $d\sigma$ has full rank $\ell$ for infinitely many points in $\mathbb{T}$. Under this condition, the right (left) orthogonal matrix polynomials $\varphi_n^R$ ($\varphi_n^L$) are uniquely determined by the standard Gram-Schmidt orthonormalization. It is important to notice that the left orthogonal matrix polynomials are obtained with respect to the left quadratic ‘form’: $$\label{matrixpoly2int}
\int_{\mathbb{T}}p(e^{i\theta}) d\sigma(\theta)p(e^{i\theta})^\dagger .$$ Every $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ is uniquely determined by the sequence of its parameters $\{\alpha_k\}_{k\geq 0}$. These parameters are contractive matrices in ${\mathcal M}_\ell$. If $\sigma$ is a matrix-valued measure with parameters $\{\alpha_k\}_{k\geq 0}$, then the parameters $\{\alpha_k^\dagger\}_{k\geq 0}$ correspond to the measure $\overline{\sigma}$ such that $$\overline{\sigma}(E)=\sigma(\overline{E}),\quad \overline{E}=\{\overline{z}:z\in E\},$$ for any Borel set $E$, where $\overline{z}$ stands for the complex conjugate of a complex number $z$. We write $\varphi_n(z,\sigma)$ for the orthogonal polynomials if the dependence on $\sigma$ is important.
For a matrix polynomial $P_n$ of degree $n$, we define the [*reversed*]{} (or [*Szegő dual)*]{} polynomial $P_n^*$ by $$\label{reverseP}
P_n^*(z) = z^n P_n(1/\bar{z})^\dagger.$$ The relationship between the left orthogonal polynomials ${\varphi}_n^L$ and the right orthogonal polynomials ${\varphi}_n^R$ is given by the formula $$\label{lefttoright45}
{\varphi}_n^L(e^{i\theta},\overline{\sigma}){=}{\varphi}_n^R(e^{-i\theta},\sigma)^\dagger$$ (see Corollary \[eqnumberfive8\]). The $n$th left normalized orthogonal polynomial $\varphi_n^L(z,\overline{\sigma})$ depends on the parameters $\alpha_0^\dagger$, $\alpha_1^\dagger$, $\ldots$, $\alpha_{n-1}^\dagger$. Hence, the $n$th right polynomial $\varphi_n^R(z,\sigma)$ can be obtained from the left $\varphi_n^L(z,{\sigma})$ by replacing each $\alpha_k^\dagger$ by $\alpha_k$, replacing $z\in\mathbb{T}$ by $\overline{z}$ and applying the conjugation $\dagger$.
The main result of this paper is the following theorem.
[**Theorem.**]{} Every matrix probability measure $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ satisfies the following matrix equality: $$\label{theMainFormula}
\lim_{n\rightarrow +\infty}\int_{0}^{2\pi}\log([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}=\int_{\mathbb{T}}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}~.$$
If the parameters $\alpha_k$ of $\sigma$ form a family of commuting normal matrices, then can be simplified to $$\label{commutcase}
\prod_{k=0}^{\infty}\left({\bf 1} -\alpha_k \alpha_k^\dagger\right)=\exp\int_{\mathbb{T}}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}.$$
[**Remark.**]{} Alternatively, the commuting case reduces to the diagonal and hence to the scalar case.
Regardless of the normality or commutativity of $\{ \alpha_k \}$, the following determinantal-trace version [@DGK:78] follows from (\[theMainFormula\]): $$\label{traceclass}
\prod_{k=0}^{\infty}\det\left({\bf 1} -\alpha_k \alpha_k^\dagger\right)=\exp\int_{\mathbb{T}}\text{tr}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}.$$ The symmetry $z\mapsto \overline{z}$ keeps the Lebesgue measure on $\mathbb{T}$ invariant. Hence, combining and a simple formula $$\int_{\mathbb{T}}\log\overline{\sigma}^{\,\prime}\,\frac{d\theta}{2\pi}=\int_{\mathbb{T}}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi},$$ we also obtain the left version of (\[theMainFormula\]): $$\label{LeftFormula4}
\lim_{n\rightarrow +\infty}\int_{0}^{2\pi}\log([\varphi_n^{L,*} (e^{i\theta}) \varphi_n^{L,*} (e^{i
\theta})^\dagger]^{-1})\frac{d\theta}{2\pi}=\int_{\mathbb{T}}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}.$$
Matrix preliminaries
====================
Recall that we denote by ${\mathcal M}_\ell$ the ring of all $\ell\times \ell$ complex-valued matrices, its identity matrix by ${\bf 1}$ and its zero matrix by ${\bf 0}$. Along with the Euclidean norm $\| \cdot \|$ on ${\mathcal M}_\ell$, we also consider the trace norm $\pmb{|\!\!|}\alpha\pmb{|\!\!|}_1=\text{tr}(\alpha^\dagger\alpha)^{1/2}$ and the Hilbert-Schmidt norm $\pmb{|\!\!|}\alpha\pmb{|\!\!|}_2=(\text{tr}(\alpha^\dagger\alpha))^{1/2}$. It is easy to see that $$\label{ineqfornorms}
\Vert{\alpha}\Vert\leq\pmb{|\!\!|}\alpha\pmb{|\!\!|}_2\leq \pmb{|\!\!|}\alpha\pmb{|\!\!|}_1\leq \ell\Vert{\alpha}\Vert.$$
We say that a self-adjoint matrix $A(=A^\dagger) \in {\mathcal M}_\ell$ is [*nonnegative*]{} ([*positive*]{}) if the corresponding quadratic form $x\mapsto x^\dagger A x$ is nonnegative definite (positive definite). We denote the class of all nonnegative self-adjoint $\ell\times\ell$ matrices by ${\mathcal M}^+_\ell$. The corresponding partial order is known as the [*Loewner ordering*]{} and is denoted by $\succ$: $A\succ B$ means that $A-B$ is positive, i.e., $A-B\succ {\mathbf 0}$, and $A\succeq B$ means that $A-B\succeq {\mathbf 0}$, or $A-B\in{\mathcal M}^+_\ell$.
Here is the first fact about the Loewner ordering that we will use later:
\[triv87\] Let ${\bf 0}\preceq A_j\preceq B_j$ for $j=1,\ldots k$. Then $
{\bf 0}\preceq A_1+\cdots+A_k\preceq B_1+\cdots+B_k.
$
Evaluate and compare the quadratic forms of both sums.
We will also need the following result connecting traces of self-adjoint matrices and their Loewner ordering:
\[triv88\] Suppose $A\succeq B$ and $\operatorname{tr}A= \operatorname{tr}B$. Then $A=B$.
By the linearity of traces, this is equalent to the statement: Suppose $A\succeq {\bf 0}$ and $\operatorname{tr}A=0$. Then $A={\bf 0}$. The latter follows from the fact that $\operatorname{tr}A = \sum_{j=1}^\ell
e_j^\dagger A e_j$, so if the trace of $A$ is zero, the action of $A$ on all standard unit vectors (hence on the entire space) must be trivial.
Another fact about traces we will need is the following:
\[triv89\] If $A\succ {\bf 0}$, then $
\log \det(A)=\text{tr}(\log A).
$
Without loss of generality, $A$ is a diagonal matrix with positive diagonal elements, since the formula is invariant under unitary similarity. But then $\log A$ is the diagonal matrix whose elements are the logarithms of the diagonal elements of $A$. The conclusion of the Lemma is thus straightforward.
\[triv90\] Let $A_1$, $\ldots$, $A_n\succ {\bf 0}$ and let $A:=A_1\cdots A_n \succ {\bf 0}$. Then $$\text{tr}\log(A_1\cdots A_n)=\log \det(A_1\cdots A_n)=\sum_{k=1}^n\text{tr}(\log A_k).$$
Apply Lemma \[triv89\] to the product $A=A_1\cdots A_n$.
Finally, we will need the following interesting characterization of the determinant via the trace (also used by Helson and Lowdenslager in [@HelsonLowdenslager], also see, e.g., [@HornJohnson Exercise 19, p.486]):
\[HL-Lemma\] Let $\mathcal{A}$ be the set of all matrices in $\mathcal{M}_\ell$ with determinant $1$. Then every positive matrix $C$ satisfies $$\label{bestin54}
\inf_{A\in\mathcal{A}}\frac{1}{\ell}\,\text{tr}\left(ACA^\dagger\right)=[\det(C)]^{1/\ell}.$$
Let $U$ be a unitary matrix such that $C=UDU^\dagger$, where $D$ is the diagonal matrix with eigenvalues $\lambda_1,\ldots\lambda_\ell$. Then $\lambda=\det(U)\in\mathbb{T}$. It follows that $ACA^\dagger$=$(A\overline{\lambda}U)D(A\overline{\lambda}U)^\dagger$, implying that we may assume without loss of generality that $C=D$. Then $$\text{tr}(ADA^\dagger)=\lambda_1\Vert a_1\Vert^2+\lambda_2\Vert a_2\Vert^2+\cdots+\lambda_\ell\Vert a_\ell\Vert^2,$$ where $a_k$ denotes the $k$th column of $A$. By the arithmetic-geometric mean inequality, $$\label{AGM749}
\frac{\lambda_1\Vert a_1\Vert^2+\lambda_2\Vert a_2\Vert^2+\cdots+\lambda_\ell\Vert a_\ell\Vert^2}{\ell}\geq \sqrt[\ell]{\lambda_1\cdots\lambda_\ell\Vert a_1\Vert^2\cdots\Vert a_\ell\Vert^2}.$$ By Hadamard’s inequality [@HornJohnson Inequality 7.8.2], $$\label{Hadamard}
\Vert a_1\Vert\cdots\Vert a_\ell\Vert\geq \det(A)=1.$$ The equality in occurs if and only if $$\lambda_1\Vert a_1\Vert^2=\cdots=\lambda_\ell\Vert a_\ell\Vert^2.$$ The equality in occurs if and only if the columns $a_k$ form an orthogonal system in $\mathbb{C}^\ell$. It follows that the equality in is attained for the diagonal matrix $A$ with $a\lambda_1^{-1/2},\ldots,a\lambda_{\ell}^{-1/2}$ on the diagonal. Here $a$ is chosen so as to make the determinant of $A$ equal $1$.
Matrix measures
===============
A matrix-valued nonnegative measure $\mu$ on the unit circle ${\mathbb T}$ is a countably additive mapping of the Borel $\sigma$-algebra $\mathfrak{B}({\mathbb T})$ on ${\mathbb T}$ into the set ${\mathcal M}_\ell^{+}$ of all nonnegative $\ell\times \ell$ matrices $\mu: B\mapsto\mu(B)\in {\mathcal M}_\ell^{+}$. It follows that for any $E\in\mathfrak{B}({\mathbb T})$ $$0\leq \mu(E)\leq \mu(E)+\mu(\mathbb{T}\setminus E)=\mu(\mathbb{T}).$$ Then $\nu(E)=\mu(\mathbb{T})^{-1/2}\mu(E)\mu(\mathbb{T})^{-1/2}$ is also a matrix-valued nonnegative measure which is called the normalization of $\mu$. As before, we assume that $\mu$ is normalized: $ \mu({\mathbb T})={\bf 1}. $
Recall that $\textsf{P}_{\ell}(\mathbb{T})$ denotes the set of all matrix probability measures, i.e., the normalized matrix-valued nonnegative measures on $\mathbb{T}$. Let $\{e_j\}_{j=1}^l$ be the standard basis in ${\mathbb C}^\ell$. Then for every $E\in\mathfrak{B}({\mathbb T})$ we obtain the matrix of $\mu(E)$ $$\label{matrixOfMeasure}
\mu(E)=\begin{pmatrix}\mu_{11}(E)&\mu_{12}(E)&\cdots&\mu_{1\ell}(E)\\
\mu_{21}(E)&\mu_{22}(E)&\cdots&\mu_{2\ell}(E)\\
\vdots&\vdots&\ddots&\vdots\\
\mu_{\ell 1}(E)&\mu_{\ell 2}(E)&\cdots&\mu_{\ell \ell}(E)
\end{pmatrix}.$$ Since $\pmb{|\!\!|}\alpha\pmb{|\!\!|}_1=\text{tr}(\alpha)$ for every $\alpha\in{\mathcal M}_\ell^{+}$, we see that $$\label{tracemeasure}
|\mu_{ij}(E)|=|(\mu(E)e_j,e_i)|\leq \Vert{\mu(E)}\Vert\leq\pmb{|\!\!|}\mu(E)\pmb{|\!\!|}_1=\text{tr}(\mu(E)).$$ It follows that the entries $\mu_{ij}(E)$ of $\mu(E)$ are finite complex measures on $\mathbb{T}$ which are absolutely continuous with respect to $\operatorname{tr}(\mu)$. Thus any element $\mu$ of $\textsf{P}_{\ell}(\mathbb{T})$ is nothing but a table of measures subject to positivity conditions and domination by the $\operatorname{tr}(\mu)$. We say that $\mu\in\textsf{P}_{\ell}(\mathbb{T})$ is [*absolutely continuous (discrete, singular)*]{} if so is its trace measure. It follows that any $\mu\in\textsf{P}_{\ell}(\mathbb{T})$ can be uniquely decomposed into the sum $$\label{LebesgueDecomp}
\mu=\mu_a+\mu_d+\mu_s,$$ where $\mu_a$ is absolutely continuous with respect to the Lebesgue measure $d\theta/(2\pi)$, $\mu_d$ is the discrete part of $\mu$ and $\mu_s$ is its singular part. Indeed, taking the Hahn-Lebesgue decomposition of the trace measure, we can associate three matrix-valued measures with it. Namely, the entries of $\mu_a$ are the absolutely continuous parts of $\mu_{ij}$ with respect to $\text{tr}(\mu)_a$, and similarly the enties of $\mu_d$ and $\mu_s$ for the discrete and singular parts, respectively. Since Borel supports of $\text{tr}(\mu)_a$, $\text{tr}(\mu)_d$, $\text{tr}(\mu)_s$ can be chosen to be disjoint, the positivity of the corresponding matrices follows immediately. Moreover $d\mu=\mathbf{M}(\mu, \zeta)\text{tr}(d\mu)$ where $\mathbf{M}(\mu,\zeta)\in{\mathcal{M}}_\ell^+$ for $\zeta\in\mathbb{T}$.
The measure $\mu_a$ can be found by Lebesgue’s differentiation theorem: $$\label{difflebesgue}
\mu^{\prime}(e^{i\theta})=\lim_{\epsilon\rightarrow 0^+}\frac{\mu(I_\epsilon)}{2\epsilon}\quad\text{a.e. on } \mathbb{T},$$ where $I_{\epsilon}$ denotes the arc of length $2\epsilon$ on $\mathbb{T}$ centered at $e^{i\theta}$. Then $$\mu_a(E)=\int_E\mu^{\prime}(e^{i\theta})\frac{d\theta}{2\pi}.$$
We say that a sequence $\{\mu^{(n)}\}_{n\geq 0}$ in $\textsf{P}_{\ell}(\mathbb{T})$ converges to $\mu\in\textsf{P}_{\ell}(\mathbb{T})$ in $*$-weak topology if $$\label{starconv47}
*{-}\lim_n \mu_{ij}^{(n)}=\mu_{ij}$$ for every pair of indices $(i,j)$. For our class $\textsf{P}_{\ell}(\mathbb{T})$, we need matrix analogues of two Helley’s lemmas as they are stated in [@Khrushchev Lemma 8.5, Theorem 8.6] for scalar measures:
\[Helley1\] If $*{-}\lim_n \mu^{(n)}=\mu$ in $\textsf{P}_{\ell}(\mathbb{T})$, then $$\label{arctest}
\lim_n \mu^{(n)}(I)=\mu(I)$$ for any open arc $I$ on $\mathbb{T}$ such that $\mu$ vanishes at the endpoints of $I$.
Let ${x}$ be an arbitrary fixed column-vector in $\mathbb{C}^\ell$; as usual, ${x}^\dagger$ denotes its conjugate transpose (row-vector). Let $t\mapsto f(t)$ be a nonnegative continuous function with values in $[0,1]$ supported on an open arc $I$. Then $$\label{mainineq538}
x^\dagger\mu^{(n)}(I)x=\int_I x^\dagger d\mu^{(n)}(t)x\geq \int_I f(t)x^\dagger d\mu^{(n)}(t)x.$$ By $$\lim_n \int_I f(t)x^\dagger d\mu^{(n)}(t)x=\lim_n\int_I f(t)\sum_{i,j=1}^\ell\overline{x_i} \, d\mu_{ij}^{(n)} x_j=\sum_{i,j=1}^\ell\lim_n\int_I f(t)\overline{x_i} \, d\mu_{ij}^{(n)} x_j=\int_I f(t)x^\dagger d\mu(t)x.$$ This and imply $$\label{firstineq245}
\limsup_n x^\dagger\mu^{(n)}(I)x\geq \liminf_n x^\dagger\mu^{(n)}(I)x\geq \sup_{f}\int_I f(t)x^\dagger d\mu(t)x=\int_Ix^\dagger d\mu(t)x=x^\dagger \mu(I) x.$$ Similarly for the complementary arc $J$ to the closure of $I$ in $\mathbb{T}$ we have $$\label{secondineq245}
\limsup_n x^\dagger\mu^{(n)}(J)x\geq \liminf_n x^\dagger\mu^{(n)}(J)x\geq \sup_{f}\int_J f(t)x^\dagger d\mu(t)x=\int_Jx^\dagger d\mu(t)x=x^\dagger \mu(J) x.$$ Since $\mu$ vanishes at the endpoints of $I$ and $\mu^{(n)}\in \textsf{P}_{\ell}(\mathbb{T})$ we obtain that $$\label{twoineq76}
\mu^{(n)}(I)+\mu^{(n)}(J)\preceq\mathbf{1}~, \qquad \mu(I)+\mu(J)=\mathbf{1}.$$ Combining and with we conclude that $$\begin{gathered}
x^\dagger x=x^\dagger\mathbf{1}x\geq \limsup_n x^\dagger(\mu^{(n)}(I)+\mu^{(n)}(J))x
\geq \liminf_n x^\dagger(\mu^{(n)}(I)+\mu^{(n)}(J))x \\ \geq \int_Ix^\dagger d\mu(t)x+\int_Jx^\dagger d\mu(t)x=x^\dagger \mu(I) x+x^\dagger \mu(J) x=
x^\dagger \mu(\mathbb{T}) x=x^\dagger x.\\\end{gathered}$$ This is only possible if equalities hold in and in . Since the vector $x$ was arbitrary, this implies the conclusion of the theorem.
\[HelleysTh\] Let $\{\mu^{(n)}\}_{n\geq 0}$ be a sequence in $\textsf{P}_{\ell}(\mathbb{T})$ and $\mu\in\textsf{P}_{\ell}(\mathbb{T})$. Then $*{-}\lim_n \mu^{(n)}=\mu$ if and only if $\lim_n \mu^{(n)}(I)=\mu(I)$ for any open arc $I$ on $\mathbb{T}$ such that $\mu$ does not have point masses at the endpoints of $I$.
One direction has been already proved in Theorem \[Helley1\]. Suppose now that $\lim_n \mu^{(n)}(I)=\mu(I)$ for any open arc $I$ on $\mathbb{T}$ such that $\mu$ does not have point masses at the endpoints of $I$. Then for every $x\in\mathbb{C}^\ell$, $\Vert x\Vert=1$, the sequence of usual probability measures $x^\dagger\mu^{(n)}x$ converges to $x^\dagger\mu x$ on any interval which does not have point masses of $\mu$ (and therefore of $x^\dagger\mu x$ since it is absolutely continuous with respect to $\mu$). Then by Theorem 8.6 of [@Khrushchev] the sequence $x^\dagger\mu^{(n)}x$ converges to $x^\dagger\mu x$ in the $*$-weak topology. Now the polarization identity implies $*{-}\lim_n x^\dagger\mu^{(n)}y=x^\dagger\mu y$ for any pair of vectors $(x,y)$. Setting $x:=e_i$, $y:=e_j$ for all pairs $i,j$, we obtain the weak limits for all entries of $\mu$.
\[Ltwoth\] Suppose that $\nu_{n}=h_{n}d\theta/(2\pi)$ where $h_{n}$ are matrix-valued functions on $\mathbb{T}$. Suppose that there is a positive constant $C$ such that $$\int_{\mathbb{T}}\Vert{h_{n}}\Vert^2\frac{d\theta}{2\pi}<C.$$ Then any $*$-weak limit point of $\{\nu_{n}\}$ is an absolutely continuous matrix-valued measure.
By norm equivalence in ${\mathcal M}_\ell$, we can replace the operator norm of $h_n$ by its Hilbert-Schmidt norm. For each matrix entry, the result of this theorem is standard, so it holds for the Hilbert-Schmidt (and hence the spectral) norm of the entire matrix as well.
Every $\mu\in\textsf{P}_{\ell}(\mathbb{T})$ defines two positive definite quadratic forms on the two-sided module $C(\mathbb{T}, {\mathcal M}_\ell)$ over ${\mathcal M}_\ell$ of all continuous functions with values in ${\mathcal M}_\ell$. They correspond to the right and left multiplication and are defined as matrix-valued ‘inner products’ by $$\begin{aligned}
{\langle\! \langlef,g\rangle\! \rangle}_R:=\int f(x)^\dagger\, d\mu(x) g(x), \label{ileft} \\
{\langle\! \langlef,g\rangle\! \rangle}_L:=\int g(x)\, d\mu(x) f (x)^\dagger. \label{iright}\end{aligned}$$
Let ${\mathcal P}$ denote the set of all polynomials in $z\in{\mathbb C}$ with coefficients from ${\mathcal M}_\ell$. For a nonnegative integer $n$, ${\mathcal P}_n$ will denote the set of polynomials in ${\mathcal P}$ of degree at most $n$. Note that, to generate an infinite sequence of orthogonal polynomials, $\mu$ must satisfy for every nonzero polynomial $p$. This is equivalent to the condition that the non-negative Borel measure $$\det\left(\mathbf{M}(\mu,\zeta)\right)\text{tr}(d\mu)$$ has infinite Borel support, see [@DPS:08; @Wiener].
Analysis of operator functions
==============================
In this section we list some properties of the logarithm as an operator function. We start with the definitions of operator monotone, convex, and concave functions defined on the half real line $(0,\infty)$. Let $\mathcal{H}$ be an infinite-dimensional (separable) Hilbert space. Let $B_+(\mathcal{H})$ denote the set of all positive operators in $B(\mathcal{H})$. A continuous real function $f$ on $(0,\infty)$ is said to be [*operator monotone*]{} (or, more precisely, [*operator monotone increasing)*]{} if $A\preceq B$ implies $f(A)\preceq f(B)$ for $A,B\in B_+(\mathcal{H})$, and [*operator monotone decreasing*]{} if $-f$ is operator monotone increasing, i.e., if $A\preceq B$ implies $f(A)\succeq f(B)$, where $f(A)$ and $f(B)$ are defined via functional calculus as usual. Also, $f$ is said to be [*operator convex*]{} if $f(\lambda A+(1-\lambda)B)\preceq \lambda
f(A)+(1-\lambda)f(B)$ for all $A,B\in B_+(\mathcal{H})$ and $\lambda\in(0,1)$, and [*operator concave*]{} if $-f$ is operator convex (see also [@Bhatia:96]).
One should not expect that the operator monotonicity and the operator convexity of $f$ follow from the same properties of the scalar function $f$. For example, a power function $t^\alpha$ on $(0,\infty)$ is operator monotone if and only if $\alpha\in[0,1]$, operator monotone decreasing if and only if $\alpha\in[-1,0]$, and operator convex if and only if $\alpha\in[-1,0]\cup[1,2]$ (see, for instance, [@Bhatia:96 Chapter V]). Moreover, the function $f(t)=\exp(t)$ is neither operator monotone nor operator convex on any (spectral) interval.
As is known, the operator monotone functions are generated by holomorphic functions that map the upper half plane into the upper half plane. Clearly, if one fixes a branch of the logarithm so that it is real on $(0,\infty)$ then the corresponding holomorophic function maps the upper half plane into the upper half plane.
\[OpMon\] The functions $\log t$ and $-1/t$ are operator monotone increasing on $(0,\infty)$.
A detailed proof can be found in [@Bhatia:96 Section V.4].
So, $1/t$ is operator monotone decreasing on $(0,\infty)$. Furthermore, it follows from [@Bhatia:96 Exercise V.3.14] that the integration of an operator monotone decreasing function gives an operator concave function.
\[OpConc\] The function $\log t$ is operator concave on $(0,\infty)$.
This statement can also be verified by means of [@AnHi:11 Theorem 3.1].
Now, we are in a position to formulate the matrix Jensen inequality for the logarithm. Namely, Proposition \[OpConc\] and [@FarZh:07 Theorem 4.2] yield the following statement.
\[prJensen\] Let $f : {\mathbb T} \to B_+({\mathcal M}_\ell)$ be a measurable function. Then the following inequality holds: $$\label{MJIN}
\int_{\mathbb{T}}\log f(\theta)
\frac{d\theta}{2\pi}
\preceq \log\left(\int_{\mathbb{T}}f(\theta)\frac{d\theta}{2\pi}\right).$$
Besides monotonicity and convexity, we will also deal with operator continuity. Recall that a function $f$ defined on $(0,\infty)$ is operator continuous if the relation $\Vert A_n-A\Vert_{\mathcal{H}}\to 0$ implies $\Vert
f(A_n)-f(A)\Vert_{{\mathcal H}}\to 0$ for any $A, A_n\in B_+({\mathcal H})$.
\[OpCont\] The function $\log t$ is operator continuous on $(0,\infty)$.
Since $\log t$ can be extended to a holomorphic function on $\mathbb{C}\setminus (-\infty, 0)$, the statement follows directly from the Dunford-Schwarz operator calculus.
Matrix orthogonal polynomials on the unit circle
================================================
We begin by recalling some basic facts from [@DPS:08] for convenience of the reader. Let $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ be a matrix probability measure such that $\det(\mathbf{M}(\sigma,\zeta))\text{tr}(d\sigma(\zeta))$ has infinite Borel support. We define right and left monic orthogonal matrix polynomials $\Phi_n^R, \Phi_n^L$ by applying the Gram–Schmidt procedure in $C(\mathbb{T}, {\mathcal M}_\ell)$ with respect to the ‘inner products’ and to the sequence $\{{\bf 1}, z {\bf 1}, z^2 {\bf 1}, \ldots \}$. In other words, $\Phi_n^R$ is the unique matrix polynomial $z^n {\bf 1} + \textit{lower order terms}$ satisfying the orthogonality conditions $$\label{ortR}
0={\langle\! \langlez^k{\bf 1}, \Phi_n^R\rangle\! \rangle}_R =\int \left(z^k{\bf 1}\right)^\dagger\, d\sigma(x) \Phi_n^R, \qquad k=0,1,\dots, n-1.$$ Similarly $\Phi_n^L$ is the unique matrix polynomial $z^n {\bf 1} + \textit{lower order terms}$ satisfying $$\label{leftmatrixpoly}
0={\langle\! \langlez^k{\bf 1},\Phi_n^L\rangle\! \rangle}_L:=\int \Phi_n^L\, d\sigma(x) \left(z^k{\bf 1}\right)^\dagger, \qquad k=0,1,\dots, n-1.$$
The normalized orthogonal matrix polynomials are defined by $$\label{monicandnorm}
\varphi_0^L=\varphi_0^R={\bf 1},\qquad
\varphi_n^L = \kappa_n^L \Phi_n^L \qquad \text{ and } \quad \varphi_n^R =
\Phi_n^R\kappa_n^R$$ where the $\kappa$’s are defined according to the normalization conditions $$\label{ortcond}
{\langle\! \langle\varphi_n^R,\varphi_m^R\rangle\! \rangle}_R=\delta_{nm}{\bf 1},
\qquad
{\langle\! \langle\varphi_n^L,\varphi_m^L\rangle\! \rangle}_L=\delta_{nm}{\bf 1},$$ along with the following positivity conditions: $$\label{NormCond}
\kappa_{n+1}^L (\kappa_n^L)^{-1} \succ {\bf 0} \quad \text{and} \quad
(\kappa_n^R)^{-1}
\kappa_{n+1}^R \succ {\bf 0}.$$ Note that the $\kappa_{n}^L$ are determined by the normalization condition up to multiplication on the left by unitary matrices. It can be shown that these unitaries can always be uniquely chosen so as to satisfy , see [@DPS:08].
Now define $$\rho_n^L :=\kappa_n^L (\kappa_{n+1}^L)^{-1} \qquad \text{and} \quad
\rho_n^R :=
(\kappa_{n+1}^R)^{-1} \kappa_n^R.$$ Being inverses of positives matrices, $\rho_n^L$ and $
\rho_n^R $ are positive definite as well. In particular, we have that $$\label{kappaformula}
\kappa_n^L = (\rho_{0}^L \cdots \rho_{n-1}^L)^{-1} \quad \text{and}
\quad \kappa_n^R =
(\rho_{n-1}^R \cdots \rho_{0}^R)^{-1}.$$
In the matrix case as well as in the scalar case we have the Szegő recursion. Before stating it, we recall that, for a matrix polynomial $P_n$ of degree $n$, we define the reversed polynomial $P_n^*$ by (\[reverseP\]): $P_n^*(z) = z^n P_n(1/\bar{z})^\dagger.$
\[szegoth\] There is a sequence of contractive matrices $\alpha_n$ in ${\mathcal M}_\ell$ such that $$\begin{aligned}
z \varphi_n^L - \rho_n^L \varphi_{n+1}^L & = \alpha_n^\dagger
\varphi_n^{R,*}, \label{S1}\\
z \varphi_n^R - \varphi_{n+1}^R \rho_n^R & = \varphi_n^{L,*}
\alpha_n^\dagger, \label{S2}\end{aligned}$$ where $\rho_n^L$ and $\rho_n^R$ are defined as follows $$\label{rhoalpha4}
\rho_n^L = ({\bf 1} - \alpha_n^\dagger \alpha_n)^
{1/2},\qquad \rho_n^R = ({\bf 1} -\alpha_n \alpha_n^\dagger)^{1/2}.$$
Setting $z=0$ in and using , we derive the following formulas for the parameters: $$\label{parametersformula}
\alpha_n=-(\kappa_n^R)^{-1}\mathbf{\Phi}^L_{n+1}(0)^\dagger (\kappa_n^L)^\dagger=
-(\kappa_n^R)^\dagger\mathbf{\Phi}^R_{n+1}(0)^\dagger(\kappa_n^L)^{-1} .$$ Alternatively, one can also set $z=0$ in formulas (3.11) of [@DPS:08].
\[LeftAndRight\] The left and right monic orthogonal polynomials of $\overline{\sigma}$ and $\sigma$ are related by $$\label{monicrel}
\mathbf{\Phi}_n^L(e^{i\theta},\overline{\sigma}){=}\mathbf{\Phi}_n^R(e^{-i\theta},\sigma)^\dagger.$$
For $k<n$ we have, by , $$\begin{gathered}
0 = \left(\int \left(z^k{\bf 1}\right)^\dagger\, d\sigma(z) \Phi_n^R(z,\sigma)\right)^\dagger =
\int \left(\Phi_n^R(z,\sigma)\right)^\dagger\, d\sigma(z) \left(z^k{\bf 1}\right)=
\int \left(\Phi_n^R(\overline{z},\sigma)\right)^\dagger\, d\sigma(\overline{z}) \left(\overline{z}^k{\bf 1}\right)
\\=\int \left(\Phi_n^R(\overline{z},\sigma)\right)^\dagger\, d\overline{\sigma}({z}) \left({z}^k{\bf 1}\right)^\dagger=\int \Phi_n^L({z},\overline{\sigma})\, d\overline{\sigma}({z}) \left({z}^k{\bf 1}\right)^\dagger,\end{gathered}$$ which implies by .
\[parametersofsymmeasure\] If $\{\alpha_k\}_{k\geq 0}$ are the parameters of $\sigma$, then $\{\alpha_k^\dagger\}_{k\geq 0}$ are the parameters of $\overline{\sigma}$.
By , the matrix coefficients of the polynomial $\mathbf{\Phi}_n^L(e^{i\theta},\overline{\sigma})$ are the matrices adjoint to the coefficients of the polynomial $\mathbf{\Phi}_n^R(e^{i\theta},\sigma)$. In particular, $$\label{4589}
\mathbf{\Phi}_{n+1}^L(0,\overline{\sigma})=\mathbf{\Phi}_{n+1}^R(0,\sigma)^\dagger.$$
Since $\kappa_0^R=\kappa_0^L=\mathbf{1}$, we see that $$\alpha_0(\overline{\sigma})=-\mathbf{\Phi}^L_{1}(\overline{0,\sigma})^\dagger=
-\mathbf{\Phi}^R_{1}({0,\sigma})=\alpha_0(\sigma)^\dagger.$$ Suppose that we already proved that $\alpha_k(\overline{\sigma})=\alpha_k(\sigma)^\dagger$ for $k<n$. Then, by the induction hypothesis and by , $$\label{leftformright5063}
\kappa_n^R(\overline{\sigma})=\kappa_n^L({\sigma})^\dagger,\qquad \kappa_n^L(\overline{\sigma})=\kappa_n^R({\sigma})^\dagger.$$ It follows that $$\begin{gathered}
\alpha_n(\overline{\sigma})=-(\kappa_n^R(\overline{\sigma}))^{-1}\mathbf{\Phi}_{n+1}^L
(0,\overline{\sigma})^\dagger(\kappa_n^L(\overline{\sigma}))^\dagger=
-(\kappa_n^L({\sigma})^\dagger)^{-1}\mathbf{\Phi}_{n+1}^R(0,\sigma)\kappa_n^R({\sigma})\\=
\left(-\kappa_n^R({\sigma})^\dagger\mathbf{\Phi}_{n+1}^R(0,\sigma)^\dagger(\kappa_n^L({\sigma}))^{-1}\right)^\dagger=
\alpha_n(\sigma)^\dagger,\end{gathered}$$ see .
\[eqnumberfive8\] The left and right orthogonal polynomials are related by formula $\eqref{lefttoright45}$.
We have $$\varphi_n^L(e^{i\theta}, \overline{\sigma})=\kappa_n^L(\overline{\sigma})\mathbf{\Phi}_n^L(e^{i\theta}, \overline{\sigma})=\kappa_n^R(\sigma)^\dagger\mathbf{\Phi}_n^R(e^{-i\theta},{\sigma})^\dagger=
\varphi_n^R(e^{-i\theta},{\sigma})^\dagger.$$
We next recall the notion of Bernstein-Szegő approximation. We begin with a list of properties of matrix orthogonal polynomials.
\[PropMOP\] The polynomials $\varphi^L$, $\varphi^R$ satisfy the following conditions:
1. For $z\in{\mathbb T}$, all of $\varphi_n^{R,*}(z)$, $\varphi_n^{L,*}(z)$, $\varphi_n^R(z)$, $\varphi_n^L(z)$ are invertible.
2. For $z\in{\mathbb D}$, $\varphi_n^{R,*}(z)$ and $\varphi_n^{L,*}(z)$ are invertible.
3. For any $z\in{\mathbb T}$, $$\label{LeftRight}
\varphi_n^R(z) \varphi_n^R(z)^\dagger =
\varphi_n^L (z)^\dagger
\varphi_n^L (z)~.$$
Given a finite sequence $\{\alpha_j\}_{j=0}^{n-1}$ of contractive matrices, we can always use the Szegő recursion to define the polynomials $\varphi_j^R, \varphi_j^L$ for $j=0,1,\dots, n$. Analogously to the scalar case, let us define a measure $d\mu_n$ on ${\mathbb T}$ by $$\label{eqszego1}
d\mu_n(\theta) = [\varphi_n^R(e^{i\theta})\varphi_n^R (e^{i\theta})^
\dagger]^{-1}
\, \frac{d\theta}{2\pi}.$$ In view of , we also see that $$\label{eqszego2}
d\mu_n(\theta)= [\varphi_n^L (e^{i\theta})^\dagger \varphi_n^L (e^{i
\theta})]^{-1} \, \frac{d\theta}{2\pi}\, .$$ Also, directly from the definition of the right orthogonal polynomials, we have $$\label{eqszego3}
d\mu_n(\theta)= [\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1} \, \frac{d\theta}{2\pi}.$$ The measure $d\mu_n$ in (\[eqszego3\]) is called the [*right Bernstein-Szegő approximation*]{} to $\sigma$. The [*left Bernstein-Szegő approximation*]{} to $\sigma$ is given by $$\label{eqszego32}
d\mu_n^L(\theta)= [\varphi_n^{L,*} (e^{i\theta})\varphi_n^{L,*} (e^{i
\theta})^\dagger ]^{-1} \, \frac{d\theta}{2\pi}.$$
Now we are in a position to formulate the main result of this section.
\[BSconv\] The matrix-valued measure $d\mu_n$ is normalized and its right matrix orthogonal polynomials for $j=0,\dots, n$ are $\{\varphi_j^R\}_{j=0}^n$. The Verblunsky coefficients for $d\mu_n$ are $$\alpha_j (d\mu_n) = \begin{cases} \alpha_j, & j\leq n, \\
\bf{0}, & j\geq n+1.
\end{cases}$$ Moreover, $*{-}\lim_{n \to \infty} d\mu_n =d \sigma$.
Following [@DPS:08], we associate the matrix $$\label{alphamatrix}
A^L(\alpha,z)=\begin{pmatrix}z(\rho^L)^{-1}&-(\rho^L)^{-1}\alpha^\dagger\\
-z(\rho^R)^{-1}\alpha&(\rho^R)^{-1}\end{pmatrix}$$ to a given matrix parameter $\alpha$. Then $$\label{it6}
\begin{pmatrix}
\varphi_{n}^L\\
\varphi_{n}^{R,*}
\end{pmatrix}=A^L(\alpha_{n-1},z)\cdots A^L(\alpha_{0},z)
\begin{pmatrix}
\mathbf{1}\\
\mathbf{1}
\end{pmatrix} .$$ Applying the adjoint $\dagger$ to both sides and taking the product over $\alpha_j$ for $j=0, \ldots, n-1$, we obtain $${\varphi_{n}^L}^{\dagger}\varphi_{n}^L+
{\varphi_{n}^{R,*}}^{\dagger}
\varphi_{n}^{R,*}=
\begin{pmatrix}
\mathbf{1}&
\mathbf{1}
\end{pmatrix}
{A^L(\alpha_{0},z)}^{\dagger}\cdots
{A^L(\alpha_{n-1},z)}^{\dagger}
A^L(\alpha_{n-1},z)\cdots A^L(\alpha_{0},z)
\begin{pmatrix}
\mathbf{1}\\
\mathbf{1}
\end{pmatrix} .$$ Note that and imply that the equality $$\label{importanteq}{\varphi_{n}^L}^{\dagger}\varphi_{n}^L=
{\varphi_{n}^{R,*}}^{\dagger}
\varphi_{n}^{R,*}$$ holds on the circle $\mathbb{T}$, implying that $$\label{idforBernstein}
{\varphi_{n}^{R,*}}^{\dagger}
\varphi_{n}^{R,*}=\frac{1}{2}\begin{pmatrix}
\mathbf{1}&
\mathbf{1}
\end{pmatrix}
{A^L(\alpha_{0},z)}^{\dagger}\cdots
{A^L(\alpha_{n-1},z)}^{\dagger}
A^L(\alpha_{n-1},z)\cdots A^L(\alpha_{0},z)
\begin{pmatrix}
\mathbf{1}\\
\mathbf{1}
\end{pmatrix} .$$
Using the fact $\rho^R \alpha = \alpha \rho^L$, the matrix in can be factored as follows: $$\label{factoralpha}
A^L(\alpha,z)=\begin{pmatrix}(\rho^L)^{-1}&0\\0&(\rho^R)^{-1}\end{pmatrix}
\begin{pmatrix}z\mathbf{1}&-\alpha^\dagger\\-z\alpha&\mathbf{1}\end{pmatrix}=
\begin{pmatrix}z\mathbf{1}&-\alpha^\dagger\\-z\alpha&\mathbf{1}\end{pmatrix}\begin{pmatrix}(\rho^L)^{-1}&0\\0&(\rho^R)^{-1}\end{pmatrix}.$$ To factor the non-diagonal matrix in , we apply Schur’s factorization $$\label{schurfact}
\begin{pmatrix}A&B\\C&D\end{pmatrix}=\begin{pmatrix}\mathbf{1}&0\\CA^{-1}&\mathbf{1}\end{pmatrix}
\begin{pmatrix}A&0\\0&D-CA^{-1}B\end{pmatrix}
\begin{pmatrix}\mathbf{1}&A^{-1}B\\0&\mathbf{1}\end{pmatrix}$$ with $$\begin{matrix}A=z\mathbf{1}&B=-\alpha^\dagger\\C=-z\alpha&D=\mathbf{1}\end{matrix} .$$ Then $$\label{4factor}
A^L(\alpha,z)=\begin{pmatrix}(\rho^L)^{-1}&0\\0&(\rho^R)^{-1}\end{pmatrix}
\begin{pmatrix}\mathbf{1}&0\\-\alpha&\mathbf{1}\end{pmatrix}
\begin{pmatrix}z\mathbf{1}&0\\0&\mathbf{1}-\alpha\alpha^\dagger\end{pmatrix}
\begin{pmatrix}\mathbf{1}&-(z)^{-1}\alpha^\dagger\\0&\mathbf{1}\end{pmatrix}.$$ Since $\det \rho^L=\det \rho^R$, we conclude that $$\label{detofA}
\det A^L(\alpha, z)=z^{\ell}.$$ It is easy to check that $$\varphi_1^L(z)=(\rho_0^L)^{-1}(z-\alpha_0^\dagger),
\qquad
\varphi_1^R(z)=(z-\alpha_0^\dagger)(\rho_0^R)^{-1}.$$ Further, forming the Szegő dual, we obtain $$\varphi_1^{L,*}(z)=({\bf 1}-z\alpha_0)(\rho_0^L)^{-1},
\qquad
\varphi_1^{R,*}(z)=(\rho_0^R)^{-1}({\bf 1}-z\alpha_0).$$ After pertinent multiplications, this produces $$\varphi_1^{R,*} (z)^\dagger \varphi_1^{R,*}(z)=
({\bf 1}-\overline{z}\alpha_0^\dagger)(\rho_0^R)^{-2}({\bf 1}-z\alpha_0),$$
$$\varphi_1^{L,*} (z) \varphi_1^{L,*}(z)^\dagger=
({\bf 1}-{z}\alpha_0)(\rho_0^L)^{-2}({\bf 1}-\overline{z}\alpha_0^\dagger).$$
\[formulaneq1\] For every $\alpha\in{\mathcal M}_\ell$ and $z\in\mathbb{T}$ $$({\bf 1}-\overline{z}\alpha^{\dag})({\bf 1}-\alpha\alpha^{\dag})^{-1}({\bf 1}-z\alpha)=\left[({\bf 1}-\overline{z}\alpha^{\dag})^{-1}+({\bf 1}-z\alpha)^{-1}-{\bf 1}\right]^{-1}.$$
Consider the matrix polynomial $$\label{poly}
p(z)=({\bf 1}-\overline{z}\alpha^{\dag})({\bf 1}-\alpha\alpha^{\dag})^{-1}({\bf 1}-z\alpha).$$ Since $z\overline{z}=1$, we have $$\label{first}
{\bf 1}-\alpha\alpha^{\dag}={\bf 1}-z\overline{z}\alpha\alpha^{\dag}=({\bf 1}-z\alpha)({\bf 1}+\overline{z}\alpha^{\dag})+z\alpha-\overline{z}\alpha^{\dag}.$$ Similarly, $$\label{second}
{\bf 1}-\alpha\alpha^{\dag}={\bf 1}-z\overline{z}\alpha\alpha^{\dag}=({\bf 1}+z\alpha)({\bf 1}-\overline{z}\alpha^{\dag})-z\alpha+\overline{z}\alpha^{\dag}.$$ The sum of the expressions and yields $$\label{main}
\begin{array}{ccl}
2({\bf 1}-\alpha\alpha^{\dag}) &= &({\bf 1}+z\alpha)({\bf 1}-\overline{z}\alpha^{\dag})+({\bf 1}-z\alpha)({\bf 1}+\overline{z}\alpha^{\dag}) \\
&=&({\bf 2}-({\bf 1}-z\alpha))({\bf 1}-\overline{z}\alpha^{\dag})+({\bf 1}-z\alpha)({\bf 2}-({\bf 1}-\overline{z}\alpha^{\dag})).
\end{array}$$ Let us denote ${\bf 1}-z\alpha$ by $B$ for brevity. From and we obtain $$\label{poly.2}
\begin{array}{ccl}
p(z) &= &B^{\dag}({\bf 1}-\alpha\alpha^{\dag})^{-1}B \; =\; 2B^{\dag}\left[B({\bf 2}-B^{\dag})+({\bf 2}-B)B^{\dag}\right]^{-1}B \\
&=&2\left[({\bf 2}-B^{\dag})B^{-\dag}+B^{-1}({\bf 2}-B)\right]^{-1} \; = \; 2\left[2B^{-\dag}-{\bf 2}+2B^{-1}\right]^{-1}\\
&=&\left[({\bf 1}-\overline{z}\alpha^{\dag})^{-1}+({\bf 1}-z\alpha)^{-1}-{\bf 1}\right]^{-1}.
\end{array}$$
The Bernstein–Szegő approximation
=================================
In this section we obtain a formula for the Bernstein–Szegő approximation of a matrix probability measure.
\[neglemma7\] Let $\beta_n$ be the matrix defined by $$\label{an89}
\beta_n=\exp\int_{0}^{2\pi}\log([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}.$$ Then $\log\beta_n$ is self-adjoint and nonpositive.
Since $[\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1}$ is a positive matrix for every $\theta$ its logarithm is self-adjoint as well as the integral $\log\beta_n$ of the logarithm of this matrix. By Proposition \[prJensen\], $$\log\beta_n=\int_{0}^{2\pi}\log([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}\preceq\log\left(\int_{0}^{2\pi}[\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}\right)=\log\mu_n(\mathbb{T})=\mathbf{0}$$ since $\mu_n$ is normalized: $\mu_n(\mathbb T)={\bf 1}$.
The standard operator calculus and Lemma \[neglemma7\] imply that the matrix $\beta_n=\exp \log\beta_n$ is self-adjoint and satisfies $$\label{bettaop9}
\mathbf{0}\prec \beta_n\preceq \mathbf{1}.$$
\[detid0\] For $\beta_n$, we have $$\label{ident65}
\log\det{\beta_n}=\text{tr}\log\beta_n=\log\prod_{k=0}^{n-1}\det\left({\bf 1} -\alpha_k \alpha_k^\dagger\right)=\sum_{k=0}^{n-1}\text{tr}\log\left({\bf 1} -\alpha_k \alpha_k^\dagger\right).$$
Applying elementary transformations and formula , we obtain $$\begin{gathered}
\text{tr}(\log\beta_n)=\int_{0}^{2\pi}\text{tr}\log([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}=\int_{0}^{2\pi}\log\det([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})\frac{d\theta}{2\pi}\\=\int_{0}^{2\pi}\log\det([\varphi_n^{R,*} (e^{i\theta})^\dagger]^{-1})\frac{d\theta}{2\pi}+\int_{0}^{2\pi}\log\det([\varphi_n^{R,*} (e^{i\theta})]^{-1})\frac{d\theta}{2\pi}=2\textbf{Re}\int_{0}^{2\pi}\log\det([\varphi_n^{R,*} (e^{i\theta})]^{-1})\frac{d\theta}{2\pi}\\=2\textbf{Re}\log\det([\varphi_n^{R,*} (0)]^{-1})=
2\log\det\left(\rho_{0}^R\cdots\rho_{n-1}^R\right)=\log\prod_{k=0}^{n-1}\det\left({\bf 1} -\alpha_k \alpha_k^\dagger\right)\end{gathered}$$ since the function $z\mapsto\log\det\left([\varphi_n^{R,*} (z)]^{-1}\right)$ is analytic in the closed unit disc.
[**Remark.**]{} In general, $\log (AB)$ cannot be written as $\log A + \log B$ if $A$ and $B$ are matrices. So, the integral in Lemma \[neglemma7\] cannot be evaluated by the mean value theorem. In other words, the function $\log([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1})$ is in general not a restriction of a harmonic function to the unit circle. Our next lemma addresses the easy case when the logarithm in question does split.
If $\{\alpha_1,\ldots,\alpha_{n-1}\}$ is a commuting family of normal matrices, then the self-adjoint matrix $\beta_n$ can be evaluated explicitly as follows:
\[directcomp\] Let $\{\alpha_1,\ldots,\alpha_{n-1}\}$ be a commuting family of normal matrices. Then $$\beta_n=\prod_{k=0}^{n-1}\left({\bf 1} -\alpha_k \alpha_k^\dagger\right).$$
The proof follows the proof of Lemma \[detid0\] since in this case $\varphi_n^{R,*} (z)$ is a normal matrix for any value of $z$.
The Matrix Szegö Theorem
========================
A matrix probability measure $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ is said to be a Szegö measure if $$\label{defSz}
\int_{\mathbb{T}}\operatorname{tr}\log \sigma^{\prime}\,\frac{d\theta}{2\pi}>-\infty.$$
\[firstineq\]For any matrix probability measure $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ and any $n\in {\mathbb N}$, $$\label{HalfEq}
\int_{\mathbb{T}}\operatorname{tr}\log \sigma^{\prime}\,\frac{d\theta}{2\pi}\leq
\operatorname{tr}\log\beta_n=\log\prod_{k=0}^{n-1}\det(1-\alpha_k^\dagger \alpha_k).$$
If $\int_{\mathbb{T}}\operatorname{tr}\log \sigma^{\prime}\,\frac{d\theta}{2\pi}=-\infty$, the conclusion of the theorem holds trivially, so assume that $\sigma$ is a Szegő measure, i.e., the corresponding integral is not $-\infty$. Jensen’s matrix inequality from Proposition \[prJensen\] implies $$\label{GPOP519}
\int_{\mathbb{T}}\log
\left(\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\varphi_n^{R,*} (e^{i
\theta})\beta_n^{1/2}]\right)
\frac{d\theta}{2\pi}
\preceq \log\left(\int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\frac{d\theta}{2\pi}\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right).$$ Now we replace $\sigma^{\prime}$ by $\sigma$, which is larger in the Loewner ordering, according to . Since $\sigma$ is absolutely continuous with respect of $\operatorname{tr}(\sigma)$, there exist two disjoint Borel sets $E$ and $F$ and a Borel matrix function $x\mapsto \texttt{M}(x)$ such that $$d\sigma =\texttt{M}\chi_E\operatorname{tr}(\sigma_a)+\texttt{M}\chi_F\operatorname{tr}(d\sigma_d+d\sigma_s)\,$$ where $E$ is a Borel support of $\operatorname{tr}(d\sigma_a)$, $F$ is a Borel support of $\operatorname{tr}(d\sigma_d+d\sigma_s)$ and $\chi_E$, $\chi_F$ are the indicators of $E$ and $F$ correspondingly. Notice that $$\label{eqer57}
\begin{aligned}
\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \texttt{M}\chi_E\operatorname{tr}(\sigma_a)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}&=\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\frac{d\theta}{2\pi}\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2};\\
\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \texttt{M}\chi_F\operatorname{tr}(d\sigma_d+d\sigma_s)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}&=\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger (d\sigma_d+d\sigma_s)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}.
\end{aligned}$$ Combining with the result of Lemma \[triv87\], we obtain $$\begin{gathered}
\int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\frac{d\theta}{2\pi}\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}=\int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \texttt{M}\chi_E\operatorname{tr}(\sigma_a)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2} \\
\preceq \int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger \texttt{M}\chi_E\operatorname{tr}(\sigma_a)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}+\int_{\mathbb{T}}[\varphi_n^{R,*} (e^{i\theta})^\dagger \texttt{M}\chi_F\operatorname{tr}(d\sigma_d+d\sigma_s)\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\\=
\int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger d\sigma\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}.\end{gathered}$$ By and by the operator monotonicity of the logarithm from Lemma \[OpMon\], we obtain $$\begin{gathered}
\label{importantineq}
\int_{\mathbb{T}}\log
\left(\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)
\frac{d\theta}{2\pi}
\preceq\log\left(\int_{\mathbb{T}}\beta_n^{1/2}
[\varphi_n^{R,*} (e^{i\theta})^\dagger{d\sigma} \varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)\\=\log\left(\beta_n^{1/2}\int_{\mathbb{T}}[\varphi_n^{R,*} (e^{i\theta})^\dagger{d\sigma} \varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)=\log\left(\beta_n^{1/2}{\bf 1}\beta_n^{1/2}\right)=\log\beta_n\,,\end{gathered}$$ in view of the orthonormality of the polynomials $\varphi^R_n$. Next, we have $$\begin{gathered}
\text{tr}\log\left(\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime} \varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)=\log\det\left(\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime} \varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)\\=
\log\left[\det(\beta_n)\det([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})])\det(\sigma^{\prime})\right]=\log\det(\beta_n)+\log\det([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})])+\log\det(\sigma^{\prime})\\=
\text{tr}\log\beta_n+\text{tr}\log[\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]+\text{tr}\log\sigma^{\prime}.\end{gathered}$$ Integrating the above equality and taking into account , and , we arrive at $$\begin{gathered}
\text{tr}\log\beta_n\geq \int_{\mathbb{T}}\text{tr}\log
\left(\beta_n^{1/2}[\varphi_n^{R,*} (e^{i\theta})^\dagger \sigma^{\prime}\varphi_n^{R,*} (e^{i
\theta})]\beta_n^{1/2}\right)
\frac{d\theta}{2\pi}\\=
\text{tr}\log\beta_n+\int_{\mathbb{T}}\text{tr}\log
\left([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]\right)\frac{d\theta}{2\pi}+\int_{\mathbb{T}}\text{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}=
\int_{\mathbb{T}}\text{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}\,.\end{gathered}$$ It remains to apply Lemma \[detid0\].
\[leftineqb\] If $\sigma$ is a Szegő measure, then $$\label{firstin56}
\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}\leq\inf_n\text{tr}\log\beta_n\leq-\sup_n\log\Vert\beta_n^{-1}\Vert \leq 0,$$ in particular, $\sup_n\Vert\beta_n^{-1}\Vert<+\infty$.
Since $\beta_n$ satisfies , all its eigenvalues $\lambda_k$, $1\leq k\leq \ell$, must lie in the interval $(0,1]$. In addition, $$\Vert\beta_n^{-1}\Vert=\max_{1\leq k\leq \ell}\lambda_k^{-1}.$$ By and Lemma \[triv89\] $$-\infty<\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}\leq\log\det\left(\beta_n\right)=\operatorname{tr}\log\beta_n\leq 0,$$ implying that $$\log\Vert\beta_n^{-1}\Vert=\max_k\log\lambda_k^{-1}<\sum_{k=1}^{\ell}\log\lambda_k^{-1}=\operatorname{tr}\log\beta_n^{-1}\leq-\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}<+\infty.$$
By compactness of closed balls in the finite-dimensional space ${\mathcal M}_\ell$, a bounded sequence of matrices has a limit point. It follows that if $\{\beta_n^{-1}\}_{n\geq 0}$ is uniformly bounded, then any of its limit points $\beta^{-1}$ in ${\mathcal M}_\ell$ satisfies $$\label{boundbeta}
\Vert\beta^{-1}\Vert\leq \sup_n\Vert\beta_n^{-1}\Vert<+\infty.$$ We denote by $C(\mathbb{T}, {\mathcal M}_\ell^{+})$ the set of all continuous matrix functions on $\mathbb{T}$ with values in ${\mathcal M}_\ell^{+}$, by $L^2(\mathbb{T}, {\mathcal M}_\ell^{+})$ the set of all square-integrable matrix functions on $\mathbb{T}$ with values in ${\mathcal M}_\ell^{+}$, and by $M(\mathbb{T}, {\mathcal M}_\ell^{+})$ the set of all finite Borel measures with values in ${\mathcal M}_\ell^{+}$.
Let $\mu$ be a finite Borel measure with values in ${\mathcal M}_\ell^{+}$. Suppose that, for any open arc $I\subset\mathbb{T}$ whose endpoints do not carry point masses of $\sigma$, the inequality $$\mu(I)\preceq\sigma(I)$$ holds. Then we write $d\mu\preceq d\sigma$.
\[main295\] Let $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ satisfy $\sup_n\Vert\beta_n^{-1}\Vert<+\infty$, with $\beta_n$ defined as above, and let $\{f_n\}_{n\geq 0}$ be a sequence in $C(\mathbb{T}, {\mathcal M}_\ell^{+})$ such that $f_n(e^{i\theta})\succ \mathbf{0}$ on $\mathbb{T}$ and let $$\begin{aligned}
\int_{\mathbb{T}}f_n&\frac{d\theta}{2\pi}\preceq \mathbf{1};\label{formula1}\\
*{-}\lim_nf_n&\frac{d\theta}{2\pi}\preceq d\sigma;\label{formula2}\\
\log\beta_n\preceq \int_{\mathbb{T}}&\log f_n\frac{d\theta}{2\pi}.\label{formula3}\end{aligned}$$ Then $$\label{matrixeq}
\lim_n\log\beta_n=\int_{\mathbb{T}}\log\sigma^{\,\prime}\frac{d\theta}{2\pi}$$ in ${\mathcal M}_\ell$.
By and , the sequence of negative operators $\log\beta_n$ is uniformly bounded. Suppose that $\log\beta$ is a limit point of this sequence of matrices in ${\mathcal M}_\ell$. Then there is an infinite subset $\Lambda$ of $\mathbb N$ such that $$\label{limitbetan}
\lim_{n\in\Lambda}\log\beta_n=\log\beta.$$
Let $$\log^{+} x=\max(\log x, 0)\;,\; \log^{-} x= \log^{+} x-\log x\;.$$ Then $\log^{+} x\leq x$ for every $x>0$. The Spectral Theorem applied to a (strictly) positive operator $A$ yields $$\label{logAa2}
\log^{+}(A)\preceq A .$$ We apply to $A:=f_n(e^{i\theta})$ pointwise in $\theta$ and obtain the operator inequality $$\label{oppointinlog1}
\log^{+}(f_n(e^{i\theta}))\preceq f_n(e^{i\theta}).$$ Integrating and taking into account , we obtain $$\label{IneqJen3.61}
\int_{\mathbb{T}}\log^{+}(f_n(e^{i\theta}))\frac{d\theta}{2\pi}\preceq \mathbf{1}.$$ Observing that $\log=\log^{+}-\log^{-}$ and using and , we see that $$\label{eqRT3.71}
\int_{\mathbb{T}}\log^{-}(f_n(e^{i\theta}))\frac{d\theta}{2\pi}\preceq \mathbf{1}+\log\beta_n^{-1}.$$ Let $$d\nu_n^{+}\eqbd \log^{+}(f_n(e^{i\theta}))\frac{d\theta}{2\pi},\qquad d\nu_n^{-}\eqbd \log^{-}(f_n(e^{i\theta}))\frac{d\theta}{2\pi}.$$ Since $\{\beta_n^{-1}\}_{n\geq 0}$ is bounded, (\[eqRT3.71\]) implies that $\{\nu_n^{-}\}_{n\geq 0}$ has a $*$-weak limit point $\nu^-\in M(\mathbb{T}, {\mathcal M}_\ell^{+})$: $$\label{limmuminus5221}
d\nu^-=*{-}\lim_{n\in\Lambda^{'}} d\nu_n^-=(\nu^-)'\frac{d\theta}{2\pi}+d\nu^-_s\quad
\text{\rm for some } \; \Lambda^{'}\subset\Lambda,$$ where $d\nu^-_s$ is the singular part of $d\nu^-$ (it may include the discrete part as well), $(\nu^-)'=d\nu^-/(\frac{d\theta}{2\pi})$. It follows from the inequality $(\log^{+} x)^2\leq x$ and that $$\label{IneqJen3.6h1}
\int_{\mathbb{T}}\left(\log^{+}(f_n(e^{i\theta})\right)^2\,\frac{d\theta}{2\pi}
\preceq{\bf 1}\;,$$ By (\[IneqJen3.6h1\]), the function $d\nu_n^+/(\frac{d\theta}{2\pi})$ is in the unit ball of $L^2(\mathbb{T}, {\mathcal M}_\ell)$, which is compact in the weak topology of $L^2(\mathbb{T}, {\mathcal M}_\ell)$, see Theorem \[Ltwoth\]. It follows that any $*$-limit point $\omega$ of $\{\nu_n^+\}_{n\geq 0}$ in $M(\mathbb{T}, {\mathcal M}_\ell)$ is absolutely continuous with respect to the Lebesgue measure and, moreover, belongs to $L^2(\mathbb{T}, {\mathcal M}_\ell^{+})$. Then there exist a subset $\Lambda^{''}\subset\Lambda^{'}$ and some $\omega^{\prime}$ in the unit ball of $L^2(\mathbb{T}, {\mathcal M}_\ell^{+})$ such that $$\label{mainpaper523}
\begin{gathered}
d\nu^+ \eqbd *{-}\lim_{n\in\Lambda^{''}}d\nu_n^+=\omega^{\prime}\frac{d\theta}{2\pi}
\;,\qquad *{-}\lim_{n\in\Lambda^{''}}d\nu^{-}_n=d\nu^{-}\;,\\
d\nu \eqbd d\nu^+-d\nu^-=(\omega^{\prime}-(\nu^-)') \frac{d\theta}{2\pi}-d\nu^-_s,
\end{gathered}$$ see . Let $I$ be an open arc on $\mathbb{T}$ such that its endpoints do not carry point masses of $d\nu^-_s$ or $d\sigma_s$. By matrix Jensen’s inequality from Proposition \[prJensen\], we get $$\label{JenSz3.101}
\frac{1}{|I|}\int_{I}\log(f_n(e^{i\theta}))\frac{d\theta}{2\pi}\preceq
\log\left\{\frac{1}{|I|}\int_I f_n(e^{i\theta})\frac{d\theta}{2\pi}\right\}\;.$$ Applying Helly’s Theorem \[Helley1\] separately to $\{\nu_n^+\}_{n\in\Lambda''}$ and to $\{\nu_n^-\}_{n\in\Lambda''}$, we obtain $$\label{mainpap5261}
\lim_{n\in\Lambda''}\frac{1}{|I|}\int_{I}\log(f_n(e^{i\theta}))\frac{d\theta}{2\pi}=\frac{\nu(I)}{|I|}\;.$$ Applying Helly’s Theorem \[Helley1\], we derive from the inequality $$\label{mainpap5271}
\lim_{n}\frac{1}{|I|}\int_If_n(e^{i\theta})\frac{d\theta}{2\pi}\preceq \frac{\sigma(I)}{|I|}\;.$$ A substitution of (\[mainpap5261\]) and (\[mainpap5271\]) into (\[JenSz3.101\]) results in the inequality $$\frac{\nu(I)}{|I|}\preceq \log\left(\frac{\sigma(I)}{|I|}\right)\;$$ (here we use the operator continuity of the logarithm, see Proposition \[OpCont\]). It follows from Lebesgue’s theorem on differentiation and the operator continuity of the logarithm that $$\label{LebDiff3.111}
\nu^{\prime}\preceq \log(\sigma^{\prime})$$ almost everywhere on $\mathbb{T}$. In view of (\[LebDiff3.111\]) and (\[formula3\]), we obtain $$\label{MainPaper5291}
\log{\beta}+\nu^-_s(\mathbb{T})\preceq\int_{\mathbb{T}}d\nu+\nu^-_s(\mathbb{T})=
\int_{\mathbb{T}}\nu^{\prime}\frac{d\theta}{2\pi}
\preceq\int_{\mathbb{T}}\log\sigma^{\prime}\frac{d\theta}{2\pi}\;.$$ Combining with (\[MainPaper5291\]), we see that $$\label{logsigmaform85941}
\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}=\operatorname{tr}\log\beta.$$ and $\operatorname{tr}\nu_s(\mathbb{T})=0$, so $\nu^-_s=0$ by the nonnegativity of the measure $\nu^-_s$. It follows that $$\log{\beta}\preceq\int_{\mathbb{T}}\log\sigma^{\prime}\frac{d\theta}{2\pi}.$$ Since the traces of the operators on both sides are equal by , we invoke Lemma \[triv88\] and conclude that $$\log{\beta}=\int_{\mathbb{T}}\log\sigma^{\prime}\frac{d\theta}{2\pi}.$$ Since $\log\beta$ is an arbitrary limit point of $\{\log\beta_n\}_{n\geq 0}$, we obtain .
\[secondth0\]Let $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ satisfy $\sup_n\Vert\beta_n^{-1}\Vert<+\infty$. Then $$\lim_n\log\beta_n=\int_{\mathbb{T}}\log\sigma^{\,\prime}\frac{d\theta}{2\pi} .$$
Set $$f_n(e^{i\theta})=[\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1}$$ in Theorem \[main295\]. Then and follow from Theorem \[BSconv\]. Finally, follows from .
\[szego\_measure\] For any $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$, $$\label{Szlogcond3.4}
\log\prod_{k=0}^{\infty}\det(1-\alpha_k^\dagger \alpha_k)=
\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}\;.$$
Since $\det(1-\alpha_k^\dagger\alpha_k)<1$ for all $k$, the sum of the series $$\sum_{k=0}^{\infty}\log\det(1-\alpha_k^\dagger\alpha_k)$$ with negative terms satisfies $$\label{pofght68re}
\sum_{k=0}^{\infty}\log\det(1-\alpha_k^\dagger\alpha_k)\geq\int_{\mathbb{T}}\operatorname{tr}\log\sigma^{\prime}\frac{d\theta}{2\pi}$$ by Theorem \[firstineq\]. We have two cases. If the series on the left-hand side of diverges, then $\sigma$ is not a Szegö measure and both sides of equal $-\infty$. If the series on the left-hand side of converges, then $$\lim_k\log\det(1-\alpha_k^\dagger\alpha_k)=\lim_k\text{tr}\log(1-\alpha_k^\dagger\alpha_k)=0.$$ Since the spectral norm $\|\cdot \|$ is the largest eigenvalue of a positive self-adjoint matrix, it follows that $$\lim_k\Vert\alpha_k^\dagger\alpha_k\Vert=0.$$ Since $-x\geq\log(1-x)$ for $0<x<1$, we see that $$-\Vert\alpha_k^\dagger\alpha_k\Vert\geq \log(1-\Vert\alpha_k^\dagger\alpha_k\Vert)\geq \text{tr}\log(1-\alpha_k^\dagger\alpha_k),$$ implying that $$\sum_{k=0}^{\infty}\Vert\alpha_k^\dagger\alpha_k\Vert<+\infty.$$ Lemma \[detid0\] implies that the $\Vert\beta_n^{-1}\Vert$ are bounded. An application of Theorem \[secondth0\] now completes the proof.
\[parametersth\] A measure $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ is a Szegö measure if and only if $$\sum_{k=0}^{\infty}\Vert\alpha_k^\dagger\alpha_k\Vert<+\infty.$$
One direction of this corollary was already proved in Theorem \[szego\_measure\]; the other direction can be obtained analogously to [@DGK:78]
\[colSz3.1\] Let $\sigma$ be a Szegö measure and let $\{\varphi_n\}_{n\geq 0}$ be the orthogonal polynomials in $L^2(d\sigma)$. Then $$\label{weakconvSz3.14}
*{-}\lim_n d\mu_n= *{-}\lim_n\log\left([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i
\theta})]^{-1}\right)\frac{d\theta}{2\pi}=\log(\sigma^{\prime})\frac{d\theta}{2\pi}$$ in the weak topology of $M(\mathbb{T, {\mathcal M}_\ell})$.
Apply the proof of Theorem \[main295\] to the measures $$\begin{aligned}
d\nu_n^+ \eqbd d\mu_n^+ \eqbd
\log^+\left([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*}
(e^{i\theta})]^{-1}\right) \frac{d\theta}{2\pi}, \\
d\nu_n^- \eqbd d\mu_n^- \eqbd
\log^-\left([\varphi_n^{R,*} (e^{i\theta})^\dagger \varphi_n^{R,*}
(e^{i\theta})]^{-1}\right) \frac{d\theta}{2\pi}.\end{aligned}$$ Taking into account, we obtain $$\label{mainpape530}
\nu^{\prime}=\log\sigma^{\prime}\quad \text{ a.e. on } \mathbb{T}\;.$$ The substitution of (\[mainpape530\]) into the last formula of (\[mainpaper523\]) results in $$d\nu=\log\sigma^{\prime}\frac{d\theta}{2\pi}=(\omega^{\prime}-(\nu^-)')\frac{d\theta}{2\pi}\;.$$ Since $\omega$ in the proof of Theorem \[main295\] was an arbitrary $*$-limit point of $\{\nu_n^{+}\}_{n\in\Lambda}$, this implies that $*{-}\lim_{n\in\Lambda'}d\nu_n^+=\omega^{\prime}\frac{d\theta}{2\pi}$. Since $\nu^-$ was an arbitrary $*$-limit point of $\{\nu_n^{-}\}_{n\in\Lambda}$, we conclude that $*{-}\lim_{n}d\mu_n=\log\sigma^{\prime}\frac{d\theta}{2\pi}$.
The Helson-Lowdenslager Theorem
===============================
Since $\mathbf{\Phi}_n^{R,*}$ is left orthogonal to $z\mathbf{1},\ldots,z^n\mathbf{1}$ (see [@DPS:08 Lemma 3.2]), it is also left orthogonal to any linear combination $p$ of these matrix functions with the coefficients in ${\mathcal M}_\ell$. Take any such combination $p$. Then $$\begin{gathered}
{\langle\! \langle\mathbf{\Phi}_n^{R,*}-p,\mathbf{\Phi}_n^{R,*}-p\rangle\! \rangle}_L=
{\langle\! \langle\mathbf{\Phi}_n^{R,*},\mathbf{\Phi}_n^{R,*}\rangle\! \rangle}_L+{\langle\! \langlep,p\rangle\! \rangle}_L-{\langle\! \langle\mathbf{\Phi}_n^{R,*},p\rangle\! \rangle}_L
-{\langle\! \langle\mathbf{\Phi}_n^{R,*},p\rangle\! \rangle}_L^\dagger=
{\langle\! \langle\mathbf{\Phi}_n^{R,*},\mathbf{\Phi}_n^{R,*}\rangle\! \rangle}_L+{\langle\! \langlep,p\rangle\! \rangle}_L.\end{gathered}$$ Since every polynomial $\mathbf{Q}$ satisfying $\mathbf{Q}(0)=\mathbf{1}$ is of the form $\mathbf{Q}=\mathbf{\Phi}_n^{R,*}-p$ we obtain the matrix inequality $$\label{varprincip}
{\langle\! \langle\mathbf{\Phi}_n^{R,*},\mathbf{\Phi}_n^{R,*}\rangle\! \rangle}_L\preceq {\langle\! \langle\mathbf{Q},\mathbf{Q}\rangle\! \rangle}_L.$$ These facts are also derived for the real line in [@DPS:08 Formula (2.10)].
It is therefore natural to call the square root of the positive matrix in the left-hand side of the [*left operator distance*]{} from $\mathbf{1}$ to $z\mathcal{P}_{n-1}$. Consequently, the usual distance in the left Hilbert space is equal to $$\label{distform}
\text{dist}_L\left(\textbf{1},z\mathcal{P}_{n-1}\right)^2=\text{tr}\left({\langle\! \langle\mathbf{\Phi}_n^{R,*},\mathbf{\Phi}_n^{R,*}\rangle\! \rangle}_L\right).$$ One easily verifies (see also [@DPS:08 Lemma 3.1]) that $$\label{eval745}
\begin{array}{lcl} {\langle\! \langle\mathbf{\Phi}_n^{R,*},\mathbf{\Phi}_n^{R,*}\rangle\! \rangle}_L & = & {\langle\! \langle\mathbf{\Phi}_n^{R},\mathbf{\Phi}_n^{R}\rangle\! \rangle}_R^\dagger={\langle\! \langle\mathbf{\Phi}_n^{R},\mathbf{\Phi}_n^{R}\rangle\! \rangle}_R
= (\kappa_n^{R})^{-\dagger}{\langle\! \langle{\varphi}_n^{R},{\varphi}_n^{R}\rangle\! \rangle}_R(\kappa_n^{R})^{-1}
\\[2mm]
&=& (\kappa_n^{R})^{-\dagger}\mathbf{1}
(\kappa_n^{R})^{-1}=(\kappa_n^{R})^{-\dagger}(\kappa_n^{R})^{-1}.
\end{array}$$ It follows from that $$\label{distformula78}
((\kappa_n^{R})^{-\dagger}(\kappa_n^{R})^{-1})^{1/2} =\rho_{n-1}^R\cdots\rho_{0}^R=({\bf 1} -\alpha_{n-1} \alpha_{n-1}^\dagger)^{1/2}\cdots
({\bf 1} -\alpha_{0} \alpha_{0}^\dagger)^{1/2}$$ is the left matrix distance from $\mathbf{1}$ to $z\mathcal{P}_{n-1}$.
The identity polynomial $\mathbf{1}$ is in the left closure of the sets of matrix polynomials $z\mathcal{P}_{n-1}$ if and only if $$\exp\int_{\mathbb{T}}\log\sigma^{\,\prime}\frac{d\theta}{2\pi}=\mathbf{0}.$$
The distance formula is useful if the parameters $\{\alpha_k\}_{k\geq0}$ of $\sigma$ are known. If this is not the case, then one can apply an estimate for $(\kappa_n^{R})^{-1}$ from below which was obtained by Helson-Lowdenslager in [@HelsonLowdenslager].
Now we are in a position to prove the main result of [@HelsonLowdenslager].
\[HL\] For every $\sigma\in\textsf{P}_{\ell}(\mathbb{T})$ $$\label{HLeq}
\exp\int_{\mathbb{T}}\frac{1}{\ell}\text{tr}\log\sigma^{\,\prime}\,\frac{d\theta}{2\pi}=
\inf\limits_{A,P}\int_{\mathbb{T}}\frac{1}{\ell}\operatorname{tr}\left[(A+P)^\dagger
d\sigma(A+P)\right],$$ where $A$ runs over all matrices with determinant one, and $P$ over all trigonometric polynomials of the form $$P(e^{i\theta})=\sum_{k>0}A_ke^{ik\theta}.$$
Combining Lemma \[HL-Lemma\] with formula , we get $$\begin{aligned}
&& \inf_{A\in\mathcal{A}}
\frac{1}{\ell}\text{tr}
A \left( (\kappa_n^R)^{-\dagger}
(\kappa_n^R)^{-1} \right) A^\dagger =
[\det((\kappa_n^{R})^{-\dagger} (\kappa_n^R)^{-1}) ]^{1/\ell} \\ && =
\exp\left\{\frac{1}{\ell}\sum_{j=0}^{n-1}\log\det(1-\alpha_j\alpha_j^\dagger)\right\}=
\exp\left\{\frac{1}{\ell}\text{tr}(\log\beta_n)\right\}.\end{aligned}$$ Combining this formula with , we obtain $$\begin{aligned}
\inf\limits_{A\in\mathcal{A},P\in z\mathcal{P}_{n-1}}\int_{\mathbb{T}}\frac{1}{\ell}\operatorname{tr}\left[(A+P)
d\sigma(A+P)^\dagger\right]=\inf\limits_{A\in\mathcal{A},P\in z\mathcal{P}_{n-1}}
\int_{\mathbb{T}}{1\over \ell}\operatorname{tr}A (\mathbf{1}+zA^{-1}P) d \sigma
(\mathbf{1}+zA^{-1}P)^\dagger A^\dagger \\
= \inf_{\mathcal{A}} {1\over \ell} \operatorname{tr}A \left( (\kappa_n^R)^{-\dagger}
(\kappa_n^R)^{-1} \right) A^\dagger =
\exp\left\{\frac{1}{\ell}\int_{0}^{2\pi} \text{tr}\log([\varphi_n^{R,*}
(e^{i\theta})^\dagger \varphi_n^{R,*} (e^{i\theta})]^{-1})\frac{d\theta}{2\pi}\right\}.\end{aligned}$$ Passing to the limit, we arrive at , which was initially proved via a different method in [@HelsonLowdenslager Theorem 8].
Acknowledgments {#acknowledgments .unnumbered}
===============
The research leading to these results was carried out at Technische Universität Berlin and has received funding from the Alexander von Humboldt Foundation under the Sofja Kovalevskaja Prize Programme and from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement ${\rm n}^\circ$ 259173.
[0]{}
Ando, T., and Hiai, F. “Operator log-convex functions and operator means.” *Mathematische Annalen*, DOI: 10.1007/s00208-010-0577-4.
Aptekarev, A. I.; Nikishin, E. M. “The scattering problem for a discrete Sturm-Liouville operator.” (Russian) *Mat. Sb. (N.S.)* 121(163) (1983), no. 3, 327–358.
Bhatia, R. *Matrix Analysis*. New York: Springer, 1996.
Delsarte, P., Genin, Y. V., and Kamp, Y. G. “Orthogonal polynomial matrices on the unit circle.” *IEEE Trans. Circuits and Systems* CAS-25 (1978), no. 3: 149–160.
Damanik, D., Pushnitski, A., and Simon, B., “The analytic theory of matrix orthogonal polynomials.” *Surveys in Approximation Theory* 4 (2008): 1–85.
Farenick, D. R., and Zhou, F. “Jensen’s inequality relative to matrix-valued measures.” *J. Math. Anal. Appl.* 327 (2007): 919–929.
Hannan, E.J. *Multiple Time Series.* New York: Wiley, 1970.
Helson H., Lowdenslager D. “Prediction theory and Fourier series in several variables.” *Acta Math.*, 99 (1958): 165–202.
Horn R., Johnson C. *Matrix Analysis*. Corrected reprint of the 1985 original. Cambridge, 1990.
Khrushchev, S. “Schur’s algorithm, orthogonal polynomials, and convergence of Wall’s continued fractions in $L^2(\mathbb{T})$.” *J. of Approx. Theorey*, 108 (2001), 161–248.
Khrushchev, S. *Orthogonal Polynomials and Continued Fractions. From Euler’s point of view*. Cambridge, 2008.
Masani, P. “Wiener’s contributions to generalized harmonic analysis, prediction theory and filter theory. ” *Bull. Amer. Math. Soc.*, 72, No 1, Part 2 (1966), 73–125.
Rozanov, Yu.A. *Stationary Random Processes*. Holden-Day, San Francisco, 1967.
Simon, B. *Orthogonal Polynomials on the Unit Circle, Part I: Classical Theory*. AMS Colloquium Publications, vol 54, Part 1, 2005.
Szegő, G. “Über den asymptotischen Ausdruck von Polynomen, die durch eine Orthogonalitätseigenschaft definiert sind”, *Math. Ann.*, 86 (1922), 114–139.
Verblunsky, S. “On positive harmonic functions.” *Proc. London Math. Soc.*, 38, No 2 (1935), 125–157.
Wiener, N., Masani, P. “The prediction theory of multivariate stochastic processes, I. The regularity conditions.” *Acta Math.*, 98 (1957), 111–150.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In 1879, Thomae discussed the relations between two generic hypergeometric $_3F_2$-series with argument $1$. It is well-known since then that there are 120 such relations (including the trivial ones which come from permutations of the parameters of the hypergeometric series). More recently, Rhin and Viola asked the following question (in a different, but equivalent language of integrals): If there exists a linear dependence relation over $\mathbf{Q}$ between two convergent ${} _{3} F _{2}$-series with argument $1$, with integral parameters, and whose values are irrational numbers, is this relation a specialisation of one of the 120 Thomae relations? A few years later, Sato answered this question in the negative, by giving six examples of relations which cannot be explained by Thomae’s relations. We show that Sato’s counter-examples can be naturally embedded into two families of infinitely many $_3F_2$-relations, both parametrised by three independent parameters. Moreover, we find two more infinite families of the same nature. The families, which do not seem to have been recorded before, come from certain $_3F_2$-transformation formulae and contiguous relations. We also explain in detail the relationship between the integrals of Rhin and Viola and $_3F_2$-series.'
address:
- 'Institut Girard Desargues, Universit[é]{} Claude Bernard Lyon-I, 21, avenue Claude Bernard, F-69622 Villeurbanne Cedex, France'
- 'Institut Fourier, CNRS UMR 5582 / Université Grenoble 1, 100 rue des Maths, BP 74, 38402 Saint-Martin d’Hères cedex, France '
author:
- 'Christian Krattenthaler$^\dagger$ and Tanguy Rivoal'
title: 'How can we escape Thomae’s relations?'
---
[^1]
Prelude: introduction and summary of the results
================================================
In this article, we are interested in two families of two-term relationships between hypergeometric ${}_{3} F_{2}$-series with argument $1$, and the possible links between them. The first family consists of 120 relations found by Thomae [@thomae], which can be interpreted as the action of the symmetric group $\mathfrak{S}_5$ on five parameters related to the parameters of a generic ${}_{3} F_{2}$-series. This action has been discovered and rediscovered many times. We shall start our article by describing two of its seemingly different incarnations: one involving series (Thomae, Whipple, Hardy and others: see Section \[sec:thomaerelations\], in particular Theorem \[prop:hardydixonandcie\]) and the other involving integrals (Dixon, Rhin–Viola: see Section \[sec:rhinviolagroup\], in particular Theorem \[prop:dixon\]), while in Section \[sec:reformulation\] we explain their equivalence.
Our main aim is to find a hypergeometric explanation of a second family of six “exotic” integral relations recently discovered by Sato [@sato] (see Theorem \[theo:sato\]). The latter provide counter-examples to a conjecture of Rhin and Viola [@rv] (see Conjecture \[conj:1\] in Section \[sec:rhinviolagroup\]), which essentially predicted the universality of Thomae’s relations in the case of integral parameters. As we shall show, this explanation is given by the following two identities in Theorems \[theo:??\] and \[thm:gamma2\], respectively, which seemingly have not been stated explicitly before.
The first one (with proof in Section \[sec:escapethomae\]) covers five of Sato’s six original relations, and we will obtain from it infinitely many explicit counter-examples to the conjecture by Rhin and Viola (see Theorem \[theo:2\] in Section \[sec:rhinviolagroup\]).
\[theo:??\] Let $\al, \be, \ga$ be complex numbers such that $2\al+\be+1$ and $2\be+\al+1$ are not non-positive integers, and such that $\Re(2\al+2\be-\ga)>0$. Then $$\label{eq:exotique2}
{} _{3} F _{2} \!\left [ \begin{matrix} {\al+1, \;\,\be+1, \;\,\gamma}
\\ {2\al+\be+1, \,2\be+\al+1}\end{matrix}
;1\right ] = \frac{ 2(\al+\be)}{2(\al+\be)-\gamma}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {\al, \quad \be, \quad\gamma}
\\ {2\al+\be+1, \,2\be+\al+1}\end{matrix} ;1\right ].$$
The second one (with proof in Section \[sec:contig\]) covers the remaining counter-example of Sato. It implies another set of infinitely many counter-examples to the conjecture by Rhin and Viola (see Theorem \[theo:2a\] in Section \[sec:rhinviolagroup\]).
\[thm:gamma2\] For any complex numbers $\al,\be,\ga$ such that $\Re\big(2-\be- \frac{\al ( \al - \ga+1 ) }{\be-1}\big)>0$, and such that $\al+1$ and $ \ga+ \frac{\al ( \al - \ga+1 ) }{\be-1}$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b12}
{} _{3} F _{2} \!\left [ \begin{matrix} { \al, \be, \ga}\\ { \al+1,\ga+
\frac{\al ( \al - \ga +1) }{\be-1}+1 }\end{matrix} ;
{\displaystyle 1}\right ]\\=
\frac { ( \al - \be +2) ( \al + \al^2 - \ga - \al \ga + \be \ga ) }
{( \al+1 ) ( 2 \al + \al^2 - \al \be - \ga - \al \ga + \be \ga ) }\,
{} _{3} F _{2} \!\left [ \begin{matrix} { \al+1, \be-1, \ga}\\ { \al+2,
\ga+ \frac{\al ( \al - \ga+1 ) }{\be-1} }\end{matrix} ;
{\displaystyle 1}\right ].\end{gathered}$$
Clearly, since Sato’s counter-examples are special cases of and , but are not consequences of Thomae’s relations (see Section \[sec:eff\]), the two identities provide an answer to the question in the title. (Let us point out that Theorems \[theo:??\] and \[thm:gamma2\] are “independent" of each other, that is, neither is it possible to derive Theorem \[thm:gamma2\] from a combination of Theorem \[theo:??\] with Thomae’s relations, nor is this possible in the other direction.)
Of course, there may exist many more ways of escaping Thomae’s relations. For example, a rather simple-minded one consists in examining for which integral values of the parameters a ${} _{3} F _{2}$-series with argument $1$ can be a rational number. In fact, a complete characterisation for the latter problem is available, see Theorem \[theo:3\] in Section \[sec:simple\]. Leaving this simple possibility aside, in our proofs of identities and in Sections \[sec:escapethomae\] and \[sec:contig\], we make use of two fundamentally different ways to escape Thomae’s relations:
\(1) One applies a transformation formula transforming a $_3F_2$-series with argument $1$ into a hypergeometric series with a larger number of parameters (in our case, this is the transformation formula transforming a $_3F_2$-series into a very-well-poised $_7F_6$-series) in order to “exit" the “$_3F_2$-domain," and then one “re-enters" the “$_3F_2$-domain" in a different way (in our case, we use the same transformation formula in the other direction, but after a permutation of the parameters of the $_7F_6$-series has been carried out before).
\(2) One starts with a $_3F_2$-series in which one lower parameter exceeds one upper parameter by a positive integer. Subsequently, one applies contiguous relations to obtain a sum of several series, in which for all but one the use of the contiguous relations has made these two parameters equal, and thus these $_3F_2$-series with argument $1$ reduce to a $_2F_1$-series (with argument $1$), which can then be summed by means of the Gau[ß]{} summation formula . The various results of these evaluations are then combined into one expression, thereby generating a (possible huge) polynomial term, which is then equated to zero. In order to make this work, this polynomial must have integral solutions. (See the Remark after the proof of Proposition \[prop:gamma2\] in Section \[sec:contig\] for more precise explanations, and, in particular, for an explanation of the term “contiguous relation").
Whereas we failed to find results other than Theorem \[theo:??\] by using recipe (1), we show in Section \[sec:contig2\] that recipe (2) can be used in many more ways than the one yielding Theorem \[thm:gamma2\] (see Theorems \[thm:beta\] and \[thm:gamma1\]), thus producing many more counter-examples to the conjecture by Rhin and Viola. In fact, there are certainly many more relations that can be found in that way. We report on a curious phenomenon in that context at the end of the “round-up" Section \[sec:post\], where we indicate the ideas that we used to find the hypergeometric results in Theorems \[theo:??\], \[thm:gamma2\], \[thm:beta\] and \[thm:gamma1\].
So, in summary, as disappointing as this may be, our results show that the conjecture of Rhin and Viola was over-optimistic. The counter-examples by Sato are not just rare exceptions, they even embed in infinite families of counter-examples, and there are others beyond that. In view of this, and since the data that we produced do not give much guidance, we better refrain from coming up with a modified conjecture towards a generating set of transformations for the relations between $_3 F_2$-series that would correct the conjecture by Rhin and Viola. Nevertheless, finding one appears to be an interesting, and challenging, problem.
[*Acknowledgment.*]{} This work began during the second author’s visit to Nihon University, Tokyo, in October 2004. He would like to thank Noriko Hirata–Kohno for her invitation and Takayuki Oda who kindly gave him a copy of Sato’s Master Thesis. The first author would like to thank Anders Björner and Richard Stanley, and the Institut Mittag–Leffler, for inviting him to work in a relaxed and inspiring atmosphere during the “Algebraic Combinatorics" programme in Spring 2005 at the Institut, during which this article was completed.
Thomae’s relations {#sec:thomaerelations}
==================
Hypergeometric series are defined by $$\begin{aligned}
{} _{q+1} F _{q} \!\left [ \begin{matrix} {\alpha_0, \alpha_1, \ldots,\alpha_{q}}\\ {
\beta_1, \ldots, \beta_q}\end{matrix} ; {\displaystyle z}\right ] =
\sum_{k=0}^{\infty} \frac{(\al_0)_k\,(\al_1)_k\cdots (\al_q)_k}
{k!\,(\be_1)_k\cdots (\be_{q})_k} \,z^k,
\label{eq:hyper}\end{aligned}$$ where $(\al)_0=1$ and $(\al)_n=\al(\al+1)\cdots (\al+n-1)$ for $n\ge 1$. The series converges provided that the argument $z$ is a complex number with $\vert z \vert <1$, $\alpha_j\in\mathbf{C}$ and $\beta_j\in\mathbf{C}\setminus\mathbf{Z}_{\le 0}$; it also converges for $z= 1$ if in addition $\Re(\be_1+\dots+\be_q)>\Re(\al_0+\dots+\al_q)$. Any “permutation” in $\mathfrak{S}_{q+1}\times\mathfrak{S}_q$ acting on the upper parameters $\al_i$, $i=0,1,\dots,q$, and the lower parameters $\be_i$, $i=1,2,\dots,q$, on the left-hand side of does not affect the value of the right-hand side: we use the term “trivial symmetries” to indicate this fact.
As mentioned in the introduction, the symmetric group $\mathfrak{S}_5$ acts classically on the hypergeometric series ${}_{3} F_{2}$-series with argument $1$, which leads to exactly 120 formal relations between them. This group action is obtained using the following fundamental identity, due to Thomae [@thomae Eq. (12)] (given as (3.2.2) in [@Bailey]), which is valid under certain conditions on the parameters to ensure convergence of the involved series, $$\begin{aligned}
{} _{3} F _{2} \!\left [ \begin{matrix} {a, b, c}
\\ {d, e}\end{matrix};1
\right ] &=\frac {\Gamma(e)\,\Gamma(d+e-a-b-c)}
{\Gamma(e-a)\,\Gamma( d+e-b-c)}
{} _{3} F _{2} \!\left [ \begin{matrix} {a, d-b, d-c}
\\ {d, d+e-b-c}\end{matrix};1
\right ].
\label{eq:thomae1}
%\\
%{} _{3} F _{2} \!\left [ \begin{matrix} {a, b, c}
%\\ {d, e}\end{matrix};1
%\right ] &= \frac {\Gamma(d)\,\Gamma(e)\,\Gamma(d+e-a-b-c)}
%{\Gamma(b)\,\Gamma(d+e-a-b)\,\Gamma(d+e-b-c)} \notag\\
%&\kern3cm
%\times
%{} _{3} F _{2} \!\left [ \begin{matrix} {d-b, e-b,d+e-a-b-c}
%\\ {d+e-a-b, d+e-b-c}\end{matrix};1 \right ],
%\label{eq:thomae2}\end{aligned}$$ The iterative application of , together with the trivial symmetries, yields 120 relations, of which only 10 are inequivalent modulo the trivial symmetries. These were given by Thomae [@thomae Art. 4] and put in a more suitable form by Whipple [@whipple]. It is apparently Hardy [@hardy p. 499] who first gave a group theoretic interpretation: we state his observation in the striking form given in [@kratrao; @jr].
\[prop:hardydixonandcie\] Let $s=s(x_1,x_2,x_3,x_4,x_5)=x_1+x_2+x_3-x_4-x_5$. The function $$\label{eq:hardy?}
\frac{1}{\Ga(s)\,\Ga(2x_4)\,\Ga(2x_5)}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2x_1-s, 2x_2-s, 2x_3-s}
\\ {2x_4,\quad 2x_5}\end{matrix};1
\right ]$$ is a symmetric function of the five variables $x_1, x_2, x_3, x_4, x_5$.
Care is needed using this theorem, since $s$ is not a symmetric function of $x_1, x_2, x_3, x_4, x_5$ and some of the ${} _{3} F
_{2}$-series might not be convergent. This result is surprising since one could not expect [*a priori*]{} a much bigger invariance group than $\mathfrak{S}_3\times\mathfrak{S}_2$, obtained by the permutations of $\{x_1,x_2,x_3\}$ and $\{x_4,x_5\}$, which trivially leave invariant.
The Rhin–Viola group for $\zeta(2)$ {#sec:rhinviolagroup}
===================================
In 1996, Rhin and Viola introduced in [@rv] the integral $$\label{eq:I}
I(h,i,j,k,l)=\int_0^1 \!\!\int_0^1
\frac{x^h(1-x)^iy^k(1-y)^j}{(1-xy)^{i+j-l+1}}\,\dd x\, \dd y,$$ which is convergent under the assumption that $h,i,j,k,l$ are non-negative integers, which will be the case throughout the rest of this article unless otherwise stated. Their motivation was to use the fact that $I(h,i,j,k,l)\in\mathbf{Q} +\mathbf{Q}\zeta(2)$ to get a good irrationality measure for $\zeta(2)=\sum_{n\ge 1}1/n^2=\pi^2/6$, as had been done in previous work using similar but less general integrals (see the bibliography in [@rv]). They developed a beautiful new algebraic method for handling the general case above and were rewarded with the best known irrationality measure for $\pi^2$. See also [@fis; @zud] for related work.
From now on, we focus essentially on the hypergeometric structure underlying their method, which is made transparent by the identity (see Section \[sec:reformulation\], where we recall the proof for $z=1$) $$\label{eq:lossofsymmetry}
{} _{3} F _{2} \!\left [ \begin{matrix} {a, b, c}
\\ {d,e}\end{matrix} ; {\displaystyle z}
\right ] =
\frac {\Gamma(d)\,\Gamma(e)}
{\Gamma(a)\,\Gamma(d-a)\,\Gamma(b)\,\Gamma(e-b)}
\int_0^1\!\!\int_0^1
\frac{x^{a-1}(1-x)^{d-a-1}y^{b-1}(1-y)^{e-b-1}}{(1-zxy)^{c}} \,\dd x\, \dd y,$$ which is valid provided $\Re(d)>\Re(a)>0$ and $\Re(e)>\Re(b)>0$ if $\vert z\vert <1$, with the further assumption that $\Re(d+e-a-b-c)>0$ if $z=1$. To simplify, we set $B(h,i,j,k,l)=I(h,i,j,k,l)/(h!\,i!\,j!\,k!\,l!)$. The main new idea in [@rv] was to use the action of a group on the parameters of $h,i,j,k,l$ leaving the value of $B(h,i,j,k,l)$ invariant. To do this, Rhin and Viola showed that, under the two changes of variables $\{X=y, Y=x\}$ and $\{X=(1-x)/(1-xy),
Y=1-xy\}$, the value of $I(h,i,j,k,l)$ (and hence also that of $B(h,i,j,k,l)$) is not changed if the parameters are permuted by the product of transpositions $\sigma=(h\;k)(i\;j)$ and the 5-cycle $\tau=(h\;i\;j\;k\;l)$. The group ${\bf T}=\langle \sigma, \tau\rangle$ generated by $\sigma$ and $\tau$ is isomorphic to $\mathscr{D}_5$, the dihedral group of order 10: for a visual proof, place the letters $h,i,j,k,l$, in this order, at the vertices of a regular pentagon.
But a more important invariance group can be obtained by extending the action of $\sigma$ and $\tau$ by linearity to the set $\mathscr{P}=\{h,i,j,k,l, j+k-h, k+l-i, l+h-j, h+i-k, i+j-l\}$ (by “linearity”, we mean $\tau(h+i-k)=\tau(h)+\tau(i)-\tau(k)=i+j-l$, etc.). Provided the five values $j+k-h$, $k+l-i$, $l+h-j$, $h+i-k$, $i+j-l$ are non-negative (see Theorem \[theo:3\] in Section \[sec:reformulation\] for the arithmetic meaning of this hypothesis). One can then use the apparent loss of the trivial symmetries in the parameters $a,b,c$ and $d,e$ on the right-hand side of to prove that the value of $B(h,i,j,k,l)$ is invariant under the permutation on $\mathscr{P}$ defined by $\varphi=(h\;\, i+j-l)(i\;\,l+h-j)(j+k-h\;\,k+l-i)$. Rhin and Viola managed to prove that the group ${\bf \Phi}=\langle \varphi, \sigma, \tau\rangle$ acting on $\mathscr{P}$ and leaving the value of the associated integrals invariant can be viewed as the permutation group $\mathfrak{S}_5$ acting on the set $\{h+i,\;i+j,\;j+k,\;k+l,\;l+h\}$, and hence has cardinality 120. This remark was first made by Dixon [@dixon], in an even more general form.
\[prop:dixon\] Assume that the complex numbers $h,i,j,k,l, j+k-h, k+l-i, l+h-j, h+i-k, i+j-l$ have real part $>-1$. Then the integral $B(h,i,j,k,l)$ [*(*]{}where $x!$ is assumed to mean $\Gamma(x+1)$ for complex $x$[*)*]{} is a symmetric function of the five parameters $h+i$, $i+j$, $j+k$, $k+l$, $l+h.$
Finally, Rhin and Viola proposed the following conjecture.
\[conj:1\] Let $h,i,j,k,l, h',i',j',k',l'$ be non-negative integers.
1 [*(i)*]{} If $I(h,i,j,k,l)=I(h',i',j',k',l')$, then there exists $\rho\in{\bf T}$ such that $\rho(h)=h',\rho(i)=i', \rho(j)=j', \rho(k)=k'$ and $\rho(l)=l'.$
2 [*(ii)*]{} Suppose furthermore that the numbers $$\begin{aligned}
&j+k-h,\quad k+l-i,\quad l+h-j,\quad h+i-k,\quad i+j-l\label{eq:nonnegative1}\\
&j'+k'-h',\quad k'+l'-i',\quad l'+h'-j',\quad h'+i'-k',\quad i'+j'-l'\label{eq:nonnegative2}\end{aligned}$$ are all non-negative. If $I(h,i,j,k,l)/I(h',i',j',k',l')\in\mathbf{Q}$, then there exists $\rho\in{\bf \Phi}$ such that $\rho(h)=h',\rho(i)=i', \rho(j)=j', \rho(k)=k'$ and $\rho(l)=l'.$
The truth of (i) and (ii) would have shown that their method is optimal, but both have been shown to be false in 2001 by Susumu Sato [@sato], who found the following counter-examples, apparently by numerical inspection.
\[Sato\] \[theo:sato\] Both cases of Conjecture \[conj:1\] are false, as shown by the following six counter-examples: $$\begin{aligned}
I(1,1,1,1,1) &=5-3\zeta(2)=I(3,1,1,2,0),\label{eq:sato1}\\
I(3,1,2,2,1) &=79/4-12\zeta(2)=I(4,2,2,3,0),\label{eq:sato2}\\
I(3,1,2,1,1) &=3\zeta(2)-59/12= I(3,3,1,3,0),\label{eq:sato4}\\
I(3,2,2,2,1) &=10\zeta(2)-148/9=I(5,1,3,2,1),\label{eq:sato5}\\
I(3,0,3,1,1) &=9\zeta(2)-59/4=9I(3,3,1,2,1),\label{eq:sato6}\\
I(3,1,3,1,0) &=\zeta(2)-29/18= I(3,2,1,2,0).\label{eq:sato3}\end{aligned}$$
(Sato mis-stated as $I(3,0,3,1,1)=I(3,3,1,2,1)$. The reader should also note that and altogether relate four different integrals rationally.) Equation is already a counter-example to both (i) and (ii). The following questions are natural, but were not considered by Sato:
- Are these counter-examples merely numerical accidents, or do they admit a theoretical explanation?
- Do there exist infinitely many counter-examples to the conjecture of Rhin and Viola?
We give a complete answer to both questions in the two theorems below which we prove in Sections \[sec:proof\] and \[sec:proofa\], respectively.
\[theo:2\] [*(i)*]{} Sato’s counter-examples up to can be explained by purely hypergeometric means, i.e., there exists a general hypergeometric identity that generates them.
[*(ii)*]{} For each integer $\al\ge 1$, the equation $$\label{eq:infinicontreexemples}
I(2\al-1,2\al-1,\al, 2\al-1,\al)=I(2\al+1,2\al-1,\al, 2\al,\al-1)$$ provides a counter-example to the cases [*(i)*]{} and [*(ii)*]{} of Conjecture [*\[conj:1\]*]{}.
\[theo:2a\] [*(i)*]{} Sato’s counter-example can be explained by purely hypergeometric means, i.e., there exists a general hypergeometric identity that generates them.
[*(ii)*]{} For each integer $\al\ge 2$, the equation $$\label{eq:infinicontreexemplesa}
I(\al^2-1,\al-1,\al^2-\al+1,\al-1,0)=(\al-1)\,I(\al^2-1,\al,\al^2-\al-1,\al,0)$$ provides a counter-example to case [*(ii)*]{} of Conjecture [*\[conj:1\]*]{}, and also to case [*(i)*]{} if $\al=2$.
\(1) A particularly elegant instance of is the one where $\ga=\al+\be$: for any complex numbers $\al$ and $\be$ which are not non-positive integers and which satisfy $\Re(\al+\be) >0$, we have $$\label{eq:exotique}
{} _{3} F _{2} \!\left [ \begin{matrix} {\al+1, \;\,\be+1, \;\,\al+\be}
\\ {2\al+\be+1, 2\be+\al+1}\end{matrix};1
\right ] = 2\,
{} _{3} F _{2} \!\left [ \begin{matrix} {\al, \quad \be, \quad\al+\be}
\\ {2\al+\be+1, 2\be+\al+1}\end{matrix};1
\right ].$$ The action of Thomae’s relations on both sides of independently provides ten variations of , up to trivial symmetries: one example, given by in Section \[sec:proof\], will be used in the proof of Theorem \[theo:2\]. Equation follows from the case of where $\al=\be$ is a positive integer. Furthermore, we shall show in Section \[sec:proof\] that $I(2\al-1,2\al-1,\al, 2\al-1,\al)$ tends to 0 as $\al$ tends to infinity: this fact implies that provides infinitely many counter-examples to Conjecture \[conj:1\].
\(2) As we show in Section \[sec:pattern\], the “general” hypergeometric identity that generates up to is exactly identity , via the translation between integrals and hypergeometric $_3F_2$-series given in . Equation is a special case. (3) Similarly, we show in Section \[sec:pattern\] that the “general” hypergeometric identity that generates is exactly identity , again via the translation . Equation is a special case. Since we show there that the integral on the left-hand side of tends to zero as $\al$ tends to infinity, also provides infinitely many counter-examples to Conjecture \[conj:1\]. (4) It would also be interesting to look at the analogous problem arising from the group action on the triple integral $$\int_0^1\!\!\int_0^1\!\!\int_0^1 \frac{u^h(1-u)^lv^k(1-v)^sw^j(1-w)^q}
{(1-(1-uv)w)^{q+h-r+1}}\,\dd u\,\dd v\,\dd w \in\mathbf{Q}+\mathbf{Q}\zeta(3)$$ found by Rhin and Viola in [@rv2]: do there exist exotic relationships between such integrals that are not described by this group action? Note that this action admits an interpretation in terms of very-well-poised ${} _{7} F _{6}$-series exactly in the style of Theorem \[prop:hardydixonandcie\] (see [@zud Sec. 4] for the passage from the integrals to very-well-poised ${} _{7} F _{6}$-series, and [@jr Proposition 5, $q\to 1$, p. 6698] for a particularly elegant formulation of the group structure). For very clear expositions of various group actions on $(q-)$ hypergeometric series, see [@kratrao; @jr] and the references therein.
From Dixon to Thomae {#sec:reformulation}
=====================
In this section, we show more precisely how Rhin and Viola’s integrals are related to hypergeometric series. To get a new expression for the integral $I(h,i,j,k,l)$, we transform the integrand of by using the binomial series expansion $$\frac{1}{(1-xy)^{i+j-l+1}} =\sum_{n=0}^{\infty}
\frac{(i+j-l+1)_n}{n!}\, (xy)^{n},$$ and the beta integral evaluations $$\int_0^1 x^{n+h}(1-x)^{i}\dd x = \frac{(n+h)!\,i!}{(n+h+i+1)!},\quad
\int_0^1 y^{n+k}(1-y)^{j}\dd y = \frac{(n+k)!\,j!}{(n+k+j+1)!}.$$ We have $$\begin{aligned}
B(h,i,j,k,l)
&=\frac{1}{h!\,k!\,l!}\,\sum_{n=0}^{\infty}
\frac{(n+h)!\,(n+k)!\,(i+j-l+1)_n}{n!\,(n+h+i+1)!\,(n+k+j+1)!}\label{eq:fin4}\\
&=\frac{1}{l!\,(h+i+1)!\,(k+j+1)!}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {h+1, k+1, i+j-l+1}
\\ {h+i+2,k+j+2}\end{matrix};1\right]\label{eq:fin3},\end{aligned}$$ since the interchange of summation and integral is justified by Fubini’s theorem. The passage from to uses the trivial identity $(\al+n)!=\al!\,(\al+1)_n$.
Under this interpretation, it is not surprising that the group obtained by Dixon and Rhin–Viola should be a reformulation of Theorem \[prop:hardydixonandcie\], in terms of integrals rather than series. Indeed, if we define a bijection between the tuples $(x_1, x_2, x_3, x_4, x_5)$ and $(h,i,j,k,l)$ by $$\begin{aligned}
h+1&=\hphantom{-}x_1-x_2-x_3+x_4+x_5,\\
i+1&=-x_1+x_2+x_3+x_4-x_5,\\
j+1&=\hphantom{-}x_1-x_2+x_3-x_4+x_5,\\
k+1&=-x_1+x_2-x_3+x_4+x_5,\\
l+1&=\hphantom{-}x_1+x_2+x_3-x_4-x_5=s,\end{aligned}$$ then we see that $B(h,i,j,k,l)$, written as , perfectly matches and the 120 possible series are all convergent if the ten integers in the set $\mathscr{P}$ are non-negative. Since $$\label{eq:xhijkl}
2x_1=l+h+2,\;
2x_2=k+l+2,\;
2x_3=i+j+2,\;
2x_4=h+i+2,\;
2x_5=j+k+2,$$ we also see that the symmetry of in the variables $x_1, x_2, x_3, x_4, x_5$ is equivalent to the symmetry of $B(h,i,j,k,l)$ in the variables $h+i$, $i+j$, $j+k$, $k+l$, $l+h.$
When is $I(h,i,j,k,l)$ rational? {#sec:simple}
================================
In this section we answer the question of “simple-minded" counter-examples to Conjecture \[conj:1\] that was raised in the Introduction.
\[theo:3\] Let $h,i,j,k,l$ be non-negative integers. Then the following assertions are equivalent:
- The integers $j+k-h$, $k+l-i$, $l+h-j$, $h+i-k$, $i+j-l$ are all non-negative.
- The integral $I(h,i,j,k,l)$ is an irrational number.
\(1) We remarked earlier (see and ) that $I(h,i,j,k,l)$ is essentially equal to a $_3F_2$-series. If we translate Theorem \[theo:3\] into the analogous theorem for $$\label{eq:3F2}
{} _{3} F _{2} \!\left [ \begin{matrix} a,b,c\\
d,e
\end{matrix};1 \right ],$$ via the relations $a=h+1$, $b=k+1$, $c=i+j-l+1$, $d=h+i+2$, $e=k+j+2$, we get the following necessary and sufficient condition for the irrationality of the series for integral values of $a,b,c,d,e$: $$\begin{gathered}
\label{eq:cond1}
d+e\ge a+b+c+1,\quad a\ge1,\quad b\ge1,\quad c\ge1,\\
d\ge \max\{ a, b, c\}+1,\\
e\ge \max\{ a, b, c\}+1.
\label{eq:cond2}\end{gathered}$$
\(2) As a marginal consequence, Theorem \[theo:3\] proves that the analogue of the case (ii) of Conjecture \[conj:1\], where we now suppose that the non-negativity condition is not true, cannot hold either. Indeed, if one of the integers in were negative and none in were negative, then the value of $I(h,i,j,k,l)/I(h',i',j',k',l')$ would be irrational and the conjecture would be empty. And if one of the integers in and one in were negative, then the conjecture would be trivially false because, although $I(h,i,j,k,l)$ and $I(h',i',j',k',l')$ are rational, there exist many ${\bf \Phi}$-unrelated choices for $h, h'$, etc.: one may consider $I(1,1,1,1,3)$ and $I(1,1,1,1,4)$ for example.
We first show the implication (b)${}\Rightarrow{}$(a). Since the parameters $j+k-h$, $k+l-i$, $l+h-j$, $h+i-k$, $i+j-l$ are cyclically permuted by $\tau\in{\bf T}$, if one of them were negative, then without loss of generality, we may assume that it is $i+j-l$. But $i+j-l\le -1$ implies that the integrand of $I(h,i,j,k,l)$ is a polynomial with integral coefficients and hence that $I(h,i,j,k,l)\in\mathbf{Q}.$
The reverse implication (a)${}\Rightarrow{}$(b) is a little bit more complicated. Since $i ,j$ and $i+j-l$ are positive integers, we can write the expansion in the equivalent form: $$\label{eq:rationalfraction}
I(h,i,j,k,l)=\frac{i!\,j!}{(i+j-l)!}\,\sum_{n=0}^{\infty}
\frac{(n+1)_{i+j-l}}{(n+h+1)_{i+1}(n+k+1)_{j+1}}.$$ (We have used trivial identities such as $(n+h+i)!/(n+h)!=(n+h+1)_{i+1}$.) We know that $I(h,i,j,k,l)\in\mathbf{Q}+\mathbf{Q}\zeta(2)$ and it will be enough to prove that the coefficient $p(h,i,j,k,l)$ of the irrational number $\zeta(2)$ is non-zero. A standard way to find an explicit expression for this coefficient is to expand the summand of , which is a rational function of $n$, in partial fractions (see the introduction of [@ri] for details in many similar cases and references). All computations done, one finds that $$p(h,i,j,k,l)=(-1)^{h+i+j+k+l}\sum_{s=\max(h,k,i+j-l)}^{\min(h+i,k+j)}
\binom{i}{s-h}\binom{j}{s-k}\binom{s}{i+j-l},$$ with the convention that the value of the sum is $0$ if it is empty. The latter is the case if and only if $\min(h+i,k+j)<\max(h,k,i+j-l)$.
We now show that condition (a) ensures that the sum is non-empty and hence that $$(-1)^{h+i+j+k+l}p(h,i,j,k,l)>0,$$ because it is a sum of binomial coefficients. We have already used the fact that $i+j-l\ge 0$. Since the inequalities $h+i-k\ge 0$ and $k+j-h\ge 0$ imply that $\max(h,k)\le\min(h+i,k+j)$, it only remains to show that $i+j-l\le \min(h+i,k+j)$ to finally prove that $\min(h+i,k+j)\ge \max(h,k,i+j-l)$. But $\min(h+i,k+j)-(i+j-l)=\min(h+l-j, k+l-i)\ge 0$, which finishes the proof.
Effective computation of Thomae relations {#sec:eff}
=========================================
In this section, we show how to compute the complete set of (generically) 120 Thomae relations (convergent or not) for any given ${} _{3} F _{2}$-series with argument $1$. We need this to transform Sato’s counter-examples into more suitable forms. The most effective way to do this is by using the parametrisation $2x_1-s,2x_2-s,2x_3-s$ and $2x_4, 2x_5$ of the upper and lower parameters of the $_3F_2$-series from Theorem \[prop:hardydixonandcie\]. If we denote the upper parameters by $a,b,c$ and the lower parameters by $d,e$, then $x_1,x_2,x_3,x_4,x_5,s$ and $a,b,c,d,e$ are related by $$2x_1=d+e-b-c, \, 2x_2=d+e-c-a,\, 2x_3=d+e-a-b, \, 2x_4=d,\, 2x_5=e,$$ and $$s=x_1+x_2+x_3-x_4-x_5=2(d+e-a-b-c).$$ If one prefers the integral setting, then one gets the parametrisation of Theorem \[prop:hardydixonandcie\] of the integral $I(h,i,j,k,l)$ by the formulae given in Section \[sec:reformulation\].
The following simple [*Maple*]{} commands compute all possible values of the arrays of parameters $[2x_{\rho(1)}-s_{\rho},2x_{\rho(2)}-s_{\rho},
2x_{\rho(3)}-s_\rho ; 2x_{\rho(4)},2x_{\rho(1)}]$ (with $s_\rho=x_{\rho(1)}+
x_{\rho(2)}+x_{\rho(3)}-x_{\rho(4)}-x_{\rho(1)}$) over all permutations $\rho$ of $\{1,2,3,4,5\}$, with the nice feature to output only the term-wise different arrays (viewed as 5-tuples by the program):
$> {\tt with(combinat)\!:}$ $> {\tt p:=(u,v,w,x,y)-\!\!>permute([u,v,w,x,y])\!:}$ $> {\tt s:=(u,v,w,x,y)-\!\!>u\!+\!v\!+\!w\!-\!x\!-\!y\!:}$ $> {\tt A:=(u,v,w,x,y)-\!\!>[2\!*\!u\!-\!s(u,v,w,x,y),
2\!*\!v\!-\!s(u,v,w,x,y), 2\!*\!w\!-\!s(u,v,w,x,y), 2\!*\!x,2\!*\!y]\!:}$ $> {\tt T:=(u,v,w,x,y)-\!\!>\tt seq(A(op(1,op(j,p(u,v,w,x,y))),op(2,op(j,p(u,v,w,x,y))),}$ ${\tt op(3,op(j,p(u,v,w,x,y))),op(4,op(j,p(u,v,w,x,y))),op(5,op(j,p(u,v,w,x,y)))),}$ ${\tt j=1...nops(p(u,v,w,x,y)))\!:}$ $> {\tt F:=(a,b,c,d,e)-\!\!>T((d\!+\!e\!-\!b\!-\!c)/2,(d\!+\!e-\!c\!-\!a)/2,(d\!+\!e\!-\!a\!-\!b)/2,d/2,e/2)\!:}$ $> {\tt I:=(h,i,j,k,l)-\!\!>T((h\!+\!l\!+\!2)/2,(k\!+\!l\!+\!2)/2,(i\!+\!j\!+\!2)/2,(h\!+\!i\!+\!2)/2,
(j\!+\!k\!+\!2)/2)\!:}$
The function [T]{} computes all the different expressions for the value of the symmetric function in Theorem \[prop:hardydixonandcie\], [F]{} does the same for a ${} _{3} F _{2}[a,b,c;d,e]$ and $\tt I$ for ([^2]) $I(h,i,j,k,l)$. Only the Gamma-factors are not computed, but this could be easily done. For example, we obtain
$> {\tt I(1,1,1,1,1);}$ $${\tt [2,2,2,4,4]}$$ $> {\tt I(3,1,1,2,0);}$ $${\tt [4, 3, 3, 6, 5], \;[2, 1, 3, 4, 5],\; [3, 2, 3, 6, 4], \;[1, 1, 2, 4, 4], \;[2, 4, 2, 5, 5].}$$
[*Maple*]{} outputs 25 other arrays for $I(3,1,1,2,0)$ but since they correspond to the five above by the trivial symmetries, we do not list them. We can also find the Thomae relations for both sides of counter-example :
$> {\tt I(3,1,3,1,0);}$ $$\!{\tt [4, 2, 5, 6, 6], \;[1, 2, 2, 6, 3],\; [1, 4, 4, 5, 6],
\;[1, 1, 1, 3, 5]}$$ $>{\tt I(3,2,1,2,0);}$ $${\tt [4, 3, 4, 7, 5],\; [2, 1, 4, 5, 5],\; [1, 1, 3, 4, 5],\; [3, 3, 3, 4, 7].}$$
This shows that the relations $I(1,1,1,1,1)\!=\!I(3,1,1,2,0)$ and $I(3,1,3,1,0)\!=\!I(3,2,1,2,0)$ are not consequences of Thomae relations. Similar computations provide a verification of the other counter-examples.
The pattern behind Sato’s counter-examples {#sec:pattern}
===========================================
With the interpretation given in Section \[sec:reformulation\], the case (ii) of Conjecture \[conj:1\] can be reformulated as follows:
Sato’s counter-examples destroy this hope. We can formulate his counter-examples and in hypergeometric form (with simplification of the Gamma-factors) as follows: $$\,{} _{3} F _{2} \!\left [ \begin{matrix} {2, 2, 2}
\\ {4, 4}\end{matrix} ;1\right]
= \frac{3}{20} \,
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 3, 3}
\\ {6, 5}\end{matrix};1 \right]
\label{eq:3f2-1}
\quad \textup{and}\quad
\,{} _{3} F _{2} \!\left [ \begin{matrix} {4, 3, 3}
\\ {6, 6}\end{matrix};1 \right]
= \frac{2}{21}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {5, 4, 5}
\\ {8, 7}\end{matrix} ;1\right].$$ Under this form, the parameters on the left-hand sides and those on the right-hand sides seem still rather unrelated, and it is thus still unclear whether we face numerical accidents or if there is something deeper behind.
However, a natural thing to do here is to seek new numerical relations by applying Thomae’s transformations (using the [*Maple*]{} commands of the previous section) to each of the four ${} _{3} F _{2}$-series in , independently. We find that we are trying to prove that $$\label{eq:counter1}
{} _{3} F _{2} \!\left [ \begin{matrix} {2, 2, 2}
\\ {4, 4}\end{matrix};1 \right] = 2\,
{} _{3} F _{2} \!\left [ \begin{matrix} {1, 1, 2}
\\ {4, 4}\end{matrix} ;1\right] \quad \textup{and}\quad
%\label{eq:3f2-11}
{} _{3} F _{2} \!\left [ \begin{matrix} {3, 2, 3}
\\ {6, 5}\end{matrix};1 \right] = 2\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2, 1, 3}
\\ {6, 5}\end{matrix};1 \right],$$ where a pattern now emerges, explained by the earlier identity .
The hypergeometric forms of the three counter-examples , and are $$\begin{gathered}
\label{eq:counter2}
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 2, 3}
\\ {6, 5}\end{matrix};1 \right] = \frac{3}{35}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 4, 5}
\\ {8, 6}\end{matrix};1 \right],\quad
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 3, 4}
\\ {7, 6}\end{matrix};1 \right] = \frac{5}{7}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {6, 3, 4}
\\ {8, 7}\end{matrix};1 \right],\\
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 2, 3}
\\ {5, 6}\end{matrix};1 \right]=\frac{1}{21}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {4, 3, 4}
\\ {8, 5}\end{matrix};1 \right],\end{gathered}$$ which become much more illuminating when rewritten as $$\begin{gathered}
{} _{3} F _{2} \!\left [ \begin{matrix} {3, 2, 4}
\\ {6, 5}\end{matrix};1 \right] = \frac13\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2, 1, 4}
\\ {6, 5}\end{matrix};1 \right], \quad
{} _{3} F _{2} \!\left [ \begin{matrix} {3, 2, 2}
\\ {6, 5}\end{matrix};1 \right] = \frac23\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2, 1, 2}
\\ {6, 5}\end{matrix};1 \right], \\
{} _{3} F _{2} \!\left [ \begin{matrix} {3, 2, 4}
\\ {6, 5}\end{matrix};1 \right] = \frac13\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2, 1, 4}
\\ {6, 5}\end{matrix};1 \right],\end{gathered}$$ by using Thomae’s relations. (In particular, and are consequences of the same identity). The connexion with Theorem \[theo:??\] is now clear.
Finally, the hypergeometric form of the counter-example is $$\label{eq:sato3hyp}
{{} _{3} F _{2} \!\left [ \begin{matrix} { 4, 2, 5}\\ { 6,
6}\end{matrix} ; {\displaystyle 1}\right ] } =\frac {5} {9}\,
{{} _{3} F _{2} \!\left [ \begin{matrix} { 4, 3, 4}\\ { 7,
5}\end{matrix} ; {\displaystyle 1}\right ] } ,$$ which is obviously the special case $\al=4$, $\be=3$, $\ga=4$ of .
Proof of Theorem \[theo:??\] {#sec:escapethomae}
=============================
For the proof of we need the following transformation formula due to Verma and Jain (see [@GaRaAA (3.5.10), $q\to 1$, reversed], being implied by [@VeJaAD (4.1)]) between a $_3F_2$-series and a very-well-poised $_7F_6$-series: $$\begin{gathered}
\label{eq:T3240}
{} _{3} F _{2} \!\left [ \begin{matrix} { b, c,d}\\ { a, a - b + c}\end{matrix} ;
{\displaystyle 1}\right ] =
\frac{
\Gamma({ \textstyle 2 a}) \,\Gamma({ \textstyle 2 a - 2 b - d}) \,
\Gamma({ \textstyle a - b + c}) \,\Gamma({ \textstyle a - d + c}) }
{\Gamma({ \textstyle 2 a - 2 b}) \,\Gamma({ \textstyle 2 a - d}) \,
\Gamma({ \textstyle a + c}) \,\Gamma({ \textstyle a - b - d +
c}) }\\
\times
{} _{7} F _{6} \!\left [ \begin{matrix} {
a- \frac{1}{2}, \frac{a}{2}+\frac{3}{4}, b, \frac{d}{2},
\frac{d}{2}+\frac{1}{2}, \frac{a}{2} - \frac{c}{2}, \frac{a}{2} -
\frac{c}{2}+ \frac{1}{2}}\\ { \frac{a}{2}- \frac{1}{4},
a - b+\frac{1}{2}, a - \frac{d}{2}+\frac{1}{2}, a - \frac{d}{2},
\frac{a}{2} + \frac{c}{2}+\frac{1}{2}, \frac{a}{2} +
\frac{c}{2}}\end{matrix} ; {\displaystyle 1}\right ].\end{gathered}$$ If we apply this transformation to the $_3F_2$-series on the left-hand side of , then we obtain $$\begin{gathered}
\label{eq:7F6a}
\frac{ \Gamma({ \textstyle \al + 2 \be+1}) \,
\Gamma({ \textstyle 4 \al + 2 \be+2}) \,
\Gamma({ \textstyle 2 \al + 2 \be -
\ga}) \,\Gamma({ \textstyle 2 \al +
2 \be - \ga+2}) }{\Gamma({ \textstyle
2 \al + 2 \be}) \,
\Gamma({ \textstyle 2 \al + 2 \be+2}) \,
\Gamma({ \textstyle \al + 2 \be -
\ga+1}) \,\Gamma({ \textstyle 4 \al +
2 \be - \ga+2}) }\\
\times
{} _{7} F _{6} \!\left [ \begin{matrix} { 2 \al +
\be+\frac{1}{2}, \al + \frac{\be}{2}+\frac{5}{4},
\al+1, \frac{\ga}{2},
\frac{\ga}{2}+\frac{1}{2}, \al,
\al+\frac{1}{2}}\\ { \al +
\frac{\be}{2}+\frac{1}{4}, \al + \be+\frac{1}{2},
2 \al + \be -
\frac{\ga}{2}+\frac{3}{2}, 2 \al + \be -
\frac{\ga}{2}+1, \al + \be+\frac{3}{2},
\al + \be+1}\end{matrix} ; {\displaystyle 1}\right
] .\end{gathered}$$ We permute the parameters in the $_7F_6$-series to get the equivalent expression $$\begin{gathered}
\label{eq:7F6b}
\frac{ \Gamma({ \textstyle \al + 2 \be+1}) \,
\Gamma({ \textstyle 4 \al + 2 \be+2}) \,
\Gamma({ \textstyle 2 \al + 2 \be -
\ga}) \,\Gamma({ \textstyle 2 \al +
2 \be - \ga+2}) }{\Gamma({ \textstyle
2 \al + 2 \be}) \,
\Gamma({ \textstyle 2 \al + 2 \be+2}) \,
\Gamma({ \textstyle \al + 2 \be -
\ga+1}) \,\Gamma({ \textstyle 4 \al +
2 \be - \ga+2}) }\\
\times
{} _{7} F _{6} \!\left [ \begin{matrix} { 2 \al +
\be+\frac{1}{2}, \al + \frac{\be}{2}+\frac{5}{4},
\al, \frac{\ga}{2},
\frac{\ga}{2}+\frac{1}{2}, \al+\frac{1}{2},
\al+1}\\ { \al +
\frac{\be}{2}+\frac{1}{4}, \al + \be+\frac{3}{2},
2 \al + \be -
\frac{\ga}{2}+\frac{3}{2}, 2 \al + \be -
\frac{\ga}{2}+1, \al + \be+1,
\al + \be+\frac{1}{2}}\end{matrix} ;
{\displaystyle 1}\right ] .\end{gathered}$$ To this $_7F_6$-series, we apply the transformation in the backward direction, that is we apply the transformation $$\begin{gathered}
{} _{7} F _{6} \!\left [ \begin{matrix} { a, \frac{a}{2}+1, b, c,
c+\frac{1}{2}, d, d+\frac{1}{2}}\\ { \frac{a}{2}, a - b+1, a -
c+1, a - c+\frac{1}{2}, a - d+1, a - d+\frac{1}{2}}\end{matrix} ;
{\displaystyle
1}\right ] \\= \frac{
\Gamma({ \textstyle 2 a - 2 b+1}) \,
\Gamma({ \textstyle 2 a - 2 c+1}) \,
\Gamma({ \textstyle 2 a - 2 d+1}) \,
\Gamma({ \textstyle 2 a - b - 2 c - 2 d+1}) }{\Gamma({ \textstyle
2 a+1}) \,\Gamma({ \textstyle 2 a - 2 b - 2 c+1}) \,
\Gamma({ \textstyle 2 a - b - 2 d+1}) \,
\Gamma({ \textstyle 2 a - 2 c - 2 d+1}) }\\
\times
{} _{3} F _{2} \!\left [ \begin{matrix} { 2 c, b,
a - 2 d+\frac{1}{2}}\\ { 2 a - b - 2 d+1,
a+\frac{1}{2}}\end{matrix} ; {\displaystyle 1}\right ].\end{gathered}$$ Thus we directly arrive at the right-hand side of .
Proof of Theorem \[theo:2\] {#sec:proof}
===========================
As already mentioned in Section \[sec:pattern\], the cases $(\al, \be, \ga)=(1,1,2)$, $(2,1,3)$, $(2,1,4)$, $(2,1,2)$, $(2,1,4)$ of identity are simply reformulations of Sato’s counter-examples up to and (i) is proved.
For (ii), the idea is to prove that no specialisation of both sides of can follow from the 120 Thomae relations, at least when $\al=\beta$. One may note that cannot formally be a consequence of any of Thomae’s relations since two (one would be enough) of its specialisation are not such consequences. But this does not rule out the possibility that some other specialisations would follow from Thomae’s relations. However, we show that this is never the case when $\al=\be$ is a positive integer.
We first determine the 120 Thomae relations for the left-hand side of when $\al=\be$ and to do this painlessly, we express $$\label{eq:membre de gauche}
{} _{3} F _{2} \!\left [ \begin{matrix} {\al+1, \al+1, 2\al}
\\ {3\al+1, 3\al+1}\end{matrix};1
\right ]$$ in the symmetric form in Theorem \[prop:hardydixonandcie\], which gives $2x_1=2x_2=2x_4=2x_5=3\al+1$ and $2x_3=4\al.$ The permutations of $x_1$, $x_2$, $x_3$, $x_4$ and $x_5$ show that is related only to the ${} _{3} F_{2}$-series $$_3F_2\!\!\left[ \begin{matrix} {\al+1, \al+1, 2\al}
\\ {3\al+1, 3\al+1}\end{matrix};1
\right ], \quad
_3F_2\!\!\left [ \begin{matrix} {2\al, 2\al, 2\al}
\\ {3\al+1, 4\al}\end{matrix};1
\right ],\label{eq:fin1}$$ and those obtained by the trivial symmetries. The same process applied to the right-hand side of (for $\al=\be$), $$\label{eq:membre de droite}
{} _{3} F _{2} \!\left [ \begin{matrix} {\al, \al, 2\al}
\\ {3\al+1, 3\al+1}\end{matrix};1
\right ],$$ gives $2x_1=2x_2=3\al+2$, $2x_3=4\al+2$ and $2x_4=2x_5=3\al+1$. The permutations of $x_1$, $x_2$, $x_3$, $x_4$ and $x_5$ show that is related to the five ${} _{3} F _{2}$-series $$\begin{gathered}
_3F_2\!\!\left[ \begin{matrix} {\al, \al, 2\al}
\\ {3\al+1, 3\al+1}\end{matrix};1
\right ],\;
_3F_2\!\!\left [ \begin{matrix} {2\al+2, 2\al+1, 2\al+1}
\\ {4\al+2, 3\al+2}\end{matrix};1
\right ],\;
_3F_2\!\!\left [ \begin{matrix} {2\al+1, \al+1, \al}
\\ {3\al+2, 3\al+1}\end{matrix};1
\right ],\; \nonumber\\
_3F_2\!\!\left [ \begin{matrix} {2\al+1, 2\al+1, 2\al}
\\ {4\al+2, 3\al+1}\end{matrix};1
\right ],\;
_3F_2\!\!\left [ \begin{matrix} {2\al+2, \al+1, \al+1}
\\ {3\al+2, 3\al+2}\end{matrix};1
\right ],
\label{eq:fin2}\end{gathered}$$ and those obtained by the trivial symmetries.
Inspection quickly reveals the impossibility of any numerical coincidence between one of the two arrays of parameters in and one of the five arrays in , even with trivial symmetries. However, each such coupling provides a variation of and, for example, we have that $$\label{eq:couplage}
{} _{3} F _{2} \!\left [ \begin{matrix} {2\al, 2\al, 2\al}
\\ {4\al, 3\al+1}\end{matrix};1
\right ]= \frac{\al(2\al+1)}{(3\al+1)(4\al+1)}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2\al+2, 2\al+1, 2\al+1}
\\ {4\al+2, 3\al+2}\end{matrix};1
\right ],$$ which will be used below.
We are now in a position to prove the claim about the infinity of counter-examples to the cases (i) and (ii) of Conjecture \[conj:1\]. First, thanks to , we have that $$I(2\al-1,2\al-1,\al, 2\al-1,\al)=
\frac{(2\al-1)!^3\,\al!}{(4\al-1)!\,(3\al)!} \,
{} _{3} F _{2} \!\left [ \begin{matrix} {2\al, 2\al, 2\al}
\\ {4\al, 3\al+1}\end{matrix} ;1
\right ]$$ and $$I(2\al+1,2\al-1,\al, 2\al,\al-1)=
\frac{(2\al+1)!\,(2\al)!\,(2\al-1)!\,\al!}{(4\al+1)!\,(3\al+1)!}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {2\al+2, 2\al+1, 2\al+1}
\\ {4\al+2, 3\al+2}\end{matrix};1
\right ].$$ We can relate these two equations by , and the simplification of the Gamma-factors yields $$I(2\al-1,2\al-1,\al, 2\al-1,\al)= I(2\al+1,2\al-1,\al, 2\al,\al-1),$$ which is exactly the identity we are looking for. For both integrals, the non-negativity conditions and in case (ii) of Conjecture \[conj:1\] are verified and the above discussion proves that there exists no permutation $\rho$ in the group ${\bf \Phi}$ (and, [*a fortiori*]{}, also none in ${\bf T}$) such that $\rho(2\al-1)=2\al+1$, $\rho(2\al-1)=2\al-1$, $\rho(\al)=\al$, $\rho(2\al-1)=2\al$, $\rho(\al)=\al-1.$ Thus, for each value of the positive integer $\al$, we obtain a counter-example to the cases (i) and (ii) of Conjecture \[conj:1\] at the same time. That this provides infinitely many counter-examples is a consequence of the fact that $I(2\al-1,2\al-1,\al, 2\al-1,\al)$ tends to 0 as $\al$ tends to infinity, because $$\begin{aligned}
\lim_{\al\to +\infty}
I(2\al-1,2\al-1,&\al, 2\al-1,\al)^{1/\al}\\
&=\max_{(x,y)\in [0,1]^2}
\bigg(\frac{x^2(1-x)^2y^2(1-y)}{(1-xy)^2}\bigg)=17-12\sqrt{2}<1.\end{aligned}$$
We could do the same thing with $\alpha$ not necessarily equal to $\beta$. To find all Thomae relations for the left-hand side of , one should use Theorem \[prop:hardydixonandcie\] with $$2x_1=2x_4=2\al+\be+1, \;2x_2=2x_5=2\be+\al+1,\; 2x_3=2\al+2\be,$$ leading to five different arrays up to trivial symmetries, and for the right-hand side with $$\begin{aligned}
2x_1=2\al+\be+2, \;2x_2&=2\be+\al+2,\\ &2x_3
=2\al+2\be+2,\; 2x_4=2\al+\be+1, \;2x_5=2\be+\al+1,\end{aligned}$$ leading to a complete set of 120 different arrays for generic $\al$ and $\beta$ (in fact, only 10 arrays, up to trivial symmetries). This explains why we consider only the case $\al=\be$, which is much simpler to deal with.
Proof of Theorem \[thm:gamma2\] {#sec:contig}
===============================
In order to derive Theorem \[thm:gamma2\], we require the following proposition, relating two “contiguous" $_3F_2$-series in a way that the “rest" is a closed form expression. (See the Remark after the proof of the proposition for an explanation of the term “contiguous.")
\[prop:gamma2\] For any complex numbers $a,b,c$ such that $\Re(d-b-c+1)>0$, and such that $a+1$ and $d$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b11}
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c}\\ { a+1, d}\end{matrix} ;
{\displaystyle 1}\right ]=
\frac{ ( a - b+1 ) ( a - b+2 )
( a-c+1 ) ( d -1)}
{( a+1 ) ( b-1 ) ( a - d +2 )
( a - d +1 ) }
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b-1, c}\\ { a+2,
d-1}\end{matrix} ; {\displaystyle 1}\right ]
\\+\frac {(1 - a - a^2 - b + c + a c - b c - d + b d)}
{ ( b-1 ) ( a - d +2 )
( a - d +1 )}
\frac{
\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d - b - c + 1}) }{
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) }.\end{gathered}$$
We start by applying the contiguous relation $$\begin{gathered}
\label{eq:C55}
{} _{3} F _{2} \!\left [ \begin{matrix} { A_1, A_2, A_3}\\ { B_1, B_2}\end{matrix} ;
{\displaystyle z}\right ] =
{\frac{\left( 1 - A_1 + A_2 \right) \left( B_1 - 1 \right)
}
{\left( A_1 - 1 \right) \left( 1 + A_2 - B_1 \right) }}
{} _{3} F _{2} \!\left [ \begin{matrix} { A_1 - 1, A_2, A_3}\\ { B_1 - 1,
B_2}\end{matrix} ; {\displaystyle z}\right ] \\+
{\frac {A_2 \left( B_1 - A_1 \right)
}
{\left( A_1 - 1 \right) \left( B_1 - A_2 - 1 \right) }}
{} _{3} F _{2} \!\left [ \begin{matrix} { A_1 - 1, A_2 + 1, A_3}\\ { B_1,
B_2}\end{matrix} ; {\displaystyle z}\right ],\end{gathered}$$ with $A_1=b$, $A_2=a$, and $B_1=d$, to the $_3F_2$-series on the left-hand side. This yields the expression $$\frac{( a-b+1 ) ( d-1 )
}{( b-1 )
( a-d+1 ) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a,b-1, c}\\ { d-1,
a+1}\end{matrix} ; {\displaystyle 1}\right ]
-\frac{a ( d-b )
}{( b-1 )
( a-d+1 ) }\,
{} _{2} F _{1} \!\left [ \begin{matrix} { b-1, c}\\ {
d}\end{matrix} ; {\displaystyle 1}\right ] .$$ We sum the $_2F_1$-series by means of the Gau[ß]{} summation formula (see [@SlatAC (1.7.6); Appendix (III.3)]) $$\label{eq:Gausz}
{} _{2} F _{1} \!\left [ \begin{matrix} { a, b}\\ { c}\end{matrix} ; {\displaystyle
1}\right ] = \frac {\Gamma ( c)\,\Gamma( c-a-b)} {\Gamma( c-a)\,
\Gamma( c-b)}.$$ Thus, we obtain $$\frac{( a-b+1 ) ( d-1 )
}{( b-1 )
( a-d+1 ) } \, {} _{3} F _{2} \!\left [ \begin{matrix} { a, b-1,c}\\ {
a+1,d-1}\end{matrix} ;
{\displaystyle 1}\right ] -
\frac{a\,\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d-b-c+1}) }
{( b-1 ) ( a-d+1 ) \,
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) }.$$ Next we apply the contiguous relation $$\label{eq:C15}
{} _3 F _2 \!\left [ \begin{matrix} { A_1, A_2,A_3}\\ { B_1,B_2}\end{matrix} ; {\displaystyle
z}\right ] = {} _3 F _2 \!\left [ \begin{matrix} { A_1 + 1, A_2,A_3}\\ {
B_1,B_2}\end{matrix} ; {\displaystyle z}\right ] -
{z }
{\frac{A_2A_3 }
{B_1B_2}}
{} _3 F _2 \!\left [ \begin{matrix} { A_1 + 1,A_2+1,A_3+1}\\ {B_1+1,B_2+1}\end{matrix} ;
{\displaystyle z}\right ],$$ with $A_1=a$. This gives $$\begin{gathered}
\frac{( a-b+1 ) ( d-1 )
}{( b-1 )
( a-d+1 ) } \, {} _{2} F _{1} \!\left [ \begin{matrix} { b-1, c}\\ {
d-1}\end{matrix} ; {\displaystyle 1}\right ]-
\frac{( a-b+1 ) c
}{( a+1 )
( a-d+1 ) }\,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b, c+1}\\ { a+2,
d}\end{matrix} ; {\displaystyle 1}\right ] \\+
\frac{a\,\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d-b-c+1}) }
{( b-1 ) ( -1 - a + d ) \,
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) }.\end{gathered}$$ Of course, the $_2F_1$-series can be summed by means of the Gau[ß]{} summation formula . After some simplification, we arrive at $$\label{eq:b13a}
- \frac{( a-b+1 ) c
}{( a+1 )
( a-d+1 ) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b, c+1}\\ { 2 +
a, d}\end{matrix} ; {\displaystyle 1}\right ] +
\frac{( 1 + a + c - d ) \,\Gamma({ \textstyle d}) \,
\Gamma({ \textstyle d-b-c}) }{( a-d+1 ) \,
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) }.$$ Now we apply another time the contiguous relation , this time with $A_1=b$, $A_2=a+1$, and $B_1=d$. We obtain $$\begin{gathered}
-
\frac{( a-b+1 ) ( a-b+2 ) c
( d-1 ) }{( a+1 ) ( b-1 )
( a-d+1 ) ( a-d+2 ) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} {
a+1, b-1, c+1}\\ { a+2, d-1}\end{matrix} ; {\displaystyle
1}\right ] \\+ \frac{( a-b+1 ) c ( d-b )
}{( b-1 )
( a-d+1 ) ( a-d+2 ) } \,
{} _{2} F _{1} \!\left [ \begin{matrix} { b-1, c+1}\\ {
d}\end{matrix} ; {\displaystyle 1}\right ] \\+
\frac{( a + c - d +1) \,\Gamma({ \textstyle d}) \,
\Gamma({ \textstyle d-b-c}) }{( a-d+1 ) \,
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) },\end{gathered}$$ and after evaluation of the $_2F_1$-series by means of Gau[ß]{}’ summation formula , $$\begin{gathered}
- \frac{( a-b+1 ) ( a-b+2 ) c
( d-1 )
}{
( a+1 ) ( b-1 )
( a-d+1 ) ( a-d+2 ) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b-1, c+1}\\ {
a+2, d-1}\end{matrix} ; {\displaystyle 1}\right ] \\ +
\frac{P(a,b,c,d)\,
\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d-b-c}) }
{( b-1 ) ( a-d+1 )
( a-d+2 ) \,\Gamma({ \textstyle d-b}) \,
\Gamma({ \textstyle d-c}) },\end{gathered}$$ where $$\begin{gathered}
P(a,b,c,d)=-2 - 3 a - a^2 + 2 b + 3 a b + a^2 b - 3 c - 2 a c +
3 b c + a b c - c^2 - a c^2 \\+ b c^2 + 3 d + 2 a d -
3 b d - 2 a b d + 2 c d + a c d - 2 b c d - d^2 + b d^2.\end{gathered}$$ The final contiguous relation that we apply is $$\begin{gathered}
\label{eq:C27}
{} _3 F _2 \!\left [ \begin{matrix} { A_1, A_2, A_3}\\ { B_1,B_2}\end{matrix} ;
{\displaystyle z}\right ] =
{\frac{ A_1 - A_2 - 1
} {A_1 - 1}}\,
{} _3 F _2 \!\left [ \begin{matrix} { A_1 - 1, A_2, A_3}\\ { B_1,B_2}\end{matrix} ;
{\displaystyle z}\right ] \\+
{\frac{A_2 } {A_1 - 1}}\,
{} _3 F _2 \!\left [ \begin{matrix} { A_1 - 1, A_2 + 1, A_3}\\ {
B_1,B_2}\end{matrix} ; {\displaystyle z}\right ],\end{gathered}$$ with $A_1=c+1$ and $A_2=a+1$. The result is $$\begin{gathered}
\frac{( a-b+1 ) ( a-b+2 )
( a-c+1 ) ( d-1 )
}{( a+1 )
( b-1 ) ( a-d+1 )
( a-d+2 ) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b-1, c}\\ { a+2,
d-1}\end{matrix} ; {\displaystyle 1}\right ] \\
- \frac{( a-b+1 ) ( a-b+2 )
( d-1 )
}{( b-1 )
( a-d+1 ) ( a-d+2 ) } \,
{} _{2} F _{1} \!\left [ \begin{matrix} { b-1, c}\\ {
d-1}\end{matrix} ; {\displaystyle 1}\right ]
\\+
\frac{P(a,b,c,d)\,\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d-b-c}) }
{( b-1 ) ( a-d+1 )
( a-d+2 ) \,\Gamma({ \textstyle d-b}) \,
\Gamma({ \textstyle d-c}) }.\end{gathered}$$ A last use of Gau[ß]{}’ summation formula and some simplification then leads to .
Since we shall re-use it in Section \[sec:contig2\], it will be beneficial if we briefly summarise the idea of the proof of the above proposition: it is crucially based on the fact that the $_3F_2$-series on the left of has the parameter $a$ on top and the parameter $a+1$ at the bottom. Now we apply elementary contiguous relations (such as the one in ). In principle, it expresses our $_3F_2$-series as a sum of two other $_3F_2$-series in which the parameters are “contiguous" to the original $_3F_2$-series, meaning that they differ from the parameters of the original series by small integer amounts. (In , these differences are $0$ and $\pm1$.) However, in one of the two $_3F_2$-series on the right-hand side of the relation, the top parameter $A_2=a$ is raised by $1$, while the bottom parameter $B_2=a+1$ is left invariant. Thus, the two $(a+1)$’s cancel, and the $_3F_2$-series reduces to a $_2F_1$-series, to which the Gau[ß]{} summation formula can be applied to express it in closed form. This partial simplification happens as well when we apply the contiguous relations and . Thus, each time, we obtain a $_3F_2$-series plus an additional expression in closed form. These additional expressions are put together, and they finally form the expression containing the gamma functions on the right-hand side of . However, since several similar, but not identical, such expressions were put together, when factoring the resulting term, a polynomial factor built up. Hence, in order to obtain a relation between two $_3F_2$-series without any additional term, this polynomial factor must vanish. While, normally (i.e., if one plays the above described game in a random fashion), equating this polynomial factor to zero will not have any nice solutions (in particular, no [*integral*]{} solutions, which we would however need to construct counter-examples to the conjecture by Rhin and Viola, in the cases of Propositions \[prop:gamma2\]–\[prop:gamma3\]), the contiguous relations have been carefully selected so that at least one of the variables $a,b,c,d$ is contained only linearly in the polynomial factor. This makes it possible to have many non-trivial solutions when equating the polynomial factor to zero.
In view of the above remark, the proof of Theorem \[thm:gamma2\] is now straight-forward.
If we now choose $d$ such that the polynomial factor on the right-hand side of vanishes, that is, $$d= c+\frac{a( a - c+1 ) }{b-1} +1,$$ and subsequently do the replacements $a\to\al$, $b\to\be$, $c\to\ga$, then we obtain exactly .
Proof of Theorem \[theo:2a\] {#sec:proofa}
============================
As already mentioned in Section \[sec:pattern\], the case $(\al, \be, \ga)=(4,3,4)$ of identity is a simple reformulation of Sato’s counter-example . Thus, (i) is proved.
For (ii), we proceed in a similar fashion as in Section \[sec:proof\]. First of all, we observe that identity is the special case of in which $\al$ is replaced by $\al^2$, and in which $\be=\al+1$ and $\ga=\al^2$, again via the translation . To wit, this is $$\label{eq:b3}
{} _{3} F _{2} \!\left [ \begin{matrix} { \al^2,\al^2,
\al+1}\\ { \al ^2+1,
\al^2+\al +1}\end{matrix} ; {\displaystyle 1}\right ]
=
\frac{\al ^3+1}{\al ^2+1} \;
{} _{3} F _{2} \!\left [ \begin{matrix} { \al^2
+1, \al^2, \al}\\ {
\al^2+2, \al^2+\al}\end{matrix} ; {\displaystyle 1}\right ] .$$ In the sequel we concentrate on this special case, always assuming that $\al$ is a positive integer strictly greater than $1$.
Using Thomae’s relations, we can generate three other series which are related to the $_3F_2$-series on the left-hand side of , namely $$\label{eq:rel1}
{} _{3} F _{2} \!\left [ \begin{matrix}
\al+1,\al+1,\al+1\\\al+2, \al^2+\al+1
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
1,1,\al+1\\\al+2,\al^2+1
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
1, \al^2, \al^2-\al\\\al^2+1, \al^2+1
\end{matrix}; {\displaystyle 1}\right ] .$$ On the other hand, there are six series related to the $_3F_2$-series on the right-hand side of , $$\begin{gathered}
\label{eq:rel2}
{} _{3} F _{2} \!\left [ \begin{matrix}
\al-1,\al^2,\al^2\\\al^2+1,\al^2+\al
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
\al-1,\al,\al\\\al+1,\al^2+\al
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
2,\al^2+1,\al^2-\al+2\\\al^2+2,\al^2+2
\end{matrix}; {\displaystyle 1}\right ] ,\\
{} _{3} F _{2} \!\left [ \begin{matrix}
1,\al^2,\al^2-\al+2\\\al^2+1,\al^2+2
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
1,2,\al\\\al+1,\al^2+2
\end{matrix}; {\displaystyle 1}\right ] ,\quad
{} _{3} F _{2} \!\left [ \begin{matrix}
1,1,\al-1\\\al+1,\al^2+1
\end{matrix}; {\displaystyle 1}\right ] .\end{gathered}$$ None of these match with the series on the left-hand side of or with one of the series in . Thus, indeed, for any positive integer $\al$, is a counter-example to the conjecture by Rhin and Viola.
Finally, in order to see that produces infinitely many counter-examples, we show again that the involved integral tends to zero when $\al$ tends to infinity. Indeed, for $\al\ge 1$, we have $$\begin{aligned}
\notag
I(\al^2-1,\al-1,\al^2-\al+1,\al-1,0)
&=\int_0^1 \!\!\int_0^1
\frac{x^{\al^2-1}(1-x)^{\al-1}y^{\al-1}(1-y)^{\al^2-\al+1}}
{(1-xy)^{\al^2+1}}\,\dd x\, \dd y\\
\notag
&\le \int_0^1 \!\!\int_0^1 x^{\al -1}y ^{\al-1} \frac{\dd x\, \dd y}{1-xy} =
\sum_{k=0} ^{\infty} \frac{1}{(k+\al)^2},\end{aligned}$$ from which the claim follows. (In the second line, we used the trivial facts that $x^{\al^2}\le x ^\al$ and $(1-x)^{\al-1}(1-y)^{\al ^2-\al+1}\le (1-xy)^{\al^2}$ for $0\le x, y\le 1$.) This completes the proof of Theorem \[theo:2a\].
It is obvious that Theorem \[thm:gamma2\] will generate many more counter-examples to the conjecture by Rhin and Viola, by choosing the parameters $\al,\be,\ga$ to be positive integers in other ways such that $\al ( \al - \ga +1)/({\be-1}+1 )$ is as well a positive integer (and such that the conditions – are satisfied). To have a convenient parametrisation, one would replace $\ga$ by $\al+1-\ga$, subsequently $\al$ by $a_1a_2$, $\ga$ by $c_1c_2$, and $\be$ by $a_1c_1+1$. The resulting relation is $$\begin{gathered}
\label{eq:b14}
{} _{3} F _{2} \!\left [ \begin{matrix} { a_1a_2,
a_1c_1+1, a_1a_2 -
c_1 c_2+1}\\ { a_1a_2+1,
a_1a_2 + a_2 c_2 -
c_1 c_2+2}\end{matrix} ; {\displaystyle 1}\right ] \\=
\frac{( a_1a_2 -
a_1c_1 +1)
( a_1a_2 +
a_2 c_2 - c_1 c_2
+1 ) }
{( a_1a_2+1 )
( a_2 c_2 -
c_1 c_2+1 ) }
\,{} _{3} F _{2} \!\left [ \begin{matrix} {
a_1a_2+1, a_1c_1,
a_1a_2 - c_1 c_2+1}\\ {
a_1a_2+2, a_1a_2 +
a_2 c_2 -
c_1 c_2+1}\end{matrix} ; {\displaystyle 1}\right ] .\end{gathered}$$
More exotic contiguous relations {#sec:contig2}
================================
In this section, we present two more relations of the kind of Theorem \[thm:gamma2\] (which itself followed from the more general Proposition \[prop:gamma2\]), see Theorems \[thm:beta\] and \[thm:gamma1\]. These are obtained along the lines described in the Remark after the proof of Proposition \[prop:gamma2\]. The two theorems imply further counter-examples to the conjecture by Rhin and Viola.
\[prop:beta\] For any complex numbers $a,b,c$ such that $\Re(d-b-c+1)>0$, and such that $a+1$ and $d$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b5}
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c}\\ { a+1,
d}\end{matrix} ; {\displaystyle 1}\right ]=
\frac{b c \left( a-d-1 \right) \left( a - d \right)
}{\left( a-b
\right) \left( a - c \right) d \left( d+1 \right) }
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b+1, c+1}\\ {
a+1, d+2}\end{matrix} ; {\displaystyle 1}\right ]\\
+\frac {a ( b c + a d - b d - c d ) }
{ ( a - b ) ( a - c ) }
\frac{
\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d-b-c+1}) }{
\Gamma({ \textstyle d-b+1}) \,\Gamma({ \textstyle d-c+1}) }.\end{gathered}$$
To the left-hand side, we apply the contiguous relation $$\begin{gathered}
\label{eq:C54}
{} _3 F _2 \!\left [ \begin{matrix} { A_1, A_2, A_3}\\ { B_1, B_2}\end{matrix} ;
{\displaystyle z}\right ] =
{\frac {A_2 \left( B_1 - A_1 \right) }
{\left( A_2 - A_1 \right) B_1}}
{} _3 F _2 \!\left [ \begin{matrix} { A_1, A_2 + 1,
A_3}\\ { B_1 + 1, B_2}\end{matrix} ; {\displaystyle z}\right ] \\+
{\frac {A_1 \left( B_1 - A_2 \right)
}
{\left( A_1 - A_2 \right) B_1}}
{} _3 F _2 \!\left [ \begin{matrix} { A_1 + 1, A_2, A_3}\\ { B_1 + 1,
B_2}\end{matrix} ; {\displaystyle z}\right ]\end{gathered}$$ with $A_1=b$, $A_2=c$, and $B_1=d$. As a result we obtain $$\frac{c( d-b )
}{( c-b ) d}
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c+1}\\ { a+1,
d+1}\end{matrix} ; {\displaystyle 1}\right ]
+ \frac{b ( d-c )
}{( b - c)d }
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b+1, c}\\ { a+1,
d+1}\end{matrix} ; {\displaystyle 1}\right ] .$$ We apply the contiguous relation again, to the first series with $A_1=a$, $A_2=b$, and $B_1=d+1$, to the second with $A_1=a$, $A_2=c$, and $B_1=d+1$. After some simplification, this leads to the expression $$\begin{gathered}
-
\frac{b c \left( a-d-1 \right) \left( a - d \right)
}{\left( -a + b \right)
\left( a - c \right) d \left( d+1 \right) }
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b+1, c+1}\\ { a+1,
d+2}\end{matrix} ; {\displaystyle 1}\right ]\\
+ \frac{a c \left( d-b \right) \left( d-b+1 \right)
}{\left( a - b \right)
\left( c-b \right) d \left( d+1 \right) }
{} _{2} F _{1} \!\left [ \begin{matrix} { b, c+1}\\ {
d+2}\end{matrix} ; {\displaystyle 1}\right ]+
\frac{a b \left( d-c \right) \left( d-c+1 \right)
}{\left( a - c \right)
\left( b - c \right) d \left( d+1 \right) }
{} _{2} F _{1} \!\left [ \begin{matrix} { b+1, c}\\ {
d+2}\end{matrix} ; {\displaystyle 1}\right ] .\end{gathered}$$ Finally, we use Gau[ß]{}’ summation formula to evaluate the two $_2F_1$-series. Some manipulation then yields the claimed result on the right-hand side of .
We may now choose $a$ so that the second term on the right-hand side of vanishes, that is, we choose $$a=b+c-\frac {b c}d.$$ After the additional replacements of $b$ by $\be$, of $c$ by $\ga$, and of $d$ by $\be\ga/\de$, we arrive at the following result.
\[thm:beta\] For any complex numbers $\al,\be,\ga$ such that $\Re\(\frac {\be\ga} {\de}-\be-\ga+1\)>0$, and such that $\be + \ga - \de+1$ and $\frac {\be\ga} {\de}$ are not non-positive integers, we have the identity $$\label{eq:b6}
{} _{3} F _{2} \!\left [ \begin{matrix} { \be + \ga - \de, \be,
\ga}\\ { \be + \ga -
\de+1, \frac{\be \ga}{\de}}\end{matrix} ; {\displaystyle 1}\right ] =
\frac{ \be \ga + \de - \be \de - \ga \de + \de^2
}{\be \ga + \de} \,
{} _{3} F _{2} \!\left [ \begin{matrix} { \be + \ga - \de, \be+1, \ga+1}\\ {
\be + \ga - \de+1, \frac{\be \ga}{\de}+2}\end{matrix} ; {\displaystyle
1}\right ] .$$
Again, if one wants a more convenient parametrisation for generating counter-examples to the conjecture by Rhin and Viola, then one would replace $\be$ by $b_1b_2$, $\ga$ by $c_1c_2$, and $\de$ by $b_1c_1$. The resulting relation then is $$\begin{gathered}
\label{eq:b15}
{} _{3} F _{2} \!\left [ \begin{matrix} { b_1b_2 + c_1c_2
- b_1c_1, b_1b_2, c_1c_2}\\ { b_1b_2 + c_1c_2 -
b_1c_1+1, {b_2 c_2}}\end{matrix}
; {\displaystyle 1}\right ] \\=
\frac{ b_1b_2 c_1c_2 + b_1c_1 -
b_1b_2 b_1c_1 - c_1c_2 b_1c_1 + b_1^2c_1^2
}{b_1b_2 c_1c_2 + b_1c_1} \\
\times
{} _{3} F _{2} \!\left [ \begin{matrix} { b_1b_2 +
c_1c_2 - b_1c_1, b_1b_2+1, c_1c_2+1}\\ {
b_1b_2 + c_1c_2 - b_1c_1+1,
{b_2 c_2}+2}\end{matrix} ; {\displaystyle
1}\right ] .\end{gathered}$$
\[prop:gamma3\] For any complex numbers $a,b,c$ such that $\Re(d-b-c+1)>0$, and such that $a+1$ and $d$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b13}
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c}\\ { a+1,
d}\end{matrix} ; {\displaystyle 1}\right ]
=
\frac{( a - b +1) (a-c+1 )
}{( a+1 )
( a - d +1) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b,c}\\ { a+2,
d}\end{matrix} ; {\displaystyle 1}\right ]\\ -
\frac{\Gamma({ \textstyle d}) \,\Gamma({ \textstyle d - b - c + 1}) }
{( a - d +1) \,\Gamma({ \textstyle d-b}) \,
\Gamma({ \textstyle d-c}) }.\end{gathered}$$
The first few steps of this proof are identical with the one of Proposition \[prop:gamma2\]. More precisely, we use that the series on the left-hand side is equal to the expression . There, we apply now instead the contiguous relation with $A_1=c+1$ and $A_2=a+1$. As a result, we obtain $$\begin{gathered}
\frac{\left( a - b+1 \right) \,\left(a-c+1 \right) \,
}{\left( a+1 \right) \,
\left( a - d +1\right) } \,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+1, b,c}\\ { a+2,
d}\end{matrix} ; {\displaystyle 1}\right ]
- \frac{ a - b +1
}{a - d+1} \,
{} _{2} F _{1} \!\left [ \begin{matrix} {b, c}\\ { d}\end{matrix} ;
{\displaystyle 1}\right ]\\+
\frac{\left( a + c - d+1 \right) \,\Gamma({ \textstyle d}) \,
\Gamma({ \textstyle d-b-c}) }{\left( a - d+1 \right) \,
\Gamma({ \textstyle d-b}) \,\Gamma({ \textstyle d-c}) },\end{gathered}$$ which, by another use of the Gau[ß]{} summation formula and some simplification, turns out to be equal to the right-hand side of .
An iterative use of Proposition \[prop:gamma3\] produces the following formula.
\[prop:gamma1\] For any complex numbers $a,b,c$ such that $\Re(d-b-c+1)>0$, and such that $a+1$ and $d$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b9}
{} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c}\\ { a+1,
d}\end{matrix} ; {\displaystyle 1}\right ]
=
\frac{ ( a - b +1) ( a - b+2 )
( a - c+1 ) ( a - c+2 )}
{( a+1 ) ( a+2 )
( a-d+2)(a-d+1 ) }\,
{} _{3} F _{2} \!\left [ \begin{matrix} { a+2, b, c}\\ { a+3,
d}\end{matrix} ; {\displaystyle 1}\right ]\\
-\frac {(3 + 5 a + 2 a^2 - b - a b - c - a c + b c - d - a d)}
{( a+1 )
( a-d+2)(a-d+1 )}
\frac{ \Gamma({ \textstyle d}) \,
\Gamma({ \textstyle d - b - c + 1}) }{
\Gamma({ \textstyle d-b}) \,
\Gamma({ \textstyle d-c}) }.\end{gathered}$$
If we now choose $d$ such that the polynomial factor on the right-hand side of vanishes, that is, $$d=2 a - b - c +3+ \frac{b c}{ a+1},$$ then we obtain the following theorem.
\[thm:gamma1\] For any complex numbers $\al,\be,\ga$ such that $\Re\(2\al-2\be-2\ga+\frac {\be} {\al+1}+4\)>0$, and such that $\al+1$ and $ 2 \al - \be - \ga+ \frac{\be \ga}{\al+1}+3$ are not non-positive integers, we have the identity $$\begin{gathered}
\label{eq:b10}
{} _{3} F _{2} \!\left [ \begin{matrix} { \al, \be, \ga}\\ { \al+1,
2 \al - \be - \ga + \frac{\be \ga}{\al+1}+3}\end{matrix} ; {\displaystyle
1}\right ] \\=
\frac {(\al+1)( \al - \be +2) ( \al - \ga+2 ) } {( \al +2)
( 3 \al + \al^2 - \be - \al \be - \ga - \al \ga + \be \ga+2 ) }
{} _{3} F _{2} \!\left [ \begin{matrix} { \al+2, \be, \ga}\\ { \al+3,
2 \al - \be - \ga+ \frac{\be \ga}{\al+1}+3}\end{matrix} ; {\displaystyle 1}\right
].\end{gathered}$$
If one wants a more convenient parametrisation for generating counter-examples to the conjecture by Rhin and Viola, then one would replace $\be$ by $b_1b_2$, $\ga$ by $c_1c_2$, and $\al$ by $b_1c_1-1$. The resulting relation then is $$\begin{gathered}
\label{eq:b16}
{} _{3} F _{2} \!\left [ \begin{matrix} {
b_1c_1-1,b_1b_2,
c_1c_2}\\ {b_1c_1, 2b_1c_1 -
b_1b_2
+b_2c_2 -
c_1c_2+1}\end{matrix} ; {\displaystyle 1}\right ] \\ =
\frac{( b_1c_1-b_1b_2
+1 )
( b_1c_1 -
c_1c_2+1 )
}{( b_1c_1+1 )
( b_1c_1 -b_1b_2
+b_2c_2 -
c_1c_2+1 ) } \\
\times {} _{3} F _{2} \!\left [ \begin{matrix} {
b_1c_1+1,b_1b_2,
c_1c_2}\\ { b_1c_1+2, 2b_1c_1
-b_1b_2
+
b_2c_2 -
c_1c_2+1}\end{matrix} ; {\displaystyle 1}\right ].\end{gathered}$$
Postlude: how were these identities found? {#sec:post}
==========================================
The reader may wonder how we found the identities in Theorem \[theo:??\] and Propositions \[prop:gamma2\]–\[prop:gamma3\] (the latter implying Theorems \[thm:gamma2\], \[thm:beta\] and \[thm:gamma1\]) and their proofs. This section describes some of the ideas that led us to their discovery, with some of them being interesting in their own right, as we believe. Since we shall make reference to it several times, we mention right away that all the hypergeometric calculations were carried out using the first author’s [*Mathematica*]{} package HYP [@hyp].
The counter-examples – of Sato, in their original form, do not give any hints for a general result that may be behind them. However, as we explain in Section \[sec:pattern\], if we bring them into different, but equivalent, forms using Thomae’s relations, patterns emerge. More precisely, by staring at the forms and of –, we extracted the wild guess that should hold. The first proof that we found (which is not presented here) showed first the special case $\ga=\al+\be$ of , given in , by using elementary contiguous relations. A somewhat involved analytic continuation argument, using the Gosper–Zeilberger algorithm (see below) and Carlson’s theorem then extended to .
However, it was “obvious" to us that one should be able to prove by a combination of several classical transformation formulae for hypergeometric series. Clearly, since we know that is not a consequence of Thomae’s relations, the classical $_3F_2$-transformations are not of any use. So we asked HYP to tell us which (of the built-in) transformations can be applied to the left-hand side of . (This is done by using [TListe]{}; see [@hyp].) The only “non-standard" transformation that HYP came up with was . (This is [T3240]{} in HYP.) So we applied it and quickly realized that we could exchange $\al$ and $\al+1$ in the obtained $_7F_6$-series (cf. and ) and apply in the other direction, in order to obtain a result different from the original $_3F_2$-series, which then turned out to be exactly the right-hand side of .
Having found an explanation for $83.33333\dots$ percent of Sato’s counter-examples did not completely satisfy us. We also wanted an explanation for . Since this is just one single identity, there is only very little guidance where to look for. What caught our eyes was that, in the hypergeometric form , both $_3F_2$-series were balanced (that is, the sum of the lower parameters exceeds the sum of the upper parameters by exactly 1). Not only that, in both series there is a lower parameter which exceeds an upper parameter by exactly 1. So, we made our computer work out the values of all series of the form $${} _{3} F _{2} \!\left [ \begin{matrix} { a, b, c}\\ { a+1,
b+c}\end{matrix} ; {\displaystyle 1}\right ]$$ for $1\le a,b,c\le 40$, and then compared which series were rational multiples of each other. By staring at the results, we extracted identities such as $$\label{eq:id1}
{} _{3} F _{2} \!\left [ \begin{matrix} { \al^2,
\al+1,\al^2}\\ { \al ^2+1,
\al^2+\al +1}\end{matrix} ; {\displaystyle 1}\right ]
=
\frac
{\al^3+1}
{\al^2 +1}\,
{} _{3} F _{2} \!\left [ \begin{matrix} { \al^2
+1, \al, \al^2}\\ {
\al^2+2, \al^2+\al}\end{matrix} ; {\displaystyle 1}\right ]$$ (this is identity , the special case $\al\to\al^2$, $\be=\al+1$, $\ga=\al^2$ of ), or $$\label{eq:id2}
{} _{3} F _{2} \!\left [ \begin{matrix} {
{\al}^ 2-\al+1, \al,
{\al}^ 2-\al}\\ {
{\al}^ 2-\al+2, {\al}^ 2}\end{matrix} ;
{\displaystyle 1}\right ] =
\frac{\al }{{\al}^2+1}\,
{} _{3} F _{2} \!\left [ \begin{matrix} {
{\al}^ 2-\al+1, \al+1, {\al}^ 2-\al+1}\\ {
{\al}^ 2-\al+2, {\al}^2+2
}\end{matrix} ; {\displaystyle 1}\right ]$$ (this is the special case $\be=\al$, $\ga=\al^ 2-\al$, $\de=\al-1$ of ) or $$\label{eq:id3}
{} _{3} F _{2} \!\left [ \begin{matrix}
{6\al+1, 4\al+2,3\al+1}\\ {
6\al+2,7\al+3}\end{matrix} ;
{\displaystyle 1}\right ] =
\frac{3\al+2 }{3{\al}+3}\,
{} _{3} F _{2} \!\left [ \begin{matrix}
{6\al+3, 4\al+2,3\al+1}\\ {
6\al+4,7\al+3}\end{matrix} ;
{\displaystyle 1}\right ]$$ (this is the special case $\al\to6\al+1$, $\be=4\al+2$, $\ga=3\al+1$ of ).
We then attempted to prove these identities. It seems sort of “obvious" that one should be able to prove them by using known contiguous relations. Indeed, in HYP there are approximately 100 such contiguous relations built-in. We played with those, but we were not able to arrive at the right-hand sides of the conjectured identities. At some point, we had the idea to “cheat" and to make recourse to the “modern" way of treating hypergeometric series, namely applying the Gosper–Zeilberger algorithm (see [@ek; @PeWZAA; @ZeilAP; @ZeilAM; @ZeilAV]; what we do below is in the spirit of [@PaulAZ]). For example, aiming to prove (a generalisation of) , we considered the series $$\label{eq:abc1}
{} _{3} F _{2} \!\left [ \begin{matrix}
a+n,b-n,c\\ a+n+1,b+c-n
\end{matrix} ;
{\displaystyle 1}\right ]$$ and tried to find a first-order recurrence for it (which is what is). Thus, we put the summand of this series, $$F(n,k)=\frac {(a+n)_k\,(b-n)_k\,(c)_k} {(a+n+1)_k\,(b+c-n)_k\,k!}$$ into the Gosper–Zeilberger algorithm, and we got $$\begin{gathered}
\label{eq:rek1}
( a + n+1) (b-n-1) (a - b - c + 2 n+1)
(a - b - c + 2 n+2) F(n,k) \\+
(a-b+2n+2) (a-b+2n+1) (a-c+n+1)
( b + c - n-1) F(n+1,k) =\De_k F(n,k) R(n,k),\end{gathered}$$ where $$\begin{gathered}
R(n,k)= \frac {k (b + c + k - n-1 )} {( a + n
) ( b + c - n-1 ) ( b + k - n-1 )}
\\\times
( k ( b + c - n-1 )
( a + n +1)
( 1 - a - a^2 - 2 b + b^2 + a c + n - 2 a n - 2 b n +
c n )\\ +\text{terms not containing $k$}),\end{gathered}$$ and where $\De_k$ is the forward difference operator, $(\De_k
f)(k)=f(k+1)-f(k)$. If we now sum both sides of over $k$ from $0$ to $N$, then we obtain $$\begin{gathered}
( a + n+1) (b-n-1) (a - b - c + 2 n+1)
(a - b - c + 2 n+2)
\sum _{k=0} ^{N} F(n,k) \\+
(a-b+2n+2) (a-b+2n+1) (a-c+n+1)
( b + c - n-1)
\sum _{k=0} ^{N} F(n+1,k)\\ = F(n,N+1) R(n,N+1),\end{gathered}$$ since the terms on the right-hand side telescope. Subsequently, the limit $N\to\infty$ yields $$\begin{gathered}
\label{eq:m=0}
( a+n+1 ) ( b-n-1 ) ( a - b-c+2n +2 )
( a - b-c +2n+1 )\\
\times
{} _{3} F _{2} \!\left [ \begin{matrix} { a+n, b-n, c}\\ { a+n+1,
b+c-n}\end{matrix} ;
{\displaystyle 1}\right ]\\=
{ ( a - b+2n+1 ) ( a - b+2n+2 )
( a-c+n+1 ) ( b+c-n -1)}
\\
\times
{} _{3} F _{2} \!\left [ \begin{matrix} { a+n+1, b-n-1, c}\\ { a+n+2,
b+c-n-1}\end{matrix} ; {\displaystyle 1}\right ]
\\+( a+n+1 )
(1 - a - a^2 - b + c + a c - b c - d + b d + n - 2 a n - b n +
2 c n - d n )
\frac{
\Gamma({ \textstyle b+c-n}) }{
\Gamma({ \textstyle c}) \,\Gamma({ \textstyle b-n}) }.\end{gathered}$$ (The reader should notice that this is with $a$ replaced by $a+n$, $b$ replaced by $b-n$, and $d$ replaced by $b+c-n$.)
At this point, we became greedy. Why should this be something special for balanced series? So, we replaced the bottom parameter $b+c-n$ in by $d-n$, — and we were disappointed to learn that the Gosper–Zeilberger algorithm is unable to find a two-term recurrence for this more general series. (It finds only a three-term recurrence.) However, it [*does*]{} find a two-term recurrence for every $d$ of the form $d=b+c+m$, where $m$ is a non-negative integer. From the data for $m=0$ (given in ) and for $m=1,2,3$, one is then easily able to work out a (at this point, conjectural) formula for the output of the algorithm, namely if $$F(n,k)=\frac {(a+n)_k\,(b-n)_k\,(c)_k} {(a+n+1)_k\,(b+c+m-n)_k\,k!},$$ then $$\begin{gathered}
\label{eq:rek2}
(b-n)_m\,(c)_m\,( a + n+1) (b-n-1)\\
\times (a - b - c-m + 2 n+1)
(a - b - c-m + 2 n+2) F(n,k) \\+
(b-n)_m\,(c)_m\, (a-b+2n+2) (a-b+2n+1)\kern5cm\\
\times (a-c+n+1)
( b + c+m - n-1) F(n+1,k] =\De_k F(n,k) R(n,k),\end{gathered}$$ where $$\begin{gathered}
R(n,k)= \frac {k (b + c+m + k - n-1 )} {( a + n
) ( b + c +m- n-1 ) ( b + k - n-1 )}
\\\times
( k^{m+1} ( b + c +m- n-1 )
( a + n +1) \\
\cdot
( 1 - a - a^2 - 2 b + b^2 + a c + n - 2 a n - 2 b n +
c n -m(n+1-b) )\\ +\text{terms with lower powers in $k$}),\end{gathered}$$ If one, for simplicity, replaces $b+c+m$ by $d$ in , sums both sides over $k$ from $0$ to $N$, and finally lets $N$ tend to infinity, one arrives exactly at , with $a$ replaced by $a+n$, $b$ replaced by $b-n$, and $d$ replaced by $d-n$.
As we pointed out, this is at best a half rigorous derivation of in the case that the difference $d-b-c$ is a non-negative integer, but there is no guarantee at all that this formula should also hold for [*any*]{} $d$. (To explain two of the possible pitfalls: first, there are always two ways to translate expressions such as $(c)_m$ into gamma functions: $(c)_m=\Ga(c+m)/\Ga(c)=(-1)^m\Ga(1-c)/\Ga(1-c-m)$. These lead to different formulae if $m$ is replaced by $d-b-c$, where $d$ is arbitrary. Second, sometimes one may even miss whole additional terms in a formula, which one does not see if some parameter is specialised to a non-negative integer because this additional term happens to vanish for this specialisation.) However, one can now [*prove*]{} continuing along the above lines: first, one verifies that is valid for $d=b$ by using Gau[ß]{}’ summation formula . Next, one replaces $d$ by $b+n$ in , and one uses the Gosper–Zeilberger algorithm to find recurrences in $n$ for the left-hand and the right-hand sides of . Thus, one knows that holds with $d=b+n$ for any non-negative integer $n$. Since both sides of are analytic in $d$ in a neighbourhood of $\infty$, one can use the principle of analytic continuation to deduce that holds for any complex $d$ where both sides are defined.
We did that, but finally we did succeed to work out a proof using known contiguous relations. Since this is completely elementary and, as we believe, more instructive, this is the proof that we have included in Section \[sec:contig\]. For obtaining the general identities which are behind and , given in Propositions \[prop:beta\] and \[prop:gamma3\], we proceeded similarly. In fact, the contiguous relations , , , , which we used in the proofs, are [C55]{}, [C15]{}, [C27]{}, and [C54]{}, respectively, in HYP.
Our computer experiments suggest that the above procedure produces a relation of the type for [*any*]{} series $$% \label{eq:abc1}
{} _{3} F _{2} \!\left [ \begin{matrix}
a+a_1n,b+b_1n,c+c_1n\\ a+a_1n+a_2,d+d_1n
\end{matrix} ;
{\displaystyle 1}\right ],$$ as long as $a_1,a_2,b_1,c_1,d_1$ are integers, $a_2$ a positive integer, and $b_1+c_1=d_1$. However, most of the time none of $a,b,c,d$ appears linearly in the big polynomial factor on the right-hand side. This makes it difficult to extract a general solution of the Diophantine equation which arises when one equates the polynomial factor to zero. Nevertheless, experimentally, there are many solutions for various choices of $a_1,b_1,c_1,d_1$.
[1]{}
W. N. Bailey, [*Generalized Hypergeometric Series*]{}, Cambridge University Press, Cambridge, 1935.
S. B. Ekhad, [*Maple*]{} program available at [http://www.math.rutgers.edu/\~zeilberg/tokhniot/EKHAD8]{}.
G. Gasper and M. Rahman, [*Basic hypergeometric series*]{}, Encyclopedia of Mathematics And Its Applications 35, Cambridge University Press, Cambridge, 1990.
A. C. Dixon, [*On a certain double integral*]{}, Proc. London Math. Soc [**2**]{}.2 (1905), 8–15.
S. Fischler, [*Groupes de Rhin–Viola et intégrales multiples*]{}, J. Théor. Nombres Bordeaux [**15**]{}.2 (2003), 479–534. Available at [http://almira.math.u-bordeaux.fr/jtnb/2003-2/Fischler.ps]{}.
G. H. Hardy, [*A chapter from Ramanujan’s note-book*]{}, Proc. Camb. Phil. Soc. [**21**]{} (1923), 492–503. C. Krattenthaler, [*HYP and HYPQ — [*Mathematica*]{} packages for the manipulation of binomial sums and hypergeometric series respectively $q$-binomial sums and basic hypergeometric series*]{}, J. Symbol. Comput. [**20**]{} (1995), 737–744; the [*Mathematica*]{} programs and their manuals are available at [http://igd.univ-lyon1.fr/\~kratt]{}.
C. Krattenthaler and T. Rivoal, [*Hypergéométrie et fonction zêta de Riemann*]{}, preprint (2004). Available at [http://front.math.ucdavis.edu/math.NT/0311114]{}.
P. Paule, [*Contiguous relations and creative telescoping*]{}, in preparation.
M. Petkovšek, H. Wilf and D. Zeilberger, [*A=B*]{}, A. K. Peters, Wellesley, 1996.
G. Rhin and C. Viola, [*On a permutation group related to $\zeta(2)$*]{}, Acta Arith. [**77**]{}.1 (1996), 23–56.
G. Rhin and C. Viola, [*The group structure for $\zeta(3)$*]{}, Acta Arith. [**97**]{}.3 (2001), 269–293.
S. Sato, [*On a generalisation of Beukers’ integrals*]{} (in japanese), Master Thesis, University of Tokyo, 2001.
L. J. Slater, [*Generalized hypergeometric functions*]{}, Cambridge University Press, Cambridge, 1966.
K. Srinivasa Rao and Christian Krattenthaler, [*On group theoretical aspects and symmetries of angular momentum coefficients*]{}, in: Symmetries in Science XI, B. Gruber, G. Marmo, N. Yoshinaga, eds., Springer, Berlin, New York, 2004. Available at [http://igd.univ-lyon1.fr/\~kratt/artikel/bregenz1.html]{}.
J. Thomae, [*Ueber die Functionen, welche durch Reihen von der Form dargestellt werden: $ 1+\frac p1\ \frac {p'}{q'}\ \frac {p''}{q''} +\frac p1\ \frac
{p+1}2\ \frac {p'}{q'}\ \frac {p'+1}{q'+1}\ \frac {p''}{q''}\ \frac
{p''+1}{q''+1} +\dotsm $*]{}, Borchardts J. für Math. (J. reine angew. Math.) [**87**]{} (1879), 26–73.
J. Van der Jeugt and K. Srinivasa Rao, [*Invariance groups of transformations of basic hypergeometric series*]{}, J. Math. Phys. [**40**]{}.12 (1999), 6692–6700.\
Available at [http://allserv.ugent.be/\~jvdjeugt/files/tex/grouphyJMP.tex]{}.
A. Verma and V. K. Jain, [*Transformations of nonterminating basic hypergeometric series, their contour integrals and applications to Rogers–Ramanujan identities*]{}, J. Math. Anal. Appl. [**87**]{} (1982), 9–44.
J. F. Whipple, [*A group of generalized hypergeometric series: relations between 120 allied series of type $F[a,b,c;d,e]$*]{}, Proc. London Math. Soc. [**23**]{}.2 (1925), 104–114.
D. Zeilberger, [*Three recitations on holonomic systems and hypergeometric series*]{}, Séminaire Lotharingien Combin. [**24**]{} (1990), Article B24a, 28 pp.
D. Zeilberger, [*A fast algorithm for proving terminating hypergeometric identities*]{}, Discrete Math. [**80**]{} (1990), 207–211.
D. Zeilberger, [*The method of creative telescoping*]{}, J. Symbolic Comput. [**11**]{} (1991), 195–204.
W. Zudilin, [*Arithmetic of linear forms involving odd zeta values*]{}, J. Th[é]{}or. Nombres Bordeaux [**16**]{} (2004), 251–291.
[^1]: $^\dagger$ Research partially supported by EC’s IHRP Programme, grant HPRN-CT-2001-00272, “Algebraic Combinatorics in Europe", and by the “Algebraic Combinatorics" Programme during Spring 2005 of the Institut Mittag–Leffler of the Royal Swedish Academy of Sciences.
[^2]: The letter “[I]{}” denotes the complex number $i$ in [*Maple*]{} and one must use another symbol. But since LaTeX is not [*Maple*]{}, there is no problem here.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- |
Theoretical Physics, The Blackett Laboratory, Imperial College,\
Prince Consort Road, London SW7 2BZ, U.K..
- |
Department of Astronomy and Astrophysics, Tata Institute of Fundamental Research,\
Homi Bhabha Road, Colaba, Mumbai 400005, India\
and\
Department of Physics, Case Western Reserve University,\
10900 Euclid Avenue, Cleveland, OH 44106-7079, USA.
author:
- Levon Pogosian
- Tanmay Vachaspati
title: 'Space of kink solutions in $SU(N)\times Z_2$'
---
epsf
Introduction
============
It is relatively easy to determine if a field theory with spontaneous symmetry breaking admits topological defects. If the asymptotic field configuration is topologically non-trivial, the interior field configuration must have a topological defect. However, there can be a large class of asymptotic field configurations, all having the same topological characteristics. Which of the many different boundary conditions with given topology should one use when trying to find a topological defect solution?
We will restrict our attention to the simplest kind of topological defects, namely kinks in one spatial dimension. However the field theories we will consider are rather general, having symmetry groups $SU(N)\times Z_2$ with $N$ being an odd integer. The field content will be a scalar field $\Phi$ transforming in the adjoint representation of $SU(N)$, and the $Z_2$ takes $\Phi$ to $- \Phi$. The potential of the field theory is taken to be such that it gives a vacuum expectation value of $\Phi$ that breaks the symmetry spontaneously to $H = [SU((N+1)/2)\times
SU((N-1)/2)\times U(1)]/C$, where $C=Z_{(N+1)/2}\times Z_{(N-1)/2}$ is the center of $SU((N+1)/2)\times SU((N-1)/2)$; other than having this property the potential is not restricted in any way. The vacuum manifold of the theory is disconnected because the $Z_2$ is broken down completely by the vacuum expectation value. Hence there are topological kinks in the theory.
Suppose we want to find the explicit solution for these kinks. Let $\Phi (x=-\infty ) = \Phi_-$ and $\Phi (x=+\infty ) =
\Phi_+$. Then, to obtain a topological defect, the only constraint is that $\Phi_+$ and $\Phi_-$ should lie in distinct topological sectors of the vacuum manifold. In fact, if $\Phi_+$ is a choice, $U\Phi_+ U^\dagger$ for $U\in
SU(N)$ is also a valid choice. In [@PogVac00] it was shown that the $SU(5)\times Z_2$ kink with $\Phi_+=-\Phi_-$ is unstable to small perturbations and that there exists a stable domain wall solution of lower energy corresponding to a different choice of $\Phi_+$. These results were generalized to $SU(N)\times Z_2$ in [@Vac01] where the concept of different classes of kink solutions was introduced. Given a kink solution, the rest of the solutions from the same class can be constructed by applying global gauge transformations from the coset space $H/I$ where $H$ is the unbroked symmetry group and $I \subseteq H$ is the “internal” symmetry group that leaves the original kink solution invariant. One such class of solutions was constructed in [@Vac01], however, several questions of relevance were left unanswered. Will there exist a kink solution for any choice of $\Phi_+$? Are the different solutions really distinct? How many distinct solutions can one obtain? Are these solutions stable? We will answer these questions in this paper.
In Sec. \[kinksolutions\] we will show that not all choices of $\Phi_+$ lead to kink solutions and we find that we must have $[\Phi_+, \Phi_-]=0$ in order for a solution to exist. This leads to a finite, discrete set of topological boundary conditions that can yield distinct kink solutions. Each boundary condition determines a class of continuously degenerate kink solutions in the model. Surprisingly, we also find that there are non-topological kink solutions for which the boundary conditions do not lie in distinct topological sectors. These solutions can also be classified and counted. We then find the manifold that describes the continuous degeneracy of every class. This manifold has non-trivial topological properties which suggests that certain closed domain walls are incontractable. In Sec. \[su5kinks\] we consider the specific example of an $SU(5)$ model with a quartic potential and construct the topological and non-topological kink solutions explicitly. In this case we also analyze the stability of the kink solutions in the three different classes. There is one globally stable class of solutions; another is locally stable for some parameters; the remaining classes are unstable for our choice of potential.
In Sec. \[su5mm\] we discuss the extension of our results on domain walls to $SU(5)$ magnetic monopoles. With fixed asymptotic field configurations, our findings suggest that there should exist three generations of fundamental $SU(5)$ magnetic monopole solutions. We summarize our results in Sec. \[conclusions\].
Kink boundary conditions {#kinkbc}
========================
The Lagrangian of our (1+1 dimensional) model is: $$L = {\rm Tr} (\partial_\mu \Phi )^2 - V(\Phi ) \ .
\label{lagrangian}$$ $V(\Phi )$ is a potential invariant under $$G\equiv SU(N) \times Z_2 \ , \label{originalsymm}$$ $N$ is taken to be odd, and the parameters in $V$ are such that $\Phi$ has an expectation value that can chosen to be $$\Phi_0 = \eta \sqrt{2 \over {N(N^2-1)}}
\pmatrix{n{\bf 1}_{n+1}&{\bf 0}\cr
{\bf 0}&-(n+1){\bf 1}_n\cr} \ ,
\label{phi0}$$ where ${\bf 1}_p$ is the $p\times p$ identity matrix and $\eta$ is an energy scale determined by the minima of the potential $V$. Such an expectation value spontaneously breaks the symmetry down to: $$H = [SU(n+1)\times SU(n)\times U(1)]/C \label{unbrokensymm} \ ,$$ where we have defined $$N\equiv 2n+1 \ , \label{littlen}$$ with $n \ge 1$ being an integer. The exact form of $V(\Phi )$ will not be important for most of our analysis. However, it does play a role in the stability of solutions and then we will choose it to be a quartic polynomial in $\Phi$.
If $\Phi (x=-\infty )=\Phi_-$, then $\Phi (x=+\infty) = \Phi_+ = -
U\Phi_- U^\dagger$ for any $U \in SU(N)$ implies that the boundary conditions are topologically non-trivial. For example, if $U\in
H$, the symmetry group that leaves $\Phi_-$ invariant, then $\Phi_+ = -\Phi_-$. The first question we ask is: for a fixed $\Phi_-$, for what choices of $\Phi_+$ can we obtain kink solutions? As we shall now see, for a solution to exist, we must necessarily choose $\Phi_+$ such that $[\Phi_+ , \Phi_- ]=0$.
In Appendix \[appendixa\] we will prove the stronger result that if $\Phi_k (x)$ is a solution then $[\Phi_\pm , \Phi_k (x)] =0$. Here we will give a qualitative argument in support of this statement. Once the boundary condition at $x=-\infty$ is fixed, the various small excitations of the field $\Phi$ around $\Phi_-$ can be classified as massless or massive. The only components of $\Phi$ that can be non-trivial in the kink solution are the massive modes since the massless modes, also called the Nambu-Goldstone modes, if non-vanishing inside the kink, will not decay as we go further away from the kink. The massive modes are given precisely by the generators that commute with $\Phi_-$ while the Nambu-Goldstone modes are those that do not commute. Hence $[\Phi_-, \Phi_k (x)]=0$ and, in particular, $[\Phi_-, \Phi_+]=0$.
Therefore to construct a kink solution, one needs to fix $\Phi_-$ to a vacuum expectation value and consider all possible commuting vacuum expectation values for $\Phi_+$. $\Phi_-$ can be chosen to be diagonal and by performing rotations that leave $\Phi_-$ invariant ([*i.e.*]{} lie in the unbroken group $H$ at $x=-\infty$) $\Phi_+$ can also be brought to diagonal form.
Now we can explicitly list all the possible boundary conditions (up to gauge rotations) that can lead to kink solutions. At $x=-\infty$, we fix $\Phi_-=\Phi_0$ given in eq. (\[phi0\]). Then we can have $$\begin{aligned}
\Phi_+&{}&=
\epsilon_T \eta \sqrt{2 \over {N(N^2-1)}} \times \nonumber \\
&{}& {\rm diag} ( n{\bf 1}_{n+1-q}, -(n+1){\bf 1}_{q},
n{\bf 1}_{q}, -(n+1){\bf 1}_{n-q}) \ ,
\label{phi+choices}\end{aligned}$$ where we have introduced a parameter $\epsilon_T =\pm 1$ and another $q=0,...,n$. The label $\epsilon_T$ is $+1$ when the boundary conditions are topologically trivial and is $-1$ when they are topologically non-trivial. $q$ tells us how many diagonal entries of $\Phi_-$ have been permuted in $\Phi_+$. The case $q=0$ is when $\Phi_+ = \epsilon_T \Phi_-$. The case $q=n$ was considered in detail in Ref. [@Vac01].
Kink solutions {#kinksolutions}
==============
We now find kink solutions for any allowed boundary conditions $\Phi_\pm$. As a starting point we take the following ansatz: $$\Phi_k = F_+ (x) {\bf M_+} + F_- (x) {\bf M_-} + g(x) {\bf M}\ ,
\label{kinkexplicit2}$$ where $${\bf M}_+ = {{\Phi_+ + \Phi_-}\over {2}} \ , \ \ {\bf M}_- =
{{\Phi_+ - \Phi_-}\over {2}} \label{M+M-} \ ,$$ $g(\pm \infty)=0$ and $M$ is yet to be found. Explicitly, for $\epsilon_T=-1$, we have $$\begin{aligned}
{\bf M}_+ = \eta &N& \sqrt{1 \over {2N(N^2-1)}} \nonumber \\
&{}& {\rm diag} ( 0_{n+1-q}, {\bf 1}_q, -{\bf 1}_q, 0_{n-q} )
\label{M+} \ ,\end{aligned}$$ $$\begin{aligned}
{\bf M}_- = &\eta& \sqrt{1 \over {2N(N^2-1)}} \nonumber \\
&{}&{\rm diag} ( -2n {\bf 1}_{n+1-q}, {\bf 1}_q,
{\bf 1}_q, 2(n+1){\bf 1}_{n-q} ) \ .
\label{M-}\end{aligned}$$ Note that the matrices ${\bf M}_\pm$ are orthogonal: $${\rm Tr}({\bf M}_+{\bf M}_-) = 0 \ , \label{trM+M-}$$ but are not normalized to 1/2. The boundary conditions for $F_\pm$ are: $$\begin{aligned}
F_- (- \infty ) &=& -1 \ , \ \ F_- (+\infty ) =+1 \ , \nonumber \\
F_+ (- \infty ) &=& +1 \ , \ \ F_+ (+\infty ) =+1 \ .
\label{Fpmbc}\end{aligned}$$ The advantage of this form of the ansatz is that, for particular values of the parameters of a quartic potential in the $q=n$ topological ($\epsilon_T=-1$) case, one finds the explicit and simple solution $F_- (x) = \tanh (\sigma x)$, $F_+ (x) =1$ and $g(x)=0$, where $\sigma$ is the kink width which can be written in terms of the parameters [@PogVac00; @Vac01]. Also, for $q=0$, $\epsilon_T=-1$, the solution is the embedded $Z_2$ kink [*i.e.*]{} $F_+(x)=g(x)=0$, $F_- (x) = \tanh (\sigma x)$.
Now we would like to find the unknown matrix ${\bf M}$ in the ansatz (\[kinkexplicit2\]). This can be done by treating $g(x) {\bf M}$ as a small perturbation to $$\Phi_k^{(0)}\equiv F_+ (x) {\bf M_+} + F_- (x) {\bf M_-} \, .
\label{phi-try}$$ The perturbation is restricted to generators that are orthogonal to $\Phi_k^{(0)}$: $${\rm Tr}(\Phi_k^{(0)} {\bf M}) = 0 \ . \label{phikM}$$ We need to check if the energy density contains any terms that are linear in $g(x)$, otherwise we could always construct a stable kink solution with $g(x)=0$. The quadratic terms in the energy density clearly will not have such terms since ${\rm
Tr}(\Phi_k^{(0)} {\bf M}) =0$. The only terms that may be linear in $g(x)$ will be from terms in the potential such as ${\rm Tr}
(\Phi^s )$ for even $s \ge 4$. ($s$ has to be even since the potential is taken to have a $Z_2$ symmetry under $\Phi
\rightarrow -\Phi$.) There will be no terms linear in $g(x)$ only if $${\rm Tr} ((\Phi_k^{(0)})^{s-1} {\bf M}) = 0 \label{trphikM}$$ for every possible choice of ${\bf M}$ satisfying the conditions: $${\rm Tr} ({\bf M}) = 0 \ , \ \ {\rm Tr} ({\bf M}_- {\bf M}) = 0 \
, \ \ {\rm Tr} ({\bf M}_+ {\bf M}) = 0 \ . \label{conditionsonM}$$ If ${\bf M}$ is off-diagonal, eq. (\[trphikM\]) is satisfied because the trace of the product of a diagonal and an off-diagonal matrix vanishes. ($\Phi_k^{(0)}$ is diagonal.) The non-trivial part is to check the condition for diagonal ${\bf M}$ and we shall now concentrate on this case.
Let us write ${\bf M}$ as: $${\bf M} = {\rm diag} ({\bf U}_{n+1-q}, {\bf V}_q ,
{\bf W}_q, {\bf X}_{n-1}) \ ,
\label{Mmatrices}$$ where ${\bf U}_{n+1-q}$, ${\bf V}_q$, ${\bf W}_q$ and ${\bf
X}_{n-1}$ are diagonal matrices of order given by their subscripts. Implementation of the conditions in eq. (\[conditionsonM\]) leads to: $${\rm Tr} {\bf U}_{n+1-q} = - {\rm Tr} {\bf V}_q = -{\rm Tr} {\bf
W}_q = {\rm Tr} {\bf X}_{n-q} \ . \label{conditionsonUVWX}$$ Note that if $q=0$ or if $q=n$, this condition enforces each matrix to be traceless.
Now, to check if eq. (\[trphikM\]) is satisfied, we insert the form of $\Phi_k^{(0)}$ from eq. (\[phi-try\]). From the boundary conditions in eq. (\[Fpmbc\]), it is clear that the functions $F_\pm (x)$ are linearly independent and so eq. (\[trphikM\]) can only be satisfied if: $${\rm Tr} ({\bf M}_+^\alpha {\bf M}_-^\beta {\bf M}) = 0
\label{conditiononta}$$ for integers $\alpha$, $\beta$ such that $0 \le \alpha+\beta \le
s-1$. Explicit evaluation of this trace, together with the relations in eq. (\[conditionsonUVWX\]) shows that the condition is satisfied by all ${\bf M}$ with ${\rm Tr}{\bf V}_q=0$. However, for ${\bf M}$ with ${\rm Tr}{\bf V}_q\ne 0$, the condition is not met if $\alpha$ is an even integer.
How many generators are there for which ${\rm Tr}{\bf V}_q \ne 0$ and that satisfy the conditions in eq. (\[conditionsonUVWX\])? There are a total number of $N-1$ diagonal $SU(N)$ generators. Of these, the number of generators satisfying the conditions in eq. (\[conditionsonUVWX\]) together with ${\rm Tr}{\bf V}_q=0$ are $$(n+1-q-1)+(q-1)+(q-1)+(n-q-1)=N-4 .$$ Hence there are $(N-1)-(N-4)=3$ choices of ${\bf M}$ for which the condition in eq. (\[conditionsonUVWX\]) plus ${\rm Tr}{\bf
V}_q=0$ is not met. However this number includes the two possibilities ${\bf M}={\bf M}_\pm$. Hence there is only one remaining possible choice of ${\bf M}$ and this is: $$\begin{aligned}
{\bf M} = &\mu& \, {\rm diag} ( q(n-q){\bf 1}_{n+1-q}, \nonumber \\
&-&(n-q)(n+1-q){\bf 1}_q , -(n-q)(n+1-q){\bf 1}_q, \nonumber \\
&& \hskip 1.0 truecm q(n+1-q){\bf 1}_{n-q} )
\label{Mresult}\end{aligned}$$ with $\mu$ being a normalization factor in which we also include the energy scale $\eta$ for convenience: $$\mu = \eta [ 2q(n-q)(n+1-q)\{ 2n(n+1-q)-q\} ]^{-1/2} \ .
\label{muvalue}$$ Note that the matrix ${\bf M}$ is not normalizable if $q=0$ or if $q=n$. For these values of $q$, we can set $g(x)=0$ and $\Phi_k^{(0)}$ coincides with the ansatz $\Phi_k$.
It is easy to see that $\Phi_k$ is a valid ansatz. Any perturbations that are orthogonal to $\Phi_k$ would have to satisfy eq. (\[conditionsonUVWX\]) as well as be orthogonal to ${\bf M}$. Such perturbations necessarily have ${\rm Tr}{\bf V}_q
=0$. Further, all traces of the kind in eq. (\[trphikM\]) are proportional to ${\rm Tr}{\bf V}_q$ and hence vanish. This justifies the ansatz in eq. (\[kinkexplicit2\]).
The functions $F_\pm (x)$ and $g(x)$ can be found by solving their equations of motion derived from the Lagrangian together with the specified boundary conditions. There is no guarantee that a solution will exist and so we find the solutions explicitly for $N=5$ with a quartic potential in Sec. \[su5kinks\].
An interesting point to note is that the ansatz is valid even if $\Phi_\pm$ are not in distinct topological sectors [*i.e.*]{} even if $\epsilon_T=+1$. These imply the existence of non-topological kink solutions in the model. If we include a subscript $NT$ to denote “non-topological” and $T$ to denote “topological”, we have $$\Phi_{NTk} = F_+ (x) {\bf M}_{NT+} + F_- (x) {\bf M}_{NT-}
+ g(x) {\bf M}_{NT} \ .
\label{NTfinalansatz}$$ Since $\Phi_{NT+}=-\Phi_{T+}$, we find $${\bf M}_{NT+} = {\bf M}_{T-} \ , \ \ {\bf M}_{NT-} = {\bf M}_{T+}
\ , \ \ {\bf M}_{NT} ={\bf M}_{T} \ . \label{NTtoT}$$ Hence $$\Phi_{NTk} = F_- (x) {\bf M}_{T+} + F_+ (x) {\bf M}_{T-}
+ g(x) {\bf M}_T \ .
\label{NTsolution}$$ So to get $F_-$ ($F_+$) for the non-topological kink we have to solve the topological $F_+$ ($F_-$) equation of motion with the boundary conditions for $F_-$ ($F_+$). To obtain $g$ for the non-topological kink, we need to interchange $F_+$ and $F_-$ in the topological equation of motion. The boundary conditions for $g$ are unchanged.
In Sec. \[su5kinks\] we will find the topological and the non-topological kinks explicitly for $N=5$. Generally the non-topological solutions, if they exist, will be unstable. However, the possibility that some of them may be locally stable for certain potentials cannot be excluded.
Kink classes {#kinkclasses}
============
In Sec. \[kinkbc\] we showed that there is a discrete set of boundary conditions that lead to different topological kink solutions. The discrete set is labeled by the integer $q$ which runs from $0$ to $n$. Hence there are $n+1$ distinct classes of kink solutions in the $SU(N)\times Z_2$ model under consideration [@Vac01].
The explicit construction of the $n+1$ classes of kinks has already been described in Sec. \[kinksolutions\]. Eq. (\[kinkexplicit2\]) describes the form of the solution for a fixed value of $q$. A solution of this form is one member of the class of kinks labeled by $q$. What are the other members of the class?
The members of a class of kinks is given by the set of boundary conditions that will lead to gauge equivalent kinks. In other words, there is a set of transformations belonging to the unbroken symmetry group, $H_-$ in eq. (\[unbrokensymm\]) defined by the vacuum expectation value $\Phi_-$, that will leave $\Phi_-$ invariant but will rotate $\Phi_+$ non-trivially. The kink solutions obtained by these global gauge transformation will appear different from the original kink at the level of field configurations but are degenerate and belong to the same class. If $K_q$ is the subgroup of $H_-$ that leaves the $q$-kink solution, $\Phi_k$, invariant, then $$\Sigma_q \equiv H_-/K_q$$ describes the class of $q$-kinks.
Another way to describe $\Sigma_q$ is in terms of all perturbative modes that do not change the energy of the solution [*i.e.*]{} the zero modes on the solution background. This will include modes that give spatial translations and internal space rotations. The translations have not been included in $\Sigma_q$, while the internal space rotations have been included just as in the case of a “moduli space”. However, the internal zero modes may not vanish at $x=+\infty$ and hence are not required to be normalizable.
Now we will find $\Sigma_q$ for various $q$.
When $q=0$, $\Phi_k$ is proportional to $\Phi_-$ and $K_q =H_-$ [*i.e.*]{} the symmetry group that leaves the kink invariant is the entire unbroken symmetry group. Therefore $\Sigma_0 = 1$ and there is only one element in the $q=0$ kink class.
When $0< q <n$, it is clear from eq. (\[phi+choices\]) that the elements of $H_-$ that leave $\Phi_+$ invariant are $SU(n+1-q)$ in the first block, $SU(q)$ in the second block, $SU(q)$ in the third block, and $SU(n-q)$ in the fourth block. In addition, the diagonal generators of $H_-$ commute with $\Phi_+$ and these yield another three $U(1)$ factors. Hence the boundary condition at $x=+\infty$ is invariant under $$[SU(n+1-q) \times (SU(q))^2 \times SU(n-q) \times U(1)^3]/Z_K \ ,
\label{qkinksymm}$$ where we have modded out the continuous group by its center, symbolically denoted by $Z_K$. (This is necessary since the center of $SU(n+1-q)$ for example, is also contained in the $U(1)$ factors.) From the form of ${\bf M}$ in eq. (\[Mresult\]), it is clear that the group in (\[qkinksymm\]) is also the symmetry group that leaves ${\bf M}$ invariant. Hence it is also the symmetry group that leaves the entire kink solution $\Phi_k$ invariant and so: $$\begin{aligned}
K_q = [SU(n+1-q) \times (SU(q))^2 &\times& SU(n-q)\nonumber \\
&\times& U(1)^3]/Z_K \ .
\label{qkinksymmfinal}\end{aligned}$$ Therefore $\Sigma_q = H/K_q$ where $H$ is given in eq. (\[unbrokensymm\]) and $K_q$ in eq. (\[qkinksymmfinal\]).
When $q=n$, the analysis is modified a little bit since now $n-q=0$ and the last block in $\Phi_+$ is absent. So now we have $$K_n = [ (SU(n))^2 \times U(1)^2 ]/Z_K \ . \label{nkinksymm}$$
Note that the above classification scheme holds for both topological ($\epsilon_T=-1$) and non-topological ($\epsilon_T=+1$) kink solutions.
The space $\Sigma_q$ ($q\ne 0$) has interesting topological properties. For example, it has a non-trivial second homotopy group. This suggests that certain spherical configurations of domain walls (in three spatial dimensions) will be topologically non-trivial and may not be able to contract. We postpone a detailed investigation of the interpretation of the non-trivial topology of $\Sigma_q$ and its consequences for future work.
Kink solutions for $N=5$ {#su5kinks}
========================
In this section we will explicitly construct the kink solutions when $N=5$ and when the potential is quartic: $$V(\Phi ) = - m^2 {\rm Tr}[ \Phi ^2 ]+ h ( {\rm Tr}[\Phi ^2 ])^2 +
\lambda {\rm Tr}[\Phi ^4] + V_0 \ .
\label{quarticV}$$ The desired symmetry breaking to $$H= [SU(3)\times SU(2)\times U(1)]/[Z_3\times Z_2] \label{His321}$$ is achieved in the parameter range $${h \over \lambda} > - {{N^2+3}\over {N(N^2-1)}}\biggl |_{N=5}
= - {7\over {30}} \ .
\label{symmbreakparam}$$ The vacuum expectation value, $\Phi_-$ is $$\Phi_- = \eta {1\over \sqrt{60}} (2,2,2,-3,-3) \label{su5phi-}$$ with $$\eta \equiv {{m} \over {\sqrt{\lambda '}}} \label{eta}$$ and $$\lambda ' \equiv
h + {{N^2+3}\over {N(N^2-1)}}\biggl |_{N=5} \lambda
= h + {7\over {30}} \lambda \ .
\label{lambdaprime}$$
The $q=0$ topological kink ($\Phi_+=-\Phi_-$) has been found in Ref. [@PogVac00] and is simply an embedded $Z_2$ kink for all parameters: $$\Phi_k^{q=0} = \tanh \left ( {{mx}\over {\sqrt{2}}} \right )
\Phi_- \ .
\label{su5qis0}$$ As discussed in Sec. \[kinkclasses\], there is only one kink solution in this class.
To find the $q=1$ topological kink solution, we use the ansatz found in Sec. \[kinksolutions\] $$\Phi_k^{q=1} = F_+{\bf M}_+ + F_- {\bf M}_- + g {\bf M}$$ with $${\bf M}_+ = \eta \sqrt{5\over {48}} {\rm diag}(0,0,1,-1,0) \ ,
\label{bfM+su5}$$ $${\bf M}_- = \eta {1\over {\sqrt{240}}} {\rm diag}(-4,-4,1,1,6) \ ,
\label{bfM-su5}$$ $${\bf M} = \eta {1\over {2\sqrt{7}}} {\rm diag}(1,1,-2,-2,2) \ .
\label{bfMsu5}$$ Inserting the ansatz in the Lagrangian we can derive the equations of motion for the functions $F_\pm$ and $g$. (These are given in Appendix \[appendixb\].) The boundary conditions on these functions are: $$F_+ (\pm \infty )=1\ , \ \ F_- (\pm \infty ) = \pm 1 \ , \ \ g(\pm
\infty ) = 0 \ . \label{su5bcs}$$ If we assume that $|g''| << m^2 |g| << 1$ and $|F_+''| << m^2 |F_+|$, an approximate analytic solution can be obtained when $h = -3\lambda /70$. (The assumptions can later be checked for self-consistency.) The approximate solution is: $$F_- \simeq \tanh \left ( {m\over {\sqrt{2}}} x \right ) \ ,
\label{approxF-}$$ $$g \simeq - {{\gamma_6 F_- (\alpha_1 + \alpha_2 F_-^2)} \over
{(\alpha_2 \gamma_1 -\alpha_1 \gamma_3) +
(\alpha_2 \gamma_4 + \alpha_5 \gamma_6) F_-^2}} \ ,
\label{approxg}$$ $$F_+ \simeq \alpha_2^{-1/2}
[ -\alpha_1 - \alpha_5 g F_- ]^{1/2} \ ,
\label{approxF+}$$ where the coefficients $\alpha_i$ and $\gamma_i$ are given in Appendix \[appendixb\]. This approximate solution can be extended to other near-by parameters and a comparison with the numerically obtained solutions shows that the approximation is reasonably good except at the turning points of $F_+$ and $g$. However, the qualitative features of the numerical solution are captured by the approximation. We show the numerical solution for $h=-3\lambda /70$ in Fig. \[qeq1top\]. A numerical investigation for other values of $h/\lambda$ shows that a solution always exists for the $q=1$ topological kink.
The class of $q=1$ kinks is described by the space $$\Sigma_1 = H/K_1 \label{sigma1} \ ,$$ where $$K_1 = [SU(2) \times U(1)^3]/Z_2 \label{k1forsu5} \ .$$
The $q=2$ kink has been found in Ref. [@PogVac00] (also see [@Vac01]). In the case when $${h\over \lambda} = - {3\over {20}} \label{hlambdaqeq2}$$ the solution can be written down simply as: $$\Phi_k^{q=2} = {{1-\tanh(\sigma x)}\over 2} \Phi_- +
{{1+\tanh(\sigma x)}\over 2} \Phi_+
\label{qeq2solution}$$ with $$\Phi_+ = - \eta {1\over \sqrt{60}} (2,-3,-3,2,2) \ .
\label{qeq2su5phi+}$$ ($\Phi_-$ is given by eq. (\[su5phi-\]) and $\sigma
=m/\sqrt{2}$.)
A more general ansatz, valid for all values of $h/\lambda$, is $$\Phi_k^{q=2}={{F_+(x)-F_-(x)}\over 2} \Phi_- +
{{F_+(x)+F_-(x)}\over 2} \Phi_+ \, ,
\label{q2ansatz}$$ where functions $F_+$ and $F_-$ satisfy the same boundary conditions as in (\[su5bcs\]). The equations of motion for the $q=2$ kink along with a numerical solution were presented in [@PogVac00].
-0.25 truein
0.5 truecm = 0.95 0.5 truecm
The class of $q=2$ kinks is described by the space $$\Sigma_2 = H/K_2 \label{sigma2} \ ,$$ where $$K_2 = [SU(2)^2 \times U(1)^2]/Z_2^2 \ . \label{k2forsu5}$$
Now we will also construct the non-topological ($\epsilon_T=+1$) kinks in the model.
The $q=0$ non-topological kink is simply the vacuum $\Phi_{NTk} = \Phi_+$ and there is only one member in this class.
As discussed at the end of Sec. \[kinksolutions\], to construct the $q=1$ non-topological kink we can use the same equations as for the topological case but we should switch the boundary conditions on $F_+$ and $F_-$ (eq. (\[su5bcs\])). The system of equations has been solved numerically for a few choices of parameters. For $h=-14\lambda /70$, the profile functions are shown in Fig. \[q1nontop14o70\]. For $h=-3\lambda /70$ we find that the $q=1$ non-topological kink breaks up into two $q=2$ topological kinks. Specifically the $q=1$ kink interpolating between $\Phi \propto (2,2,2,-3,-3)$ and $(2,2,-3,2,-3)$ breaks up into one $q=2$ kink interpolating between $(2,2,2,-3,-3)$ and $-(-3,-3,2,2,2)$ and another interpolating between $-(-3,-3,2,2,2)$ and $(2,2,-3,2,-3)$ This suggests that there is a repulsive force between different $q=2$ kinks for parameters close to $h=-3\lambda /70$ and so there will be no non-topological $q=1$ kink solution in a certain range of parameters. Numerically we have determined the critical parameter where the $q=1$ non-topological boundary conditions lead to two well-separated topological $q=2$ kinks instead of one bound object. Hence we find that there are no $q=1$ non-topological kink solutions for $h >
-0.18 \lambda$ .
-0.25 truein
0.5 truecm = 0.95 0.5 truecm
The $q=2$ non-topological kink can be found by solving the same equations of motion as for the topological $q=2$ kink after switching the boundary conditions on $F_+$ and $F_-$ ($g=0$ in this case). Then, for the parameter $h=-3\lambda/20$, one has $$\Phi_{NTk}^{q=2} = {{1-\tanh(\sigma x)}\over 2} \Phi_- +
{{1+\tanh(\sigma x)}\over 2} \Phi_+ \ ,
\label{NTqeq2solution}$$ where $$\Phi_+ = + \eta {1\over \sqrt{60}} (2,-3,-3,2,2) \ .
\label{NTqeq2su5phi+}$$ For general values of parameters the profile functions can be found by numerical relaxation.
Kink stability {#kinkstability}
==============
To analyze the stability of the various kink solutions, we have to expand the energy density to second order in perturbations and then look for unstable modes. This would have to be done on a case by case basis for every different choice of potential. Here we will analyze the stability of the $SU(5)$ kinks constructed in the previous section.
The $q=0$ topological kink is known to be unstable [@PogVac00]. To see this, note that the Nambu-Goldstone modes are massless at $x=\pm \infty$ and have a negative mass squared at the origin where $\Phi_k =0$. Furthermore, it can be checked that the mass squared for the Nambu-Goldstone modes is everywhere negative for any choice of parameters. We know that an everywhere negative potential in one dimension always admits a bound state. Therefore the $q=0$ topological kink is unstable towards the growth of the Nambu-Goldstone modes for all parameters.
The $q=1$ topological kink is perturbatively unstable. The unstable modes correspond to the four generators of $SU(5)$ which commute with $\Phi^{q=1}_k (0) \propto {\bf M}_+$ and do not commute with $\Phi_-$ and $\Phi_+$. These modes are massless at $x=\pm \infty$ and have a non-zero mass at the origin. The corresponding potential is given by $$\begin{aligned}
\nonumber U^{q=1}(x) = &-& m^2+{7\over 12}(h+{2\lambda\over
5})\eta^2 F_-^2 + {5\over 12} \eta^2 h F_+^2 \\ &+&
(h+{\lambda\over 2})\eta^2 g^2 + \sqrt{7 \over 60} \eta^2 \lambda
F_- g \, . \label{q1potential}\end{aligned}$$ We have evaluated $U^{q=1}(x)$ numerically and found that it is everywhere negative for any choice of parameters.
As shown in Ref. [@PogVac00], the $q=2$ topological kink is perturbatively stable, at least for a range of parameters around the choice in eq. (\[hlambdaqeq2\]).
Next we discuss the perturbative stability of non-topological kinks.
The $q=0$ non-topological kink is simply the vacuum and is trivially stable.
We have seen that the $q=1$ non-topological kink solution may not exist for some parameter values. In other words, the $q=1$ configuration may split and become two $q=2$ topological kinks. When the $q=1$ non-topological kink does not split into two well-separated $q=2$ topological kinks, we find that it is locally stable. The potentially unstable modes are the two generators of $SU(5)$ that commute with $\Phi_k^{q=1NT}(0) \propto {\bf M}_-$ and do not commute with $\Phi_-$ and $\Phi_+$. The corresponding potential has a particularly simple form: $$U_{NT}^{q=1}(x) = {F_+'' \over F_+} \, . \label{q1ntpotential}$$ The plot of $U_{NT}^{q=1}(x)$ versus $x$ for $h/\lambda=-14/70$ is shown in Fig. \[fig:potq1nt\]. We have checked that the value of the potential at $x=0$ remains positive for all parameters for which the $q=1NT$ kink solution exists.
-0.25 truein
0.5 truecm = 0.95 0.5 truecm
The $q=2$ non-topological kink is perturbatively unstable for all parameter choices. The unstable modes are the eight Nambu-Goldstone modes for which the potential is given by the same expression as in eq. (\[q1ntpotential\]). Numerically we find that $U_{NT}^{q=2}(x) < 0$ for all $x$.
A general statement we can make is that the topological kinks in one of the classes will be globally stable. This just follows from the fact that the kinks are topological and so there must be a lowest energy kink. In the analysis done for the $SU(5)$ case in Sec. \[su5kinks\], the $q=n$ kink is the least energetic while the $q=0$ kink has the largest number of unstable modes. This suggests that perhaps the $q=n$ topological kink is the globally stable kink for any choice of potential and not just the quartic potential considered in this section. Another argument in support of this conjecture is that the change in the values of the field components in going from $x=-\infty$ to $+\infty$ is the least for the $q=n$ kink. Only one component need vanish inside the core of the $q=n$ kink while a greater number of components vanish inside the core for $q\le n-1$. The situation with the non-topological kinks is precisely the opposite. Here we know that the $q=0$ non-topological kink is the vacuum and hence is the least energy state.
$SU(5)$ magnetic monopoles {#su5mm}
==========================
A possible ansatz for a spherically symmetric $SU(5)$ fundamental magnetic monopole solution is [@DokTom80; @WilGol77]: $$\Phi_M \equiv \sum_{a=1}^{3} P(r) {\hat r}^a T^a + M(r) T^4 +
N(r) T^5 \ ,
\label{monopolesolution}$$ where the subscript $M$ denotes the monopole field configuration, $$\begin{aligned}
T^a = \frac{1}{2}{\rm diag}(0,0,\sigma^a,0)\ &,& \ \
T^4 = \frac{1}{2\sqrt{3}}(1,1,0,0,-2) \ , \nonumber \\
T^5 = {1\over {2\sqrt{15}}}&(&2,2,-3,-3,2) \ ,
\label{Tagenerators}\end{aligned}$$ $\sigma^a$ are the Pauli spin matrices, $r=\sqrt{x^2+y^2+z^2}$ is the spherical radial coordinate, and ${\hat r}^a$ denotes the unit radial vector. The ansatz for the gauge fields for the monopole can also be written down $$W^a_i = \epsilon^a_{~ij}
\frac{{\hat r}^j}{er}(1-K(r)) \ , \
(a=1,2,3) \ ,$$ $$W^b_i = 0 \ , \ \ (b \ne 1,2,3) \ , \label{gaugeansatz}$$ where $e$ is the gauge coupling. $P(r)$, $M(r)$, $N(r)$ and $K(r)$ are profile functions.
In the BPS case, when the $SU(5)$ potential vanishes, the exact, minimal energy solution is known [@Mec99]: $$P(r) = \frac{1}{er}(\frac{Cr}{\tanh(Cr)}-1) \ , \
K(r) = \frac{Cr}{\sinh(Cr)} \ ,
\label{bpsphiprofile}$$ $$M(r) = \frac{2}{\sqrt{3}}\frac{C}{e} \ , \ \
N(r) = {1\over {\sqrt{15}}}\frac{C}{e} \ .
\label{bconmandn}$$ where $C$ is a constant.
We can also write the monopole aysmptotic field configuration in more transparent form as $\Phi_M (r=\infty ) = U^\dagger_{34} \Phi_+ U_{34}$ where $$U_{34} (\theta , \phi )=
e^{-i\phi T^3} e^{-i\theta T^2} e^{+i\phi T^3} \ ,$$ $\theta$, $\phi$ are spherical angular coordinates and the generators $T^a$ are given in eq. (\[Tagenerators\]). Note that the winding of the monopole lies entirely in the $(3,4)$ block of $\Phi$. We are now – in contrast to the earlier sections – also choosing $$\Phi_+ = \eta {1\over \sqrt{60}} (2,2,2,-3,-3) \ .$$ Any other choice can be transformed to this choice by a global $SU(5)$ rotation.
The existence of the BPS solution does not preclude the existence of other higher energy magnetic monopole solutions even for fixed asymptotics since the boundary conditions at the origin can be chosen in different ways. (Ansatze with other asymptotics can be found in [@DokTom80].) One possible route to determining the different monopole boundary conditions at $r=0$ is to assume that the cores of magnetic monopoles are like the cores of domain walls. Then we would like to find the different spherical domain walls that have the asymptotics of the BPS solution. This will provide the spherical domain walls with monopole topology. If these spherical domain walls can shrink to zero size, the collapse will produce a monopole whose core is the same as that of the spherical domain wall that we started out with. In this way we might hope to determine the different possibilities for the boundary condition $\Phi_M (0)$.
We have three classes $q=0,1,2$ each of topological and non-topological walls. Let us consider each of these classes one by one.
The $q=0NT$ ($q=0$, non-topological) kink is trivial and we need not discuss it any further. The $q=0T$ ($q=0$, topological) kink has $$\Phi^{q=0}_k (x) = \tanh (\sigma x) \Phi_+ \ .$$ Using the kink solution, we can write down a field configuration corresponding to a spherical $q=0T$ domain wall: $$\Phi^{q=0T} (r, \theta ,\phi ) \approx \tanh (\sigma (r-R)) \Phi_+
\ ,$$ where $R$, the radius of the spherical domain wall, is taken to be very large. Next we would like to introduce monopole topology as a boundary condition to get an object that is a monopole in which all the energy resides in a shell made of a domain wall. (We call this object a “monopole-wall” (MW).) To do this we need to apply an $SU(5)$ rotation $U_{34}$ on $\Phi$. This will generally be ill-defined at the center ($r=0$) of the spherical domain wall since the field there will then become multi-valued. However, we are ultimately interested in letting the radius of the spherical domain wall go to zero and hence we need only apply the gauge transformation on $\Phi$ for $r\ge R$. Therefore $\Phi$ for the monopole-wall is: $$\Phi_{MW}^{q=0T} (r,\theta ,\phi) \approx \tanh (\sigma (r-R))
U^\dagger_{34} \Phi_+ U_{34} \ , \ \ \ r > R \ .$$ Note that the value of the field in the core of the wall is the same everywhere on the wall, that is, $\Phi_{MW}^{q=0} (R,\theta , \phi)=0$ regardless of the spherical angular coordinates. Therefore the monopole-wall can collapse to a point and the field will remain single-valued. The resulting monopole will have $\Phi_M (r=0) = 0$. That is, the new boundary conditions on $M(r)$ and $N(r)$ suggested by this argument are: $M(0)=0=N(0)$.
Next consider the $q=1NT$ kink. Here $\Phi^{q=1NT} (0) \propto (4,4,-1,-1,-6)$ in the core of the domain wall. Once again we may construct the monopole-wall by applying the transformation $U_{34}$. Since $$U^\dagger_{34} \Phi^{q=1NT} (0) U_{34} \propto \Phi (0) \ ,$$ the monopole-wall can collapse into a monopole. This suggests that we should be able to find a monopole solution with $\Phi_M (r=0) \propto (4,4,-1,-1,-6)$. This is precisely the monopole with boundary conditions given in eq. (\[bconmandn\]).
The $q=1T$ kink has $\Phi^{q=1T} (0) \propto (0,0,1,-1,0)$ and this is not invariant under rotations by $U_{34}$. Therefore once we impose monopole boundary conditions on a spherical domain wall of this type, the field in the core of the domain wall will depend on the angular coordinates. Such a wall cannot simply collapse to zero radius since that would violate single-valuedness of the field. Hence we do not expect to find a monopole whose center has $\Phi$ proportional to $(0,0,1,-1,0)$.
The $q=2T$ kink has $\Phi^{q=2T} (0) \propto (0,1,1,-1,-1)$ and, as this is not invariant under $U_{34}$, a monopole with $\Phi_M (0) \propto (0,1,1,-1,-1)$ is not possible.
The $q=2NT$ kink as described in Sec. \[su5kinks\] has $\Phi^{q=2NT} (0) \propto (-4,1,1,1,1)$ and this is invariant under $U_{34}$. This suggests that a monopole with $\Phi_M (0)
\propto (-4,1,1,1,1)$ is possible. However, this monopole-wall does not quite fit the form of the monopole solution given in eq. (\[monopolesolution\]). We find that if we choose $M(0) =
-\sqrt{5} N(0)$, the center of the monopole has $\Phi_M (0)
\propto (1,1,1,1,-4)$ and not $(-4,1,1,1,1)$. A global $SU(5)$ rotation on the monopole solution could be used to make $\Phi_M
(0) \propto (-4,1,1,1,1)$, however this would then rotate the asymptotic field to $\Phi_M (z=\infty ) \propto (-3,2,2,-3,2)$, once again providing a mismatch between the monopole-wall and the monopole ansatz in eq. (\[monopolesolution\]). In spite of this mismatch, the monopole-wall has the same topologically non-trivial asymptotic field configuration as the BPS solution and can also contract to a point without any conflict with single-valuedness. Hence we think that a monopole solution with $\Phi_M (0) \propto
(-4,1,1,1,1)$ should exist.
The above discussion, suggesting that there could be several monopole solutions corresponding to different boundary conditions on the scalar field at $r=0$, clearly applies to global monopoles. In the case of gauge monopoles, the only non-trivial gauge fields are the three fields associated with the $SU(2)$ group of the embedded monopole, as in the BPS case above. These fields still satisfy the form in eq. (\[gaugeansatz\]) and the only quantity that will depend on the “monopole generation” is the profile function $K(r)$.
This completes an analysis of all the cases. Three of the five non-trivial cases led to the possibility of a monopole solution. This suggests the existence of three classes of fundamental monopoles in $SU(5)$ with the same asymptotics as the BPS monopole.
Conclusions
===========
We have shown that the kink solutions in $SU(N)\times Z_2$ occur in $(N+1)/2$ classes. All the kink solutions, regardless of class, have the same topological charge. Borrowing the terminology of the standard model where particles come in “generations” (or “families”), we dub the kink classes “kink generations”. We have determined the continuous degeneracy associated with every kink generation. The degeneracy is described by certain manifolds which themselves have interesting topological properties. In particular, the manifolds have non-trivial second homotopy, suggesting that certain configurations of closed domain walls in three spatial dimensions may be incontractable.
We have also examined the stability of the various classes of kinks in an $SU(5)$ model with quartic potential. Our analysis shows that two classes of solutions are perturbatively stable (for some parameters) while the other non-trivial kinks are unstable.
The generation structure of domain walls suggests a generation structure for the magnetic monopoles in the gauged version of the model - a possibility that seems worth exploring further in the context of the dual standard model [@Vac96]. We have found that spherical domain walls of the $q=0T, 1NT, 2NT$ classes can collapse into monopoles that al have the same asymptotic field configurations. Hence monopole solutions with $\Phi_M (0)=0$ and $\Phi_M(0) \propto (-4,1,1,1,1)$ should be possible to construct in addition to the known case where $\Phi_M (0) \propto
(4,4,-1,-1,-6)$. If all these different boundary conditions lead to magnetic monopole solutions and there are none others[^1], it would indicate that there are exactly three generations of $SU(5)$ magnetic monopole solutions. To confirm this statement would require an explicit construction of the $SU(5)$ monopole solutions with the various possible boundary conditions.
We anticipate that a survey of the space of $SU(N)$ magnetic monopole solutions will show novel features, similar to those we have discovered in the case of kinks.
Conversations with Gautam Mandal and Spenta Wadia are gratefully acknowledged. This work was supported by the DoE.
L. Pogosian and T. Vachaspati, Phys. Rev. [**D62**]{}, 123506 (2000).
T. Vachaspati, Phys. Rev. [**D63**]{}, 105010 (2001).
C. P. Dokos and T. N. Tomaras Phys. Rev. [**D21**]{}, 2940 (1980)
D. Wilkinson and A. Goldhaber, Phys. Rev. [**D16**]{}, 1221 (1977).
M. Meckes, private communication (1999).
T. Vachaspati, Phys. Rev. Lett. [**76**]{}, 188 (1996).
Proof that solutions require $[\Phi_+,\Phi_-]=0$ {#appendixa}
================================================
Let $\Phi_k (x)$ be a kink solution. We can expand the solution in an orthonormal set of $SU(N)$ generators $T^a$ (${\rm Tr} (T_a
T_b) = \delta_{ab}/2$): $$\Phi_k (x) = \sum_a \phi_a (x) T^a \ . \label{Phi0expansion}$$ Here an alternate expansion will be more convenient: $$\Phi_k (x) = \sum_a \psi_a (x) R^a \ ,
\label{Phi0alternateexpansion}$$ where $$R^1 \equiv {1\over \eta} \Phi_- \equiv R_- \ , \ \ \ R^2 \equiv
{1\over \eta} \Phi_+ \equiv R_+ \ , \label{R1R2definition}$$ where $\eta$ is a normalization factor so that ${\rm Tr} (R_\pm
^2) = 1/2$ and the remaining $R^a$ complete the set of generators. Depending on the boundary conditions, it may well turn out that ${\rm Tr}(R_+ R_- ) \ne 0$ and so these generators are not orthogonal. However, we shall choose the other generators, [*i.e.*]{} $R^a$ with $a \ne 1,2$, to satisfy the orthogonality conditions ${\rm Tr} (R_+ R^a) = 0 = {\rm Tr} (R_- R^a)$ and also normalize them to satisfy ${\rm Tr} (R_a R^a)=1/2$. We define new structure constants $r_{abc}$ by $$[ R^a , R^b ] = i r_{abc} R^c \ . \label{rstructureconstants}$$
Next we need to state certain properties of the functions $\psi_a
(x)$. Due to the boundary conditions $\Phi (x\rightarrow \pm
\infty ) \rightarrow \Phi_\pm$, we have $$\psi_1 (-\infty ) = \eta \ , \ \ \psi_a (-\infty ) = 0 \ (a \ne 1)
\ , \label{minusinfinity}$$ $$\psi_2 (+\infty ) = \eta \ , \ \ \psi_a (+\infty ) = 0 \ (a \ne 2)
\ . \label{plusinfinity}$$ (Just as for the generators, $\psi_- \equiv \psi_1$ and $\psi_+
\equiv \psi_2$.) These boundary conditions ensure that there is no non-trivial solution of the kind $\psi_a (x) = {\rm constant}$.
Let us now perturb the kink solution $\Phi_k (x)$. For this, consider the field configuration $$\Phi_1 (x) = U(x) \Phi_k U^\dagger (x) \label{Phi1} \ ,$$ where $U(x) \in SU(N)$. Note that $V(\Phi_1 )=V(\Phi_k)$ since the potential is invariant under $SU(N)$ local gauge transformations. Then the energy of the configuration $\Phi_1$ is: $$\begin{aligned}
E[\Phi_1] = E[\Phi_k ] +
2 {\rm Tr}( \partial_x \Phi_k [&&U^\dagger \partial_x U, \Phi_k ]
) \nonumber \\
&& + {\rm Tr} ( [U^\dagger \partial_x U ,\Phi_k ]^2 ) \ .
\label{energy}\end{aligned}$$ If we now consider infinitesimal rotations, the second term is linear in these while the last term is quadratic. If $\Phi_k$ is to be a solution, the linear variation must vanish. Therefore, $${\rm Tr}( \partial_x \Phi_k [U^\dagger \partial_x U, \Phi_k ] ) =0
\label{solutioncondition}$$ for all $U(x)$ infinitesimally close to unity and for all $x$.
The condition in eq. (\[solutioncondition\]) can also be rewritten as: $${\rm Tr}( [\Phi_k , \partial_x \Phi_k ] U^\dagger \partial_x U )
=0 \ , \label{conditionrewrite}$$ which should hold for any $U(x) \in SU(N)$. (For infinitesimal rotations this condition is $${\rm Tr}( [\Phi_k , \partial_x \Phi_k ] T^a ) =0 \ , \ \forall x,
a \ , \label{smallconditionrewrite}$$ where $T^a$ form a complete set of $SU(N)$ generators.) Hence the solution must necessarily satisfy $$[\Phi_k , \partial_x \Phi_k ] = 0 \label{commcondition}$$ for all $x$.
Next use the expansion of $\Phi_k$ of eq. (\[Phi0alternateexpansion\]) in eq. (\[commcondition\]) and that gives us: $$\sum_{b>a} r_{abc}
[\psi_a (x) \psi_b '(x) - \psi_b (x) \psi_a ' (x) ] = 0 \ ,
\ \forall c, x \ .\label{condition2}$$
If the functions $$F_{ab} \equiv \psi_a (x) \psi_b '(x) - \psi_b (x) \psi_a ' (x)$$ are linearly independent, eq. (\[condition2\]) implies that $r_{abc} =0$ whenever $F_{ab}\ne 0$. It is easy to see that $F_{ab}\ne 0$ provided both $\psi_a$ and $\psi_b$ are non-trivial and linearly independent. Hence the (assumed) linear independence of $F_{ab}$ implies that $r_{abc}=0$ whenever $\psi_a$ and $\psi_b$ are non-trivial and linearly independent. It is sufficient to assume that all the $\psi_a$ are linearly independent since if two components are linearly dependent, the basis of generators, $R^a$, can be redefined so that only linearly independent functions occur in the expansion in eq. (\[Phi0alternateexpansion\]). This shows that if $F_{ab}$ are linearly independent then $[R^a, R^b] =0$ if $\psi_a$ and $\psi_b$ are non-trivial. Therefore the solution $\Phi_k$ can be expanded in a Cartan basis and in particular $[\Phi_+, \Phi_-]=0$.
Without assuming the linear independence of the functions $F_{ab}$, we can still show the desired result $[\Phi_\pm ,
\Phi_k]=0$ by examining the condition in eq. (\[condition2\]) as $x\rightarrow +\infty$. In this spatial region, the only non-vanishing function is $\psi_+ (x) \rightarrow \eta$. The term $\psi_+ \psi_a '$ is small because all derivatives vanish at infinity. The terms $\psi_a \psi_b'$ with $\psi_a \ne \psi_+$ are also small since both $\psi_a$ and $\psi_b '$ tend to zero at $x=+\infty$. In this region, where the field is nearly at its vacuum value, we can examine the behavior of the fields by perturbing the potential around the vacuum. This tells us that $\psi_a$ ($a\ne +$) falls off exponentially as $x \rightarrow
\infty$. Therefore, $$\psi_+ \psi_a ' >> \psi_a \psi_b'$$ for all $a\ne +$. So the condition in eq. (\[condition2\]) in the large, positive $x$ region yields $$\sum_{b\ne 2} r_{2bc} \psi_b '(x) = 0 \ . \label{condition3}$$ An integration over the interval $(x,+\infty )$ then gives $$\sum_{b\ne 2} r_{2bc} \psi_b (x) = 0 \ , \label{condition4}$$ where we have used the boundary conditions $\psi_b (+\infty )=0$ except for $b=2$ (which does not appear in the sum). As discussed above, it is sufficient to consider the case when the set of functions $\psi_b (x)$ are linearly independent. Therefore, if $\psi_b$ is non-trivial, we get $$r_{2bc} = 0 \ , \ \ \ \forall b , c \ . \label{r2b1}$$ Similarly, by considering the region with $x \rightarrow -\infty$, $$r_{1bc} = 0 \ , \ \ \ \forall b , c \ . \label{r1b2}$$ This shows that $[R_+ , R^a]=0=[R_- , R^a]$ if $\psi_a \ne 0$ for any choice of $a$ and hence $[\Phi_\pm , \Phi_k (x)]=0$. In particular, we can only get a kink solution if $[R_+,R_-]=0$ which is equivalent to $[\Phi_+,\Phi_-]=0$.
Equations of motion for the $\lowercase{q}=1$ kink in $SU(5)$ {#appendixb}
=============================================================
The equations of motion for the topological $q=1$ kink functions $F_\pm$ and $g$ are: $$\begin{aligned}
-F_+''+ \alpha_1 F_+ + \alpha_2 F_+^3 &+& \alpha_3 F_+ F_-^2
+\alpha_4 g^2 F_+
\nonumber \\
&+&\alpha_5 gF_- F_+=0 \ ,
\label{F+eq}\end{aligned}$$ $$\begin{aligned}
-F_-''+ \beta_1 F_- &+& \beta_2 F_-^3 + \beta_3 F_+^2 F_-
+ \beta_4 g^2 F_-
\nonumber \\
&+&\beta_5 g(3F_-^2-F_+^2) +\beta_6 g^3 =0 \ ,
\label{F-eq}\end{aligned}$$ $$\begin{aligned}
-g''+\gamma_1 g &+&\gamma_2 g^3 +\gamma_3 gF_+^2
+ \gamma_4 gF_-^2
\nonumber \\
&+& \gamma_5 g^2F_- + \gamma_6 F_-(F_-^2-F_+^2) =0 \ ,
\label{geq}\end{aligned}$$ where $$\begin{aligned}
\alpha_1 &=& \beta_1 = \gamma_1 = -m^2 \, ,
\nonumber \\
\alpha_2 &=& \eta^2 {5\over 12} (h + {1 \over 2}\lambda) \, ,
\nonumber \\
\alpha_3 &=& \eta^2 {7\over 12} (h + {3 \over 70}\lambda) \, ,
\nonumber \\
\alpha_4 &=& \eta^2 (h + {6 \over 7}\lambda) \, ,
\nonumber \\
\alpha_5 &=& - \eta^2 \lambda \sqrt{3 \over 35} \, ,
\nonumber \\
\beta_2 &=& \eta^2 {7\over 12} (h + {181 \over 490}\lambda) \, ,
\nonumber \\
\beta_3 &=& \eta^2 {5\over 12} (h + {3 \over 70}\lambda) \, ,
\nonumber \\
\beta_4 &=& \eta^2 (h + {138 \over 245}\lambda) \, ,
\nonumber \\
\beta_5 &=& \eta^2 \lambda {5 \over 14} \sqrt{3 \over 35}\, ,
\nonumber \\
\beta_6 &=& \eta^2 \lambda {12 \over 49} \sqrt{3 \over 35}\, ,
\nonumber \\
\gamma_2 &=& \eta^2 (h + {25 \over 98}\lambda) \, ,
\nonumber \\
\gamma_3 &=& \eta^2 {5 \over 12} (h + {6 \over 7}\lambda) \, ,
\nonumber \\
\gamma_4 &=& \eta^2 {7 \over 12}(h + {138 \over 245}\lambda) \, ,
\nonumber \\
\gamma_5 &=& \eta^2 \lambda {3 \over 7} \sqrt{3 \over 35}\, ,
\nonumber \\
\gamma_6 &=& \eta^2 \lambda {5 \over 24} \sqrt{3 \over 35}\, .\end{aligned}$$
[^1]: While the only $SU(5)$ monopole solution known to us is the BPS solution, an exhaustive list of spherically symmetric ansatze consistent with monopole topology is given in [@DokTom80]. Some of these could possibly lead to other monopole solutions with different asymptotic field configurations.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is shown that the order property of pure bipartite states under SLOCC (stochastic local operations and classical communications) changes radically when dimensionality shifts from finite to infinite. In contrast to finite dimensional systems where there is no pure incomparable state, the existence of infinitely many mutually SLOCC incomparable states is shown for infinite dimensional systems even under the bounded energy and finite information exchange condition. These results show that the effect of the infinite dimensionality of Hilbert space, the “infinite workspace” property, remains even in physically relevant infinite dimensional systems.'
address: |
$^1$[*Department of Physics, The University of Tokyo, Tokyo 113-0033, Japan*]{}\
$^2$[*Quantum Computation Group, The National Institute of Information, Tokyo 101-8430, Japan*]{}\
$^3$[*Imai Quantum Computation and Information Project, JST, Tokyo 113-0033, Japan*]{}\
$^4$[*PRESTO, JST, Kawaguchi, Saitama 332-0012, Japan*]{}
author:
- 'Masaki Owari$^{1,3}$, Keiji Matsumoto$^{2,3}$ and Mio Murao$^{1,4}$'
title: Entanglement convertibility for infinite dimensional pure bipartite states
---
Quantum information encoded in quantum systems offers possibility to provide a new and outstanding class of the information processing. Quantum two-level systems (qubit systems), which correspond to the binary ‘bit’ systems of classical information processing, are often used for quantum information processing, but quantum information processing can also be performed in finite multi-level systems and even infinite dimensional systems, such as bosonic systems.
Infinite dimensional systems, sometimes called continuous variable systems, have been expected to offer high potential for quantum information processing. One of the advantages of infinite dimensional systems is the possibility of implementation using quantum optical systems, as shown by the successful demonstration of teleportation [@teleportation]. Properties of quantum information under gaussian operations, which can be implemented by linear optical systems, have been investigated [@gaussian; @gaussian; @distill]. Another advantage is in its infinite dimensionality of Hilbert space. Since finite multi-level systems can be reduced to (finitely) many two-level systems, the essential power of quantum information processing is not different from qubit systems, although there may be advantages of tractability in implementation. However, infinite dimensional systems cannot be reduced to two-level systems in general. Therefore, there is the possibility of (yet unknown) new types of quantum information processing schemes, which do not exist in finite dimensional systems.
It has been widely believed that the fundamental properties of finite and infinite dimensional systems are similar or, at least, that in physically relevant infinite dimensional systems with bounded energy conditions and with a finite number of measurement outcomes, the properties can be well approximated to finite dimensional systems. In this letter, we will show that these beliefs are not well-founded for quantum systems by investigating entanglement properties of infinite dimensional systems.
Entanglement is regarded as the key resource which allows many quantum information processing schemes out-perform than their classical counterparts. Entanglement can be classified by transformability under local operations, such as LOCC (deterministic local operations and classical communications), SLOCC (stochastic LOCC). The convertibility properties of two different entangled states (in a single copy or multi-copy situation) under local operations are important for the qualitative and quantitative understanding of entanglement. For finite dimensional bipartite systems, we now have a better understanding of LOCC and SLOCC convertibility based on intensive work in recent years. For example, the condition for the convertibility of two pure entangled states in the single copy situation is given by Nielsen’s majorization theorem [@nielsen] for LOCC, and is given by Vidal’s theorem [@vidal] for SLOCC. It is also found that incomparable pure states only exist in multipartite systems, such as the GHZ state $({{\left \vert}000 {\right \rangle}}+{{\left \vert}111 {\right \rangle}})/\sqrt{2}$ and the W state $({{\left \vert}001 {\right \rangle}}+{{\left \vert}010 {\right \rangle}}+{{\left \vert}100 {\right \rangle}})/\sqrt{3}$ in three qubit systems [@dur].
In this letter, we will show that a fundamental property of entanglement, the order properties of pure bipartite states under SLOCC are changed significantly from total order to non-total order with the shift in dimensionality from finite to infinite. Further, we will show that there are infinitely many mutually SLOCC incomparable pure bipartite states even under the bounded energy condition. These results show that the effect of infinite dimensionality of Hilbert space, the “infinite workspace” property, remains in physically relevant infinite dimensional systems.
First, we consider ordering sets, ordered by the convertibility of two general bipartite quantum states under general operations. In the language of set theory [@settheory], convertibility can be describe by an order denoted by “$\rightarrow$”. For a set of physical states $S$, the order indicating the existence of physical transformation satisfies the reflective law $a
\rightarrow a$ and the transitive law $a \rightarrow b$ and $b
\rightarrow c$ imply $a \rightarrow c$), where $a, b, c \in S$. This ordering property is a pseudo partial ordering. We consider that two states are in the same equivalence class, if they transform each other $a \rightarrow b$ and $b \rightarrow a$. We denote this situation as $a \leftrightarrow b$. The quotient set of $S$ by the equivalence class $\leftrightarrow$ is denoted as $(S/\leftrightarrow , \rightarrow)$ and represents the classification based on the given transformation. For the set $(S/\leftrightarrow , \rightarrow)$, we can redefine the ordering $\rightarrow$ which satisfies the additional condition $a
\rightarrow b$ and $b \rightarrow a$ imply $a = b$. This ordering property is partial ordering.
If the set has the additional property that $a \nrightarrow b$ implies $b \rightarrow a$ ($a \nrightarrow b$ denotes $a
\rightarrow b$ is not true), the set is totally ordered. Total ordering is a convenient property to analyze the convertibility of a system, since there exists a unique measure of ordering for a totally ordered set. For example, the convertibility of pure state ${{\left \vert}\phi {\right \rangle}}$ under LOCC transformation in the asymptotic situation is total ordering. Therefore, there is a unique measure, or a monotone, given by the von Neumann entropy of entanglement $E({{\left \vert}\phi {\right \rangle}})$ in this case [@thermodynamics]. On the other hand, a pair of monotones, like distillable entanglement and entanglement cost [@bennett], are useful tools to distinguish the ordering properties (total ordering or partial ordering) of the system [@morikoshi].
The key idea of our formulation is that we consider a totally ordered subset $\{ \xi _r \}$ parameterized by a real number $r$ for a pseudo partial ordering set $S$. Then we can always define a pair of functions $R^-(\psi)$ and $R^+(\psi)$ for a state $\psi
\in S$, where $R^-(\psi)$ is the supremum of $r$ at which $\psi$ can be transformed to $\xi _r$, and $R^+(\psi)$ is the infimum of $r$ at which $\Psi$ can be transformed to $\xi _r$. Mathematically they are expressed as the following: For a pseudo partial ordering set, if there exists a real parameterized total ordering subset $\{ \xi _r \} _{r \in A} \subset S$ where $A \subset \mathbb{R}$ such that $r_1 \le r_2$ if and only if $\xi _{r_1} \rightarrow \xi
_{r_2}$, we can define a pair of functions on $S$ to $\overline{A}
\subset \mathbb{R}$ as $$\begin{aligned}
R^-(\psi) &=& \sup \{ r \in A | \psi \rightarrow \xi _r \}
\label{rminus}
\\
R^+(\psi) &=& \inf \{ r \in A | \xi _r \rightarrow \psi \}
\label{rplus}\end{aligned}$$ where we define $R^-({{\left \vert}\psi {\right \rangle}}) = \inf \{ A \}$ for $\{ r\ \in A
| \psi \rightarrow \xi _r \} =\emptyset $, and $R^+({{\left \vert}\psi {\right \rangle}}) =
\sup \{ A \}$ for $\{ r\ \in A | \xi _r \rightarrow \psi \}
=\emptyset $.
Although there are many ways to define a monotone for a partial ordering set from a totally ordered subset $\{ \xi _r \} _A$, our definition of functions $R^-(\psi)$ and $R^+(\psi)$ are preferable for analyzing entanglement convertibility. It is easily proven by contradiction that they are the [*unique monotones*]{} which give lower and upper bounds of any monotones defined from $\{ \xi _r
\}$ for a given pseudo partial ordered set $S$, that is, $\psi
\rightarrow \phi$ implies $R^-(\psi) \ge R^-(\phi)$ and $R^+(\psi)
\ge R^+(\phi)$, $R^+(\phi) < R^-(\psi)$ implies $\psi \rightarrow
\phi$, and also any other monotone $R_0 (\psi)$ defined from $\{
\xi _r \}$ satisfies $R^- (\psi) \le R_0 (\psi) \le R^+ (\psi)$ for all $\psi \in S$.
From the properties of $R^-(\psi)$ and $R^+(\psi)$, we can immediately derive the following important results: If the quotient set $(S/\leftrightarrow , \rightarrow)$ is totally ordered, namely, $\psi \nrightarrow \phi$ implies $\phi
\rightarrow \psi$ is satisfied, then for all $\psi \in S$, we have $R^-(\psi) = R^+(\psi)$. On the other hand, if there exists $\psi
\in S$ such that $R^-(\psi) < R^+(\psi)$, then $(S/\leftrightarrow
, \rightarrow)$ is not totally ordered, and $\psi$ is incomparable to all $\xi _r$ with $R^-(\psi) < r < R^+(\psi)$.
The advantage of our formulation is that it can be applied for many different situations: single-copy or asymptotic (infinitely many-copy) cases, for finite or infinite dimensional systems, mixed or pure states, and under LOCC, SLOCC or PPT [@ppt] operations. Many important known results of entanglement theory can be re-derived only from simple ordering properties and the existence of the real parameterized total ordering subset. For example, in the case of convertibility under LOCC operations for mixed states in the asymptotic situation, we consider $\{ {{\left \vert}\xi
_s {\right \rangle}} {{\left \langle}\xi _s {\right \vert}} \} _{s=0}^{\infty}$ to be a subset of pure states with $E({{\left \vert}\xi _s {\right \rangle}}) =s$ [@bennett] where $E({{\left \vert}\psi {\right \rangle}})$ is the amount of entanglement for pure states. Then we have $R^-(\rho)= E_d(\rho)$ and $R^+(\rho)=E_c(\rho)$ where $E_d(\rho)$ and $E_c(\rho)$ are distillable entanglement and entanglement cost, respectively. We see that the sets are not totally ordered and $R^-$ (distillable entanglement) and $R^+$ (entanglement cost) are limits of other monotones [@uniqueness], and also that there is a set of states such that $R^-=0$ but $R^+> 0$, the bound entangled states [@boundentangle].
Now we concentrate on the investigation of SLOCC convertibility (with non-zero probability) of infinite dimensional pure states in the single-copy situation. In general, the entanglement of a pure state ${{\left \vert}a {\right \rangle}}$ is characterized by the sequence of Schmidt coefficients $\{ \lambda^a_I \}$ ($0 \leq I \leq d$ and $0 \leq I
\leq \infty$ for finite $d$ and infinite dimensional systems, respectively). For finite dimensional systems, the two monotones coincide with the Schmidt rank, the number of non-zero Schmidt coefficients (Vidal’s theorem [@vidal]). On the other hand, the Schmidt rank itself cannot be a monotone for infinite dimensional systems for analyzing convertibility between “genuine” infinite dimensional states (with infinite Schmidt ranks). Further, if we consider exact conversion of genuine infinite dimensional states under SLOCC, we generally need infinite many information exchanges, which is not physical.
To extend Vidal’s theorem to physical infinite dimensional systems, we define SLOCC convertibility in infinite dimensional systems as the following: ${{\left \vert}\psi {\right \rangle}}$ is convertible to ${{\left \vert}\phi {\right \rangle}}$ by SLOCC if and only if ${{\left \vert}\psi {\right \rangle}}$ is convertible to any neighborhood of ${{\left \vert}\psi {\right \rangle}}$ by finite steps of LOCC with probability larger than some positive number. Then Vidal’s theorem can be extended to infinite dimensional systems as the following: ${{\left \vert}\psi {\right \rangle}} \in {\mathcal{H}}$ can be converted to ${{\left \vert}\phi {\right \rangle}}
\in {\mathcal{H}}$ by SLOCC with non-zero probability in the single-copy situation if and only if there exists $\epsilon > 0$, $g_\psi (n)
/ g_\phi (n) \ge \epsilon$ for all $n \in N$, where $g_a (n) =
\sum _{I=n}^{\infty} \lambda^a_I$ is a function defined in terms of Schmidt coefficients $\{ \lambda^a_I \}_{I=0}^\infty$ of a genuine infinite dimensional state ${{\left \vert}a {\right \rangle}}$ [**(Vidal’s Theorem)**]{}. Mathematically rigorous arguments for the extended Vidal’s theorem are given in [@owari-locc] together with an extension of Nielsen’s theorem to infinite dimensional systems.
The function $g_a (n)$ plays the central role in the construction of the monotones $R^-$ and $R^+$ for infinite dimensional states. By definition, a sequence of the function $\{g_a (n)\}_{n \in N}$ satisfies four conditions, strict positivity $g_a (n)>0$, strict monotonicity $g_a(n) > g_a(n+1)$, convexity $g_a (n+1) \le
\{g_a(n) + g_a (n+2)\}/2$, and normalization $g_a (0)=1$. Conversely, for a given sequence of functions $\{ g(n)
\}_{n=0}^{\infty}$, there exist a genuine infinite dimensional state ${{\left \vert}a {\right \rangle}}$, where the Schmidt coefficients are give by $\lambda^a_I =g(n) -g(n+1)$ if and only if $\{g(n)\}_{n \in N}$ satisfies the strict positivity, strict monotonicity, convexity and normalization conditions. Such a function $g(n)$ is called Vidal’s monotone.
According to Vidal’s theorem, if a real parameterized subset $\{
{{\left \vert}\xi _r {\right \rangle}} \} _{r \in A } \subset {\mathcal{H}}$ is totally ordered, $g_{\xi_r}(n)$ must satisfies $\underline{\lim} _{n \rightarrow
\infty} (g_{\xi_{r_1}} (n) / g_{\xi_{r_2}}(n)) > 0$ if and only if $r_1 \le r_2$ for all $r_1$ and $r_2$. From the property of $g_{\xi_r}(n)$, we can construct the monotones $R^-$ and $R^+$ $$\begin{aligned}
R^-(\psi) &=& \inf \{ r \in A | \underline{\lim} _{n
\rightarrow \infty}
g_\psi (n)/ g_{\xi_r}(n) = 0 \}
\label{R^-} \\
R^+(\psi) &=& \inf \{ r \in A | \overline{\lim} _{n
\rightarrow \infty}
g_\psi (n)/ g_{\xi_r}(n) < + \infty
\}. \label{R^+}\end{aligned}$$ for all $\{ g_{\xi_r}(n) \} _{r \in A, n \in N}$ satisfying the three conditions: I. Strict monotonicity for all $r \in A$, II. Convexity for all $r \in A$ and $n \in N$, and III. $\underline{\lim} _{n \rightarrow \infty} (g_{\xi_{r_1}(n)} /
g_{\xi_{r_2}(n)}) > 0$ is equivalent to $r_1 \ge r_2$. The proof is given by the following: If $\{ g_{\xi_r}(n) \} _{r \in A, n
\in N}$ satisfies the conditions I, II, and III, the corresponding set of states $\{ {{\left \vert}\xi_r {\right \rangle}} \} _{r \in A}$ for $\{ g_{\xi_r}(n)
\} _{r \in A, n \in N}$ exists and is totally ordered. Since $A
\in \mathbb{R}$ is assumed to be an interval, the two functions $R^-(\psi)$ and $R^+(\psi)$ defined by Eqs. (\[rminus\]) and (\[rplus\]), can be represented by Eqs. (\[R\^-\]) and (\[R\^+\]) by using Vidal’s Theorem.
Next, we show that there exists a pairs of genuine infinite dimensional states which are incomparable to each other. We prove that the two monotones $R^-({{\left \vert}\psi {\right \rangle}})$ and $R^+(\psi)$ given by Eqs. (\[R\^-\]) and (\[R\^+\]) do not necessarily coincide with each other for infinite dimensional systems, by constructing an example. We consider a twice continuously differentiable function $g(x)$, which is the continuous counterpart of $g(n)$, since a continuous function is more convenient for analytical investigation. The conditions for $g(x)$ to relate to a genuine infinite dimensional state is now given by $g(x)>0$ (strict positivity), $g^{'}(x) <0$ (strict monotonicity), $g^{''} (x) \ge
0$ (convexity), and $g(0) =1$ (normalization), for all $x$. If $g(x)$ satisfies the above conditions except the normalization condition, we can easily normalize $g(x)$. Thus we omit the normalization condition for simplicity. Since convertibility is determined only by the ratio of functions, we introduce another function $d(x)$, which is given by $d(x)=p(x)g(x)$. Let $g(x)$ satisfy the same conditions as $g(x)$. Then $p(x)> 0$, $g^{'}(x)
p(x) + g(x)p^{'} (x)<0$, $g^{''} (x) p(x) + 2g^{'}(x)p^{'}(x)+g(x)
p^{''}(x) \ge 0$ and $p(1) = 1$ are to be satisfied.
We set our function to be $g(x)= e^{-x}$. Actually, our choice of $g(x)$ represents one of the most tractable genuine infinite dimensional entangled states, the two mode squeezed state ${{\left \vert}\psi_q {\right \rangle}} =\frac{1}{c_q} \sum _{n=0}^{\infty} q^n {{\left \vert}n {\right \rangle}}
\otimes {{\left \vert}n {\right \rangle}}$, where $q$ is a squeezing parameter. We give a construction of a function $d(x)$ which indicates the existence of incomparable genuine infinite dimensional states. In this case, the conditions for $p(x)$ become simple, $p(x) > 0 $, $p(x) -
p^{'}(x)>0$, and $p(x) -2p^{'}(x) +p^{''}(x) \ge0$. We choose $p(x)$ to be parameterized by $r$ as $p(x)=p_{r}(x) = (\log
x)^{r}\{ \sin (\log x)+1\} +({\log x})^{-1}$ where $0 < r <
+\infty$. We define two functions $m_r(x) \equiv p_r(x) -
p_r^{'}(x)$ and $c_r(x)\equiv p_r(x) -2p_r^{'}(x) +p_r^{''} (x)$ for evaluating monotonicity and convexity, respectively.
For all $0 < r_1 < r_2 < \infty$, there exists $x_{r_1, r_2}
> 0$ such that $m_r(x)>0$ and $c_r(x) \ge 0$ for all $x \ge x_{r_1, r_2}$, and $r \in [r_1, r_2]$. That is, the function $p_r(x+x_{r_1,
r_2})$ satisfies the positivity, monotonicity and convexity conditions. Therefore we can consider a state ${{\left \vert}\xi_r {\right \rangle}}$ represented by the function $d_r(x)=p_r(x+x_{r_1, r_2}) g(x)$. The ratio of the functions $d_r(x)/g(x)=p_r(x+x_{r_1, r_2})$ determines convertibility between the two states ${{\left \vert}\psi_q {\right \rangle}}$ and ${{\left \vert}\xi_r {\right \rangle}}$. To evaluate the ratio, we rewrite $p_r(x+x_{r_1, r_2})$ in the discrete form: $p_r(n')=p_r(\Delta n+
x_{r_1, r_2})$ where $\Delta=-\log q$. Then we can easily show that $\underline{\lim} _{n \rightarrow \infty} p_r (n) = 0$ and $\overline{\lim} _{n \rightarrow \infty} p_r (n) = \infty$. Defining $R^-$ and $R^+$ from $\{ {{\left \vert}\xi_r {\right \rangle}} \}_{r \in
(r_1,r_2)}$, we obtain $R^-(\psi) = r_1$ and $R^+(\psi) = r_2$. The two states ${{\left \vert}\psi {\right \rangle}}$ and ${{\left \vert}\xi_r {\right \rangle}}$ for all $r \in
[r_1,r_2]$ are now shown to be incomparable under SLOCC.
Further, we show that there are [*infinitely many*]{} mutually incomparable states by proving that if $\{ {{\left \vert}\Psi _k {\right \rangle}} \}
^{\infty}_{k=0}$ is defined by the following sequence $\{ d_k
(\Delta n) \} ^{\infty }_{k=0}$ and $d_k(x) = C_k e^{-x} \{p
(x+a_k) \}^k$, where $C_k$ is a normalization constant preserving $d_k(0) =1$ and $a_k >0$ is a real number, then ${{\left \vert}\Psi _k {\right \rangle}}$ is SLOCC incomparable each other for all $k \in N$. It is easy to prove that for all $k \in N$, there exists a real number $a_k$ such that for all $\{ d_k (\Delta n) \} ^{\infty }_{k=0}$ satisfies the conditions for Vidal’s monotone (strict positivity, strict monotonicity, convexity and normalization) for all $x \ge
a_k$.
To prove the mutual incomparability of ${{\left \vert}\Psi _k {\right \rangle}}$ for different $k$s, it is sufficient to show $\overline{\lim} _{n
\rightarrow \infty } d_k (\Delta n ) / d_l(\Delta n) = + \infty$ and $\underline{\lim} _{n \rightarrow \infty } d_k (\Delta n ) /
d_l(\Delta n) = 0$ for all $k > l$. The first condition is evaluated by $$\begin{aligned}
& \quad & \overline{\lim} _{x \rightarrow \infty}
d_k(x)/d_l(x) \\
& \ge & \overline{\lim} _{p \rightarrow \infty}
\frac{ \{ 2 h(p,k)+\frac{1}{h(p,k)} \}^k }
{[ h(p,l) \{ \sin (h(p,l)) + 1 \}
+ \frac{1}{h(p,l)} ]^l } \\
& \ge & \overline{\lim} _{p \rightarrow \infty }
\frac{2^n h(p,k)^n}
{[ h(p,l) \{ \sin (h(p,l)) + 1 \} + \frac{1}{h(p,l)} ]^l} \\
& \ge & \overline{\lim} _{p \rightarrow \infty }
\frac{2^n h(p,k)^n}
{\{ 2 h(p,l) +\frac{1}{h(p,l)} \}^l} \\
& = & + \infty,\end{aligned}$$ where $h(p,k)=\log(x_p + a_k)$ and $\{ x_p \} _{p=1}^{\infty}$ is a sequence of positive real values such that $\sin (\log x_p) = 1$ and $x_p < x_{p+1}$. The second condition is evaluated by $$\begin{aligned}
& \quad & \underline{\lim} _{x \rightarrow \infty}
d_k(x)/d_l(x)\\
& \le & \underline{\lim} _{q \rightarrow \infty}
\frac{{\{ h(q,k) \}^{-k}}}
{ [h(q,l) \{ \sin (h(q,l) +1 \}+
\frac{1}{h(q,l)}]^l} \\
& \le & \underline{\lim} _{q \rightarrow \infty}
\frac{\{ 1/h(q,k)\}^k}{\{ 1/h(q,l)\}^l }\\
& \le & 0 ,\end{aligned}$$ where $\{ x_q \}_{q=1}^{\infty}$ is a positive sequence such that $\sin (\log x_q) =0$ and $x_q < x_{q+1}$. Therefore, it is proven that ${{\left \vert}\Psi _k {\right \rangle}}$ is mutually SLOCC incomparable for all $k \in
N$ using Vidal’s theorem.
Since ${{\left \vert}\Psi_0 {\right \rangle}}$ represent a two mode squeezed state, any ${{\left \vert}\Psi _k {\right \rangle}}$ with $k>0$ is incomparable to not only the two mode squeezed state, but also to all ${{\left \vert}\Psi _{k^{'}} {\right \rangle}}$ with $k^{'} \neq k$. These family of states can be taken to have all bounded energy due to the existence of the exponential term $e^{-x}$ in the Vidal’s monotones. In physically relevant situations, we should restrict measurement obtaining only finitely many results for local operations. Shown incomparability is under stronger local operations including infinitely many measurement results, the measurement described by the countable infinite positive operator valued measure. Therefore, the incomparability property of infinite dimensional states remains under the finite measurement condition.
We have to note that our functions $R^-({{\left \vert}\psi {\right \rangle}})$ and $R^+({{\left \vert}\psi {\right \rangle}})$ are discontinuous for the usual topology of Hilbert space due to the discontinuity of the SLOCC convertibility itself. However, we can say that the maximum probability to convert ${{\left \vert}\psi ' {\right \rangle}}$ where $\| {{\left \vert}\psi {\right \rangle}} - {{\left \vert}\psi ' {\right \rangle}} \| <
\epsilon $ for small $\epsilon $ to ${{\left \vert}\phi {\right \rangle}}$ and the probability of the inverse process are both very small, if both ${{\left \vert}\psi {\right \rangle}}$ and ${{\left \vert}\phi {\right \rangle}}$ have infinite Schmidt ranks. The maximum probability for conversion between the states is continuous for the topology of Hilbert space.
In this letter, we have developed a general formulation for constructing a pair of convertibility monotones using order properties. The monotones are considered as generalizations of distillable entanglement and entanglement cost. This formulation can be applied to many different situations to analyze entanglement convertibility. We have applied the formulation to SLOCC convertibility for genuine infinite dimensional pure states in the single-copy situation. By constructing an example, we have proved the existence of infinitely many mutually SLOCC incomparable pure bipartite states even under the bound energy condition.
One of the important remarks in this letter is that the ordering property under SLOCC convertibility is changed radically, from total ordering to non-total (partial) ordering, with the shift in dimensionality from finite to infinite. Even under the restriction of finite energy and finite measurement conditions, the infinite dimensionality of Hilbert space offers fundamentally different entanglement properties from finite dimensional systems. Although we are only able to treat finite amount of classical information, the workspace for information processing should not be considered to be finite for infinite dimensional quantum systems. Our result encourages the search for other properties of the “infinite workspace”, fundamental differences between finite or infinite dimensional quantum systems, and their use for quantum information processing.
The authors are grateful to M. Hayashi, M. Ozawa, M.B. Plenio, W.J. Munro and K. Nemoto for helpful comments. This work is supported by the Sumitomo Foundation, the Asahi Glass Foundation, and the Japan Society of the Promotion of Science.
[99]{}
S.L. Braunstein and H.J. Kimble, Phys. Rev. Lett. [**80**]{}, 869 (1998); A. Furusawa, J.L. Sørensen, S.L. Breaunstein, C.A. Fuchs, H.J. Kimble and E.S. Polzik, Sience [**282**]{}, 706 (1998).
G. Giedke, J, Eisert, J.I. Cirac and M.B. Plenio, Quant. Inf. Comp. [**3**]{}, 211 (2003); J. Eisert, D. Browne, S. Scheel and M.B. Plenio, quant-ph/0307106 (2003); G. Giedke, M.M. Wolf, O. Kruger, R.F. Werner and J.I. Cirac, Phys. Rev. Lett. [**91**]{}, 107901 (2003); M.M. Wolf, G. Giedke, O. Kruger, R.F. Werner and J.I. Cirac, quant-ph/0306177 (2003).
G. Giedke and J.I. Cirac Phys. Rev. A [**66**]{}, 032316 (2002); J. Fiurasek Phys. Rev. Lett. [**89**]{}, 137904 (2002); J. Eisert, S. Scheel and M.B. Plenio, Phys. Rev. Lett. [**89**]{}, 137903 (2002).
E.M. Rains, IEEE T. Inform. Theory [**47**]{},2921 (2001); K. Audenaert, M.B. Plenio and J. Eisert, Phys. Rev. Lett. [**90**]{}, 027901.
M.A. Nielsen Phys. Rev. Lett. [**83**]{} 436 (1999).
G.Vidal, Phys. Rev. Lett. [**83**]{}, 1046 (1999).
W. Dür, G. Vidal, and J. I. Cirac, Phys. Rev. A [**62**]{}, 062314 (2000).
A.N. Kolmogorov and S.V. Fomin, [*Introductory real analysis*]{}, Dover (1970).
V. Vedral and E. Kashefi Phys. Rev. Lett. [**89**]{}, 037903 (2002).
C.H. Bennett, D.P. DiVincenzo, J.A. Smolin and W.K. Wootters, Phys. Rev. A [**54**]{}, 3824 (1996); C.H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J.A. Smolin and W.K. Wootters, Phys. Rev. Lett. [**76**]{}, 722 (1996); C.H. Bennett, H.J. Bernstein, S. Popescu and B. Schumacher, Phys. Rev. A [**53**]{}, 2046 (1996).
F. Morikoshi, M.F. Santos and V. Vedral quant-ph/0306032 (2003).
E.M. Rains, IEEE T. Inform. Theory [**47**]{},2921 (2001); K. Audenaert, M.B. Plenio and J. Eisert, Phys. Rev. Lett. [**90**]{}, 027901.
M.J. Donald, M. Horodecki and O. Rudolph, J. Math. Phys. [**43**]{}, 4252 (2002); M. Horodecki, P. Horodecki and R. Horodecki, Phys. Rev. Lett. [**84**]{}, 2014 (2000).
M. Horodecki, P. Horodecki, R. Horodecki, Phys. Rev. Lett. [**80**]{}, 5239 (1998).
M. Owari, S.L. Braunstein, K. Nemoto and M. Murao, in preperation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'At the heart of structural engineering research is the use of data obtained from physical structures such as bridges, viaducts and buildings. These data can represent how the structure responds to various stimuli over time when in operation. Many models have been proposed in literature to represent such data, such as linear statistical models. Based upon these models, the health of the structure is reasoned about, e.g. through damage indices, changes in likelihood and statistical parameter estimates. On the other hand, physics-based models are typically used when designing structures to predict how the structure will respond to operational stimuli. These models represent how the structure responds to stimuli under idealised conditions. What remains unclear in the literature is how to combine the observed data with information from the idealised physics-based model into a model that describes the responses of the operational structure. This paper introduces a new approach which fuses together observed data from a physical structure during operation and information from a mathematical model. The observed data are combined with data simulated from the physics-based model using a multi-output Gaussian process formulation. The novelty of this method is how the information from observed data and the physics-based model is balanced to obtain a representative model of the structures response to stimuli. We present our method using data obtained from a fibre-optic sensor network installed on experimental railway sleepers. The curvature of the sleeper at sensor and also non-sensor locations is modelled, guided by the mathematical representation. We discuss how this approach can be used to reason about changes in the structures behaviour over time using simulations and experimental data. The results show that the methodology can accurately detect such changes. They also indicate that the methodology can infer information about changes in the parameters within the physics-based model, including those governing components of the structure not measured directly by sensors such as the ballast foundation.'
address:
- 'Lloyd’s Register Foundation’s Programme for Data-Centric Engineering, Alan Turing Institute'
- 'Department of Mathematics, Imperial College London'
- 'Department of Engineering, University of Cambridge'
- 'Lassonde School of Engineering, York University'
- 'Civil and Architectural Engineering Department, College of Engineering, Qatar University'
author:
- Alastair Gregory
- 'F. Din-Houn Lau'
- Mark Girolami
- 'Liam J. Butler'
- 'Mohammed Z. E. B. Elshafie'
bibliography:
- 'refs.bib'
title: 'The synthesis of data from instrumented structures and physics-based models via Gaussian processes'
---
structural health monitoring ,data-centric engineering ,Gaussian processes ,damage detection
Introduction
============
The engineering research fields of structural health monitoring (SHM) and structural identification (SI) typically involve the collection and analysis of data from instrumented structures [@farrar2012structural; @farrar2007introduction; @ccatbacs2013structural; @Aktan]. The primary objectives of SHM are to better understand the behaviour of structures (e.g. bridges, buildings, railway tracks, pipelines, tunnels etc.) over time and to identify and localise damage [@Cawley]. Novel sensing technologies such as fibre-optic and piezoelectric sensors have enabled the detailed collection of performance data from structures [@Sun]. These data are opening up the possibility for engineers to better reason about the structural condition of structures through statistical analysis (e.g. linear models [@Lau]). Typical examples of response data collected from sensor networks on instrumented structures are in the form of vibration/acceleration, strain and acoustic responses [@Ye]. The data provides underlying information about how these structures respond to stimuli, and also includes noise due to various factors such as temperature effects and instrumentation error. Data is rarely measured across the whole domain of the structure either due to it being infeasible or too expensive [@Sazonov]. Therefore the response of the structure is only known at the sensor locations.
Most structures are designed and analyzed based on physical assumptions about how they should respond under various stimuli; these assumptions can be described using idealised mathematical models. These models typically are in the form of partial/ordinary differential equations [@Friswell; @Sinha; @Zeinali]. Partial differential equations (PDEs) often do not admit closed form solutions. Therefore they are often solved using e.g. finite element methods (FEM) [@Doebling]. Quantifying the uncertainty in the physically modelled response of structures from PDEs, where there exists uncertainty in the input parameter, is a well-studied field in applied mathematics (e.g. for example data-assimilation, stochastic Galerkin methods, multilevel Monte Carlo [@Blondeel] and Markov chain Monte Carlo [@Dodwell]). There are several discrepancies between the observed data and the idealised physics-based model. First, the data represents the response of the sensor network not the structure: the underlying response of the structure is corrupted by the sensors. Second, the physics-based model is based on assumptions such as constant temperature and simplified structural properties. Further, the physics-based models do not take into account the instrumentation error (the error induced by the sensors). In the SHM literature it is not clear how to account for these discrepancies in a well-principled fashion that leads to a representative model of the data.
Hybrid approaches are commonplace in literature, utilising data accrued from instrumented structures to improve predictions of structural response to stimuli from these physics-based models. A commonly used example of a data/physics hybrid approach is model updating [@Rocchetta; @Grafe; @Schommer; @Alkayem; @Vigliotti]. In model updating, the parameters of the physics-based model are estimated using the measured response data; predicted responses of the structure are then given by the model with the estimated parameters. The parameters are typically estimated by optimising an objective function, such as the likelihood, which quantifies the discrepancy between the model and data [@Schommer]. The predictive quality of the model updating approach relies on the physics-based model not incorporating too many simplifying assumptions and being closely representative of the actual response. In model updating, the physics-based model is treated as the true underlying data generating process [@Liu]. Here however, we assert that the observed data is the closest representation of the true structural response. Therefore the data should have an important role in the overall modelling procedure and prediction of structural response beyond estimating parameters.
In this paper, we introduce a novel modelling approach for structural response that balances the information from the observed data and the physics-based model. This will be referred to as a type of ‘data-centric engineering’ (DCE) model, inline with the discussion in [@LauProb]. The information is obtained from the physics-based model by simulating data from it. These simulations are combined with observed data from the physical structure using a multi-output Gaussian process joint model [@OHagan; @Raissi]. This model is constructed on the level of the geometric relationship between the physical quantities of the observed data and model simulations. Predictions of the structures response can be obtained via this joint model. Unlike model updating, not only is observed data utilised to directly improve the physics-based model, but also the physics-based model is used to guide the inferred posterior from the data in unmeasured regions of the structural domain. A multi-output Gaussian process model is also utilised in [@Zhou]; this differs from the work presented in this paper as it combines simulated data from two physics-based models with different fidelities (accuracies), instead of combining simulation from a physics-based model and observed data. Gaussian processes have also recently been used for other aspects of structural health monitoring in [@Neves; @Worden; @Teimouri; @Fuentes].
![A schematic depicting different modelling approaches to instrumented structures. Data-driven and model-driven approaches can both produce predictions for the response of the structure independently. The popular model-updating procedure uses data from the structure to estimate the input parameters within the physics-based model; this updated model then produces response predictions. In our approach, data and (updated) physics-based model simulations are combined in a joint probabilistic model that then produces response predictions.[]{data-label="figure:schematic3"}](schematic3){width="150mm"}
The benefits of the presented work are two-fold: First, through the proposed joint model one obtains aposteriori estimates for the response of the structure, inferred from both observed data and simulations from an analytical physics-based model. These two sources of information are balanced in order to improve the predictive performance of the posterior in regions where there is no measurement data available. Second, the modelling approach used estimates structural parameters within the physics-based model, even parameters governing components of the modelled system that cannot directly be measured by sensors. These capabilities make this work an important stride forward in the construction and evaluation of DCE models.
The response of railway sleeper beams (the components which carry the rail track and transfer the train axle forces into the supporting ballast and subgrade materials) will be used as a running example throughout this paper to demonstrate the proposed methodology. These sleepers can be modelled by the Euler-Bernoulli equation. This provides a simplified analytical physics-based model for the vertical deflection of the actual railway sleeper, although this can be generalised to richer models such as those that require the use of FEM. A core objective of SHM is to be able to reason about and detect structural change over time. The proposed multi-output Gaussian process model for the response of the railway sleeper can achieve this when used alongside change-point detection schemes. This is demonstrated in Sec. \[sec:results\] using simulated and experimental data sets.
The paper proceeds as follows. Section \[sec:sensordata\] concentrates on the data from sensor networks that is used throughout this work. The specifications of the instrumented railway sleeper used in the experimental aspect of this work is also discussed here. Following this in Sec. \[sec:physics\], an analytical physics-based model for the vertical deflection of the sleeper is introduced. The main methodology of this paper is outlined throughout Sec. \[sec:machinelearning\]; this includes the estimation of the parameters within the physics-based model and the construction of a posterior for the response of the sleeper conditioned on data obtained from the sensor networks and physics-based model. This section also discusses by how much the physics-based model should inform the joint model, and proposes a principled approach to this problem. Finally in Sec. \[sec:results\], simulated data and experimental data from an instrumented railway sleeper provide a demonstration on the effectiveness of the proposed methodology.
Data from instrumented structure {#sec:sensordata}
================================
This study considers the modelling of a single horizontal prestressed concrete sleeper beam supported on compacted railway ballast. Data is obtained from the sleeper by instrumenting it with a network of fibre-optic sensors (FOS) consisting of Bragg gratings (FBG). FBG sensors interrogate light spectra and reflect them with a specific wavelength known as the ‘Bragg wavelength’; as the fibre optic cable containing the FBG is placed under strain at a given time $t_n$ this Bragg wavelength shifts linearly (from $\lambda_{t_0}$ to $\lambda_{t_n}$). The corresponding strain shift $\epsilon_{t_n}$ can then be computed via a linear transformation of this relative wavelength shift, $$\epsilon_{t_n} = 10^6 \left(\frac{\lambda_{t_n}-\lambda_{t_0}}{0.78 \lambda_{t_0}}\right).$$ This strain is measured in units of microstrain, and measurements are recorded at discrete time indices $t_0,(t_0+\Delta t),\ldots,(t_0+n\Delta t)=t_n$ where $\Delta t=1/50$ seconds. Each FBG measures strain to an approximate accuracy of $\pm 4$ microstrain. A single fibre-optic cable can be inscribed with multiple FBGs. For a comprehensive review of these sensors see [@Kreuzer], which also discusses how FBG sensors are affected by external factors such as temperature.
{width="110mm"} \[figure:schematic\]
{width="100mm"} \[figure:sensordata\]
The railway sleeper considered and used in the experimental results was instrumented during its fabrication with three FBG sensors attached along the top and bottom steel prestressing strands, which are embedded within the concrete sleeper. Coordinates on the sleeper are measured in millimeters. The sleeper is $d=2500$ by 200 (length/depth) with the FBG sensors located at $X_f=[X_{f_1},X_{f_2},X_{f_3}]=[500,1250,2000]$, on both the top and bottom prestressing strands embedded within the sleeper. The length of $X_f$ is denoted by $N_f$ for the benefit of notation later in the paper. In this particular application $N_f=3$. The distance between the top and bottom rows of FBG sensors on the instrumented sleeper is $91.5$. In experiments the sleeper was subjected to loading forces to simulate the axle spacing of the train wheels. Two equal forces were exerted by pistons at $x_1=(d-2c)/2$ and $x_2=d-x_1$, with $c=750$. The force magnitude was varied over the course of the experiment (see later Sec. \[sec:experiment\] for details). A schematic of the sensor network instrumented on the sleeper beam is shown in Figure \[figure:schematic\]. A more detailed specification of the instrumented sleepers considered in this study is given in [@Butler]. For illustrative purposes only, Figure \[figure:sensordata\] shows a 7.52 second representative data set of strain measurements from an FBG (located at $x=2000$ along the top of an instrumented sleeper). This captures the moment when a train passes over the sleeper, resulting in the visibly large peaks of strain.
Given strain measurements $\epsilon_{t}(x)$ at the time index $t$ and one coordinate $x \in \mathcal{D}=[0,d]$ on both the top (denoted by the superscript $T$) and bottom (denoted by the superscript $B$) of the sleeper, the *curvature* associated with these measurements can be computed by $$z_t(x) = \frac{\epsilon^{(T)}_t(x)-\epsilon^{(B)}_t(x)}{91.5}.
\label{equation:sensorcurvature}$$ Therefore the curvature here represents the gradient of strain over the depth of the sleeper. This combines strain measurements from the top and bottom of the sleeper into a single one-dimensional quantity $z_t(x)$. A one-dimensional domain simplifies the synthesis of this observed data with a one-dimensional physics-based model (presented in the next section). Curvature can be a useful tool in damage detection for beam structures [@Dawari]. In practice, curvature data are only recorded at the sensor locations $X_f \in
\mathcal{D}$. Denote the curvature data at locations $X_f$ and time index $t$ as $z_t(X_f) =
[z_t(X_{f_1}),z_t(X_{f_2}),z_t(X_{f_3})]^T$. We assume the $M$ curvature observations $\big\{z_i(X_f)\big\}_{i=1}^{M}$ are independent realisations from $z(X_f)$ and $$z(X_f) = \mu_f + W, \quad W \sim \mathcal{N}\left(0
,\Sigma_{f}\right),
\label{equation:noisemodel}$$ where $\mu_f := f(X_f)$ is the true curvature and $\Sigma_f\in\mathbb{R}^{3\times 3}$ is the covariance matrix of the Gaussian noise $W$. We assume herein that $\Sigma_f$ has a diagonal form, meaning that the noisy curvature observations at the different coordinates in $X_f$ are independent of one another.
We will obtain an alternative model for this curvature, conditioned on both the observed noisy data and information from a physics-based model in Sec. \[sec:machinelearning\]. Curvature, $f(x)$, is geometrically related to other useful quantities for engineers, one being the vertical deflection, $y(x)$, of an idealized beam. These quantities are related via, $$-\frac{\operatorname{d}^2 y(x)}{\operatorname{d}x^2}=f(x).
\label{equation:governingequation}$$ This differential relationship is used in [@Xu] to estimate the deflection of beam structures from strain measurements. This relationship will be the basis of how curvature data from the instrumented sleeper will be synthesised with a physics-based model for vertical deflection in a joint model for the response of the sleeper under load (described in Sec. \[sec:machinelearning\]). A physics-based model for the vertical deflection of the instrumented sleeper considered in this paper is described in the following section.
Physics-based model description {#sec:physics}
===============================
Engineers typically use physics-based models to assist in understanding how a structure behaves during excitation and at rest. These physics-based models are framed as solutions to (systems of) ordinary and partial differential equations, where often numerical discretization through finite element methods is required [@Doebling; @Sinha]. This section describes an analytical physics-based model for the vertical deflection, $y(x)$, of a railway sleeper supported on compacted ballast [@Tran]. The Euler-Bernoulli equation describes how a one-dimensional beam on the domain $\mathcal{D}$, with an elastic foundation, responds under a forcing $p(x)$. Such a response is utilised within the present work as a simplified representation of the response of the considered railway sleeper. The (static-time) Euler-Bernoulli equation is given by, $$\frac{\operatorname{d}^2 \left(EI(x)\left(\frac{\operatorname{d}^2 y(x)}{\operatorname{d}x^2}\right)\right)}{\operatorname{d}x^2}=p(x),
\label{equation:governingEB}$$ where the prime notation denotes the derivative with respect to $x$, $I(x)$ is the second moment of cross-sectional area and $E$ is the Young’s modulus of the beam material (e.g. concrete). The product of $E$ and $I(x)$ is known as the flexural rigidity of the beam. We make the assumption that the flexural rigidity is constant over $x$, i.e. $I(x):=I\in (0,\infty)$. See Sec. \[sec:conclusion\] for a discussion on a possible way to relax this assumption.
We obtain the physics-based model for the deflection of a railway sleeper by solving (\[equation:governingEB\]), using a specific form for the forcing $p(x)$. In this case, an analytic solution is available. Physics-based models do not necessarily describe how the physical system operates in practice, e.g. due to structural simplification such as assuming a constant flexural rigidity. To derive the physics-based model for the deflection of a railway sleeper, we assume that during a train passing over the sleeper the forcing $p(x)$ has non-zero value at the two points $x_1$ and $x_2$ (experimentally simulated locations of train wheel axles), and $p(x_1)=p(x_2)=p$ (Newtons). A schematic of the sleeper system (including the instrumented sensor network discussed in Sec. \[sec:sensordata\]) is shown in Figure \[figure:schematic\]. A solution to (\[equation:governingEB\]) and an analytical model for the vertical deflection is [@Hetenyi], $$y(x,p,k,\lambda) = \frac{p\lambda}{k}w(x,\lambda),
\label{equation:unitsolution}$$ where $$w(x,\lambda) =
\begin{cases}
v(x, \lambda) & x \in [0,x_1]\\
\begin{split}
v(x, \lambda) +\big[&\tilde{c}(\lambda (x-x_1))s(\lambda(x-x_1))+...\\
\quad &\tilde{s}(\lambda(x-x_1))c(\lambda(x-x_1))\big]
\end{split} & x \in [x_1, x_2],
\end{cases}
\label{equation:analyticdeflection}$$ and $$\begin{split}
v(x, \lambda) =\left[\tilde{s}(\lambda d)+ s(\lambda d)\right]^{-1}\Big\{&2\tilde{c}(\lambda x)c(\lambda x)\big[\tilde{c}(\lambda x_1)c(\lambda x_2)+\tilde{c}(\lambda x_2)c(\lambda x_1)\big]+...\\
\qquad &\big[\tilde{c}(\lambda x) s(\lambda x) - \tilde{s}(\lambda x)c(\lambda x)\big]\big[\tilde{c}(\lambda x_1)s(\lambda x_2) -...\\
\quad &\tilde{s}(\lambda x_1)c(\lambda x_2) + \tilde{c}(\lambda x_2)s(\lambda x_1) - \tilde{s}(\lambda x_2)c(\lambda x_1)\big]\Big\}.
\end{split}$$ Here $c(\cdot)=\cos(\cdot)$, $s(\cdot)=\sin(\cdot)$, $\tilde{c}(\cdot)=\cosh(\cdot)$, $\tilde{s}(\cdot)=\sinh(\cdot)$. Finally $w(x,\lambda)=w(d-x,\lambda)$, for $x \in [x_2,d]$. Also $\lambda = \left(k/(4EI)\right)^{\frac{1}{4}}
$ is the flexibility of the sleeper and $k>0$ is the ballast stiffness. Let the parameters for this formulation of $y$, namely $p$, $k$ and $EI$ (which together lead to $\lambda$), be the components of the vector $\phi=(p,k,EI)$. Using (\[equation:unitsolution\]), one can rewrite the relationship between curvature and vertical deflection in (\[equation:governingequation\]) as $$\mathcal{L}_{x,\phi}w(x,\lambda)=f(x),
\label{equation:curvature}$$ with the linear differential operator $\mathcal{L}_{x,\phi}=-\frac{p\lambda}{k}\frac{\operatorname{d}^2}{\operatorname{d}x^2}$. In Sec. \[sec:machinelearning\], we introduce a joint probabilistic model that uses (\[equation:curvature\]) to combine observed curvature data (through strain data obtained from sensors at $X_f$; see Sec. \[sec:sensordata\]) and the physics-based model in (\[equation:unitsolution\]). Figure \[figure:data\_and\_physics\] shows $\mathcal{L}_{x,\phi}w(x,\lambda)$ computed numerically (using finite differences) alongside an example 5 simulated curvature data points at each coordinate in $X_f$, where the values of $k=\exp(5)$ and $EI=\exp(28)$ used in $\phi$ are given by prior beliefs in (\[equation:functionforregularization\]). The forcing used is $p=125000$. In our method, we do not use the full analytic solution $w(x,\lambda)$ for all $x \in \mathcal{D}$; instead we simulate from $w(x,\lambda)$ at $X_u \in [0,d]^{N_u}$ and combine it with observed data in the joint model. The only requirement here is that one is required to prescribe a value of $\lambda$ (estimation is used to circumvent this later in Sec. \[sec:parameterestimation\]). We use simulation from $w(x,\lambda)$ instead of using the full analytic solution (or infinitely many simulation coordinates) to create leverage on how (and where - unmeasured/measured areas of the domain) the physics-based model influences the joint model. If the full solution was used, the joint model would be fully informed by physics, and not by the observed data. In the case of a highly simplified model that poorly represents the observed response of the sleeper, this is detrimental. On the other hand, in the case of a physics-based model that represents the observed response well, it would be desireable to incorporate many simulations into the joint model. These examples inspire the process of ‘tuning’ the choice of $X_u$; see Sec. \[sec:adaptivenu\] for a discussion on this.
The analytical physics-based model presented in (\[equation:unitsolution\]) represents a simplified response of the railway sleeper. It is used because it is inexpensive to simulate from and incorporates parameters of interest which govern the sleeper/rail trackbed system, including the ballast stiffness. In other cases, a sophisticated FEM model for $y(x)$ (or ‘digital twin’) might be more appropriate; however, the number of times one can simulate from this may be limited by computational expense. The next section describes how the physical simulations discussed in this section and the strain measurement-derived curvature data from the instrumented railway sleeper can be synthesised in a joint Gaussian process model.
![An example 5 points of simulated curvature data $z_t(X_f)$ ($t=1,\dots,5$) at each coordinate $X_f$ along the sleeper. Also shown is $\mathcal{L}_{\phi,x}w(x,\lambda)$ numerically computed from the physics-based model of the vertical deflection $y(x)$ in a dashed black line. The parameters used in the model are given by prior beliefs in (\[equation:functionforregularization\]). The forcing used is $p=125000$.[]{data-label="figure:data_and_physics"}](data_and_physics){width="100mm"}
A data-centric engineering model based on Gaussian processes {#sec:machinelearning}
============================================================
This section will describe the methodology synthesising the physics-based model and the curvature data. The modelling of differential equations, such as $$\tag{\ref{equation:curvature} revisited}
\mathcal{L}_{x,\phi}w(x,\lambda)=f(x),$$ is an important field of research for physical and engineering applications. In the case considered in this paper, it allows one to model the response of a railway sleeper under forcing by fusing curvature data and physical simulation for vertical deflection on the level of the geometric relationship between them. We will now describe the components of (\[equation:curvature\]). First, recall that $f(x)$ represents the true curvature, which we do not have direct access to. Instead we observe the noisy curvature data $\big\{z_i(X_f)\big\}_{i=1}^{M}$ located at $X_f$ (see model \[equation:noisemodel\]). We will describe later in this section how the observations are used to estimate the true curvature. Second, we define $u(x)=w(x,\lambda)$ for a prescribed $\lambda$; hence $u(x)$ represents the physics-based model for vertical deflection. A joint model for $u(x)$ and $f(x)$ using (\[equation:curvature\]) is now described. A recent landmark paper [@Raissi] proposes to learn about differential systems, such as (\[equation:curvature\]), and the parameters within them by training multi-output joint Gaussian process models [@OHagan] based on observations from both $u(x)$ and $f(x)$. The multiple ‘outputs’ in the model correspond to $u(x)$ and $f(x)$, respectively. This joint model is described presently. Assuming a mean-zero Gaussian process prior for $u(x)$, $$u(x)|\sigma^2 \sim \mathcal{GP}(0, k_{u,u;\sigma^2}(x,x')) ,$$ one can also construct a Gaussian process model for $f(x)$ via the differential operator $\mathcal{L}_{x,\phi}$. The infinite-dimensional stationary covariance kernel used for the prior is $$k_{u,u;\sigma^2}(x,x') = \exp \left\{-\frac{\sigma^2}{2}\left(\frac{x-x'}{d}\right)^2 \right\},$$ where $\sigma^2$ is an unknown reciprocal length-scale parameter. This is known as the squared exponential covariance function and is a commonly used covariance kernel in many statistical applications given it’s attractive properties, e.g. see [@Zhou]. For example, as $(x-x')^2$ increases from 0 at $x=x'$, the covariance between the two points decreases from it’s maximum at 1 to approximately 0 when the two points are sufficiently ‘far away’ from each other. This local dependence behaviour is found in many physical systems. Given the form of the covariance kernel, let $\theta=(\phi,\sigma^2)$ be the extended parameter set of the system. For $X=[X_1,\ldots, X_n] \in [0,d]^n$ and $X'=[X_1',\ldots, X_m'] \in [0,d]^m$, we define the matrix $k_{u,u;\sigma^2}(X,X') \in \mathbb{R}^{n \times m}$ as $$[k_{u,u;\sigma^2}(X,X')]_{i,j}= k_{u,u;\sigma^2}(X_i,X_j')\quad
i=1,\dots,n; j=1,\dots,m.
$$ The linear operator $\mathcal{L}_{x,\phi}$ can be used to obtain the covariance kernel of the Gaussian process that models $f(x)$, through $$f(x)|\theta = \mathcal{L}_{x, \phi} u(x)|\sigma^2 \sim \mathcal{GP}(0, k_{f,f;\theta}(x,x')),$$ where $$k_{f,f;\theta}(x,x')=\mathcal{L}_{x,\phi}\mathcal{L}_{x',\phi}k_{u,u;\sigma^2}(x,x').$$ Also define the cross-covariance kernels $$k_{u,f,\theta}(x,x') = \mathcal{L}_{x,\phi}k_{u,u;\sigma^2}(x,x'), \quad k_{f,u;\theta}(x,x')=\mathcal{L}_{x',\phi}k_{u,u;\sigma^2}(x,x').$$ In this paper, we concentrate on the case where $\mathcal{L}_{x,\phi}=-(p\lambda\operatorname{d}^2)/(k\operatorname{d}x^2)$ from (\[equation:curvature\]), and therefore the kernels $k_{u,f;\theta}$, $k_{f,u;\theta}$ and $k_{f,f;\theta}$ are given by $$k_{u,f;\theta}(x,x') = k_{f,u;\theta}(x,x')= -\frac{\sigma^2p\lambda}{d^2k}\left\{\left(\frac{\sigma(x-x')}{d}\right)^2 -1\right\}k_{u,u;\sigma^2}(x,x'),$$ and $$k_{f,f;\theta}(x,x') = \frac{\sigma^4p^2\lambda^2}{d^4 k^2}\left\{3-6\left(\frac{\sigma(x-x')}{d}\right)^2+\left(\frac{\sigma(x-x')}{d}\right)^4\right\}k_{u,u;\sigma^2}(x,x').$$ Given the physics-based model simulation coordinates $X_u$ and the coordinates at which noisy strain measurement-derived curvature data is available at $X_f$, one can write the joint distribution $p(u(X_u), z(X_f)|X_u,X_f,\theta)$, where $z(X_f)$ is as defined in (\[equation:noisemodel\]), as $$\begin{bmatrix}
u(X_u)\\
z(X_f)
\end{bmatrix} \sim
\mathcal{N}\left(
\begin{bmatrix}
0\\
0
\end{bmatrix},
K_{\theta}
\right),
\label{equation:likelihood}$$ where $$K_\theta=
\begin{bmatrix}
k_{u,u;\sigma^2}(X_u,X_u) & k_{u,f;\theta}(X_u,X_f)\\
k_{f,u;\theta}(X_f,X_u) & k_{f,f;\theta}(X_f,X_f) + \Sigma_f
\end{bmatrix},$$ from [@OHagan; @Raissi]. This model allows one to model the differential system in (\[equation:curvature\]) probabilistically, estimate the parameters $\theta$ and construct point-wise posteriors for either $f(x)$ or $u(x)$. The next two sections describe how simulations from the physics-based model for vertical deflection $u(x)$ and noisy curvature data $\big\{z_i(X_f)\big\}_{i=1}^{M}$ can be used to estimate the parameters within the physics-based model and obtain aposteriori estimates for the sleeper response. The model in (\[equation:likelihood\]) is a DCE model [@LauProb], combining information about two different physical quantities, one from physical simulation and the other from measurement data obtained from instrumented structures.
Parameter estimation
--------------------
In this section, we describe the procedure used to estimate the parameters from the observation model in (\[equation:noisemodel\]); $\Sigma_f$ and $\mu_f$ and the model in (\[equation:likelihood\]); $\theta=(p,k,EI,\sigma^2)$. In the case where the forcing term $p$ is known, the parameter vector $\theta$ reduces to $(k,EI,\sigma^2)$. This is the case considered in the remainder of the paper. The estimation procedure is performed in two steps. First, the observation model parameters are estimated empirically using the observed curvature data to give $\widehat{\Sigma}_f$ and $\widehat{\mu}_f=[\widehat{\mu}_{f_1},\widehat{\mu}_{f_2},\widehat{\mu}_{f_3}]$. The type of estimation used for this first step is commonplace for data involved in Gaussian process regression [@Rasmussen]. Second, the parameters $\theta$ are estimated using $\widehat{\Sigma}$ and $\widehat{\mu}_f$ and simulations from the physics-based model for $u(x)$. The full estimation procedure is described in Algorithm \[alg:general\].
Number of simulations $N_u$ from physics-based model; number of observed locations $N_f$; observed locations $X_f$; observed curvature data $\big\{z_i(X_f)\big\}_{i=1}^{M}$; regularization function $g(\cdot)$; physics-based model $w(\cdot,\cdot)$;
Compute the estimates $$\widehat{\mu}_f=\frac{1}{M}\sum^{M}_{i=1}z_i(X_{f}),
\label{equation:muestimates}$$ and $$\widehat{\Sigma}_f =\text{diag}\left(\frac{1}{M}\sum^{M}_{i=1}\left[z_i(X_f) - \widehat{\mu}_f \right]^T \left[ z_i(X_f) - \widehat{\mu}_f\right]\right).$$
Define $X_u$ to be $N_u$ evenly spaced coordinates on $\mathcal{D}$. By defining $\lambda(\theta):=\lambda=(k/4EI)^{1/4}$ for $k,EI \in \theta$, set $Y_\theta=[w(X_{u},\lambda(\theta)), \widehat{\mu}_f]
\in\mathbb{R}^{N_u+N_f}$ and solve $$\widehat{\theta}=\text{arg}\max_\theta \left[L(\theta;Y_{\theta})+\log g(\theta)\right],
\label{equation:regularizedoptimization}$$ where $$L(\theta;Y_{\theta})=\frac{1}{2}\log \abs{K_{\theta}}+\frac{1}{2}Y_\theta^T K_{\theta} Y_\theta +\frac{(N_u+N_f)}{2}\log(2\pi).
\label{equation:optimizablefunction}$$ $\widehat{\theta} =
(\widehat{k},\widehat{EI},\widehat{\sigma^2})$ and $\widehat{\lambda}
= \left(\widehat{k} / 4 \widehat{EI} \right)^{1/4}$
\[sec:parameterestimation\]
The input $g(\theta)$ is a regularization function which allows for prior belief about $\theta$ to inform the optimization. From a Bayesian perspective, (\[equation:regularizedoptimization\]) corresponds to maximum aposteriori estimation (MAP), where the objective function $L(\theta;Y_{\theta})$ is the logarithm of the marginal likelihood $p(Y_{\theta}|X_u,X_f,\theta)$ in (\[equation:likelihood\]). For parameters such as the flexural rigidity of a sleeper, $EI$, and ballast stiffness, $k$, there are extensive guidelines that specify confidence intervals for them [@PWI] and could be used to decide $g(\theta)$. Define the following log-normal distribution function, $$q(x,m,s) = \frac{1}{x\sqrt{2s^2\pi}}\exp\left\{-\frac{(\log(x)-m)^2}{2s^2}\right\},$$ then we assume that the regularization function $g(\theta)$ takes the form $$g(\theta)=q(\sigma^2, 2.5, 0.125)q(k, 5, 0.25)q(EI, 28, 0.25),
\label{equation:functionforregularization}$$ based on order-of-magnitude estimates from [@PWI]. This is the form of $g(\theta)$ that we consider for the remainder of the paper.
In the case of the simple physics-based model for $u(x)$ utilised in this study, the optimization problem can be solved using various iterative methods. Simulated annealing [@Brooks] is used for the numerical examples presented in the remainder of the paper. Let the estimated parameter vector for the model in (\[equation:likelihood\]) be $\widehat{\theta}=(\widehat{k},\widehat{EI},\widehat{\sigma^2})$, where $\widehat{\phi}=(\widehat{k},\widehat{EI})$ and $\widehat{\lambda}=(\widehat{k}/4\widehat{EI})^{1/4}$. Then $Y_{\widehat{\theta}}=[w(X_u,\widehat{\lambda}), \widehat{\mu}_f]$ is the concatenation of the optimized physics-based model simulation and the estimated curvature. The estimation procedure above outlines the training of the proposed multi-output Gaussian process model using the batch of noisy curvature observations $\big\{z_i(X_f)\big\}_{i=1}^{M}$. Figure \[figure:data\_and\_estimated\_physics\] shows the same as Figure \[figure:data\_and\_physics\] only with the physics-based model for the vertical deflection computed using the estimated parameters $\widehat{k}=209$ and $\widehat{EI}=6\times 10^{11}$. They were estimated using $N_u=6$ simulations from the physics-based model in (\[equation:optimizablefunction\]). These parameter estimates lead to physics-informed curvature (via the numerically computed $\mathcal{L}_{x,\phi}w(x,\lambda)$) that is more representative of the data than that from the prior parameter beliefs in (\[equation:functionforregularization\]).
Generating a posterior {#sec:posterior}
----------------------
Once obtaining $\widehat{\theta}$ from Algorithm \[alg:general\], a point-wise posterior at $x \in \mathcal{D}$ for the curvature of the sleeper given $Y_{\widehat{\theta}}$ can be computed. The posterior can be interpreted as the physics-based model guiding the inference from the sensor data in regions of the domain which are not instrumented with sensors. Consider the joint distribution (the dependence on $X_u$, $X_f$, $\widehat{\theta}$ and $x$ is dropped for notational convenience), $$\begin{bmatrix}
Y_{\widehat{\theta}}\\
f(x)
\end{bmatrix} \sim \mathcal{N}\left(
\begin{bmatrix}
0\\
0
\end{bmatrix},
\begin{bmatrix}
K_{\widehat{\theta}} & k_{u,f;\widehat{\theta}}(X_u,x)\\
k_{f,u;\widehat{\theta}}(x,X_u)&k_{f,f;\widehat{\theta}}(x,x)
\end{bmatrix}
\right).$$ Then the posterior is given by [@Rasmussen; @Robert], $$f(x)|Y_{\widehat{\theta}},\widehat{\theta} \sim \mathcal{N}\left(q_{f}^{T}K_{\widehat{\theta}}^{-1}Y_{\widehat{\theta}},k_{f,f;\widehat{\theta}}(x,x)-q_{f}^{T}K_{\widehat{\theta}}^{-1}q_{f} \right),
\label{equation:posteriorcurvature}$$ where $$q_{f}^{T}=\left(k_{f,u;\widehat{\theta}}(x,X_{u}), k_{f,f;\widehat{\theta}}(x,X_{f})\right).$$ This posterior can be used to predict curvature (via sampling from this normal distribution), detecting system changes and to select the number of simulated data points for $u(x)$ (see next two sections). In a similar way, the posterior for the vertical deflection of the sleeper, $u(x)|Y_{\widehat{\theta}},\widehat{\theta}$, can also be obtained due to the incorporation of the physics-based model, and therefore the observed curvature data can be used to infer a physical quantity that it cannot directly measure. Notice that the posterior variance interestingly does not depend on the simulated values from the physics-based model, but only on the simulation coordinates $X_u$. As aforementioned, one could manipulate this observation by selecting the position of the simulation coordinates such that the posterior variance is minimized for future prediction; this is known as active learning [@Seo]. Figure \[figure:posterior\] shows the same as Figure \[figure:data\_and\_estimated\_physics\] in addition to the posterior mean and 95% confidence intervals in (\[equation:posteriorcurvature\]).
{width="80mm"} \[figure:data\_and\_estimated\_physics\]
{width="80mm"} \[figure:posterior\]
{width="70.00000%"}
{width="70.00000%"}
Tuning the quantity of simulation from $u(x)$ {#sec:adaptivenu}
---------------------------------------------
This section investigates how the similarity between the observed data and physics-based model can be assessed using the point-wise posterior for sleeper curvature. This leads to a principled method of selecting $N_{u}$, the number of simulated data-points from the analytical model for $u(x)$, that most improves the predictive performance of the posterior. In the case of an over-simplified physics-based model, we propose using more observed curvature data than the simulations from the crude physics-based model. We now demonstrate the effect that $N_u$, the number of simulations from $u(x)$, has on the posterior $p(f(x)|Y_{\widehat{\theta}},\widehat{\theta})$; this is done via simulated curvature data. Consider the simulated curvature data $\big\{z_i(X_f)\big\}_{i=1}^{5}$ sampled from $$z(X_f) = \begin{bmatrix}
-1\times10^{-5}\\
2.45\times 10^{-6}
\\-1\times10^{-5}
\end{bmatrix}
+ W,\quad W \sim \mathcal{N}\left(
0,1\times 10^{-6} I_3 \right).$$ We take the forcing as $p=125000$N. The parameters of the multi-output Gaussian process in (\[equation:likelihood\]) are estimated using $\big\{z_i(X_f)\big\}_{i=1}^{5}$ and Algorithm \[alg:general\], along with the number of simulated data-points set to $N_u=2$, $9$ and $14$. The pointwise curvature posteriors over a range of $x \in \mathcal{D}$ are then given for each model by , and the corresponding means and 95% confidence intervals are shown in Figure \[figure:goldilocks\_principle\]. The blue points show the simulated curvature samples at each point in $X_f$. For larger values of $N_u$ the posterior is more akin to the physics-based model than the observed data. On the other hand, in the extreme case of $N_u=0$, the proposed method becomes the same as standard Gaussian process regression (kriging) on the observed curvature data; the posterior is based solely on the observed data. As $N_u$ increases, the parameter estimates from the associated models change accordingly, for example resulting in the posterior diverting away from the data when $N_u=14$. In this case, the posterior is not representative of the observed data since it incorporates too many simulations from the physics-based model. To obtain a posterior more representative of the observed data, fewer simulations from the physics-based model are required, but enough that the posterior has less variance than in the case of $N_u=2$. This indicates that a trade-off exists between the number of simulations from $u(x)$, $N_u$, and the predictive performance of the posterior with respect to the observed data.
![The posteriors of curvature $p(f(x)|Y_{\widehat{\theta}},\widehat{\theta})$ corresponding to the Gaussian process models trained around $N_u=2$ (top), $N_u=9$ (middle) and $N_u=14$ (bottom) simulation coordinates from $u(x)$. The points show the simulated samples of curvature data at each point in $X_f$, and the dashed lines show the 95% confidence interval for the posterior.[]{data-label="figure:goldilocks_principle"}](goldilocks_principle){width="100mm"}
The mean squared error is now employed as a metric to evaluate the data/physics information trade-off highlighted here, and to tune the number of simulations from $u(x)$. The objective here is to find the value of $N_u$ that maximises the predictive performance of the posterior in (\[equation:posteriorcurvature\]), for a prescribed $N_u$, with respect to a test data-point. Here we suggest to use $\widehat{\mu}_{f_3}$, derived from the noisy data at $X_{f_3}$, as the test data-point. Equally this could be exchanged with either $\widehat{\mu}_{f_2}$ or $\widehat{\mu}_{f_3}$. Let $Z_{N_u}\sim f(X_{f_{3}})|Y_{N_u},\widehat{\theta}$ using (\[equation:posteriorcurvature\]), where $Y_{N_u}=[w(X_u,\widehat{\lambda}),\widehat{\mu}_{f_1},\widehat{\mu}_{f_2}]$, and $X_u$ are $N_u$ evenly spaced coordinates along the domain. Here, $\widehat{\mu}_{f_1}$, $\widehat{\mu}_{f_2}$, $\widehat{\lambda}$ and $\widehat{\theta}$ are estimated by using Algorithm \[alg:general\] with the inputs $N_f:=2$ and $X_f:=[X_{f_1},X_{f_2}]$ (therefore excluding the noisy curvature observations at the last sensor location). Then the mean squared error of $Z_{N_u}$ with respect to the test data-point $\widehat{\mu}_{f_3}$, $\mathbb{E}\left[(Z_{N_u}-\widehat{\mu}_{f_3})^2\right]$, is to be minimized, $$\text{arg}\min_{N_u} \left\{\mathbb{E}\left[(Z_{N_u}-\widehat{\mu}_{f_3})^2\right]\right\}=\text{arg}\min_{N_u}\left\{\left(\mathbb{E}[Z_{N_u}]-\widehat{\mu}_{f_3}\right)^2+\mathbb{V}(Z_{N_u})\right\}.
\label{equation:wasserstein}$$ The mean squared error is the sum of the squared bias of $Z_{N_u}$, with respect to $\widehat{\mu}_{f_3}$, and the variance of $Z_{N_u}$; the hope here is that the posterior $Z_{N_u}$ corresponding to the minimum in (\[equation:wasserstein\]) will allow the physics-based model to guide it in the region where data is missing, but not be overconfident. In practice, an interval search can be implemented between two end-points $N_u^{\min}$ and $N_u^{\max}$ to find the minimum in (\[equation:wasserstein\]), for suitable values of $N_u^{\min}$ and $N_u^{\max}$. A summary of how this procedure interacts with the modelling framework set out in Sec. \[sec:parameterestimation\] and \[sec:posterior\] is given here:
- For $N_u=N_u^{\min},N_u^{\min}+1,\ldots,N_u^{\max}$, implement Algorithm \[alg:general\] with $N_f:=2$ and $X_f:=[X_{f_1},X_{f_2}]$ as inputs to step 1 to estimate $\widehat{\theta}_{N_u}$.
- For each $\widehat{\theta}_{N_u}$ compute the posterior in (\[equation:posteriorcurvature\]) and record the mean squared error $\mathbb{E}\left[(Z_{N_u}-\widehat{\mu}_{f_3})^2\right]$ where $Z_{N_u}\sim f(X_{f_3})|Y_{N_u},\widehat{\theta}_{N_u}$.
- Choose the value of $N_u$ corresponding to the lowest mean squared error.
- Implement Algorithm \[alg:general\] with this value of $N_u$, and by using $N_f:=3$ and $X_f:=[X_{f_1},X_{f_2},X_{f_3}]$ as inputs to step 1, to obtain the posterior in (\[equation:posteriorcurvature\]).
Figure \[figure:goldilocks\_principle\_2\] shows the posteriors $p(f(x)|Y_{N_u},\widehat{\theta})$ for the same simulated curvature data as used in Figure \[figure:goldilocks\_principle\], with $N_u=2$, $N_u=9$ and $N_u=14$ again. The simulated curvature samples at the last coordinate in $X_f$, that are excluded from the training of the models, are shown here. One can see the effect of the additional simulation coordinates, for the posteriors corresponding to $N_u=9$ and $N_u=14$, in imparting structure from the physics-based model. Figure \[figure:mean\_squared\_error\] shows the mean squared error for the posteriors corresponding to the range of $N_u$ values between 2 and 14 (again computed using the same simulated curvature data as above); this quantifies the predictive performance of the posteriors shown in Figure \[figure:goldilocks\_principle\]. By using the mean squared error to tune $N_u$ in the way described above, the value of $N_u$ could be used as a proxy to evaluate how well the data fits the physics-based model.
Once $N_u$ has been prescribed, the choice of the coordinates $X_u$ has to be decided. For the remainder of the paper they are chosen to be distributed uniformly along the entire domain $\mathcal{D}$. For a more advanced scheme, one could use active learning techniques [@Seo], which prescribes simulation coordinates that minimise the variance of aposteriori estimates. In addition to this, knowledge of where the curvature data are available, $X_f$, could lead to a wiser decision on the choice of simulation coordinates $X_u$. For example if all data were known to come from only a single region of the domain, simulations should be taken at points outside this region as to guide the inference from the data in these areas as much as possible.
![The posteriors $p(f(x)|Y_{N_u},\widehat{\theta})$ corresponding to the same simulated curvature data as used in Figure \[figure:goldilocks\_principle\], for $N_u=2$ (top), $N_u=9$ (middle) and $N_u=14$ (bottom).[]{data-label="figure:goldilocks_principle_2"}](goldilocks_principle_2){width="100mm"}
![The mean squared error of the posteriors $p(f(X_{f_3})|Y_{N_u},\widehat{\theta})$ with respect to $\widehat{\mu}_{f_3}$, for the range of $N_u$ values between 2 and 14.[]{data-label="figure:mean_squared_error"}](mean_squared_error){width="100mm"}
Detection of system changes {#sec:damagedetection}
---------------------------
This section will discuss the use of the methodology presented in the previous sections for detecting changes in the modelled system and the parameters within the physics-based model. In terms of the application of this paper, an underlying change in the properties of the sleeper (and it’s foundation) could signify potential damage. Therefore the methods discussed in this section can be used for structural damage detection. The aim of detecting system changes is to signal when the underlying physics (e.g. model parameters) has changed. To achieve this we use batches (length $B$) of curvature data $\mathcal{Z}_i=\big\{z_j(X_{f})\big\}_{j = iB+1}^{i(B+1)}$, for $i=0,1,2,3,\ldots$. Then $\widehat{\mu}_f^{(i)}$ are the sample means, in (\[equation:muestimates\]), of the batches. This can be implemented alongside the proposed methodology in this paper via two general approaches:
- **Change in estimated parameters:** Fit separate Gaussian process models for the curvature data $\mathcal{Z}_i$, for $i=0,1,2,3,\ldots$ using Algorithm \[alg:general\], and find estimates for $\theta$ denoted by $\widehat{\theta}^{(i)}$. Then statistically compare $\widehat{\theta}^{(i)}$ with $\widehat{\theta}$.
- **Change in log marginal likelihood:** Let $Y^{(i)}=[w(X_u,\widehat{\lambda}),\widehat{\mu}^{(i)}_f]$ and find the log marginal likelihood $\log p(Y^{(i)}|X_u,X_f,\widehat{\theta}):=L(\widehat{\theta};Y^{(i)})$. Statistically compare this to $\log p(Y_{\widehat{\theta}}|X_u,X_f,\widehat{\theta})$.
The former technique is commonly utilised in model updating, for example by using the relative error of $\widehat{\theta}^{(i)}$ with respect to $\widehat{\theta}$ as a metric to detect significant parameter changes over time [@Schommer]. On this note, many popular unsupervised change-point detection algorithms can be used alongside the two approaches mentioned to identify a sudden change in parameter estimates/marginal likelihood, such as those presented in [@Lau; @Gregory]. There are many aspects of these methods to consider, such as the uncertainty in the hypothesis test statistics and the choice of any hard-thresholds. For the results presented in the next section, we will use the metrics discussed here to assess changes in the modelled system over time.
Results {#sec:results}
=======
The following sections contain example implementations and results of the proposed methodology, for simulated and experimental data under laboratory conditions. These examples both consist of system change detection problems discussed in the previous section.
Prior knowledge about parameters {#sec:priors}
--------------------------------
To implement the estimation procedure in Sec. \[sec:parameterestimation\], for the construction of the Gaussian process models in the following numerical demonstrations, the regularization term $g(\theta)$ in (\[equation:regularizedoptimization\]) needs to be specified. In practice, this should be constructed using expert knowledge of the particular parameters, for example according to the manufacturing guidelines for a particular component of the instrumented sleeper system.
Simulated curvature data {#sec:simulatedresults}
------------------------
This section will demonstrate the effectiveness of the proposed methodology for detecting changes in the modelled system. A data set of 100 data-points are simulated, the first 50 curvature data-points are generated using the values $EI_0$ and $k_0$ for the flexural rigidity and the ballast stiffness respectively, and the remaining 50 are generated using the values $EI_1$ and $k_1$. Further we set the force value $p=125000$N. The simulated curvature data, $z_j(X_f)$, for $j=1,\ldots,100$, is generated via perturbations to a finite-difference approximation to (\[equation:curvature\]), namely for all $x \in X_f$, $$z_j(x)=
\begin{cases}
\begin{split}
\frac{p\lambda_0}{k_0}\big\{&w((x+1),\lambda_0)-2w(x,\lambda_0)+...\\
\quad & w((x-1),\lambda_0)\big\}
+ \xi,\end{split} &\text{for } j= 1,\ldots,50 \\
\begin{split}
\frac{p\lambda_1}{k_1}\big\{&w((x+1),\lambda_1)-2w(x,\lambda_1)+...\\
\quad & w((x-1),\lambda_1)\big\}
+ \xi,\end{split} &\text{for } j= 51,\ldots,100,
\end{cases}
\label{equation:simulateddataresults}$$ where $\xi \sim \mathcal{N}(0, 5\times10^{-7})$ and $\lambda_0=(k_0/(4EI_0))^{\frac{1}{4}}$ and $\lambda_1=(k_1/(4EI_1))^{\frac{1}{4}}$. The proposed multi-output Gaussian process model is trained using the first batch (length $B=5$) of data, $\mathcal{Z}_{0}$, in Algorithm \[alg:general\]. In order to quantify the uncertainty in detecting simulated parameter changes, given the uncertainty in the data-points, the log marginal likelihood and parameter estimates in the following sections will be computed over 20 independently sampled data sets.
### Changing the flexural rigidity {#sec:flexuralrigidity}
In the first case we set $k_0=k_1=450$, $EI_0=8 \times 10^{11}$ and $EI_1=6 \times 10^{11}$; these values are of the same order of magnitude to those prior beliefs (see \[equation:functionforregularization\]) for the sleeper system considered in this paper. Thus the flexural rigidity is reduced at the midpoint of the simulation and the ballast stiffness is kept fixed. A simulated reduction, as supposed to an increase, in the flexural rigidity is used to represent the cracking of concrete. The tuning procedure outlined in Sec. \[sec:adaptivenu\] is used to specify the value of $N_u$ used for the training of the proposed model and this results in $N_{u}=9$. Using the training data $\mathcal{Z}_0$ and Algorithm \[alg:general\], the estimates $\widehat{\theta}=[\widehat{k},\widehat{EI},\widehat{\sigma^{2}}]=[286.3, 5.0 \times 10^{11}, 13.6]$ are obtained. Due to the given variance in the simulated data-points from (\[equation:simulateddataresults\]), it is not expected that the estimates will be exactly the values prescribed, $k_0$ and $EI_0$. Next, the log marginal likelihood $\log p(Y^{(i)}|\widehat{\theta})$ is computed, recalling that $Y^{(i)}=[w(X_u,\widehat{\lambda}),\widehat{\mu}_f^{(i)}]$ from Sec. \[sec:damagedetection\], where $\widehat{\mu}^{(i)}_f$ is the sample mean in (\[equation:muestimates\]) of the batch $\mathcal{Z}_i$, for $i=1,...,19$. Figure \[figure:curvature\_likelihood\_sim\_data\] shows these log marginal likelihoods over all batches of data (and all 20 sampled data sets). Also shown in this figure are values of $\widehat{k}^{(i)}$ and $\widehat{EI}^{(i)}$ (over all data sets) obtained by fitting separate Gaussian process models on each batch of simulated data points $\mathcal{Z}_i$, for $i=1,\ldots,19$, using Algorithm \[alg:general\]. Interestingly here, the results correctly detect that $EI$ is reduced after the true change occurs at the 50’th data point and that $k$ stays approximately constant throughout. Also note the clear decrease in likelihood after the true change in $EI$ occurs.
![The curvature posterior mean and 95% confidence interval at points across the domain $\mathcal{D}$ for the Gaussian process model trained on the first batch of data $\mathcal{Z}_0$ (with $EI_0=8 \times 10^{11}$ and $k_0=450$). The first 5 simulated curvature data points for each coordinate in $X_f$ are shown by the blue dots.[]{data-label="figure:curvature_posterior_trained_sim_data"}](curvature_posterior_trained_sim_data){width="100mm"}
![The box-plots (over 20 independently sampled data sets) of the log marginal likelihood $\log p(Y^{(i)}|\widehat{\theta})$ and estimates $\widehat{k}^{(i)}$ and $\widehat{EI}^{(i)}$ for each batch $\mathcal{Z}_i$ ($i=1,...,19$) of simulated curvature data points. As the true change in $EI$ occurs after the 50’th data point, the likelihood decreases.[]{data-label="figure:curvature_likelihood_sim_data"}](curvature_likelihood_sim_data){width="100mm"}
### Changing the ballast stiffness {#sec:ballaststiffness}
In the second case we set $EI_0=EI_1=8 \times 10^{11}$, $k_0=450$ and $k_1=300$. Thus the ballast stiffness now changes at the midpoint of the simulation and the flexural rigidity is kept fixed. As done in the previous case with varying flexural rigidity, the log marginal likelihoods $\log p(Y^{(i)}|\widehat{\theta})$ in addition to the estimates $\widehat{k}^{(i)}$ and $\widehat{EI}^{(i)}$ are obtained for each batch of simulated curvature data points (over 20 independently sampled data sets) $\mathcal{Z}_i$, for $i=1,\ldots,19$. These quantities are all shown in Figure \[figure:curvature\_likelihood\_k\_change\_sim\_data\]. Once again the likelihoods and the values of $\widehat{k}^{(i)}$ significantly decrease after the true change in $k$ occurs. In addition to this behaviour, the new estimates $\widehat{EI}^{(i)}$ also stay approximately constant as expected.
![Same as Figure \[figure:curvature\_likelihood\_sim\_data\] only for a change in $k$ for the second set of 50 simulated data points.[]{data-label="figure:curvature_likelihood_k_change_sim_data"}](curvature_likelihood_k_change_sim_data){width="100mm"}
[ \[[[**I would introduce the signaling procedure here and add its results for this particular set of parameters - this will add formality to the procedure we are using and link with the next section better. The next section should be sorely reserved for the ROC analysis results.**]{}]{}\] ]{}
### Receiver operating characteristic analysis
A receiver operating characteristic (ROC) analysis [@Fawcett] will now be carried out to demonstrate the sensitivity of the log marginal likelihood shown in both Figures \[figure:curvature\_likelihood\_sim\_data\] and \[figure:curvature\_likelihood\_k\_change\_sim\_data\] to any changes in the parameters $EI$ and $k$ during the second set of 50 data points. This analysis evaluates how large the change in the parameters is required to be in order to produce a significant shift in the log marginal likelihood. The parameters are changed from $k_0=450$ and $EI_0=8 \times 10^{11}$ to $k_1$ and $EI_1$ respectively after the 50’th simulated data point as done in Sec. \[sec:simulatedresults\], and are varied as follows: $EI_1 \in [6 \times 10^{11}, 6.2 \times 10^{11}, 6.4 \times 10^{11},
\ldots, 8 \times 10^{11}]$ and $k_1 \in [300, 315, 330, \ldots,
450]$. Therefore there are $11^2$ different parameter combinations after the change in this analysis. We now define the signalling procedure, based on the log marginal likelihood $p(Y^{(i)}|\widehat{\theta})$, to flag changes in the model parameters. Fitting a normal distribution to the log marginal likelihood $\log p(Y_{\widehat{\theta}}|\widehat{\theta})$ from using the first batch of data points, $\mathcal{Z}_0$, to train the Gaussian process through Algorithm \[alg:general\], we define $\mu_0$ and $\sigma_0$ to be the empirical mean and standard deviation over 1000 independent data sets in (\[equation:simulateddataresults\]). Then the test-statistic for the signalling procedure we use is based on the log marginal likelihood $\log p(Y^{(i)}|\widehat{\theta})$, for $i=1,...,19$, and is given by $$p_i=\Phi\left( \log p(Y^{(i)}|\widehat{\theta}) |\mu_0, \sigma_0\right),$$ where $\Phi(\cdot|\mu_0,\sigma_0)$ is a Gaussian distribution function with mean $\mu_0$ and standard deviation $\sigma_0$. Then using a threshold $\gamma$ to signal a change in a log marginal likelihood, define the true positive detections as an actual change in the true parameters when one is signalled; i.e. $TP(\gamma)=\lvert \big\{i \in [10, 19]; p_i < \gamma\big\} \rvert/10$. Similarly, define the false positive detections as no actual change in the true parameters when one is signalled; i.e. $FP(\gamma)=\lvert \big\{i \in [0, 9]; p_i < \gamma\big\} \rvert/10$, where $p_0=\Phi\left(\log p(Y_{\widehat{\theta}}|\widehat{\theta})|\mu_0,\sigma_0\right)$. For a fixed parameter combination the detection rates are computed over 1000 independent data sets, $TP_j(\gamma)$ and $FP_j(\gamma)$, for $j=1,\dots,1000$: $$TPR(\gamma)=\frac{\sum^{1000}_{j=1}TP_j(\gamma)}{1000}, \quad FPR(\gamma)=\frac{\sum^{1000}_{j=1}FP_j(\gamma)}{1000}.
\label{equation:rates}$$ A ROC curve is the graph of $TPR(\gamma)$ against $FPR(\gamma)$ for varying $\gamma \in [\mu_0-0.5 \sigma_0,\mu_0-0.75 \sigma_0,\mu_0-\sigma_0,\ldots,\mu_0-2 \sigma_0]$. Figure \[figure:roc\_curves\] shows these curves for the parameter change combinations $(EI_1, k_1)$ of $(7.8 \times 10^{11}, 435)$ and $(7.8 \times 10^{11}, 420)$. The area under each curve (AUC), $\sum_{\gamma} TPR(\gamma)[1-FPR(\gamma)]$, in this case represents the probability of the detecting a random instance of a change in the parameters with respect to the test statistic used. The AUC for all parameter combinations is summarized in Figure \[figure:auc\]. This plot shows the magnitude of parameter changes that the detection utilising the log marginal likelihood is sensitive to; this includes the values of $EI_1$ and $k_0$ prescribed in Sec. \[sec:flexuralrigidity\] and \[sec:ballaststiffness\] respectively. The ROC analysis presented here has allowed us to evaluate the sensitivity of the log marginal likelihood from the proposed multi-output Gaussian process model to parameter changes within the physics-based model. For example, the AUC suggests that a true decrease of 10% in both parameters, i.e. $EI_1=0.9EI_0$ and $k_1=0.9k_0$, is 80% likely to be detected. Figure \[figure:sim\_curvature\_data\] show boxplots for the simulated curvature data, for each location in $X_f$, in (\[equation:simulateddataresults\]) corresponding to this 10% decrease in the parameters. Note that the 10% change in the true parameters is difficult to detect by visual inspection, however the signalling procedure detects this change with high probability.
{width="85mm"} \[figure:roc\_curves\]
{width="95mm"} \[figure:auc\]
![The boxplots of the simulated curvature data in (\[equation:simulateddataresults\]) for both $j \in [1,\ldots,50]$ and $j \in [51,100]$ and the parameters $EI_0=8 \times 10^{11}$, $k_0=450$, $EI_{1}=0.9 EI_0$ and $k_1=0.9k_0$. These are shown for $X_{f_1}$ (left), $X_{f_2}$ (middle) and $X_{f_3}$ (right).[]{data-label="figure:sim_curvature_data"}](simulated_curvature_data_boxplot){width="140mm"}
Experimental curvature data {#sec:experiment}
---------------------------
This section demonstrates the effectiveness of the proposed technique when applied to the experimental data. The laboratory experiment considered was introduced in [@Butler]. The experiment consisted of a single instrumented railway sleeper resting in the centre of a 3m wide and 0.4m deep ballast test-bed. This test setup was designed to replicate the conditions under which the sleeper would exhibit within an actual railway bed. The sleeper has the same specifications and sensor locations as considered throughout this paper (see Figure \[figure:schematic\]). A 0.6m deep steel spreader beam was positioned above the sleeper that applied a vertical actuator/hydraulic jack load through two rigid loading points (at $x_1$ and $x_2$) mounted on the sleeper. The experimental setup is shown in Figure \[figure:schematic4\]. The testing procedure involved applying a set of two equal vertical forces at a linearly increasing rate up until the point where damage (in the form of concrete cracking and/or significant ballast settlement) was visually observed. The entire test was completed in approximately 1300 seconds. The actuator forcing for the testing procedure is shown in Figure \[figure:forcing\]. For full specifications of the experiment and testing procedure, turn to [@Butler].
{width="105mm"} \[figure:schematic4\]
{width="90mm"} \[figure:forcing\]
The aim of this experiment is to estimate the flexural rigidity of the sleeper and the stiffness of the ballast beneath, and to obtain a posterior model for the curvature of the sleeper beam using the first 100 seconds of curvature from the instrumented sleeper. This is done using the technique proposed in this paper. From this model, the objective is then to detect at which time throughout the testing procedure a change to the system (e.g. reduction in flexural rigidity resulting from concrete cracking or reduction in ballast stiffness due to settlement/crushing) occurs. The approach taken for this detection is the second one listed in Sec. \[sec:damagedetection\]. For this example, the actuator forcing $p$ is known (shown in Figure \[figure:forcing\]). Figure \[figure:strain\_sleeper\_experiment\] shows the strain measurements from the FBG sensor located at $x=d/2$ on the top prestressing strand embedded in the sleeper during the testing procedure; a large increase at strain between 1100 and 1150 seconds was reported to likely be because of cracking along the top of the sleeper [@Butler]. The Gaussian process model is trained on (i.e. estimating $\widehat{\mu}_f$) the observed curvature data corresponding to the first 100 seconds of FBG strain measurements (5000 data points recorded at 50Hz) using Algorithm \[alg:general\]. The tuning procedure outlined in Sec. \[sec:adaptivenu\] is used to specify the value of $N_u$ used for the training of the Gaussian process model and this results in $N_{u}=5$. $EI$ was estimated as 9.07$\times 10^{11}$, and $k$ was estimated as 219.41. The parameter for the covariance kernels, $\sigma^2$, was estimated as 24.6. These estimates make up the set $\widehat{\theta}$.
The log marginal likelihood $\log p(Y^{(i)}|\widehat{\theta})$ is computed for each 1000-wide batch of curvature data points for the entire experiment after the training period (refer to Figure \[figure:log\_likelihood\_sleeper\_experiment\]). The likelihood very slightly decreases over time, but sharply drops between 1100 seconds and 1120 seconds (55000’th and 56000’th data points) signifying a change in the structural state of the system. This is inline with the change-point observed in [@Butler] and shown by the strain measurements in Figure \[figure:strain\_sleeper\_experiment\], and so the proposed modelling method works as expected alongside this experimental setup. It uses the drop-off in log marginal likelihood as a precursor for a visibly detected structural change (e.g. concrete cracking).
![The log marginal likelihood $\log p(Y^{(i)}|\widehat{\theta})$ of each 1000-wide batch of curvature data points given the parameters $\widehat{\theta}$ estimated during the first 1000 data points. This likelihood sharply decreases between 1100 and 1120 seconds.[]{data-label="figure:log_likelihood_sleeper_experiment"}](log_likelihood_sleeper_experiment){width="100mm"}
{width="85mm"} \[figure:strain\_sleeper\_experiment\]
{width="85mm"} \[figure:log\_likelihood\_sleeper\_experiment\]
Case study with a instrumented railway bridge in the UK
-------------------------------------------------------
- Bridge 5
- Estimating $EI$ via matching force.
Conclusions {#sec:conclusion}
===========
This paper has presented an important contribution to the structural identification and health monitoring community. The contribution takes the form of a multi-output data-centric engineering Gaussian process model that fuses together information from an analytical physics-based model for the deflection and observed data for the curvature of an instrumented railway sleeper supported on compacted ballast. This fusion is implemented on the level of the geometric relationship between these different physical quantities. The multi-output Gaussian process model provides aposteriori estimates to the curvature of the sleeper, conditioned on both the observed data and physics-based model. This approach allows the physics-based model to guide the inference from the data in unmeasured regions of the domain. The resulting methodology also produces estimates to parameters within the physics-based model, such as flexural rigidity and ballast stiffness. These are used for the detection of system changes; underlying changes in the parameters of the physics-based model can indicate potential structural damage.
This work builds upon a surge of recent work within the model updating community [@Vigliotti; @Schommer; @Alkayem]. The estimation in the material properties of beams, such as flexural rigidity, has been a major area of interest in such literature. As demonstrated in this study, these methods even have the potential to shed light on material properties (and therefore any changes) in components of physics-based models that are not measured by instrumented sensors such as the ballast supporting railway sleepers. A limitation of the proposed methodology is the assumption that the flexural rigidity utilised in the physics-based model is constant over the sleeper. A future aim for this line of research is to modify the methodology to relax this assumption, perhaps by assuming a functional form for the flexural rigidity (e.g. piecewise linear) and estimating the piecewise coefficients. The physics-based model used can also be generalised to something more sophisticated (e.g. FEM [@Zhou]) in order to model a larger structure, such as an entire railway bridge rather than only the sleeper component. In addition to this, aposteriori estimates of sleeper response obtained through the presented methodology could inform the optimal placement of sensors [@Guratzsch] and the detection of faulty sensors [@Hernandez], possibly through active learning. An attractive side effect of the proposed technique is the potential to use it as a method for signal-level sensor fusion [@Soman; @Hall] (the combination of data from multiple sensor networks). For such an implementation, the simulations from the analytical physics-based model in this work would be exchanged for observed data from another sensor network measuring vertical deflection, e.g. deflection transducers [@Rodrigues].
This paper demonstrates the proposed methodology using simulated and experimental curvature data. The experimental data is the product of a recent full-scale laboratory testing procedure for a railway sleeper supported on compacted ballast, outlined in [@Xu]. The testing procedure imparts forcing onto a sleeper instrumented with fibre optic based strain sensors until damage occurs. The proposed model in this paper detects such structural change via evaluating the marginal likelihood with future sensor data. Another important development made in this work is the process of balancing the information obtained from the observed data and the physics-based model used in training the multi-output Gaussian process model. This is done in order to improve the predictive performance of a test set of observed data assumed to be from unmeasured regions of the domain. This type of methodology is aimed at inspiring future work in the development and evaluation of data-centric engineering models.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by The Alan Turing Institute under the EPSRC grant EP/N510129/1 and the Turing-Lloyd’s Register Foundation Programme for Data-Centric Engineering. The authors would also like to acknowledge EPSRC and Innovate UK (grant no. 920035) for funding this research through the Centre for Smart Infrastructure and Construction (CSIC) Innovation and Knowledge Centre. Research related to installation of the sensor system was carried out under EPSRC grant no. EP/N021614. Mark Girolami is supported by a Royal Academy of Engineering Research Chair in Data Centric Engineering.
- Obtain the curvature data $\widehat{f}(X_f)$ and compute $Y_f$.
- Prescribe the simulation coordinates $X_u$.
- Estimate the parameters $\phi$ within the physics-based model using the curvature data.
- Combine both the physics-based model simulations and the observed curvature data in a joint multi-output Gaussian process model.
- Predict future curvature response of sleeper using model posterior.
Simulated annealing algorithm {#sec:appendixparamest}
=============================
Simulated annealing is an iterative probability-based method for optimization of a particular function $H(\theta)$ over the space of $\theta \in \Theta$, to find $\text{arg}\max_{\theta \in \Theta}H(\theta)$. The algorithm is used in Sec. \[sec:parameterestimation\] to train the proposed Gaussian process model. It is similar to the Metropolis-Hastings algorithm that is used for Markov chain Monte Carlo [@Dashti], in that it accepts/rejects candidate/proposed parameter values $\theta^*$ with a certain probability. This probability is based on the ratio of $H(\theta^*)$ to a previous iterate $H(\theta^{n-1})$, which is ‘cooled’ by some temperature $T$. This temperature decreases over time so that at early iterations candidates are frequently accepted even if $H(\theta^*)<H(\theta^{n-1})$, but towards the end of all prescribed iterations the candidates will only be accepted if $H(\theta^*) \geq H(\theta^{n-1})$. The ($\log$-based) version of the algorithm, that is used in Sec. \[sec:parameterestimation\], proceeds as follows with $H(\theta)=L(\theta|Y)+\log g(\theta)$:
1. Let $n=1$ and define a starting value $\theta^0$. Set a maximum number of iterations, say $N=1000$. Then define the temperature function $T_n=\exp(-\delta(n/N))$, with $\delta \geq 1$ a constant.
2. Iterate over $n=1,...,N$, and generate a candidate value of $\theta$, denoted by $\theta^{*}$, by the Gaussian random walk (with variance vector $\beta \in \mathbb{R}^{\abs{\theta}}_{>0}$), $$\theta^{*} = \theta_{i-1}+\xi, \qquad \xi \sim \mathcal{N}(0, \text{diag}(\beta)).$$
3. Find $\alpha=\min\left\{T^{-1}_n\left(H(\theta^*)-H(\theta^{n-1})\right),0\right\}$ and sample $u\sim \mathcal{U}[0,1]$.
4. If $\log u \leq \alpha$ then take $\theta^n=\theta^*$ and if $\log u > \alpha$ then take $\theta^n=\theta^{n-1}$.
The final estimate for the parameter $\theta$ is set to $\theta^N$.
The MCMC scheme to sample from $p(\theta|Y)$ and approximate statistics of it is described below, starting with the prescribed Markov chain element $\theta_{1}$.
1. For $i=2,...,N$, generate a candidate value of $\theta$, denoted by $\theta^{*}$ by the Gaussian random walk (with variance vector $\beta$), $$\theta^{*} = \theta_{i-1}+\xi, \qquad \xi \sim \mathcal{N}(0, \text{diag}(\beta)).$$
2. Generate the data $Y$ using the $\lambda$’s associated with $\theta^{*}$ and $\theta_{i-1}$ and compute the acceptance probability $$\alpha = \min(\left(l(\theta_{i-1})+\log(p(\theta_{i-1}))\right)-\left(l(\theta^{*})+\log(p(\theta^{*}))\right),0).$$
3. Sample $\gamma \sim \log(\mathcal{U}(0,1))$, and accept $\theta_{i}=\theta^{*}$ if $\gamma \leq \alpha$ or accept $\theta_{i}=\theta_{i-1}$ if $\gamma > \alpha$.
4. Once $i=N$, estimate $\mathbb{E}[p(\theta|Y)]$ via $$\bar{\theta}=\frac{1}{N}\sum^{N}_{i=1}\theta_{i} .$$
| {
"pile_set_name": "ArXiv"
} |
epsf
Introduction
============
The statement that our Universe is baryon asymmetrical as a whole is quite firmly established observational fact of contemporary cosmology. Indeed, if large regions of matter and antimatter coexist, then annihilations would take place at the borders between them. If the typical size of such a domain was small enough, then the energy released by these annihilations would result in a diffuse $\gamma$ –ray background, in distortions of the spectrum of the cosmic microwave radiation and light element abundance, neither of which is observed (see for review e.g. [@A]). Recent analysis of this problem [@1] for baryon symmetric Universe demonstrates that the size of regions should exceed $1000$ Mpc., being comparable with the modern cosmological horizon. It therefore seems that the Universe is fundamentally matter–antimatter asymmetric. However the arguments used in [@1] do not exclude the case when the Universe is composed almost entirely of matter with relatively small insertions of primordial antimatter. Thus we may expect the existence of macroscopically large antimatter regions in the Universe, that differs drastically from the case of baryon symmetric Universe. We call the region filled with antimatter in the baryon dominated Universe, as [*antizillah*]{}. Of course the existence of antizillahs is not rigorous requirement of baryosynthesis, but some modification of baryogenesis scenarios will result in formation of domains with different sign of baryon charge (see for example [@dolg]). The only condition which is necessary to satisfy is the amount of antibaryons within antizillahs must be small comparing to the total baryon number of the Universe.
At the first glance it is not difficult to have some amount of antizillahs if we simply suppose that in the early Universe when the baryon excess is generated the C–and CP–violation have different sign in different space regions [@B]. This may be achieved, for example, in models with two different sources of CP–violation, explicit and spontaneous [@lee] one. However, any spontaneous CP– violation processes are a result of early phase transition of first or second order what implies very small size of primordial antizillahs [@dolg]. For example if the antizillahs are formed in the second order phase transition, their size at the moment of formation is determined by $l_i\simeq 1/(\lambda T_c)$, where $T_c$ is so called Ginsburg temperature (the critical temperature at which the phase transition take place) and $\lambda$ is the selfinteraction coupling constant of field which breaks CP symmetry [@lee]. In the result of expansion the modern sizes of domains would reach $l_{0}\simeq l_{i}(T_{c}/T_{0})=1/(\lambda T_{0})\simeq 10^{-21}pc/\lambda$, where $T_{0}$ is the present temperature of the background radiation.
On the other hand it has been revealed [@we] that the average displacement of the antizillah’s boundary caused by annihilation with surrounding matter is about $0.5pc$ at the end of radiation dominated (RD) epoch. Therefore any primordial antizillah having initial size up to $0.5 pc$ or more at the end of RD stage is survived to the contemporary epoch and in the case of successive homogeneous expansion has the size $\simeq 1kpc$ or more. Any primordial antizillah with scale less then critical survival size $l_c\simeq 1 kpc$ at contemporary epoch must be eaten up by annihilation process. Thus it is the serious problem which any model with thermal phase transition encounters to create primordial antizillah with the size exceeding the critical survival size $l_c$ to avoid complete annihilation.
There is an additional problem for baryosynthesis with surviving antizillah’s sizes. The point is that any phase transition is accompanied by formation of topological defects. If we blow–up the region with different signs of charge symmetry, we automatically blow–up the scale of respective topological defect structure. If the structure decays sufficiently late in the observable part of Universe, the contribution of energy density of such topological defects could be sufficiently high to contradict with observations. It can be easily estimated that the structure with the scale corresponding to the survival size enters the horizon and starts to decay at $T\le 0.1MeV$, i.e. in the period of Big bang nucleosynthesis. To remove these unwanted relics sufficiently early it is necessary to have a mechanism for symmetry restoration. This mechanism implies that the baryogenesis is going on within rather narrow time interval [@dolgsilk; @tkachev].
In the present paper we have elaborated the issue for inhomogeneous baryosynthesis without the difficulties pointed above. The proposed approach is based on the mechanism of spontaneous baryogenesis [@cohen]. This mechanism implies the existence of complex scalar field carrying baryonic charge with explicitly broken $U(1)$ symmetry. The baryon/antibaryon number excess is produced, when the phase of this additional field moves along the valley of its potential [@cohen; @dolgmain].
It is supposed that the vacuum energy responsible for inflation is driven by any scalar inflaton field, and additional complex field coexists with the inflaton. Due to the fact that vacuum energy during inflational period is too large, the tilt of potential is vanished. This implies that the phase of the field behaves as ordinary massless Nambu–Goldstone (NG) boson and the radius of NG potential is firmly established by the scale of spontaneous $U(1)$ symmetry breaking. Owing to quantum fluctuations of massless field at the de Sitter background [@alstar; @lindebig47] the phase is varied in different regions of the Universe. When the vacuum energy decreases the tilt of potential becomes topical, and pseudo NG (PNG) field starts oscillate. As the field rolls down in one direction during the first oscillation, it preferentially creates baryons over antibaryons, while the opposite is true as it rolls down in the opposite direction. Thus to have globally baryon dominated Universe one must have the phase sited in the point, corresponding to the positive baryon excess generation, just at the beginning of inflation (when the size of the modern Universe crosses the horizon). Then subsequent quantum fluctuations can move the phase to the appropriate position causing the antibaryon excess production. If it takes place not too late after the inflation begins, the size of antizillah may exceed the critical surviving size $l_c$.
The main idea of proposed issue is based on the existence of quantum fluctuations along the effectively massless angular direction of baryonic charged scalar field. Thus, more general, the considered issue of generation of antizillahs is applicable practically to all mechanisms of baryogenesis where the number density and sign of baryon asymmetry depend on the angular component of complex scalar field. The advantage of the mechanism of spontaneous baryogenesis [@cohen] considered here is the quite simple unambiguous inflation dynamics of scalar field generated baryon charge. This fact allows to establish quantitatively definite relationship between the effects of inflation and generation of baryon (antibaryon) excess in inhomogeneous baryogenesis. However, this relationship may be too rigid for the realistic model of antimatter domain formation compatible with the whole set of astrophysical constraints. The consistent picture may need more sophisticated scenarios. The principal possibility for such scenario can be considered on the base of Affleck Dine (AD) [@afflekdine] baryogenesis mechanism that still receives a lot of attention [@afflekdine; @lisa; @b1; @b2; @b3; @mr].
AD baryogenesis also involves the cosmological evolution of effective scalar field, which carries baryonic charge, being composed of supersymmetric partners of electrically neutral quark and lepton combinations. The important feature of supersymmetric extensions of standard model is the existence of “flat directions” in field space, on which the scalar potential vanishes [@nil; @afflekdine; @lisa]. We will refer for the definiteness to the flat directions of minimal standard supersymmetric model (MSSM) [@lisa; @mr1]. Thus, if the some component of scalar field lies along a flat direction, this component can be considered as a free massless complex scalar so called AD field [@afflekdine; @lisa]. At the level of renormalizable terms, “flat directions” are generic, but supersymmetry breaking and nonrenormalizable operators lift the “flat directions” and sets the scale for their potential. During the inflational period the AD field develops non–zero vacuum expectation value and subsequently when the Hubble rate becomes of the order of the curvature of AD potential, the condensate starts to oscillate around its present minimum. Baryon asymmetry can be induced in such condensate only if there exists phase shift between real and imaginary parts of the AD field. Such shift and consequently B and CP violation is provided by A–term in the potential which parameterizes MSSM “flat direction” [@afflekdine; @lisa]. The resulting sign and number density of baryon asymmetry depends on the magnitude of initial phase of AD field and on phase shift created by A-term at the relaxation period [@afflekdine; @lisa; @b1]. Therefore the de–Sitter fluctuations can generate antizillahs in the baryon asymmetric Universe in the similar way to the spontaneous baryogenesis if the angular direction of AD field is characterized the mass that is much smaller that the Hubble constant $H$ during inflation. It takes place if there are no of order $H$ corrections to the A–term [@mr].
The early dynamics of AD field is quite complicated [@b3] owing to the non–trivial background energy density driving inflation in MSSM. Moreover AD potential can get corrections from the vacuum energy that removes its minimum from the original one [@lisa; @b1; @b3; @mr]. In general there are two types of inflation in MSSM, D–term or F–term inflation (see for review [@riot]), depending on the type of vacuum contributing the energy density during de-Sitter stage. In the case of D–term inflation AD fields and inflaton slow roll coherently [@b3] (in the absence of order $H^2$ corrections to the mass squared term of AD potential). It implies that the radius of effectively massless angular AD direction is determined by the immediate value of inflaton field. For the case of F–term inflation the AD scalar will get an order $H^2$ negative mass squared term [@lisa; @b1; @b3; @mr] causing the minimum of AD potential. The AD field is closed to the minimum during the F–inflation stage [@b3] and this minimum determines the radius of circle valley of effectively massless angular direction.
The conclusion from this explicit example based on the MSSM is following. For any complicated inflation dynamics of baryon charged field it is possible to simulate appropriate massless direction that behaves similar to the circle valley of NG potential. This fact makes the proposed issue for generation of antizillahs viable not only for spontaneous baryogenesis mechanism, but for the all mechanisms dealing with effectively massless angular directions during inflation [@other].
The paper is organized as to following. In section II we discuss the quantum behavior of nondominant $U(1)$ symmetric scalar field at the inflation period. We estimate the amplitude and space scale of fluctuations of the phase for this field without PNG tilt. The size distribution of these fluctuations determines the size distribution of antizillahs. The section III contains calculations of baryon/antibaryon net excess production at the relaxation of phase when the tilt of Mexican hat potential becomes topical. We summarize our conclusions and discuss some problems of the considered scenarios in section IV.
\[PERT\] Phase Distribution for NG Field at The Inflation Period
================================================================
We start our consideration with the discussion of evolution of $U(1)$ symmetric scalar field which coexists with inflaton at the inflation epoch. The quantum fluctuations of such field during the inflation stage cause the perturbations for the phase marking the Nambu–Goldstone vacuum. In our model this phase determines the sign and value of baryon excess, so the size distribution of domains containing the appropriate phase values, caused by that fluctuations, coincide with the size distribution of antizillahs.
Thus to estimate the number density of antimatter regions with sizes exceeding the critical survival size $l_c$ in the baryogenesis model under consideration we have to deal with long – wave quantum fluctuations of the NG boson field at the inflation period. Various aspects of this question have been examined in the numerous papers [@kofm; @lindekofm; @spokyok; @lindelyth; @lindeax; @lindebook; @lyth; @lythstewart; @lindebig] in the connection with cosmology of invisible axion. Also the de–Sitter quantum fluctuations have been analyzed in the framework of AD baryogenesis [@b2; @b3].
The effective potential of the complex field is taken in the usual form
\[3phase\] V()=-m\_\^2\^\*+\_(\^\*)\^2+V\_0, where the field $\chi$ can be represented in the form
\[phase\] =
The $U(1)$ symmetry breaking implies that the radial component of the field $\chi$ acquires a nonvanishing classical part, $f=m_{\chi}/\sqrt{\lambda_{\chi}}$ and field $\alpha$ in eq. (\[phase\]) becomes a massless NG scalar field with a vanishing effective potential, $V(\alpha )=0$. In this case, $\chi$ has the familiar Mexican–hat potential, and the degenerated vacua correspond to the circle of radius $f$. Throughout present paper we deal with dimensionless angular field $\theta =\alpha /f$.
We concern here the possibility to store appropriate phase value in the domain with the size exceeding the critical survival size. Such value of phase plays the role of starting point for clockwise movement, which is going to generate antibaryon excess when the tilt of potential breaking $U(1)$ explicitly, will turn to be topical.
We assume that the Hubble constant varies slowly during inflation. Also we use well established behavior of quantum fluctuations on the de Sitter background [@lindebook]. It implies that vacuum fluctuations of every scalar field grow exponentially in the inflating Universe. When the wavelength of a particular fluctuation becomes greater than $H^{-1}$ the average amplitude of this fluctuation freezes out at some nonzero value because of the large friction term in the equation of motion of the scalar field, whereas its wavelength grows exponentially. In the other words such a frozen fluctuation is equivalent to the appearance of classical field that does not vanish after averaging over macroscopic space intervals. Because the vacuum must contain fluctuations of every wavelengths, inflation leads to the creation of more and more new regions containing the classical field of different amplitudes with scale greater than $H^{-1}$. The averaged amplitude of such NG field fluctuations generated during each time interval $H^{-1}$ is given by [@alstar] \[b11\] = During such time interval the universe expands by a factor of $e$. Since the NG field is massless during inflation period (the PNG tilt is vanish yet), one can see that the amplitude of each frozen fluctuation is not changed in time at all and the phases of each wave are random. Thus the quantum evolution of NG field looks like one–dimensional Brownian motion [@lindebook; @lindebig] along the circle valley corresponding to the bottom of NG potential. This statement implies that the values of the phase $\theta$ in different regions become different, and the corresponding variance grows as [@lindebig47] \[b12\] ()\^2= that means that dispersion grows as $\sqrt{\langle (\delta\theta )^2\rangle}=\frac{H}{2\pi f}\sqrt{N}$, where N is the number of e–folds. In the other words the phase $\theta$ makes quantum step with the scale $\frac{H}{2\pi f}$ at each e–fold, and the total number of steps during some time interval $\Delta t$ is given by $N=H\Delta t$.
Let us consider the scale $k^{-1}=H_0^{-1}=3000h^{-1}Mpc$ which is the biggest cosmological scale of interest. We suppose that Universe is baryon asymmetric in this scale which leaves the horizon at definite e–fold $N=N_{max}$. On the other side this scale is the one entering the horizon now, namely $a_{max}H_{max}=a_0H_0$ where the subscript $0$ indicates the contemporary epoch. This implies that: \[b13\] N\_[max]{}=- the subscript $end$ denotes the epoch at the end of inflation. The slow-roll paradigm tells us that the last term of (\[b13\]) is usually $\le 1$. The first term depends on the evolution of scale factor $a$ between the end of slow-roll inflation and the present epoch. Assuming that inflation ends by short matter dominated period, which is followed by RD stage lasting until the present matter dominated era begins, one has [@LythLiddlePhysRep] \[b14\] N\_[max]{}=62-- , where $\rho_{reh}^{1/4}$ is the reheating temperature when the RD stage is established. With $H_{end}\simeq 10^{13}GeV$ and instant reheating this gives $N_{max}\approx 62$, the largest possible value. However, if one has to invoke supersymmetry to prevent the flatness of the inflation potential, for example like as in the case of AD baryogenesis, the $\rho_{reh}^{1/4}$ should not exceed then $10^{10}GeV$ to avoid too many gravitino overproduction [@KhlopovLinde], and one have $N_{max}=58$, perhaps the biggest reasonable value. Through the paper we will use $N_{max}=60$. The smallest cosmological scale of antizillah that is survived after annihilation is $k^{-1}_{c}=l_{c}\approx 8h^2kpc$ [@we]. It is $9$ order of magnitude smaller then $H^{-1}_0$, that corresponds to \[b15\] N\_cN\_[max]{}-13-345 Thus the $l_c$ should left horizon at 45–folds before the end of inflation.
=5.1cm\
Let us assume that the phase value $\theta =0$ corresponds to South Pole of NG field circle valley, and $\theta =\pi$ corresponds to the opposite pole. The positive gradient of phase in this picture is routed as anticlockwise direction, and the dish of PNG potential would locate at the South Pole of circle (see fig.\[fig1\]). It will be shown below (see section III) that the antibaryon production corresponds to the regions that would contain phase values caused anticlockwise rolling of PNG field $\alpha$ during the first half period of oscillation. If the field $\alpha$ rolls clockwise towards the dish of tilted potential just after the start of first oscillation then baryon production will take place.
Now we are in the position to estimate the fraction of the Universe containing antizillahs. To ensure that the Universe would be baryon asymmetric as a whole it is necessary to suppose that the phase average value $\theta =\theta_{60}$ within biggest cosmological scale of interest emerging at the $N_{max}=60$ e–folds before the end of inflation is located in the range $[0,\pi ]$. The $\theta_{60}$ is the starting point for Brownian motion of the phase value along the circle valley during inflation. As it has been mentioned above, the phase makes Brownian step $\delta\theta =\frac{H}{2\pi f}$ at each e–fold. Because the typical wavelength of the fluctuation $\delta\theta$ generated during such timescale is equal to $H^{-1}$, the whole domain $H^{-1}$, containing $\theta_{60}$, after one e–fold effectively becomes divided into $e^3$ separate, causal disconnected domains of radius $H^{-1}$. Each domain contains almost homogeneous phase value $\theta_{60-1}=\theta_{60}\pm\delta\theta$. In half of these domains the phase evolves towards $\pi$ (the North Pole) and in the other domains it moves towards zero (the South Pole). To have antizillah with appropriate sizes to avoid full annihilation one should require that the phase value crosses $\pi$ or zero not later then after $15$ steps. Only in this case the antizillahs would have the sizes larger than $l_c$ and are conserved up to the modern era. This means that one of the two following inequality must be satisfied \[b16\] -\_[60]{} Consider initially the case of exact equalities in expression (\[b16\]) when the main part of antimatter is contained in the antizillahs of size $l_c$. The number of domains containing the equal values of phase at the $45$ e–folds before the end of inflation is given by the following expression \[b17a\] n\_[45]{}(e\^3/2)\^[15]{}10\^[15]{}. Then the probability that every domain of size $l_c$ would not be separated into $e^3$ domains with size one order of magnitude less then $l_c$ at the next e–fold is given by $P_s\approx (1/2)^{e^3}\approx 10^{-6}$. Thus the number of domains serving as the prototypes for antizillahs of size $l_c$ looks like \[b18\] |n=n\_[45]{}P\_s10\^9 There are about $10^{11}$ galaxies in the Universe. Thus, according to such simple consideration, we reveal that $1\%$ of volume boxes corresponding to each galaxy contains the region of size $l_c$ filled with antimatter of highest possible antibaryonic density if the $\theta_{60}$ coincides with left side of inequality (\[b16\]) or lowest one in the case if the opposite equality is held.
We are able also to find the size distribution for antizillahs. For this purpose it is necessary to study the inhomogeneities of phase induced by (\[b11\]). It has been well established that for any given scale $l=k^{-1}$ large scale component of the phase value $\theta$ is distributed in accordance with Gauss’s law [@alstar; @lindebig47; @lindebook; @lindebig]. The quantity which will be especially interesting for us is the dispersion (\[b12\]) for quantum fluctuations of phase with moments from $k=H^{-1}$ to $k_{min}=l^{-1}_{max}$ (where the $l_{max}$ is the biggest cosmological scale that corresponds to $60$ e–folds). This quantity can be expressed in the following manner \[b19\] \^2\_l=\_[k\_[min]{}]{}\^kd=\
, where $N_l$ is the number of e–folds which relates the biggest cosmological scale to the given scale $l$. This means that the distribution of phase has the Gaussian form \[b21\] P(\_l ,l)=
Suppose that at e–fold $N_t$ before the end of inflation the volume $V(\bar\theta ,N_t)$ has been filled with phase value $\bar\theta$. Then at the e–fold $N_{t+\Delta t}=N_t-\Delta N$ the volume filled with phase $\bar\theta$ will follow iterative expression $$\begin{aligned}
\label{iteration}
V(\bar\theta , N_{t+\Delta t})=e^3V(\bar\theta ,N_t)+\nonumber\\
+(V_U(N_t)-e^3V(\bar\theta ,N_t)P(\bar\theta , N_{t+\Delta t})h.\end{aligned}$$ Here the $V_U(N_t)\approx e^{3N_t}H^{-3}$ is the volume of the Universe at $N_t$ e–fold. Expression (\[iteration\]) makes it possible to calculate the size distributions of domains filled with appropriate value of phase numerically. In order to illustrate quantitatively the number distribution of domains, we present here the numerical results for definite values of $\theta_{60}$, $\bar\theta$ and $h=\frac{H}{2\pi f}$. The table contains the results concerning to number of domains with average phase $\bar\theta$ at e-fold number $N$,
------ ------------------------------- -------------------
$N$ $N_{antizillahs}$ $L_{antizillah}h$
$59$ $0$ $1103$Mpc
$55$ $5.005 \, \times \, 10^{-14}$ $37.7$Mpc
$54$ $7.91 \, \times \, 10^{-10}$ $13.9$Mpc
$52$ $1.291 \, \times \, 10^{-3}$ $1.9$Mpc
$51$ $0.499$ $630$kpc
$50$ $74.099$ $255$kpc
$49$ $8.966 \, \times \, 10^{3}$ $94$kpc
$48$ $8.012 \, \times \, 10^{5}$ $35$kpc
$47$ $5.672 \, \times \, 10^{7}$ $12$kpc
$46$ $3.345 \, \times \, 10^{9}$ $4.7$kpc
$45$ $1.705 \, \times \, 10^{11}$ $1.7$kpc
------ ------------------------------- -------------------
: The sample of distribution of proto–antizillahs by sizes and numbers of e–folds at $ \theta _{60}=\frac{\pi }{6}$; $\overline\theta =-0$; $h=0.026$
The fraction of the Universe filled with phase $\bar\theta$ appears to be equal to $7.694\times 10^{-9}$. Thus we see that the distribution of domains with size is very abrupt and should be peaked at smallest value of size. Adjusting the free parameters $\theta_{60}$ and $h$ we are able to achieve the situation that volume box corresponding to each galaxy contains ($1\div 10$) regions with appropriate phase $\bar\theta$. The sizes of such regions are larger or equal to critical surviving size. In spite of the sufficiently large total number of antizillahs only the small part of our Universe will be occupied by antizillahs (see the last line in the presented table).
The nontrivial question on the actual forms of astrophysical objects antizillahs can have in the modern Universe needs spacial analysis, which, in general, strongly depends on the assumed form of the nonbaryonic dark matter, dominating in the period of galaxy formation. However, based on the early analysis [@we; @khlop; @ams] the two extreme cases can be specified, when the evolution of antizillahs is not strongly influenced by the dark matter content. In the first case, the antibaryon density within the antizillah is by an order of magnitude higher than the average baryon density, so that the over-density inside this region can exceed the dark matter density and rapid evolution of such an antizillah with the size exceeding the surviving scale can provide formation of compact antimatter stellar system (globular cluster (see for review [@glob])) which can survive in galaxy [@khlop; @ams]. The other extreme case is antizillah with extremely low internal antibaryon density $\Omega_{\bar B}<10^{-5}$. Then the diffused antiworld is realized, when no compact antimatter objects are formed and antizillahs evolve into low density antiproton-positron plasma regions in voids outside the galaxies [@we; @khlop].
\[BARYONS\] Spontaneous Baryogenesis Mechanism
===============================================
The following element of our scenario of inhomogeneous baryogenesis should contain a conversion of the phase $\theta$ into baryon/antibaryon excess. We consider the ansatz of spontaneous baryogenesis mechanism [@cohen]. The basic feature of this mechanism is that the sign of baryon charge created by relaxation of energy of PNG field critically depends on the direction that the phase is rotated on the bottom of Mexican heat potential. It provides us to convert the domains containing the appropriate phase value, caused by fluctuations, to the antizillahs at the period when the NG potential gets the tilt.
The one of reasonable issue to the spontaneous baryogenesis [@cohen] has been considered in the work [@dolgmain]. Let us briefly discuss it. It was assumed that in the early Universe a complex scalar field $\chi$ coexists with inflaton $\phi$ responsible for inflation. This field $\chi$ has non vanishing baryon number. The possible interaction of $\chi$ that violates lepton number can be described by following Lagrangian density (see e.g. [@dolgmain]) $$\begin{aligned}
\label{1}
L=-\partial_{\mu}\chi^*\partial^{\mu}\chi -V(\chi )+i\bar
Q\gamma^{\mu}\partial_{\mu}Q +
i\bar L\gamma^{\mu}\partial_{\mu}L - \nonumber\\ -m_Q\bar QQ-
m_L\bar LL +
(g\chi\bar QL+h.c.)\end{aligned}$$ The fields $Q$ and $L$ could represent heavy quark and lepton, coupled to the ordinary quark and lepton matter fields. Since fields $\chi$ and $Q$ possess baryon number while the field $L$ does not, the couplings in the (\[1\]) violate lepton number [@dolgmain]. The $U(1)$ symmetry that corresponds to baryon number is expressed by following transformations \[2\] ,QQ,LL The effective Lagrangian density for $\theta$, $Q$ and $L$ eventually has the following form after symmetry breaking [@dolgmain] $$\begin{aligned}
\label{5}
L=-\frac{f^2}{2}\partial_{\mu}\theta\partial^{\mu}\theta
+i\bar Q\gamma^{\mu}\partial_{\mu}Q +
i\bar L\gamma^{\mu}\partial_{\mu}L - \nonumber\\ -m_Q\bar QQ
-m_L\bar LL +
(\frac{g}{\sqrt{2}}f\bar QL+h.c.)+\partial_{\mu}\theta\bar
Q\gamma^{\mu}Q\end{aligned}$$ At the energy scale $\Lambda <<f$, the symmetry (\[2\]) is explicitly broken and the Mexican–hat circle gets a little pseudo NG tilt described by the potential \[tilt\] V()=\^4(1-) This potential, of high $2\Lambda^4$, has a unique minimum at $\theta =0$. Of course, in the most cases, the potential (\[tilt\]) is the lowest–order approximation to a more complicated expressions emerged from particle physics models (see e.g. [@freese] and Refs. therein).
The important parameter for spontaneous baryogenesis is the curvature of (\[tilt\]) in the vicinity of its minimum, which is determined by the mass of PNG field \[mass\] m\^2\_= As it was mentioned above the field $\chi$ is an additional field with nondominant energy density contribution to the Habble constant deriving by de Sitter stage. Suppose that the tilt was formed during inflation. Then the order of magnitude estimation for fluctuations induced by large– scale inhomogeneity of oscillations of the field $\chi$ gives $\frac{\delta
T}{T}=\frac{1}{3}\frac{\delta \rho}{\rho}~(\Lambda /T)^4$. Thus, for $T=H/2\pi$ and reasonable value $\Lambda\simeq 10^{-5}H$ (see the end of this section) the thermal electromagnetic background fluctuations are within the observational limits.
Also we assume that the field $\theta$ behaves as massless NG field during inflation implying that the condition \[cond\] m\_<<H is valid, where the $H$ is the Hubble constant during the inflation. After the end of inflation condition (\[cond\]) is violated and the oscillations of field $\theta$ around the minimum of potential (\[tilt\]) are started. The energy density $\rho_{\theta}\simeq\theta^2_im_{\theta}^2f^2$ of the PNG field which has been created by quantum fluctuations of $\theta$ during the inflation converts to baryons and antibaryons [@cohen; @dolgmain]. The sign of baryon charge depends on the initial value of phase from which the oscillations are started.
Let us estimate the number of baryons and antibaryons produced by classical oscillations of field $\theta$ with an arbitrary initial phase $\theta_i$. The appropriate expression for the density of produced baryons (antibaryons) $n_{B(\overline{B})}$ is represented in [@dolgmain] \[n1\] n\_[B()]{}=\_[m\_[Q]{}+m\_[L]{}]{}\^[ ]{} d| \_[-]{}\^dt(t)e\^[2it]{}|\^[2]{}, that is valid if $\chi (t\rightarrow -\infty )=\chi (t\rightarrow +\infty )=0.$ General case can be obtained in the limits $\chi (t\rightarrow -\infty )\neq 0;\chi (t\rightarrow +\infty )=0$ without loss of generality. After integration by part expression (\[n1\]) has the form \[bar\] N\_[B(|B)]{}=\_[\_[i]{}]{} \_[m\_[Q]{}+m\_[L]{}]{}\^d | \_[-]{}\^d()e\^[ 2i]{}| \^[2]{}, where the $\Omega_{\theta_i}$ is the volume containing the phase value $\theta_i$. Here the surface terms appear to be zero at $t=\infty$ due to asymptotic of field $\chi $ and at $t=-\infty $ due to Feynman radiation conditions.
For our estimations it is enough to accept that the phase changes as \[osc\] (t)\_i(1-m\_t) during first oscillation. We also put $m_Q=m_L=0$ that is reasonable for estimations. Substituting (\[osc\]) and (\[phase\]) into (\[bar\]) we come to \[bar1\] N\_[B(|B)]{} \_[\_i]{}\^2\_i\_\^d , where the sign in the lower limit of integral corresponds to baryon or antibaryon net excess generation respectively. The reasonability of our approximation follows from comparison of (\[bar1\]) at small $\theta_i<<1$ \[barapprox\] N\_B-N\_[|B]{}= \_[\_i]{}\^3\_i with the result of [@dolgmain].
Using for spatially homogeneous field $\chi=\frac{f}{\sqrt{2}}e^{i\theta }$ the expression for baryon charge \[rub1\] Q=i(\^d/dt-d\^/dt)=-fd/dt, one can easily conclude that $Q>0$ if $\theta >0$ during classical movement of phase $\theta $ to zero. Thus the anticlockwise rotation gives rise to antibaryon excess while the clockwise rotation to the baryon excess one.
During reheating, the inflaton energy converts into the radiation. It is assumed that reheating takes place when the Mexican–hat potential is not sensitive to the PNG tilt yet. This implies that the total decay width of inflaton $\Gamma_{tot}$ into light degrees of freedom exceeds the mass $m_ {\theta}$. In the other words the reheating is going on under the condition (\[cond\]). The relaxation of $\theta$ field starts when $H\approx m_{\theta}$ and converts to the baryons or antibaryons. Baryonic charge is converted inside a comoving volume after reheating owing to very effective decay during the cosmological time. This means that the baryon–to–entropy ratio in $n_{B(\bar B)}/s=Const$ in the course of expansion. The entropy density after thermalization is given by \[entropy\] s=g\_\*T\^3 where $g_*$ is the total effective massless degrees of freedom. Here we concern with the temperature above the electroweak symmetry breaking scale. At this temperature all the degrees of freedom of the standard model are in equilibrium and $g_*$ is at least equal to $106.75$. The temperature is connected with expansion rate as follow \[habble\] T== The last part of expression (\[habble\]) takes into account that the relaxation starts under the condition $H\approx m_{\theta}$. Using the formulas (\[bar1\]), (\[entropy\]), (\[habble\]) we are able to get the baryon/antibaryon asymmetry \[asym\] =()\^[3/2]{} Y(\_i) The function $Y(\theta)=\theta ^2
\int\limits_{-\theta /2}^{\theta /2}d\omega\frac{\sin^2{\omega}}{\omega^2}$ takes into account the dependence of amplitude of baryon asymmetry and its sign on the initial phase value in the different space regions during inflation.
The expression (\[asym\]) allows us to get the observable baryon asymmetry of the Universe as a whole $n_B/s\approx 3\cdot 10^{-10}$. In the model under consideration we have supposed initially that $f\ge H\simeq 10^{-6}m_p$. The natural value of coupling constant is $g\le 10^{-2}$. We are coming to observable baryon asymmetry at quite reasonable condition $f/\Lambda\ge 10^5$ (see e.g. [@freese]).
\[DISCUSSION\] Discussion
=========================
In this paper we have proposed a model for inhomogeneous baryosynthesis on the base of the spontaneous baryogenesis mechanism [@cohen]. The model predicts the generation of [*antizillahs*]{} with sizes exceeding the critical surviving size. The antibaryon number density relative to background baryon density in the resulting antizillahs and its number depends on the value of phase established at the beginning and on the parameters of PNG field potential. It is possible to have one or several antizillahs the volume box corresponding to every galaxy depending on the parameter values. The observational consequences of existence of antizillahs and the restrictions on their number and sizes have been analyzed in papers [@we; @khlop; @ams]
Of course we may in general expect that some region with size exceeding $l_c$ would contain antibaryon excess after the annihilation of small primordial domains and antidomains contained in this region is completed. However the probability to have such region is suppressed exponentially. Therefore to have observational acceptable number of antimatter regions [@khlop] with the size exceeding the critical survival size, a superluminous cosmological expansion in the formation of primordial antimatter proto–domain seems necessary.
As we have mentioned, the additional problem for the most models of inhomogeneous baryogenesis invoking phase transitions at the inflation epoch is prediction of the large scale unwanted topological defects. Our scheme contains the premise for existence of domain walls too. Such walls are not formed when the only minimum of PNG potential exists, what corresponds in the considered model to the fluctuations around $\theta
=0$, when the North pole ($\theta =\pi$) is not crossed. But in the case, when such crossing takes place the multiple degeneracy of vacua appears (e.g. vacua with $\theta =0$ and $\theta =2\pi$). The equation of motion that correspondes to potential (\[tilt\]) admits kink–like, domain wall solution, which interpolates between two adjacent vacua. Thus, when the PNG tilt is significant, domain wall is formed along the closed surface (e.g. $\theta =\pi$) [@kim]. In the other words every antizillah with high relative antibaryon density will be encompassed by domain wall bag. The wall stress energy $\Delta\approx 8f\Lambda^2$ [@kim; @sik] leads to the oscillation of wall bag after the whole bag enters the cosmological horizon. During the oscillations the energy stored in the walls is released in the form of quanta of NG field and gravitational waves. As we are taken $0<\theta_{60}<\pi$, the wall’s bag will have the scale of the order of modern horizon, if the dispersion $\sigma_{l_{max}}$ is large as $\pi -\theta_{60}$. Owing to very large oscillation period such big wall bag would presumably survive to the present time, which would be cosmological disaster [@lyth; @lythstewart]. Thus the upper limit on the dispersion will be $\sigma_{60}<\pi$. From the other hand this condition should be valued if we want to have parameters of antizillah population that do not contradict to direct and indirect observational constraints [@1]. It means that we will have wall bags with the sizes less then cosmological horizon and that walls had to decay until present time. The mechanisms of their decay is a subject of separate paper, in which we also plan to obtain additional constraints on the model, which follow from the condition that walls do not dominate within the cosmological horizon before the bag decays. If the energy density of walls is sufficiently high to give local wall dominance in the border region before the bag enters the horizon, it is of interest to analyze the role of superluminous expansion in the border regions in the bag evolution (see e.g. [@vil]). The interesting question on the wall interaction with baryons in the course of wall contraction and decay will be also studied separately.
In general all baryogenesis models that are able to generate some amount of antimatter regions look like radical limit of models with local baryon number density fluctuations so called isocurvature fluctuations [@LythLiddlePhysRep; @iso]. It is known that the contribution of isocurvature fluctuations to the cosmic microwave background (CMB) anisotropy obeys to $\frac{\delta T}{T}=-\frac{1}{3}\frac{\Omega_B}{\Omega_0}\delta_{B_i}$, where $\delta_{B_i}$ is the amplitude of initial baryon number fluctuations and ${\Omega_0}$ (${\Omega_B}$) are the total (baryon) density (in units of critical density). As it follows from our numerical illustration (see \[PERT\] and expression (\[bar1\])) we must have quite large amplitude of initial baryon number fluctuations $\delta_{B_i}\sim h/\theta_{60}\simeq 10^{-2}$ at the biggest cosmological scales, and consequently we would have large amplitude of isocurvature fluctuations at large scales that contradicts with COBE measurements [@iso].
To be keeping away of the problem of large–scale isocurvature fluctuations, we can, for example, prevent the fluctuations of phase at largest cosmological scales. The point is that to have antizillah with size exceeding few kpc. we do not need to start phase fluctuations at the e–folds that correspond to the biggest cosmological scales. It is sufficiently to start fluctuations of phase from the moment, for instance, when the scale $8h^{-1}Mpc$ leaves Habble horizon during inflation, namely after the $6.2$ e–folds from the beginning of inflation. We took this scale, because it is known that at the scale less then $8h^{-1}Mpc$ we could be generated initial baryon number fluctuations at the level $\delta_{B_i}\simeq 10^{-2}\div 10^{-3}$ without any contradictions with observations.
One of the natural way to prevent the phase fluctuations at the early inflation is to keep $U(1)$ symmetry restored during first $7$ e–folds. The mechanism that is able to restore symmetry during inflation has been consider in the works [@lindekofm; @lindeax; @lindebook; @sakhlop]. According to that works we can introduce interaction between inflaton field $\phi$ and field $\chi$. The simple potential of such kind may be chosen as $
V(\phi ,\chi )= \frac{1}{4}\lambda_{\phi}\phi^4+V(\chi )+\nu\phi^2\chi^*\chi
$, where $\nu =m_{\chi}^2/cM_p^2$, and $c\simeq 1$. The effective mass of the field $\chi$ depends on $\phi$ directly $m_{\chi}^2(\phi )= m_{\chi}^2+\nu\phi^2$. One considers here for simplicity the case $\nu =m_{\chi}^2/cM_p^2$. This implies that the effective value of mass $m_{\chi}^2(\phi )$ during inflation is given by $\nu (\phi^2-cM_p^2)$ and is positive because of very large value of the inflation field. It means that our $U(1)$ symmetry is restored during the period when the amplitude of the inflaton field exceeds $\phi_c=\sqrt{c}M_P$, and the field $\chi$ settles into the minimum of its symmetric potential. During this period there was no NG boson valley and phase fluctuations. After the moment that inflaton field turns to be less then $\phi_c$ the symmetry breaking takes place and the NG potential has the radius $f_{eff}=\sqrt{\nu(cM_p^2-\phi^2)/\lambda_{\chi}}$ and fluctuations are started. To keep symmetry restored during first $7$ e–folds we should have $\phi_c=4M_p$. After the moment of symmetry breaking it is allowed to start the fluctuations of phase with appropriate dispersion to create antizillahs, without any contradictions with observed CMB anisotropy. Of course to evaluate the distribution of antizillahs by sizes we have to take another parameters then we have used in our numerical example, but it does not change the main result of this paper.
Another story will take place if we would like to consider the AD baryogenesis as a basis for generation of antizillahs.
As it was discussed in the introduction the dynamics of the AD field is more complicated that in the case of spontaneous baryogenesis. Moreover it depends on the fact, D– or F– term inflation takes place. Also some details depend on the dimension ($d=4,6..$) of non–renormalizable term lifting the flat direction [@b3; @lisa], but it is enough for the brief discussion to circumscribe ourself with the minimal AD baryogenesis [@b3], where $d=4$. Thus in the case of D– term inflation, when the coherent slow rolling of AD field and inflaton are already established, the maximal radius $f_{eff}^{AD(D)}\simeq
10^{16}GeV$ of effectively massless angular direction can be obtained from the requirement that radial de Sitter fluctuations of AD field would not disturb significantly the spectral index of primordial adiabatic density perturbations [@b3] measured by COBE. Thereby, it is possible to get dispersion of phase fluctuations at the level $h\simeq 10^{-2}$ that is required for successful generation of antizillahs. The similar situation we could have in the case of F– term inflation [@b3; @lisa] because the AD potential gets an order of $H^2$ negative mass squared term during inflation, which causes the effective minimum at $f_{eff}^{AD(F)}\simeq C_F\sqrt{Hm_p}\simeq 10^{16}GeV$ (the $C_F$ is a constant of order of one).
The isocurvature fluctuations in the model of inhomogeneous AD baryogenesis with dispersion of phase fluctuations appropriate for antizillahs generation should be already observed by COBE [@b3]. Moreover this fluctuations can get some amplification owing to possible transformation of fluctuations of AD condensate into the isocurvature fluctuations of neutralinos [@b2]. The exact solution of the problem of isocurvature fluctuations for the AD based antimatter generation is the subject of separate investigation. Here we can only present some speculations, how to avoid the large isocurvature fluctuations at large cosmological scales, which are based on the similar strategy that has been chosen in the case of spontaneous baryogenesis.
As it has been mentioned in the Introduction, to organize the angular effectively massless direction in the AD potential we should accept the condition of the absence of order $H$ correction to the A– term both during and after inflation [@b3]. This condition gets automatically satisfied in the case of D– term inflation [@mr], while it is not true if the inflation is F– term dominated (see for example [@lisa]). According to this observation we can hope to find the such kind of trajectory of inflaton in field space that corresponds to the F– term dominated inflation in the beginning and then goes into D– term dominated regime. It implies that during the F– term dominated inflation the angular direction gets a mass of order $H$ and imaginary component of AD field is dumped and exponentially close to the minimum caused by this effective mass term. In such situation there are no de Sitter fluctuations of the phase. The fluctuations start only at the moment when the inflation goes to the D– term dominated regime and the angular direction turns to be effectively massless, because there is no correction of order $H$ to the A– term anymore. As we estimated before, to put the maximal scale of isocurvature fluctuations far below the modern cosmological horizon the transition from F– term to D– term inflation should take place 5–10 e–folds after the beginning of inflation. How to organize such transition is the subject of separate publication, but it seems that it could appear, for example, in the context of a realistic supergravity theory deriven from the weak coupled supestring [@string], which is already beyond the MSSM. There is some possibility to generate the F– term from a Fayet–Iliopoulos D– term [@d-f]. It could preserve the flatness of F– term direction during the first 5–10 e–folds of inflation causing the F– term domination firstly and subsequent trasformation of the vacuum energy into the D– term domination mode when it is allowed to begin phase fluctuations of AD field with dispersion appropriate for generation of antizillahs and without contradictions with COBE measurements.
We would like to notice in conclusion that the regions with antimatter in matter–dominated Universe could arise naturally in the variety of models. The main issue, that is needed, is a valley of potential. It is the valleys that are responsible for formation of causally separated regions with different values of field which in its turn give rise to antimatter domains. Many extensions of standard model based on supersymmetry possess this property, what strongly extends the physical basis for cosmic antimatter searches.
Acknowledgments
===============
This work was partially performed in the framework of Section “Cosmoparticle physics” of Russian State Scientific Technological Program “Astronomy. Fundamental Space Research”, International project Astrodamus, Cosmion–ETHZ and AMS–EPCOS. MYuK and ASS acknowledge supporte from Khalatnikov–Starobinsky school (grant 00–15–96699). We thank R.Konoplich and A.Sudarikov for interesting discussions and suggestions. We are also grateful to J.Ulbricht, A.D.Linde and I.Tkachev for useful comments.
[999]{} V.M.Chechetkin, M.Yu.Khlopov and M.G.Sapozhnikov, Riv. Nuovo Cimento, [**5**]{}, N10, 1 (1982); V.M.Chechetkin and M.Yu.Khlopov, Sov.J.Part.Nucl. [**18**]{}, 267 (1987); M.Yu.Khlopov, [*Cosmoparticle physics*]{} (World Scientific, Singapore – New Jersy – London – Hong Kong, 1999)
A.G.Cohen, A.De Rujula and S.L.Glashow, Astrophys.J. [**495**]{}, 539 (1998)
A.D.Dolgov, Preprint TAC–1996–010 (1996); hep–ph/9605280 (1996); Phys.Rep. [**222**]{}, 309 (1992)
V.M.Chechetkin, M.Yu.Khlopov, M.G.Sapozhnikov and Ya.B.Zeldovich, Phys. Lett. [**118B**]{}, 329 (1982)
T.D.Lee, Phys. Rev. [**D8**]{}, 1226 (1973)
M.Yu.Khlopov, R.V.Konoplich, R.Mignani, S.G.Rubin and A.S.Sakharov, Astropart.Phys. [**12**]{}, 367 (2000)
A.Dolgov and J.Silk, Phys.Rev. [**D47**]{}, 4244 (1993)
V.Kuzmin, I.Tkachev and M.Shaposhnikov, Phys.Lett. [**105B**]{}, 167 (1981)
A.G.Cohen and D.B.Kaplan, Phys.Lett. [**B199**]{}, 251 (1987); Nucl.Phys. [**B308**]{}, 913 (1988)
A.Dolgov et al, Phys.Rev. [**D56**]{}, 6155 (1997)
H.P.Nilles, Phys.Rep. [**110**]{}, 1 (1984)
I.Affleck and M.Dine, Nucl.Phys. [**B249**]{}, 361 (1985)
M.Dine, L.Randall and S. Thomas, Nucl.Phys. [**B458**]{}, 291 (1996)
K.Enqvist and J.McDonald, Phys.Rev.Lett. [**81**]{}, 3071 (1998)
K.Enqvist and J.McDonald, Phys.Rev.Lett. [**83**]{}, 2510 (1999)
K.Enqvist and J.McDonald, hep–ph/9912478
C.Kolda and J.March–Russell, Phys.Rev. [**D60**]{}, 023504–1 (1999)
T.Cherghetta, C.Kolda and S.Martin, Nucl.Phys. [**B468**]{}, 37 (1996)
D.H.Lyth, A.Riotto, Phys.Rep. [**314**]{}, 1 (1999)
The another examples of baryogenesis mechanisms that deal with effectively massless angular directions have been considered by M.Yoshimura, Phys.Rev.Let. [**51**]{}, 439 (1983); S.Dodelson and L.Widrow, Phys.Rev.Lett. [**64**]{}, 340 (1990); Phys.Rev. [**D42**]{}, 326 (1990); see also [@spokyok; @dolg]
A.Starobinsky, JETP Lett. [**30**]{}, 682 (1979); Phys.Lett. [**91B**]{}, 99 (1980)
A.Vilenkin and L.Ford, Phys.Rev. [**D26**]{}, 1231 (1982); A.D.Linde, Phys.Lett. [**116B**]{} 335 (1982); A.Starobinsky, ibid. [**117B**]{}, 175 (1982)
L.Kofman, Phys.Lett. [**174B**]{}, 400 (1986)
L.Kofman and A.Linde, Nucl.Phys. [**B282**]{}, 555 (1987)
J.Yokoyama and Y.Suto, Astrophys.J. [**379**]{}, 427 (1991); M.Sasaki and B.L.Spokoiny, Moden.Phys.Lett. [**32**]{}, 2935 (1991)
A.Linde and D.Lyth, Phys.Lett. [**246B**]{}, 353 (1990)
A.Linde, Phys.Lett. [**259B**]{}, 38 (1991)
A.Linde, [*Particle Physics and Inflationary Cosmology*]{} (Harwood, Chur, Switzerland, 1990)
D.Lyth, Phys.Rev. [**D45**]{}, 3394 (1992)
D.Lyth and E.Stewart, Phys.Rev. [**D46**]{}, 532 (1992)
A.Linde, Phys.Rev. [**D49**]{}, 1783 (1994)
D.Lyth and A.Liddle, Phys.Rep. [**231**]{}, 1 (1993)
J. Ellis, A. Linde and D. Nanopoulos, Phys. Lett. [ **118B**]{}, 59 (1982); L.M. Krauss, Nucl.Phys. [**B227**]{}, 556 (1983); D. Nanopoulos, K. Olive and M. Srednicki, Phys. Lett. [**127B**]{}, 30 (1983); M. Yu. Khlopov and A. Linde, Phys. Lett. [**138B**]{}, 265 (1984); J. Ellis, J.Kim and D. Nanopoulos, Phys. Lett. [**145B**]{}, 181 (1984); J. Ellis, G.B. Gelmini, J.L. Lopez, D.V. Nanopoulos, and S. Sarkar, Nucl. Phys. [**B373**]{}, 399 (1992); M. Kawasaki and T. Moroi, Progr. Theor. Phys. [**93**]{}, 879 (1995).
A.R.Liddle and D.H.Lyth, Phys.Lett. [**291B**]{}, 391 (1992)
G.Meylan and D.C.Heggie, Astron.Astrophys.Rev. [**8**]{}, 1 (1997)
F.C.Adams et al, Phys.Rev. [**D47**]{}, 426 (1993)
M.Yu Khlopov, Gravitation & Cosmology [**4**]{}, 1 (1998)
K.M.Belotsky, Yu.A.Golubkov, M.Yu.Khlopov, R.V.Konoplich, S.G.Rubin, and A.S.Sakharov, Gravitation & Cosmology [**5**]{} Proc. Suppl., 47 (1999); (astro-ph/9901402)
P.Sikivie, Phys.Rev.Lett. [**48**]{}, 1156 (1982); A.Vilenkin and A.E.Everett, Phys.Rev.Lett. [**48**]{}, 1867 (1982)
S.Chang, C.Hagmann and P.Sikivie, Phys.Rev. [**D59**]{}, 023505 (1999)
A.Vilenkin, Phys.Lett. [**133B**]{}, 177 (1983); J.Ipser and P.Sikivie, Phys.Rev. [**D30**]{}, 712 (1984); G.B.Gelmini, M.Gleiser and E.W.Kolb, Phys.Rev. [**D39**]{}, 1558 (1989)
S.D.Burns, astro-ph/9711303
A.S.Sakharov and M.Yu.Khlopov, Sov.J.Nucl.Phys. [**56**]{}, 412 (1993)
P.Binetruy, M.K.Gaillard and Y.Wu, Nucl.Phys. [**B493**]{}, 27 (1997); Phys.Lett. [**B412**]{}, 228 (1997)
M.K.Gaillard, D.H.Lyth and H.Murayama, Phys.Rev. [**D58**]{}, 123505–1 (1998)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A computation scheme for solving elliptic boundary value problems with axially symmetric confining potentials using different sets of one-parameter basis functions is presented. The efficiency of the proposed symbolic-numerical algorithms implemented in Maple is shown by examples of spheroidal quantum dot models, for which energy spectra and eigenfunctions versus the spheroid aspect ratio were calculated within the conventional effective mass approximation. Critical values of the aspect ratio, at which the discrete spectrum of models with finite-wall potentials is transformed into a continuous one in strong dimensional quantization regime, were revealed using the exact and adiabatic classifications.
-numerical algorithms, boundary value problems, axial confinement potentials, oblate and prolate spheroidal quantum dot models, effective mass approximation, strong dimensional quantization
author:
- |
A.A. Gusev$^{1,2}$, O. Chuluunbaatar$^1$, V.P. Gerdt$^1$, V.A. Rostovtsev$^{1,2}$,\
S.I. Vinitsky$^1$, V.L. Derbov$^3$, V.V. Serov $^3$
title: ' Symbolic-Numeric Algorithms for Computer Analysis of Spheroidal Quantum Dot Models '
---
Introduction
============
To analyze the geometrical, spectral and optical characteristics of quantum dots in the effective mass approximation and in the regime of strong dimensional quantization following [@Harrison], many methods and models were used, including the exactly solvable model of a spherical impermeable well [@Gambaryan], the adiabatic approximation for a lens-shaped well confined to a narrow wetting layer [@Hawrylak96] and a hemispherical impermeable well [@Hayk], the model of strongly oblate or prolate ellipsoidal impermeable well [@79], as well as numerical solutions of the boundary value problems (BVPs) with separable variables in the spheroidal coordinates for wells with infinite and finite wall heights [@CNI2000; @Trani; @Lepadatu]. However, thorough comparative analysis of spectral characteristics of models with different potentials, including those with non-separable variables, remains to be a challenging problem. This situation stimulates the study of a wider class of model well potentials with application of symbolic-numerical algorithms (SNA) and problem-oriented software, developed by the authors of the present paper during years [@Prorammirovanie; @casc07; @kantbp; @POTHMF; @parobp; @casc09].
Here we analyse the spectral characteristics of the following models: a spherical quantum dot (SQD), an oblate spheroidal quantum dot (OSQD) and a prolate spheroidal quantum dot (PSQD). We make use of the Kantorovich method that reduces the problem to a set of ordinary differential equations (ODE) [@put]. In contrast to the well-known method of adiabatic representation [@BhK58], this method implies neither adiabatic separation of fast and slow variables, nor the presence of a small parameter. We present a calculation scheme for solving elliptical BVPs with axially-symmetric potentials in cylindrical coordinates (CC), spherical coordinates (SC), oblate spheroidal coordinates (OSC), and prolate spheroidal coordinates (PSC). Basing on the SNA developed for axially-symmetric potentials, different sets of solutions are constructed for the parametric BVPs related to the fast subsystem, namely, the eigenvalue problem solutions (the terms and the basis functions), depending upon the slow variable as a parameter, as well as the matrix elements, i.e., the integrals of the products of basis functions and their derivatives with respect to the parameter, which are calculated analytically by means of elaborated SNA MATRA, implementing in MAPLE, or numerically using the program ODPEVP [@parobp], implementing the finite-element method (FEM). These terms and matrix elements form the matrices of variable coefficients in the set of second-order ODE with respect to the slow variable. The BVP for this set of ODE is solved by means of the program KANTBP [@kantbp], also implementing the FEM. The efficiency of the calculation scheme and the SNA used is demonstrated by comparison of the spectra versus the ellipticity of the prolate or oblate spheroid in the models of quantum dots with different confining potentials, such as the isotropic and anisotropic harmonic oscillator, the spherical and spheroidal well with finite or infinite walls, approximated by smooth short-range potentials, as well as by constructing the adiabatic classification of the states.
The paper is organized as follows. In Section \[1\] the calculation scheme for solving elliptic BVPs with axially-symmetric confining potentials is presented. In Section \[2\] SNA MATRA for solving parametric BVP and corresponding integrals implemented in Maple is described. Section \[3\] is devoted to the analysis of the spectra of quantum dot models with three types of axially-symmetric potentials, including the benchmark exactly solvable models. In Conclusion we summarize the results and discuss the future applications of our calculation scheme and the SNA project presented.
The problem statement {#1}
=====================
Within the effective mass approximation under the conditions of strong dimensional quantization the Schrödinger equation for the slow envelope of the wave function $\tilde\Psi(\tilde{\vec r})$ of a charge carrier (electron $e$ or hole $h$) in the models of a spherical, prolate or oblate spheroidal quantum dot (SQD, PSQD or OSQD) has the form $$\begin{aligned}
\label{sp01}
\{\tilde {\hat H} -\tilde E \}\tilde \Psi(\tilde {\vec r}) =
\{
({2\mu_p })^{-1} \tilde {\hat { {\vec P}}}^2 + \tilde U ( \tilde
{\vec r}) -\tilde E
\}\tilde \Psi(\tilde {\vec r}) =0,\end{aligned}$$ where $\tilde {\vec r}\in \bf R^{3}$ is the position vector of the particle having the effective mass $\mu_{p}=\mu_{e}$ (or $\mu_{p}=\mu_{h}$), $\tilde {\hat {{\vec P}}}=-i\hbar \nabla
_{\tilde {\vec r}}$ is the momentum operator, $\tilde E$ is the energy of the particle $\tilde U ( \tilde {\vec r})$ is the axially-symmetric potential, confining the particle motion in SQD, PSQD or OSQD. In Model A $\tilde U( \tilde {\vec r})$ is chosen to be the potential of isotropic or anisotropic axially-symmetric harmonic oscillator with the angular frequency $\tilde \omega =\gamma
_{\tilde r_0 } {\hbar }/({\mu_p \tilde r_0^2 })$, $\gamma _{\tilde
r_0 }\sim \pi^2/3$ being an adjustable parameter: $$\begin{aligned}
\label{sp02}
\tilde U^{\mbox{À}} (\tilde {\vec r}) =
{\mu_p \tilde \omega ^2( \zeta_1(\tilde x^2+\tilde y^2)+\zeta_3\tilde z^2)}/{2},\end{aligned}$$ $r_0 =\sqrt{\zeta_1(\tilde x_0^2+\tilde y_0^2)+\zeta_3\tilde z_0^2}$ is the radius of a spherical QD ($\zeta_1=1$, $\zeta_3=1$) or that of a spheroidal QD ($\zeta_1= (\tilde r_0/\tilde a)^4$, $\zeta_3=(\tilde r_0/\tilde c)^4$), inscribed into a spherical one, where $\tilde a$ and $\tilde c$ are the semiaxes of the ellipse which transforms into a sphere at $\tilde a=\tilde c=\tilde r_0$. For Model B $\tilde U (\tilde {\vec r})$ is the potential of a spherical or axially-symmetric well $$\begin{aligned}
\label{sp03}
\tilde U^{\mbox{B}} (\tilde {\vec r})
=\{0,0 \le({\tilde x^2+\tilde y^2})/{\tilde a^2}+{\tilde z^2}/{\tilde c^2} < 1 ; \tilde U_0,
({\tilde x^2+\tilde y^2})/{\tilde a^2}+{\tilde z^2}/{\tilde c^2} \ge 1 \},\end{aligned}$$ with walls of finite or infinite height $1\ll \tilde U_0 < \infty$. For Model C $\tilde U ( \tilde {\vec r})$ is taken to be a spherical or axially-symmetric diffuse potential $$\begin{aligned}
\label{sp03VS}
\tilde U^{\mbox{C}} (\tilde {\vec r})
= \tilde U_0\left[1+\exp (({ (\tilde x^2+\tilde y^2)/\tilde a^2+\tilde z^2/\tilde c^2-1})/{s})\right]^{-1},\end{aligned}$$ where $s$ is the edge diffusiveness parameter of the function, smoothly approximating the vertical walls of finite height $\tilde U_0$. Below we restrict ourselves by considering Model B with infinite walls $\tilde U_0
\rightarrow\infty$ and Model C with walls of finite height $\tilde
U_0$. We make use of the reduced atomic units: $a_B^*= {\kappa \hbar
^2}/{\mu _p e^2}$ is the reduced Bohr radius, $\kappa$ is the DC permittivity, $E_R \equiv Ry^*= {\hbar ^2}/({2\mu _p {a_B^*}^2 })$ is the reduced Rydberg unit of energy, and the following dimensionless quantities are introduced: $\tilde \Psi(\tilde{\vec r})= {a_B^*}^{-3/2}\Psi( {\vec r})$, $2 {\hat H}= \tilde {\hat H}/{Ry^*}$, $2 {E}= {\tilde E}/{Ry^*}$, $2 {U(\vec r)}= {\tilde U (\tilde {\vec r})}/{Ry^*}$, $\vec r=\tilde
{\vec r} /a_B^*$, $a=\tilde a/a_B^*$, $\tilde c=c/a_B^*$, $r_0=\tilde r_0/a_B^*$, $\omega=\gamma_{r_{0}}/ r_0^{2}=\hbar
\tilde\omega/(2Ry^*)$. For an electron with the reduced mass $\mu_p \equiv \mu_e = 0.067 m_0 $ at $\kappa =13.18$ in GaAs: $a_B^*=102$Å$=10.2$ nm, $Ry^*=E_R =5.2$ meV.
Since the Hamiltonian $\hat H$ in (\[sp01\])–(\[sp03VS\]) commutes with the $z$-parity operator ($z \to - z$ or $\eta \to -\eta $), the solutions are divided into even ($\sigma = + 1$) and odd ($\sigma = - 1$) ones. The solution of Eq. (\[sp01\]), periodical with respect to the azimuthal angle $\varphi$, is sought in the form of a product $\Psi(x_{f},x_{s},\varphi) = \Psi^{m\sigma} (x_{f},x_{s}
){e^{im\varphi }}/{\sqrt {2\pi } }$, where $m = 0,\pm 1,\pm 2,...$ is the magnetic quantum number. Then the function $\Psi^{m\sigma} (x_{f},x_{s} )$ satisfies the following equation in the two-dimensional domain $\Omega=\Omega_{x_f}(x_s)\cup\Omega_{x_s}\subset {\bf R}^2\backslash
\{0\}$, $\Omega_{x_f}(x_s)=( x_f^{\min }(x_{s}),x_f^{\max
}(x_{s}))$, $\Omega_{x_s}=( x_s^{\min },x_s^{\max })$: $$\begin{aligned}
&&
\left( \hat H_1 (x_{f};x_s) + \hat H_2 (x_{s})+ V (x_{f},x_{s}) - 2
{E}\right) \Psi^{m\sigma} (x_{f},x_{s} ) = 0. \label{sp09}\end{aligned}$$ The Hamiltonian of the slow subsystem $\hat H_2(x_s)$ is expressed as $$\begin{aligned}
&&\label{sp09xs}
\hat H_2(x_s)= \check H_2(x_s)
= - \frac{1}{g_{1s}(x_{s})}\frac{\partial }{\partial x_{s}}g_{2s}(x_{s})
\frac{\partial }{\partial x_{s} }+ \check V_{s}(x_{s}),\end{aligned}$$ and the Hamiltonian of the fast subsystem $ \hat {H}_1 (x_f ;x_s )$ is expressed via the reduced Hamiltonian $\check H_f (x_f ;x_s)$ and the weighting factor $g_{3s}(x_s)$: $$\begin{aligned}
\label{sp09xf}&&
\hat H_1(x_{f};x_s)=g_{3s}^{-1} (x_s )\check H_f (x_f ;x_s),\\&&\nonumber
\check H_f (x_f ;x_s)= - \frac{1}{g_{1f}(x_{f})}\frac{\partial }{\partial x_{f}}g_{2f}(x_{f})
\frac{\partial }{\partial x_{f} }+ \check V_{f}(x_{f})
+\check V_{fs}(x_{f},x_{s}).\end{aligned}$$ Table \[ktpy\] contains the values of conditionally fast $x_{f}$ and slow $x_{s}$ independent variables, the coefficients $g_{1s}(x_s)$, $g_{2s}(x_s)$, $g_{3s}(x_s)$, $g_{1f}(x_f)$, $g_{2f}(x_f)$, and the reduced potentials $\check V_{f}(x_{f})$, $\check V_{s}(x_{s})$, $\check V_{fs}(x_{f},x_{s})$, entering Eqs. (\[sp09\])–(\[sp09xf\]) for SQD, OSQD and PSQD in cylindrical ($\vec x=(z,\rho,\varphi)$), spherical ($\vec x=(r,\eta=\cos\theta,\varphi)$), and oblate/prolate spheroidal ($\vec x=(\xi,\eta, \varphi)$) coordinates [@stigun].
In spherical coordinates the potential $\check{V}(r,\eta)$ in Table \[ktpy\], using the definitions (\[sp02\]), (\[sp03VS\]) in the reduced atomic units, for Model A is expressed as $${\check V}(r,\eta)={2}r^2 V(r,\eta)
={\omega ^2r ^4( \zeta_1(1-\eta^2)+\zeta_3\eta^2)},$$ and for Model C as $${\check V}(r,\eta)={2} r^2 V(r,\eta)
={2}r ^2U_0\left[1+\exp (({ r^2 (
(1-\eta^2)/a^2+\zeta_3\eta^2/c^2)-1})/{s})\right]^{-1},$$ both having zero first derivatives in the vicinity of the origin $r=0$ (equlibrium point). For Model B the potentials ${\check V}_{fs}$ are zero, since the potential (\[sp03\]) is reformulated below in the form of boundary conditions with respect to the variables $x_f$ and $x_s$. The solution of the problem (\[sp09\])–(\[sp09xf\]) is sought in the form of Kantorovich expansion [@put] $$\begin{aligned}
\Psi^{Em\sigma}_i( x_{f},x_{s})=\sum_{j=1}^{j_{\max}}
\Phi^{m\sigma}_j(x_{f}; x_{s})\chi_{j}^{(m\sigma i)}(E,x_{s}),
\label{sp15}\end{aligned}$$ using as a set of trial functions the eigenfunctions $\Phi^{m\sigma}_j(x_{f}; x_{s})$ of the Hamiltonian $\check {H}_f
(x_f ;x_s )$ from (\[sp09xf\]), i.e., the solutions of the parametric BVP $$\label{sp17} \left\{ \check {H}_f (x_f ;x_s ) -
\check {\lambda}_i (x_s) \right\} \Phi _i^{m\sigma} (x_f ;x_s ) = 0,$$ in the interval $x_{f}\in
\Omega_{x_f}(x_s)$ depending on the conditionally slow variable $x_{s}\in\Omega_{x_s}$ as on a parameter. These solutions obey the boundary conditions $$\label{sp17a} \lim \limits_{x_f \to x_f^{t}(x_{s})}\!\! \left(\!\!
{N_f^{(m\sigma)}(x_{s}) g_{2f} (x_f)\frac{d\Phi _j^{m \sigma} (x_f
;x_s)}{dx_f } + D_f^{(m \sigma)}(x_{s}) \Phi _j^{m \sigma} (x_f;x_s
)}\!\!\right)\!\! =\!\! 0$$ in the boundary points $\{x_f^{\min }(x_{s}),x_f^{\max}(x_{s})\}=\partial\Omega_{x_f}(x_s)$, of the interval $\Omega_{x_f}(x_s)$. In Eq. (\[sp17a\]), $N_f^{(m\sigma)}(x_{s})\equiv N_f^{(m\sigma)}$, $D_f^{(m\sigma)}(x_{s})\equiv D_f^{(m\sigma)}$, unless specially declared, are determined by the relations $N_f^{(m\sigma)} = 1$, $D_f^{(m\sigma)} = 0$ at $m = 0$, $\sigma = + 1$ (or at $\sigma =0$, i.e without parity separation), $N_f^{(m\sigma)} = 0$, $D_f^{(m\sigma)} = 1$ at $m = 0$, $\sigma = - 1$ or at $m \ne 0$. The eigenfunctions satisfy the orthonormality condition with the weighting function $g_{1f}(x_f)$ in the same interval $x_{f}\in \Omega_{x_f}(x_s)$: $$\label{sp19}
\left\langle \Phi _i^{m\sigma} \vert \Phi _j^{m\sigma} \right\rangle =
\int\nolimits_{x_f^{\min}(x_{s})}^{x_f^{\max}(x_{s})}\Phi _i^{m\sigma} (x_f ;x_s )
\Phi _j^{m\sigma} (x_f ;x_s )
g_{1f}(x_f)dx_f=\delta_{ij}.$$ Here $\check \lambda_1 (x_s)< ... < \check \lambda_{j_{\max } }(x_s) <...$ is the desired set of real eigenvalues. Corresponding set of potential curves $ 2{E}_1 (x_s ) < ... < 2{E}_{j_{\max } } (x_s ) <...$ of Eqs. (\[sp09xf\]) is determined by $2 {E}_{j} (x_s )=g_{3s}^{-1} (x_s )\check \lambda_j (x_s)$. Note, that for OSC and PSC the desired set of real eigenvalues $\check \lambda_j (x_s)$ depends on a combined parameter, $x_{s}\rightarrow p^{2}=(d/2)^{2}2E$, the product of spectral $2E$ and geometrical $(d/2)^{2}$ parameters of the problem (\[sp09\]). The solutions of the problem (\[sp17\])–(\[sp19\]) for Models A and B are calculated in the analytical form, while for Model C this is done using the program ODPEVP [@parobp].
Substituting the expansion (\[sp15\]) into Eq. (\[sp09\]) in consideration of (\[sp17\]) and (\[sp19\]), we get a set of ODE for the slow subsystem with respect to the unknown vector functions $ {\mbox{\boldmath $\chi$}}^{(m\sigma i)}(x_s ,E) \equiv
{\mbox{\boldmath $\chi$}}^{(i)}(x_s ) = (\chi _1^{(i)} (x_s
),...,\chi _{j_{\max } }^{(i)} (x_s ))^T$: $$\begin{aligned}
&&
\label{sp23}
\biggl(\!
-\!\frac{1}{g_{1s}(x_{s})}{\rm {\bf I}}
\frac{d}{d{ x_{s}}}g_{2s}(x_{s})\frac{d}{d{ x_{s}}}
+2 {\bf E} ({ x_{s}})+{\rm {\bf I}}\check{V}_{s}(x_s)
- 2{\rm {\bf I}} {E}
\biggr){\mbox{\boldmath $\chi$}}^{(i)}({ x_{s}})=
\\&&
=\!-\!\biggl(\frac{g_{2s}(x_s)}{g_{1s}(x_s)} {\rm {\bf W}} ({ x_{s}})+
\frac{1}{g_{1s}(x_{s})}\frac{dg_{2s}(x_{s}){\rm {\bf Q}}({x_s})}{d{x_s}}
\!+\!\frac{g_{2s}(x_{s})}{g_{1s}(x_{s})}{\rm {\bf Q}}({ x_{s}})
\frac{d}{d x_{s}}
\biggr)
{\mbox{\boldmath $\chi$}}^{(i)}({ x_{s}}).\nonumber\end{aligned}$$ Here $2 {\bf E} (x_s )=\mbox{diag}(g_{3s}^{-1} (x_s )\check \lambda_j
(x_s))$, ${\rm {\bf W}}( x_s)$, and ${\rm {\bf Q}}( x_s)$ are matrices of the dimension $j_{{\max} }\times j_{{\max} }$, $$\begin{aligned}
W_{ij}(x_{s})=W_{ji}( x_{s})=
\int\nolimits_{x_f^{\min}(x_{s})}^{x_f^{\max}(x_{s})}g_{1f}(x_{f})
\frac{\partial\Phi_{i}(x_{f}; x_{s})}{\partial x_{s}}
\frac{\partial\Phi_{j}(x_{f}; x_{s})}{\partial x_{s}}dx_{f},
\label{sp23a} \\
Q_{ij}(x_{s})=-Q_{ji}( x_{s})=
-\int\nolimits_{x_f^{\min}(x_{s})}^{x_f^{\max}(x_{s})}g_{1f}(x_{f})
\Phi_{i}(x_{f}; x_{s})
\frac{\partial\Phi_{j}(x_{f}; x_{s})}{\partial x_{s}}dx_{f},
\nonumber\end{aligned}$$ calculated analytically for Model B and by means of the program ODPEVP [@parobp] for Model C. Note, that for Model A in SC and CS and Model B in OSC and PSC the variables $x_f$ and $x_{s}$ are separated, so that the matrix elements $W_{ij}(x_{s})=Q_{ij}(x_{s})\equiv 0$ are put into the r.h.s. of Eq. (\[sp23\]) and $\check{V}_s(x_{s})$ are substituted from Table \[ktpy\]. The discrete spectrum solutions $2E: 2E_1 < 2E_2 < ... < 2E_t < ...$, that obey the boundary conditions in the points $x_s^t = \{x_s^{\min } ,x_s^{\max }\}=\partial\Omega_{x_s}$ bounding the interval $\Omega_{x_s}$: $$\label{eq6sar}
\lim \limits_{x_s \to x_s^t }
\left( {N_s^{( {m \sigma})} g_{2s} (x_s)
\frac{d{\mbox{\boldmath $\chi$}}^{({m \sigma}p)}(x_s )}{dx_s }
+ D_s^{({m \sigma})} {\mbox{\boldmath $\chi$}}^{({m \sigma}p)}(x_s )} \right) = 0,$$ where $N_s^{({m \sigma})} = 1$, $D_s^{({m \sigma})} = 0$ at $m = 0, \sigma = + 1$ (or at $\sigma =0$, i.e without parity separation), $N_s^{({m \sigma})} = 0$, $D_s^{({m \sigma})} = 1$ at $m = 0, \sigma = - 1$ or at $m \ne 0$, and the orthonormality conditions $$\begin{aligned}
\int\nolimits_{x_{s}^{\min}}^{x_{s}^{\max}}
({\mbox{\boldmath $\chi$}}^{(i)}({x_{s}}))^T
{\mbox{\boldmath $\chi$}}^{(j)}({ x_{s}})g_{1s}(x_{s}) dx_{s}
= \delta_{ij},
\label{sp23n}\end{aligned}$$ are calculated by means of the program KANTBP [@kantbp]. To ensure the prescribed accuracy of calculation of the lower part of the spectrum discussed below with eight significant digits we used $j_{\max}=16$ basis functions in the expansion (\[sp15\]) and the discrete approximation of the desired solution by Lagrange finite elements of the fourth order with respect to the grid pitch $\Omega^{p}_{h^{s}({x_s})}=[x^s_{\min},x^s_{k}= x^s_{k-1}+h_{k}^{s},
x^s_{\max}]$.
SNA MATRA for calculus of the BVP and integrals {#2}
================================================
To calculate effective potentials of problem (\[sp23\])–(\[sp23n\]) in each value ${x_s}=x^s_k$ of the FEM grid $\Omega^{p}_{h^{s}({x_s})}=[x^s_{\min},
x^s_{\max}]$, we consider a discrete representation of solutions $\Phi({x_f};{x_s})\equiv\Phi^{m\sigma}({x_f};{x_s})$ of the problem (\[sp17\]) by means of the FEM on the grid, $\Omega^{p}_{h^f({x_f})}=[x^f_{0}\!=\!x^f_{\min}(x_{s}),$ $x^f_{k}=x^f_{k-1}+h_{k}^{f}, x^f_{\bar n}\!=\!x^f_{\max}(x_{s})]$, in a finite sum: $$\begin{aligned}
\label{nnn}
\Phi({x_f};{x_s})=\sum_{\mu=0}^{\bar np}
\Phi_{\mu}^h({x_s})N_{\mu}^p({x_f}) =\sum_{k=1}^{\bar n}\sum_{r=0}^{p}
\Phi_{r+p(k-1)}^h({x_s})N^p_{r+p(k-1)}({x_f}),
\end{aligned}$$ where $N_{\mu}^p({x_f})$ are local functions and $\Phi^h_{\mu}({x_s})$ are node values of $\Phi(x^f_\mu;{x_s})$. The local functions $N_{\mu}^p({x_f})$ are piece-wise polynomial of the given order $p$ equals one only in the node $x^f_{\mu}$ and equals zero in all other nodes $x^f_{\nu}\neq x^f_{\mu}$ of the grid $\Omega^{p}_{h^f({x_f})}$, i.e., $N_{\nu}^p(x^f_\mu)=\delta_{\nu\mu}$, $\mu,\nu=0,1,\ldots,\bar np$. The coefficients $\Phi_{\nu}({x_s})$ are formally connected with solution $\Phi(x^{fp}_{k,r};{x_s})$ in a node $x^f_{\nu}=x^{fp}_{k,r}$, $k=1,\ldots,\bar n$, $r=0,\ldots,p$: $$\begin{aligned}
\Phi^h_{\nu }({x_s})=
\Phi^h_{r+p(k-1)}({x_s}) \approx \Phi(x^{fp}_{k,r};{x_s}),
\quad x^{fp}_{k,r}=x^f_{k-1}+\frac{h_k^{f}}{p}r.\nonumber
\end{aligned}$$ The theoretical estimate for the ${\bf H^0}$ norm between the exact and numerical solution has the order of $$\begin{aligned}
\vert \check \lambda_j({x_s}) -\! \check \lambda^h_j({x_s})\! \vert \!\leq \!
c_1 h^{2p} ,~
\left \Vert \! \Phi_j(x_{f}; x_{s}) \! - \mbox{\boldmath$\Phi$}^h_j({x_s})\right \Vert_0 \!\leq\!
c_2 h^{p+1},\label{eq13}
\end{aligned}$$ where $h^{f} = \max_{1<j<\bar n}h_{j}^{f}$ is maximum step of grid and constants $c_1>0$, $c_2>0$ do not depend on step $h^{f}$ [@r27]. It has been shown that we have a possibility to construct schemes for solving the BVPs and integrals with high order of accuracy comparable with the computer one in according with the following estimations [@parobp] $$\begin{aligned}
&&\left| \frac{\partial\check \lambda_j(x_{s})} {\partial x_{s}}-
\frac{\partial\check \lambda ^h_j(x_{s})}{\partial x_{s}}\right| \leq c_3h^{2p}, \,
\left\| \frac{\partial\Phi_j(x_{f};x_{s})}{\partial x_{s}} -
\frac{\partial\mbox{\boldmath$\Phi$}^h_j(x_{s})} {\partial
x_{s}}\right\|_0 \leq c_4h^{p+1},\\&&
\left|Q_{ij}(x_{s})-Q_{ij}^h(x_{s})\right| \leq c_5h^{2p}, \,
\left|W_{ij}(x_{s})-W_{ij}^h(x_{s})\right| \leq c_6h^{2p},
\label{eq13aa}\end{aligned}$$ where $h^{f}$ is the grid step, $p$ is the order of finite elements, $i$, $j$ are the number of the corresponding solutions, and constants $c_3$, $c_4$, $c_5$ and $c_6$ do not depend on step $h^{f}$. Proof is straightforward following the scheme of proof of estimations (\[eq13\]) in accordance with [@r27; @shultz]. The verification of the above estimations are examined by numerical analysis on condensed grids and by comparison with examples of exact solvable models A and B.
Let us consider the reduction of BVP (\[sp17\]), (\[sp19\]) on the interval $\Delta:~x^f_{\min}(x_s)< {x_f}< x^f_{\max}(x_s)$ with boundary conditions (\[sp17a\]) in points $x^f_{\min}(x_s)$ and $x^f_{\max}(x_s)$ rewriting in the form $$\begin{aligned}
{\bf A}({x_s})\Phi_j({x_f};{x_s})= \check \lambda_j(x_{s}){\bf B}({x_s})\Phi_j({x_f};{x_s}),
\label{3231}
\end{aligned}$$ where ${\bf A}({x_s})$ is differential operator and ${\bf B}({x_s})$ is multiplication operator are differentiable by parameter $x_s\in\Omega_{x_s}$. Substituting expansion (\[nnn\]) to (\[3231\]) and integration with respect to ${x_f}$ by parts in the interval $\Delta=\cup_{k=1}^{\bar n}\Delta_k,$ we arrive to a system of the linear algebraic equations $$\begin{aligned}
{\bf a}_{\mu\nu}^p({x_s})\Phi_{j,\mu}^h({x_s})=
\check \lambda ^h_j(x_{s}){\bf b}_{\mu\nu}^p({x_s})\Phi_{j,\mu}^h({x_s}),
\label{3231mat}
\end{aligned}$$ in framework of the briefly described FEM. Using $p$-order Lagrange elements [@r27], we present below an Algorithm 1 for construction of algebraic problem (\[3231mat\]) by the FEM in the form of conventional pseudocode. It MAPLE realization allow us show explicitly recalculation of indices $\mu$, $\nu$ and test of correspondent modules of parametric matrix problems, derivatives of solutions by parameter and calculation of integrals.
[**Algorithm 1**]{} Generation of parametric algebraic problems\
\[0mm\] [**Input:**]{}\
$\Delta=\cup_{k=1}^{\bar n}\Delta_k=[x^f_{\rm min}(x_{s}),x^f_{\rm
max}(x_{s})]$, is interval of changing of independent variable ${x_f}$, that boundaries depending on parameter $x_{s}=x^s_{k'}$;\
$h_k^{f}=x^f_k-x^f_{k-1}$ is a grid step;\
$\bar n$ is a number of subintervals $\Delta_k= [x^f_{k-1},x^f_k]$;\
$p$ is a order of finite elements;\
${\bf A}({x_s}),{\bf B}({x_s})$ are differential operators in Eq. (\[3231\]);\
[**Output:**]{}\
$N_\mu^p(x_{f})$ is a basis functions in (\[nnn\]);\
${\bf a}_{\mu\nu}^p({x_s}),~{\bf b}_{\mu\nu}^p({x_s})$ are matrix elements in system of algebraic equations (\[3231mat\]);\
[**Local:**]{}\
$x^{fp}_{k,r}$ are nodes; $\phi^p_{k,r}({x_f})$ are Lagrange elements; $\mu,\nu=0,1,\ldots,\bar np$ ;\
\[0mm\]\
1: for $k$:=1 to $\bar n$ do\
$\quad$ for $r$:=0 to $p$ do\
$\quad \quad x^{fp}_{k,r}=x^f_{k-1}+\frac{h_k^{f}}{p}r$\
$\quad$ end for;\
end for;\
2:$ \phi^p_{k,r}({x_f})=
\prod_{r'\neq r}[({x_f}-x^{fp}_{k,r'})(x_{k,r}^{fp}-x^{fp}_{k,r'})^{-1}]$\
3:$ N_0^p({x_f}):=\{\phi^p_{1,0}({x_f}),{x_f}
\in \Delta_1;0,{x_f}\not\in \Delta_1 \};$\
for $k$:=1 to $\bar n$ do\
$\quad$ for $r$:=1 to $p-1$ do\
$\quad \quad N_{r+p(k-1)}^p({x_f}):
=\{\phi^p_{k,r}({x_f}),{x_f} \in \Delta_k; 0, {x_f} \not\in \Delta_k,\}$\
$\quad$ end for;\
$\quad
N_{kp}^p({x_f}):=\{\phi^p_{k,p}({x_f}), {x_f} \in \Delta_k;
\phi^p_{k+1,0}({x_f}),{x_f}\in\Delta_{k+1};
0,{x_f} \not\in \Delta_k\bigcup\Delta_{k+1}\};$\
end for;\
$ N_{\bar np}^p({x_f}):=\{ \phi^p_{\bar
n,p}({x_f}), {x_f} \in \Delta_{\bar n}; 0,
{x_f} \not\in \Delta_{\bar n}\};$\
4:for $\mu,\nu$:=0 to $\bar np$ do\
$\quad$$\quad$ $ {\bf
a}_{\mu\nu}^p({x_s}):=\int\limits_\Delta g_1(x_f) N_{\mu}^p({x_f})
{\bf A}({x_s})
N_{\nu}^p({x_f}) d{x_f}; $\
$\quad$$\quad$ $ {\bf b}_{\mu\nu}^p({x_s}):
=\int\limits_\Delta g_1(x_f) N_{\mu}^p({x_f}) {\bf B}({x_s}) N_{\nu}^p({x_f}) d{x_f};$\
end for;\
\
[**Remarks:**]{}
1\. For equation (\[sp17\]) matrix elements of the operator (\[sp09xf\]), and $V({x_f};{x_s})=\check V_{fs}(x_f,x_s)+\check V_{f}(x_f)$ between local functions $N_{\mu}(x_f)$ and $N_{\nu}(x_f)$ defined in same interval $\Delta_j$ calculated by formula, using ${x_f}=x^f_{k-1}+0.5h^f_k(1+\eta_f),$ $q,r=\overline{0,p}$: $$\begin{array}l
\left({\bf a}({x_s})\right)_{\mu,\nu}= \int\limits_{-1}^{+1} \left
\{ {4\over (h^f_k)^2}g_{2f}(x_f)(\phi_{k,q}^p)^\prime
(\phi_{k,r}^p)^\prime +g_{1f}(x_f)V({x_f};{x_s})
\phi_{k,q}^p\phi_{k,r}^p \right
\} {h^f_k \over 2} d \eta_f,\\
\left({\bf b}({x_s})\right)_{\mu,\nu}
=\int\limits_{-1}^{+1}g_{1f}(x_f)\phi_{k,q}^p\phi_{k,r}^p
{h^f_k\over 2} d \eta_f,~~ \mu=q+p(k-1),~~ \nu=r+p(k-1).
\end{array}$$
2\. If integrals do not calculated analytically, for example, see section \[3\], then they are calculated by numerical methods [@r27], by means of the Gauss quadrature formulae of the order $p+1$.
3\. For OSQD&PSQD model C the problem (\[sp17\])–(\[sp19\]) has been solved using a grid $\Omega^{p}_{h^f({x_f})}[x^f_{\min},x^f_{\max}]=-1(20)1$ (the number in parentheses denotes the number of finite elements of order $p = 4$ in each interval).
Generally, 10-16 iterations are required for the subspace iterations to converge the subspace to within the prescribe tolerance. If matrix ${\mathbf
a}^p\equiv{\bf a}^p({x_s})$ in Eq. (\[3231mat\]) is not positively defined, problem (\[3231mat\]) is replaced by the following problem: $$\begin{aligned}
\tilde{\mathbf a}^p \> \mbox{\boldmath$\Phi$}^h =
\tilde \lambda^h\> {\mathbf b}^p \>
\mbox{\boldmath$\Phi$}^h,\quad
\tilde {\mathbf a}^p ={\mathbf a}^p -\alpha{\mathbf b}^p.\label{31b}
\end{aligned}$$ The number $\alpha$ (the shift of the energy spectrum) is chosen in such a way that matrix $\tilde {\mathbf a}^p$ is positive. The eigenvector of problem (\[31b\]) is the same, and $\check\lambda ^h= \tilde \lambda ^h+\alpha$, where shift $\alpha$ is evaluated by the Algorithm 2.
**Algorithm 2** [Evaluating the lower bound for the lowest eigenvalue of the generalized eigenvalue problem]{}
In general case it is impossible to define the lower bound for the lowest eigenvalue of Eq. (\[31b\]), because the eigenvalues $\check \lambda ^h_{1}(x_s)<...<\check \lambda ^h_{i}(x_s)<...< \check \lambda ^h_{j_{max}}(x_s)$ is depended on the parameter $x_s$. But, we can use the following algorithm to find the lower bound for the lowest eigenvalue $\check \lambda ^h_{1}(x_s)$ at fixed value of $x_s$:
[**Step 1.**]{} Calculate $\mathbf{ L\,D\,L^ T}$ factorization of $\mathbf{A}^p-\alpha\mathbf{B}^p$.
[**Step 2.**]{} If some elements of the diagonal matrix $\mathbf{D}$ are less than zero\
then put $\alpha=\alpha-1$ and go to [**Step 3**]{}, else go to [**Step 5**]{}.
[**Step 3.**]{} Calculate $\mathbf{ L\,D\,L^ T}$ factorization of $\mathbf{A}^p-\alpha\mathbf{B}^p$.
[**Step 4.**]{} If some elements of the diagonal matrix $\mathbf{D}$ are less than zero\
then put $\alpha=\alpha-1$ and go to [**Step 3**]{}, else put $\alpha=\alpha-0.5$\
and go to [**Step 8**]{}.
[**Step 5.**]{} Put $\alpha=\alpha+1$ and calculate $\mathbf{L\,D\,L^ T}$ factorization of $\mathbf{A}^p-\alpha\mathbf{B}^p$.
[**Step 6.**]{} If all elements of the diagonal matrix $\mathbf{D}$ are greater than zero\
then put $\alpha=\alpha+1$ and repeat [**Step 5**]{}.
[**Step 7.**]{} Put $\alpha=\alpha-1.5$.
[**Step 8.**]{} End.
After using the above algorithm one should find the lower bound for the lowest eigenvalue, and always $\check \lambda ^h_{1}(x_s)-\alpha\leq 1.5$
Spectral characteristics of spheroidal QDs {#3}
==========================================
{width="44.00000%" height="40.00000%"}a) {width="44.00000%" height="40.00000%"}b)\
{width="98.00000%"}
{width="98.00000%"}
### Models B and C for Oblate Spheroidal QD
At fixed coordinate $x_{s}$ of the slow subsystem the motion of the particle in the fast degree of freedom $x_{f}$ is localized within the potential well having the effective width $$\label{eq70}
\tilde L\left( x_s \right) = 2c\sqrt {1 - {x_s^2}/{a^2}} ,$$ where $L=\tilde L/a_B^*$. The parametric BVP (\[sp17\])–(\[sp19\]) at fixed values of the coordinate $x_{s}$, $x_{s}\in (0,a)$, is solved in the interval $x_f\in(- L\left( x_s \right)/2, L\left( x_s \right)/2)$ for Model C using the program ODPEVP, and for Model B the eigenvalues $\tilde E_{n_{o}}\left( x_s \right)/E_R\equiv 2 E_{i}\left( x_s \right)$, $n_{o}=i = 1,2,...$, and the corresponding parametric eigenfunctions $\Phi^\sigma_{i} \left( {x_f;x_s}\right)$, obeying the boundary conditions (\[sp17a\]) and the normalization condition (\[sp19\]), are expressed in the analytical form: $$\label{eq71}
\!\!\!\!\!\!\!\!2 E_{i}\left( x_s \right)\!=\!\frac{\pi ^2n_{o}^2}{
L^2\left( x_s \right)},\quad \Phi_{i}^{\sigma} \left(
{x_f;x_s}\right)\!=\!\sqrt{\frac{2}{ L\left( x_s \right)}}
\sin\left(\frac{\pi n_{o}}{2} \left(\frac{x_f}{ L\left( x_s
\right)/2}-1\right)\right),$$ where the even solutions $\sigma=+1$ are labelled with odd $n_{o}=n_{zo}+1=2i-1,$ and the odd ones $\sigma=-1$ with even $n_{o}=n_{zo}+1=2i$, $i=1,2,3,...$ . The effective potentials (\[sp23a\]) in Eq. (\[sp23\]) for the slow subsystem are expressed analytically via the integrals over the fast variable $x_f$ of the basis functions (\[eq71\]) and their derivatives with respect to the parameter $x_s$ [including states with both parities $\sigma=\pm1$]{}: $$\begin{aligned}
&& 2 E_{i}(x_{s})=
\frac{a^2\pi^2n_{o}^2}{4c^2(a^2-x_s^2)},\quad
W_{ii}(x_{s})=\frac{3+\pi^2n_{o}^2}{12}\frac{x_s^2}{(a^2-x_s^2)^2},\label{sp23s}\\
&&W_{ij}(x_{s})=\frac{2n_{o}n_{o}'(n_{o}^2+n_{o}'{}^2)(1+(-1)^{n_{o}+n_{o}'})}{(n_{o}^2-n_{o}'{}^2)^2}\frac{x_s^2}{(a^2-x_s^2)^2},
\nonumber\\
&&Q_{ij}(x_{s})=\frac{n_{o}n_{o}'(1+(-1)^{n_{o}+n_{o}'})}{(n_{o}^2-n_{o}'{}^2)^2}\frac{x_s}{a^2-x_s^2},\quad n_{o}'\neq n_{o}.\nonumber\end{aligned}$$
For Model B at $c=a=r_{0}$ the OSQD turns into SQD with known analytically expressed energy levels $E_{t}\equiv E_{nlm}^{sp}$ and the corresponding eigenfunctions $$2 E_{nlm}^{sp}\!=\!\frac{\alpha_{n_r+1,l+1/2}^2}{r_{0}^2},~
\Phi_{nlm}^{sp}(r,\theta,\varphi)
\!=\!\frac{\sqrt{2}J_{l+1/2}(\sqrt{2 E_{nlm}^{sp}}r)}
{r_{0}\sqrt{r}|J_{l+3/2}(\alpha_{n_r+1,l+1/2})|}
Y_{lm}(\theta,\varphi),$$ where $\alpha_{n_r+1,l+1/2}$ are zeros of the Bessel function of semi-integer index $l+1/2$, numbered in ascending order $0<\alpha_{11}<\alpha_{12}< ...<\alpha_{iv}<...$ by the integer $i,v=1,2,3,...$. Otherwise one can use equivalent pairs $iv\leftrightarrow\{n_r,l\}$ with $n_r=0,1,2,...$ numbering the zeros of Bessel function and $l=0,1,2,...$ being the orbital quantum number that determines the parity of states $\hat{\sigma}=(-1)^l=(-1)^{m}\sigma$, $\sigma=(-1)^{l-m}=\pm1$. At fixed $l$ the energy levels $\tilde
E_{nlm}/E_{R}=2 E_{t}$, degenerate with respect to the magnetic quantum number $m$, are labelled with the quantum number $n=n_r+1=i=1,2,3,...$ , in contrast to the spectrum of a spherical oscillator, degenerate with respect to the quantum number $\lambda=2n_{r}+l$. Figs. \[enekts\], \[cwe\], and \[cste\] show the lower part of non-equidistant spectrum $\tilde
E(\zeta_{ca})/E_{R}=2 E_t$ and the eigenfunctions $\Psi^{m\sigma}_t$ from Eq. (\[sp15\]) for even states OSQD Models B and C at $m=0$. There is a one-to-one correspondence rule $n_{o}=n_{zo}+1=2n-(1+\sigma)/2, n=1,2,3,...$, $n_\rho=(l-|m|-(1-\sigma)/2)/2$, between the sets of spherical quantum numbers $(n,l,m,\hat\sigma)$ of SQD with radius $r_0=a=c$ and spheroidal ones $(n_\xi=n_r,n_\eta=l-|m|,m,\sigma)$ of OSQD with the major $a$ and the minor $c$ semiaxes, and the adiabatic set of cylindrical quantum numbers $(n_{zo},n_\rho,m,\sigma)$ at continuous variation of the parameter $\zeta_{ca}=c/a$. The presence of crossing points of the energy levels of similar parity under the symmetry change from spherical $\zeta_{ca}=1$ to axial, i.e., under the variation of the parameter $0<\zeta_{ca}<1$, in the BVP with two variables at fixed $m$ for Model B is caused by the possibility of variable separation in the OSC [@stigun], i.e. the r.h.s. of Eq. (\[sp23\]) equals zero. The transformation of eigenfunctions, occurring in the course of a transition through the crossing points in Fig. \[enekts\], is shown in Fig. \[cwe\] for model B and in Fig. \[cste\] for model C (marked by arrows). One can see that number nodes [@CurantGilbert] the eigenfunctions ordered in according to increasing eigenvalues is not changed under the transition from spherical to oblate spheroidal form. So, at small value of deformation parameter ($\zeta_{ca}$ for OSQD or $\zeta_{ac}$ for PSQD) there are nodes only along corresponding major axis. For Model C at each value of the parameter $a$ their is a finite number of discrete energy levels, limited by the value $2U_0$ of the well walls height. As shown in Fig. \[enekts\]b, the number of levels of OSQD, equal to that of SQD at $a=c=r_0$, is reduced with the decrease of the parameter $c$ (or $\zeta_{ca}$), in contrast to Models A and B that have countable spectra, and avoided crossings appear just below the threshold.
{width="44.00000%" height="40.00000%"} a) {width="44.00000%" height="40.00000%"}b)\
### Models B and C for Prolate Spheroidal QD
In contrast to OSQD, for PSQD at fixed coordinate $x_s$ of the slow subsystem the motion of the particle is confined to a 2D potential well with the effective variable radius $$\label{eq99}
\rho_{0} \left( x_{s} \right) = a\sqrt {1 - {x_s^2} / {c^2}} ,$$ where $ \rho_{0}\left( x_{s} \right) = \tilde \rho_{0}\left( x_{s} \right)/ {a_B } $. The parametric BVP (\[sp17\])–(\[sp19\]) at fixed values of the coordinate $x_{s}$ from the interval $x_{s}\in (-c,c)$ is solved in the interval $x_f\in(0, \rho_{0}\left( x_{s} \right))$ for Model C using the program ODPEVP, while for Model B the eigenvalues $\tilde E_{n_{\rho p}+1}\left( x_s \right)/E_R\equiv
2E_{i}\left( x_s \right)$, $n_{\rho p}+1=i = 1,2,...$, and the corresponding parametric basis functions $\Phi^{m\sigma=0}_{i} \left( {x_f;x_s}\right)\equiv\Phi^{m}_{i} \left( {x_f;x_s}\right)$ without parity separation, obeying the boundary conditions (\[sp17a\]) and the normalization condition (\[sp19\]), are expressed in the analytical form: $$\label{eq100}
2 E _i \left( x_s \right) = \frac{\alpha _{n_{\rho p} + 1,|m|}^2}
{ \rho_{0}^2\left( x_s \right)},\quad
\Phi^{m}_{n_{\rho p}}(x_{s})=\frac{\sqrt{2}}{ \rho_0\left( x_s
\right)} \frac{ J_{|m|}(\sqrt{2 E _{n_{\rho p}+1,|m|} \left( x_s
\right)}x_{f}) }{|J_{|m|+1}(\alpha_{n_{\rho p}+1,|m|})|},$$ where $\alpha _{n_{\rho p}+ 1,|m|}=\bar J^{n_{\rho p}+ 1}_{|m|}$ are positive zeros of the Bessel function of the first kind $J_{|m|}(x_f)$, labeled in the ascending order with the quantum number $n_{\rho p}+1
=i= 1,2,...$. The effective potentials (\[sp23a\]) in Eq.(\[sp23\]) for the slow subsystem are calculated numerically in quadratures via the integrals over the fast variable $x_f$ of the basis functions(\[eq100\]) and their derivatives with respect to the parameter $x_s$ using SNA MATRA from Section 2.
Fig. \[enektv\] illustrates the lower part of the non-equidistant spectrum $E(\zeta_{ac})/E_{R}=2 \tilde E_t$ of even states of PSQD Models B and C. There is a one-to-one correspondence rule $n_{\rho
p}+1=n_p=i=n=n_r+1$, $i=1,2,...$ and $n_{z p}=l-|m|$ between the sets of quantum numbers $(n,l,m,\hat\sigma)$ of SQD with the radius $r_0=a=c$ and spheroidal ones $(n_\xi=n_r,n_\eta=l-|m|,m,\sigma)$ of PSQD with the major $c$ and the minor $a$ semiaxes, and the adiabatic set of quantum numbers $(n=n_{\rho p}+1,n_{z p},m,\sigma)$ under the continuous variation of the parameter $\zeta_{ac}=a/c$. The presence of crossing points of similar-parity energy levels in Fig. \[enektv\] under the change of symmetry from spherical $\zeta_{ac}=1$ to axial, i.e., under the variation of the parameter $0<\zeta_{ac}<1$, in the BVP with two variables at fixed $m$ for Model B is caused by the possibility of variable separation in the PSC [@stigun], i.e. r.h.s. of Eq. (\[sp23\]) equals zero. For Model C at each value of the parameter $c$ there is also only a finite number of discrete energy levels, limited by the value $2 U_0$ of the well walls height. As shown in Fig. \[enekts\]b the number of energy levels of PSQD, equal to that of SQD at $a=c= r_0$, which is determined by the product of mass $\mu_e$ of the particle, the well depth $\tilde
U_0$, and the square of the radius $ \tilde r_0$, is reduced with the decrease of the parameter $\tilde a$ (or $\zeta_{ac}$) because of the promotion of the potential curve (lower bound) into the continuous spectrum, in contrast to Models A and B, having countable spectra. Note, that the spectrum of Model C for PSQD or OSQD should approach that of Model B with the growth of the walls height $U_0$ of the spheroidal well. However, at critical values of the ellipsoid aspect ratio it is shown that in the effective mass approximation both the terms (lower bound) and the discrete energy eigenvalues in models of the B type move into the continuum. Therefore, when approaching the critical aspect ratio values, it is necessary to use models such as the lens-shaped self-assembled QDs with a quantum well confined to a narrow wetting layer [@Hawrylak96] or if a minor semiaxis becomes comparable with the lattice constant to consider models (see,e.g.[@Harper]), different from the effective mass approximation.
Conclusion
==========
By examples of the analysis of energy spectra of SQD, PSQD, and OSQD models with thee types of axially symmetric potentials, the efficiency of the developed computational scheme and SNA is demonstrated. Only Model A (anisotropic harmonic oscillator potential) is shown to have an equidistant spectrum, while Models B and C (wells with infinite and finite walls height) possess non-equidistant spectra. In Model C there is a finite number of energy levels. This number becomes smaller as the parameter $a$ or $c$ ($\zeta_{ac}$ or $\zeta_{ca}$) is reduced, because the potential curve (lower bound) moves into the continuum. Models A and B have countable discrete spectra. This difference in spectra allows verification of SQD, PSQD, and OSQD models using experimental data [@Gambaryan], e.g., photoabsorption, from which not only the energy level spacing, but also the mean geometric dimensions of QD may be derived [@79; @Trani; @Lepadatu]. It is shown that there are critical values of the ellipsoid aspect ratio, at which in the approximation of effective mass the discrete spectrum of models with finite-wall potentials turns into a continuous one. Hence, using experimental data, it is possible to verify different QD models like the lens-shaped self-assembled QDs with a quantum well confined to a narrow wetting layer [@Hawrylak96], or to determine the validity domain of the effective mass approximation, if a minor semiaxis becomes comparable with the lattice constant and to proceed opportunely to more adequate models, such as [@Harper].
Note *a posteriori*, that the diagonal approximation of the slow-variable ODE (\[sp23\]) without the diagonal matrix element $W_{ii}$ (so called rude adiabatic approximation) provides the lower estimate of the calculated energy levels. With this matrix element taken into account (adiabatic approximation) the upper estimate of energy is provided, unless in the domain of the energy level crossing points. Therefore, the Born-Oppenheimer (BO) approximation is, generally, applicable only for estimating the ground state at an appropriate value of the small parameter. For Model B in the first BO approximation $2 E_{i}\approx
E_i^{(0)}+E_{i}^{(1)}$ is given by the minimal value of the slow subsystem energy $E_1^{\min}(x_{s})$ in the equilibrium points $x_s=0$ (namely, $E_i^{(0)}=\pi^2n_{o}^2/(2c)^2$ from Eq. (\[eq71\]) for OSQD and $E_i^{(0)}= \alpha_{n_{\rho p}+1}^2/a^2$ from Eq. (\[eq100\]) for PSQD), and by the corresponding energy values $2 E_{i}^{(1)}=\pi(ac)^{-1} n_{o}(2n_{\rho}+|m|+1)$ and $E_{i}^{1}= 2(ac)^{-1}\alpha_{n_{\rho p}+1,|m|} (n_{z}+1/2)$ of the 2D and 1D harmonic oscillator, respectively. In [@Hayk] it is shown that the terms $E_i(x_{s})$ allow high-precision approximation by the Hulten potential. This can be accomplished by means of computer algebra software, e.g., Maple, Mathematica, which allows (in the rude adiabatic approximation) to obtain the lower bound of the spectrum by solving transcendent equations, expressed analytically in terms of known special functions, and to use this approach for further development of our SNA project.
The software package developed is applicable to the investigation of impurity and exciton states in semiconductor nanostructure models. Further development of the method and the software package is planned for solving the quasi-2D and quasi-1D BVPs with both discrete and continuous spectrum, which are necessary for calculating the optical transition rates, channeling and transport characteristics in the models like quantum wells and quantum wires.
Authors thank Profs. K.G. Dvoyan, E.M. Kazaryan and H.A. Sarkisyan for collaboration in the field and Profs. T. Sturm and C. Philips for support of our SNA project. This work was done within the framework of the Protocol No.3967-3-6-09/11 of collaboration between JINR and RAU in dynamics of finite-dimensional quantum models and nanostructures in external fields. The work was supported partially by RFBR (grants 10-01-00200 and 08-01-00604), and by the grant No. MK-2344.2010.2 of the President of Russian Federation.
[4]{} Harrison, P.: Quantum Well, Wires and Dots. Theoretical and Computational Physics of Semiconductor Nanostructures. Wiley, New York (2005) Gambaryan, K.M.: Interaction and Cooperative Nucleation of InAsSbP Quantum Dots and Pits on InAs(100) Substrate. Nanoscale Res Lett., DOI 10.1007/s11671-009-9510-8 (2009) Wojs A., Hawrylak, P., Fafard S., Jacak, L: Electronic structure and magneto-optics of self-assembled quantum dots Phys. Rev. B 54, 5604-5608 (1996) Juharyan, L.A., Kazaryan, E.M., Petrosyan, L.S.: Electronic states and interband light absorption in semi-spherical quantum dot under the influence of strong magnetic field. Solid State Comm. 139, 537–540 (2006) Dvoyan, K.G. Hayrapetyan, D.B., Kazaryan, E.M., Tshantshapanyan, A.A.: Electron States and Light Absorption in Strongly Oblate and Strongly Prolate Ellipsoidal Quantum Dots in Presence of Electrical and Magnetic Fields. Nanoscale Res. Lett. 2, 601–608 (2007) Cantele, G., Ninno, D., Iadonisi, G. Confined states in ellipsoidal quantum dots. J. Phys. Condens. Matt. 12, 9019–9036 (2000). Trani, F., Cantele, G., Ninno, D., and Iadonisi, G.: Tight-binding calculation of the optical absorption cross section of spherical and ellipsoidal silicon nanocrystals. Phys. Rev. B 72, 075423 (2005) Lepadatu, A.-M., Stavarache, I., Ciurea, M.L., Iancu, V.: The influence of shape and potential barrier on confinement energy levels in quantum dots. J. Appl. Phys. 107, 033721 (2010).
Vinitsky, S.I., Gerdt, V.P., Gusev, A.A., Kaschiev, M.S., Rostovtsev, V.A., Samoilov, V.N., Tupikova, T.V., Chuluunbaatar, O.: A symbolic-numerical algorithm for the computation of matrix elements in the parametric eigenvalue problem Programming and Computer Software, 33, 105–116 (2007) Chuluunbaatar, O., Gusev, A., Gerdt, V., Kaschiev, M., Rostovtsev, V., Samoylov, V., Tupikova, T., Vinitsky, S.: A Symbolic-numerical algorithm for solving the eigenvalue problem for a hydrogen atom in the magnetic field: cylindrical coordinates. Lect. Notes in Computer Science, 4770, 118-133 (2007) Chuluunbaatar, O., Gusev, A.A., Abrashkevich, A.G., Amaya-Tapia, A., Kaschiev, M.S., Larsen, S.Y. and Vinitsky, S.I.: KANTBP: A program for computing energy levels, reaction matrix and radial wave functions in the coupled-channel hyperspherical adiabatic approach. Comput. Phys. Commun. 177, 649–675 (2007) Chuluunbaatar, O., Gusev A.A., Gerdt, V.P., Rostovtsev, V.A., Vinitsky, S.I., Abrashkevich, A.G., Kaschiev, M.S., Serov, V.V.: POTHMF: A program for computing potential curves and matrix elements of the coupled adiabatic radial equations for a hydrogen-like atom in a homogeneous magnetic field Comput. Phys. Commun. 178, 301–330, (2008) Chuluunbaatar, O., Gusev, A.A., Vinitsky, S.I., Abrashkevich, A.G.: ODPEVP: A program for computing eigenvalues and eigenfunctions and their first derivatives with respect to the parameter of the parametric self-adjoined Sturm-Liouville problem. Comput. Phys. Commun. 180, 1358–1375 (2009) Vinitsky, S.I., Chuluunbaatar, O., Gerdt, V.P., Gusev, A.A., Rostovtsev, V.A.: Symbolic-numerical algorithms for solving parabolic quantum well problem with hydrogen-like impurity Lect. Notes in Computer Science 5743, 334-349 (2009) Kantorovich, L.V., and Krylov, V.I.: Approximate Methods of Higher Analysis. Wiley, New York (1964) Born, M., Huang, X.: Dynamical Theory of Crystal Lattices. The Clarendon Press, Oxford (1954) Abramowitz, M., Stegun, I.A.: Handbook of Mathematical Functions, Dover, New York, (1965) Courant, R., Hilbert, D.: Methods of Mathematical Physics. V. 1, Wiley (1989) Strang, G., Fix, G.J.: An Analysis of the Finite Element Method, Prentice-Hall, Englewood Cliffs, New York, (1973). Schultz, M.H.: $L^2 $ Error Bounds for the Rayleigh–Ritz–Galerkin Method, SIAM J. Numer. Anal. 8, 737–748 (1971) Harper, P.G.: Single Band Motion of Conduction Electrons in a Uniform Magnetic Field. Proc. Phys. Soc. A 68, 874–878 (1955)
| {
"pile_set_name": "ArXiv"
} |
[T. P. Shestakova]{}[Department of Theoretical and Computing Physics, Rostov State University,\
Sorge Str. 5, Rostov-on-Don 344090, Russia\
E-mail: [email protected]]{}
In this talk I briefly describe the results of the work done by G. M. Vereshkov, V. A. Savchenko and me[@Ours]. Our aim was to explore the possibility of constructing quantum geometrodynamics of a closed universe by a strict mathematical method without using any assumption not permitting detailed mathematical proofs. The proposed version of quantum geometrodynamics turns out to be a gauge-noninvariant theory [*radically*]{} distinguished from the Wheeler – DeWitt QGD by its content. The foundations for this investigation lay in the peculiarities of the Wheeler – DeWitt QGD among which worth mentioning are mathematical problems of gauge invariance and problems of interpretation.
One possible starting point for constructing QGD is to introduce the Batalin – Fradkin – Vilkovisky (BFV) transition amplitude $$\label{amp}
\langle\,f|i\,\rangle=\!\int\!D\mu\;\exp\,\left(iS_{eff}\right),$$ where $S_{eff}$ is a BRST-invariant effective action, $D\mu$ is a measure in extended phase space[@Hennaux]. Independence of the amplitide (\[amp\]) on a gauge condition is ensured by asymptotic boundary conditions. However, there is no asymptotic states in a closed universe and an observer cannot be removed to infinity from the investigated object which is the Universe as a whole. So appealling to asymptotic boundary conditions does not seem to be justified when constructing QGD of a closed universe. As for the transition amplitude (\[amp\]) considered without the asymptotic boundary conditions, there is no strict mathematical way to prove its gauge invariance. On the contrary, it may be demonstrated by explicit calculations that the ampitude (\[amp\]) inevitably contains gauge-noninvariant effects (see [@Ours]).
Another way is based on the Wheeler – DeWitt equation[@DeWitt]. Let us emphasize that the Wheeler – DeWitt equation is not deducible by correct mathematical methods from a path integral or somehow else: it can be just postulated. The principle of gauge invariance is commonly thought to be a motivation for postulating the Wheeler – DeWitt equation. On the other hand, the Wheeler – DeWitt equation is known to be noninvariant under choice of a gauge variable, the lapse function $N$ being usually considered as such a variable[@HP; @Halliwell]. However, the transition to another gauge variable is formally equivalent to imposing a new gauge condition, and vice versa. The latter reflects an obvious fact that the choice of gauge variables and the choice of gauge conditions have an unified interpretation: they both determine equations for the metric components $g_{0\mu}$, fixing a reference system. So, as a matter of fact, the parametrisation noninvariance of the Wheeler – DeWitt equation is ill-hidden gauge noninvariance.
It is well-known that in the Wheeler – DeWitt theory there is no quantum evolution of state vector in time. A wave function satisfying the Wheeler – DeWitt equation describes the past of the Universe as well as its future with all observers being inside the Universe in different stages of its evolution, and all observations to be made by these observers. Thus the Wheeler – DeWitt theory does not use the postulate about the reduction of a wave packet. One cannot appeal to the Copenhagen interpretation of quantum theory within the limits of the Wheeler – DeWitt QGD. One has to turn to the many-worlds interpretation of the wave function proposed by Everett[@Everett] and applied to QGD by Wheeler[@Wheeler]. The wave function of the Universe satisfying the Wheeler – DeWitt equation and certain boundary conditions is thought to be a branch of a many-worlds wave function that corresponds to a certain universe; other branches being selected by other boundary conditions. So the information about the continuous reduction of the wave function in the process of evolution of the Universe including certain observers inside is contained in the boundary conditions for the wave function only.
Bearing in mind all the mentioned above, we propose to investigate a more general theory. The features of the theory are:
[$\bullet$]{}
A basic equation for a wave function of the Universe is derived by a well-defined mathematical procedure.
The assumption about asymptotic states is not taken into account.
The theory admits of Copenhagen interpretation.
To illustrate our results it is convenient to consider the Bianchi IX model for its mathematical simplicity and physical meaningfulness. I shall remind that the interval in the Bianchi IX model looks like $$\label{ds}
ds^2=N^2(t)\,dt^2-\eta_{ab}(t)e^a_ie^b_kdx^idx^k;$$ $$\label{eta_ab}
\eta_{ab}(t)=\mathop{\rm diag}\nolimits\left(a^2(t),b^2(t),c^2(t)\right).$$ We use the parametrization $$\label{abc}
a=\displaystyle\frac12 r\exp\left[
\frac12\left(\sqrt{3}\,\varphi+\chi\right)\right];\quad
b=\displaystyle\frac12 r\exp\left[
\frac12\left(-\sqrt{3}\,\varphi+\chi\right)\right];\quad
c=\displaystyle\frac12 r\exp\left(-\chi\right);$$ $$\label{zeta}
Q^a=(q,\,\varphi,\,\chi,\,\phi,\,\ldots);\quad
q = 2\ln r;\quad
\zeta(\mu,Q)=\ln\frac{r^3}N,$$ where $\phi$ stands for scalar fields, $\zeta(\mu,Q)$ is an arbitrary function defining a gauge variable $\mu$ through the lapse function $N$. The Bianchi IX model can be considered as a model of a Friedman – Robertson – Walker closed universe with $r(t)$ being a scale factor, on which a transversal nonlinear gravitational wave $\varphi(t),\chi(t)$ is superposed.
The derivation of an equation for a wave function of the Universe implies going over to a path integral with the effective action in a Lagrange form. Since the algebra of transformations generated by constraints is closed for the model, the transition amplitude (\[amp\]) can be reduced to the path integral over extended configurational space involving ghost and gauge variables with the Faddeev – Popov effective action $$S_{ef\!f}=
\!\int\!dt\,\biggl\{\displaystyle\frac12\exp\left[
\zeta\left(\mu,Q^a\right)\right]\gamma_{ab}\dot{Q}^a\dot{Q}^b
-\exp\left[-\zeta\left(\mu,Q^a\right)\right]U\left(Q^a\right)+$$ $$\label{Seff}
+\lambda\left(\dot{\mu}-f_{,a}\dot{Q}^a\right)
+\frac i{\zeta_{,\mu}}\dot{\bar{\theta}}\dot{\theta}\biggr\}.$$ We confine attention to the special class of gauges not depending on time $$\label{mu,f,k}
\mu=f(Q)+k;\quad
k={\rm const},$$ or, in a differential form, $$\label{diff_form}
\dot{\mu}=f_{,a}\dot{Q}^a,\quad
f_{,a}\stackrel{def}{=}\frac{\partial f}{\partial Q^a};$$ $\zeta_{,\mu}=\!\partial \zeta(\mu,Q)/\partial\mu;\;\theta,\bar{\theta}$ are the Faddeev – Popov ghosts after replacement $\bar{\theta}\to-i\bar{\theta}$; indices $a,b,\ldots$ are raised and lowered with the “metric" $$\label{3-metric}
\gamma_{ab}=\mathop{\rm diag}\nolimits(-1,\,1,\,1,\,1,\,\ldots);$$ $$\label{U,Ug,Us}
U(Q)=e^{2q}\,U_g(\varphi,\chi)+e^{3q}\,U_s(\phi),$$ $$\begin{aligned}
U_g(\varphi,\chi)&=&
\frac23\left\{\exp\left[2\left(\sqrt{3}\,\varphi+\chi\right)\right]
+\exp\left[2\left(-\sqrt{3}\,\varphi+\chi\right)\right]
+\exp(-4\chi)-\right.\nonumber\\
&-&\left.2\exp\left[-\left(\sqrt{3}\,\varphi+\chi\right)\right]
-2\exp\left(\sqrt{3}\,\varphi-\chi\right)
-2\exp(2\chi)\right\}.\end{aligned}$$
Variation of the action (\[Seff\]) yields a Lagrangian set of equations mathematically equivalent to canonical equations in EPS. This set of equations can be called [*conditionally-classical*]{} for the presence of Grassmannian variables. It is gauge-noninvariant.
Any gauge condition fixes a reference system, the latter representing the observer in the theory of gravity. So the action (\[Seff\]) describes the integrated system “the physical object + observation means". According to Landau and Lifshitz[@Lan], a continual medium with broken symmetry under diffeomorphism group transformation must be considered as a reference system in the theory of gravity. Inside the medium a periodic process is going, its characteristic being used for choosing metric measurements standards. Let us note that within the limits of the classical theory one is not able to point to an object with the above properties. However, quantum theory gives us the notion about such an object – it is the vacuum condecate. Thus the gauge-fixing term in (\[Seff\]) corresponds a specific subsystem being referred to as “the gravitational vacuum condensate". The investigation of the conditionally-classical set of equations reveals the existence of a conserved quantity $E$ describing the subsystem. As a result, the Hamiltonian constraint $H_{ph}=0$ of general relativity is replaced by the constraint $H=E$, where $H_{ph}$ is a Hamiltonian of gravitational and matter fields, $H$ is a Hamiltonian in EPS.
The latter means that a Hamiltonian spectrum in the appropriate quantum theory is not limited by the unique zero eigenvalue. So the main part in this version of QGD is given to [*the Schrödinger equation in extended configurational space*]{} for a wave function of the Universe, and finding a spectrum of $E$ becomes one of the main tasks of quantum geometrodynamics in EPS. We would like to emphasize that the problem of time, so typical of the Wheeler – DeWitt QGD, does not arise here. Moreover, one gains the opportunity to appeal to the Copenhagen interpretation.
The Schrödinger equation derived from the path integral with the effective action (\[Seff\]) by the standard method[@Cheng] originated from Feinman reads $$\label{SE1}
i\,\frac{\partial\Psi(Q^a,\mu,\theta,\bar{\theta};\,t)}{\partial t}
=H\Psi(Q^a,\,\mu,\,\theta,\,\bar{\theta};\,t),$$ where $$\label{H}
H=-i\,\zeta_{,\mu}\frac{\partial}{\partial\theta}
\frac{\partial}{\partial\bar{\theta}}
-\frac1{2M}\frac{\partial}{\partial Q^{\alpha}}MG^{\alpha\beta}
\frac{\partial}{\partial Q^{\beta}}+{\rm e}^{-\zeta}(U-V);$$ $$\label{M}
M={\rm const}\cdot \zeta_{,\mu}\exp\left(\frac{K+3}2\,\zeta\right);$$ $$\begin{aligned}
V&=&-\frac3{12}\frac{(\zeta_{,\mu})^a(\zeta_{,\mu})_a}{\zeta_{,\mu}^2}
+\frac{(\zeta_{,\mu})^a_a}{3\zeta_{,\mu}}
+\frac{K+1}{6\zeta_{,\mu}}\,\zeta_a(\zeta_{,\mu})^a+\nonumber\\
&+&\frac1{24}\left(K^2+3K+2\right)\,\zeta_a\zeta^a+\frac{K+2}6\,\zeta_a^a;\end{aligned}$$ $$\label{Galpha_beta}
\zeta_a=\frac{\partial \zeta}{\partial Q^a}
+f_{,a}\frac{\partial \zeta}{\partial\mu};\quad
G^{\alpha\beta}={\rm e}^{-\zeta}\left(
\begin{array}{cc}
f_{,a}f^{,a}&f^{,a}\\
f^{,a}&\gamma^{ab}
\end{array}
\right),$$ $\alpha=(0,a),\;Q^0=\mu,\;K$ is a number of scalar fields included in the model; the wave function is defined on extended configurational space with the coordinates $Q^a,\,\mu,\,\theta,\,\bar{\theta}$.
It is worth noting that no ill-definite mathematical expression arises when deriving Eq.(\[SE1\]). It is due to using the nondegenarate conditionally-classical set of equations to approximate the path integral instead of degenerate gauge-invariant equations. Alternatively, Eq.(\[SE1\]) can be obtained from quantum canonical equations in EPS.
The general solution to the Schrödinger equation (\[SE1\]) has the folloowing structure: $$\label{time-depend.WF}
\Psi(Q^a,\,Q^0,\,\theta,\,\bar{\theta};\,t)
=\int\Psi(Q^a)\exp(-iEt)(\bar{\theta}+i\theta)\,
\delta(\mu-f(Q^a)-k)\,dE\,dk.$$ where $\Psi(Q^a)$ is a solution to the stationary equation $$\label{station.phys.SE}
H^0\,\Psi(Q^a)=E\,\Psi(Q^a),$$ $$\label{H0}
H^0=\left.\left[-\frac1{2M}\frac{\partial}{\partial Q^a}
M{\rm e}^{-\zeta}\gamma^{ab}\frac{\partial}{\partial Q^b}
+{\rm e}^{-\zeta}(U-V)\right]\right|_{\mu=f(Q^a)+k}.$$ The wave function (\[time-depend.WF\]) carries the information on 1) a physical object, 2) observation means (a gravitational vacuum condensate), 3) correlations between the physical object and observation means. Observation means are represented by the factored part of the wave function – by the $\delta$-function of a gauge and by the ghosts; the physical object is described by the function $\Psi(Q^a)$; the correlations are manifested in the effective potential $V$ and in the spectrum $E$, the gravitational vacuum condensate thus being a cosmolgical evolution factor. The dependence of the wave function (\[time-depend.WF\]) on ghosts is determined by the demand of norm positivity.
The question remains if it possible to go over to some particular solution satisfying the Wheeler – DeWitt equation from the general solution (\[time-depend.WF\]). To do it, one has to eliminate correlations between the properties of the physical object and those of observation means. Then, one puts $E=0$ and fixes the gauge $\mu=k$, making use of the formal possibility to go over to any given gauge by means of transformation of the parametrization function $\zeta$. In this case under some limitations on the measure (\[M\]) one gets that the physical part of the wave function $\Psi(Q^a)$ satisfies the Wheeler – DeWitt equation. However, it is of importance to emphasize that the correlations between the physical object and observation means cannot be eliminated completely: the information about a chosen reference system is contained in the parametization function and the parametrization function essentially determines the effective potential $V$. The same picture arises when one obtains the Wheeler – DeWitt equation as a corollary of the superselection rules for BRST- and anti-BRST-invariant quantum states.
In conclusion I shall describe the role of the gravitational vacuum condensate as a cosmological evolution factor. As I mentioned above, the gravitational vacuum condensate is a continual medium, a state equation of the medium depending on a chosen gauge condition. The state of the condensate is characterized by the parameter $E$, and the relation between $E$ and other parameters of the theory determines a cosmological scenario. For example, in a simplified model with one of two gravitational waves, $\varphi(t)$, being frozen out, taking the parametrization function and the gauge condition to be $\zeta(\mu,Q^a)=\mu=k$, one obtains an ultrastiff state equation of the condensate: $$\label{state_eq}
p=\epsilon.$$ The scalar field $\phi$ and the gravitational vacuum condensate together form the two-component medium with a positive or negative energy density depending on the parameter $E$. A cosmological evolution scenario may contain the following phenomena[@Ours]:
[$\bullet$]{}
cosmological expansion and contraction of space;
cosmological singularity;
compactification of space dimensions;
asymptotically stationary space of less dimensions;
inflation of the Universe.
The goal of the future investigation is to work out a full cosmological scenario, in the sense as it was understood by Grishchuk and Zeldovich[@GZ], based on the proposed version of quantum geometrodynamics. After Grishchuk and Zeldovich we think that the Universe was created from the state “Nothing” where there exist neither space with its geometry nor time. In this state there is no gravitational wave or matter field, and no gravitational vacuum condensate either. After creation, however, the Universe occurs in a state with broken symmetry under diffeomorphism group, and the presence of the gravitational vacuum condensate is a characteristic feature of this picture. A state spectrum can be found by solving a Schrödinger equation for a given model. Then, one may suppose that in the course of cosmological evolution the Universe appears to be in the state with $E=0$, where the correlations of physical fields with the condensate are minimal. In this state the Universe in a large scale is described by general relativity.
[99]{} =-5pt V.A. Savchenko, T.P. Shestakova and G.M. Vereshkov, [**gr-qc/9809086, gr-qc/9810035**]{}. M. Hennaux, [*Phys. Rep.*]{}, [**126**]{}, 1 (1985). B. S. DeWitt, [*Phys. Rev.*]{} [**160**]{}, 1113 (1967). S. W. Hawking and D. N. Page, [*Nucl. Phys.*]{} [**B264**]{}, 185 (1986). J. J. Halliwell, [*Phys. Rev.*]{} [**D38**]{}, 2468 (1988). H. Everett, [*Rev. Mod. Phys.*]{} [**29**]{}, 454 (1957). J. A. Wheeler, ibid., 463. L. D. Landau and E. M. Lifshitz, [*Theory of Fields*]{} (Nauka, Moscow, 1988). K. S. Cheng, [*J. Math. Phys.*]{}, [**13**]{}, 1723 (1986). L. P. Grishchuk and Y. B. Zeldovich [*in:*]{} “Quantum Structure of Space and Time”, eds. M. Duff, C. Isham (Cambridge University Press, Cambridge, 1982).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we provide a criterion of essential self-adjointness for operators in the tensor product of a separable Hilbert space and a Fock space. The class of operators we consider may contain a self-adjoint part, a part that preserves the number of Fock space particles and a non-diagonal part that is at most quadratic with respect to the creation and annihilation operators. The hypotheses of the criterion are satisfied in several interesting applications.'
author:
- Marco Falconi
title: 'Self-Adjointness criterion for operators in Fock spaces'
---
Introduction {#sec:introduction}
============
Let $\mathscr{H}_1$, $\mathscr{H}_2$ be separable Hilbert spaces. We consider the following space: $$\label{eq:2}
\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)\; ;$$ where $\Gamma_s(\mathscr{K})$ is the symmetric Fock space based on $\mathscr{K}$ [see @MR0493420; @MR3060648; @MR0044378 for mathematical presentations of Fock spaces and second quantization]. The symmetric structure of the Fock space does not play a role in the argument: in principle it is possible to formulate the same criterion for anti-symmetric Fock spaces $\mathscr{H}_1\otimes\Gamma_a(\mathscr{H}_2)$. We focus on symmetric spaces, the corresponding antisymmetric results should be deduced without effort.
We are interested in proving a criterion of essential self-adjointness for densely defined operators of the form: $$\label{eq:1}
H= H_{01}\otimes 1 + 1\otimes H_{02}+H_I\; ;$$ with suitable assumptions on $H_{01}$, $H_{02}$ and $H_I$. Operators based on these spaces and with such structure are crucial in physics, to describe the quantum dynamics of interacting particles and fields.
Self-adjointness of operators in Fock spaces has been widely studied, in particular in the context of Constructive Quantum Field Theory [e.g. @MR947959; @MR0215585; @MR0210416; @MR0270671; @MR0266533; @Se] and Quantum ElectroDynamics [e.g. @nelson:1190; @MR1891842; @MR2097788; @BFS; @BFS2; @BFSS; @MR797278; @MR1809881; @MR2436496]. A variety of advanced tools has been utilized, for even “simple” systems present technical difficulties to overcome: many questions still remain unsolved.
In some favourable situations, however, it is possible to take advantage of the peculiar structure of the Fock space and prove essential self-adjointness with almost no effort. The idea first appeared in a paper by @MR0266533; and the author utilized it in [@falconi-phd; @Ammari:2014aa] for the Nelson model with cut off: essential self-adjointness can be proved with less assumptions than using the Kato-Rellich Theorem (and that becomes particularly significative in dimension two), see Section \[sec:nels-type-hamilt\]. Another remarkable application is the Pauli-Fierz Hamiltonian describing particles coupled with a radiation field. For general coupling constants, essential self-adjointness has been first proved in a probabilistic setting, using stochastic integration [@MR1773809; @MR1891842]. In this paper we prove the same result directly in Section \[sec:pauli-fierz-hamilt\], applying the criterion formulated in Assumptions \[ass:1\], \[ass:2\] and Theorem \[thm:1\].
In the literature, self-adjointness of operators in Fock spaces has been studied using various tools of functional analysis: the Kato-Rellich and functional integration arguments mentioned above are two examples, as well as the Nelson commutator theorem [@MR1814991]. For each particular system, a strategy is utilized ad hoc: the more complicated is the correlation between $\mathscr{H}_1$ and $\Gamma_s(\mathscr{H}_2)$, the more difficult is the strategy. We realized that, if we take suitable advantage of the fibered structure of the Fock space, the type of interaction between the spaces is not so relevant. This was a strong motivation to study the problem from a general perspective. Due to the variety of possible applications, an effort has been made to formulate the necessary assumptions in a general form. Roughly speaking, the essential requirement is that the part of $H_I$ that does not commute with the number operator of $\Gamma_s(\mathscr{H}_2)$ is at most quadratic with respect to the creation and annihilation operators. As anticipated, the space $\mathscr{H}_1$ does not play a particular role, as long as $H_I$ behaves sufficiently well with respect to $H_{01}$.
Paper organization. {#sec:paper-organization}
-------------------
In Section \[sec:defin-notat\] we introduce the notation, and recall some basic definitions of operators in Fock spaces. In Section \[sec:assumptions-h\] we formulate the necessary assumptions on the operator $H$. In Section \[sec:essent-self-adjo\] we prove the criterion. In Section \[sec:applications\] we outline some of the most interesting applications. Finally in Section \[sec:conclusive-remarks-1\] we give some conclusive remarks, and an extension of the criterion to semi-bounded quartic operators.
Definitions and notations. {#sec:defin-notat}
--------------------------
- Let $\mathscr{K}$ be a separable Hilbert space. Then the symmetric Fock space $\Gamma_s(\mathscr{K})$ is defined as the direct sum: $$\Gamma_s(\mathscr{K})=\bigoplus_{n=0}^{\infty}\mathscr{K}^{\otimes_s n}\; ,$$ where $\mathscr{K}^{\otimes_s n}$ is the $n$-fold symmetric tensor product of $\mathscr{K}$, and $\mathscr{K}^{\otimes_s 0}:=\mathds{C}$.
- Let $h:\mathscr{K}\supseteq D(h)\to \mathscr{K}$ be a densely defined self-adjoint operator on a separable Hilbert space $\mathscr{K}$. Its second quantization $d\Gamma(h)$ is the self-adjoint operator on $\Gamma_s(\mathscr{K})$ defined by $$d\Gamma(h)\rvert_{D(h)^{\otimes_s n}}=\sum_{k=1}^n 1\otimes\dotsm\otimes \underbrace{h}_{k}\otimes \dotsm\otimes 1\; .$$ Let $u$ be a unitary operator on $\mathscr{K}$. We define $\Gamma(u)$ to be the unitary operator on $\Gamma_s(\mathscr{K})$ given by $$\Gamma(u)\rvert_{\mathscr{K}^{\otimes_s n}}=\bigotimes_{k=1}^n u\; .$$ If $e^{it h}$ is a group of unitary operators on $\mathscr{K}$, $\Gamma(e^{it h})=e^{it d\Gamma(h)}$.
- $N:=d\Gamma (1)$ the number operator of $\Gamma_s(\mathscr{H}_2)$.
- $H_0:= H_{01}\otimes 1 + 1\otimes H_{02}$; the free Hamiltonian.
- If $X$ is a self-adjoint operator on a Hilbert space, we denote by $D(X)$ its domain, by $q_X(\cdot ,\cdot)$ the form associated with $X$ and by $Q(X)$ the form domain.
- Let $\mathscr{K}$ be a Hilbert space; $\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$ a collection of disjoint subspaces of $\mathscr{K}$; $X$ an operator densely defined on $\mathscr{K}$. We say that $\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$ is invariant for $X$ if $\forall j\in \mathds{N}$, $X$ maps $D(X)\cap \mathscr{K}^{(j)}\to \mathscr{K}^{(j)}$, and $D(X)\cap
\mathscr{K}^{(j)}$ is dense in $\mathscr{K}^{(j)}$.
- Let $\mathscr{K}$ be a Hilbert space; $\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$ a collection of disjoint closed subspaces of $\mathscr{K}$ such that $\bigoplus_{j\in\mathds{N}}\mathscr{K}^{(j)}=\mathscr{K}$. Then we call the collection complete, and we define the dense subset $f_0(\mathscr{K}^{(\cdot)})$ of $\mathscr{K}$ as: $$\label{eq:3}
f_0(\mathscr{K}^{(\cdot)})=\Bigl\{\phi\in\mathscr{K}, \exists n\in\mathds{N}\text{ s.t. }\phi\in\bigoplus_{j=0}^n\mathscr{K}^{(j)} \Bigr\}\; .$$ Also, we denote by $\mathds{1}_j(\mathscr{K}^{(\cdot)})$ the orthogonal projection on $\mathscr{K}^{(j)}$, by $\mathds{1}_{\leq
n}(\mathscr{K}^{(\cdot)})$ the orthogonal projection on $\bigoplus_{j=0}^n\mathscr{K}^{(j)}$.
- Let $\mathscr{K}\ni f,g$ be two elements of a separable Hilbert space. We define the creation $a^{*}(f)$ and annihilation $a(f)$ operators on $\Gamma_s(\mathscr{K})$ by their action on $n$-fold tensor products (with $a(f)\phi_0=0$ for any $\phi_0\in
\mathscr{K}^{\otimes_s 0}=\mathds{C}$): $$\begin{aligned}
a(f)g^{\otimes n}&=\sqrt{n} \; \langle f , g\rangle_{\mathscr{K}} \; g^{\otimes (n-1)}\\
a^{*}(f)g^{\otimes n}&=\sqrt{n+1} \; f\otimes_s g^{\otimes n}\; .
\end{aligned}$$ They extend to densely defined closed operators and are adjoint of each other: we denote again by $a^{\#}(f)$ their closures. For any $f\in \mathscr{K}$, $D(a^{*}(f))=D(a(f))$ with $$D(a(f))=\Bigl\{\phi\in \Gamma_s(\mathscr{K})\;:\; \sum_{n=0}^{\infty}(n+1)\lVert \langle f(x) , \phi_{n+1}(x,X_n)\rangle_{\mathscr{K}(x)} \rVert_{\mathscr{K}^{\otimes_s n}(X_n)}^2<+\infty \Bigr\}\; ,$$ where $\phi_{n+1}=\phi\bigr\rvert_{\mathscr{K}^{\otimes_s n+1}}$; also $D(a(f))\supset D(d\Gamma(1)^{1/2})$, $D(a(f))\supset
f_0(\mathscr{K}^{(\cdot)})$. They satisfy the Canonical Commutation Relations $[a(f_1),a^{*}(f_2)]=\langle f_1 ,
f_2\rangle_{\mathscr{K}}$ on suitable domains (e.g. $f_0(\mathscr{K}^{(\cdot)})$).
- We decompose $\Gamma_s(\mathscr{H}_2)$ in its subspaces with fixed number of particles as usual: $\forall n\in \mathds{N}$, define $\mathscr{H}_2^{(n)}:=\mathscr{H}_2^{\otimes_s n}$, with the convention $\mathscr{H}_2^{(0)}=\mathds{C}$. Then $\{\mathscr{H}^{(n)}_{2}\}_{n\in\mathds{N}}$ is a complete collection of closed disjoint subspaces of $\Gamma_{s}(\mathscr{H}_2)$ invariant for $N$.
- Let $X$ be an operator on $\mathscr{H}$. We say that $X$ is diagonal if $\{\mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}\}_{n\in\mathds{N}}$ is invariant for $X$; $X$ is non-diagonal if for all $n\in\mathds{N}$ and $\phi\in D(X)\cap \mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}$, $X\phi \notin \mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}$.
Assumptions on $H$ {#sec:assumptions-h}
==================
In this section we discuss Assumptions \[ass:1\] and \[ass:2\](\[ass:3\]). In Section \[sec:applications\] below they are checked in concrete examples.
We recall that our Hilbert space $\mathscr{H}$ has the form $$\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)\; ;$$ while the operator is $$H= H_{01}\otimes 1 + 1\otimes H_{02}+H_I\; .$$
We separate the assumptions on $H_0$ from the ones on $H_I$, to improve readability. On $H_I$ we require either Assumption \[ass:2\] or Assumption \[ass:3\]. In \[ass:2\] the non-diagonal part of $H_I$ can be more singular: that restricts the diagonal part to be at most quadratic in the creation and annihilation operators. In \[ass:3\] on the other hand is assumed more regularity on the non-diagonal part of $H_I$, allowing for a more singular diagonal part.
[A$_{0}$]{} \[ass:1\] $H_{01}$ and $H_{02}$ are semi-bounded self-adjoint operators. We denote respectively by $-M_1$ and $-M_2$ their lower bounds. Furthermore, $\forall t\in \mathds{R}$, $\{\mathscr{H}_2^{(n)}\}_{n\in\mathds{N}}$ is invariant for $e^{it
H_{02}}$.
This is quite natural. In physical systems the Hamiltonian is often split in a part describing the free dynamics (usually a self-adjoint and positive unbounded operator), and an interaction part. The invariance of the $n$-particles subspaces is also a usual feature of free quantum theories: let $h_{02}$ be a semi-bounded self-adjoint operator on the one-particle space $\mathscr{H}_2$; then the second quantization $d\Gamma(h_{02})$ is self-adjoint, and the group $\Gamma(e^{it h_{02}})$ generated by it satisfies the assumption.
[A$_I$]{} \[ass:2\] $H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in $\mathscr{H}$. Furthermore $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$, $$\label{eq:5}
H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1\otimes \mathscr{H}_2^{(n+i)}\; .$$ Also, $H_I$ satisfies the following bound: $\forall
n\in\mathds{N}$ $\exists C>0$ such that $\forall\psi\in
\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$: $$\label{eq:6}
\begin{split}
\lvert \langle \psi , H_I\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n+i)}}^2 \Bigl[(n+1)^2\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\Bigr]\; ;
\end{split}$$ where we define $\psi_n:=1\otimes
\mathds{1}_n(\mathscr{H}_2^{(\cdot)})\psi$.
Consider Assumption \[ass:2\]. First of all, $H_I$ has to be sufficiently regular, i.e. relatively bounded by $H_0$ (in some sense) when restricted to the subspaces $\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$. Essentially, we require that $H_I$ is at most quadratic in the annihilation and creation operators, as reflected by the $n$-dependence in .
[A$_I^{\prime}$]{} \[ass:3\] $H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in $\mathscr{H}$. Furthermore $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$, $$\label{eq:24}
H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1\otimes \mathscr{H}_2^{(n+i)}\; .$$ Also, $H_I=H_{diag}+H_2$ with the following properties:
1. \[item:1\] $H_{diag}$ is diagonal; $H_2$ is non-diagonal.
2. \[item:2\] $H_{diag}$ satisfies the following bound. $\forall n\in \mathds{N}$ $\exists C(n)>0$ such that $\forall\psi \in\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes
1)\cap Q(1\otimes H_{02})\cap\mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$: $$\label{eq:16}
\begin{split}
\lvert \langle \psi , H_{diag}\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(n)\lVert \psi_{n} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^2 \Bigl(q_{H_{01}\otimes 1}(\phi,\phi)+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}\\+1)\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\; .
\end{split}$$
3. \[item:3\] $H_2$ satisfies the following bound. $\forall
n\in \mathds{N}$ $\exists C>0$ such that $\forall\psi
\in\mathscr{H}$, $\forall\phi\in \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$: $$\label{eq:19}
\begin{split}
\lvert \langle \psi , H_2\phi \rangle_{\mathscr{H}}\rvert\leq C(n+1) \lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}\sum_{\substack{i=-2 \\ i\neq 0}}^2\lVert \psi_{n+i} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n+i)}} \; .
\end{split}$$
Assumption \[ass:3\] is similar to Assumption \[ass:2\]. However since the non-diagonal quadratic part $H_2$ is more regular than before, we can be less demanding on the diagonal part $H_{diag}$: it has still to be bounded in a suitable sense by $H_0$, but it can be non-quadratic with respect to the creation and annihilation operators.
In some applications, there is a decomposition of $\mathscr{H}_{1}$ invariant for $H$. For example, it may happen that $\mathscr{H}_1$ is also a Fock space but $H$ leaves invariant each sector with fixed number of particles. In this situation, we can prove essential self-adjointness with little less regularity on the assumptions. In particular, Assumption \[ass:2\] would be changed in:
$H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in $\mathscr{H}$. Furthermore there exists a complete collection $\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}$ invariant for $H_0$ and $H_I$ such that: $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$, $$H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}\; .$$ Also, $H_I$ satisfies the following bound: $\forall j,n\in\mathds{N}$ $\exists C(j)>0$ such that $\forall\psi\in \mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$: $$\begin{split}
\lvert \langle \psi , H_I\phi\rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(j)\sum_{i=-2}^2\lVert\psi_{j,n+i}\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}}^2\Bigl[(n+1)^2\lVert \phi\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}\Bigr)\Bigr]\; ;
\end{split}$$ where we define $\psi_{j,n}:=\mathds{1}_j(\mathscr{H}_1^{(\cdot)})\otimes
\mathds{1}_n(\mathscr{H}_2^{(\cdot)})\psi$.
Theorem \[thm:1\] would then read:
Assume \[ass:1\] and \[ass:2\](\[ass:3\]). Then $H$ is essentially self adjoint on $D(H_{01}\otimes 1)\cap D(H_{02}\otimes 1)\cap
f_0(\mathscr{H}_1^{(\cdot)}\otimes \mathscr{H}_2^{(\cdot)})$.
Direct proof of self-adjointness {#sec:essent-self-adjo}
================================
In this section we present the criterion of essential self-adjointness . The strategy is to prove that $\mathrm{Ran}(H\pm i)$ is dense in $\mathscr{H}$, by an argument of reductio ad absurdum. As already discussed, the non-diagonal part of $H_I$ is at most quadratic with respect to the annihilation and creation operators of $\Gamma_s(\mathscr{H}_2)$, and that plays a crucial role in the proof. We prove Theorem \[thm:1\] assuming \[ass:2\]; the other case being analogous.
\[thm:1\] Assume \[ass:1\] and \[ass:2\](\[ass:3\]). Then $H$ is essentially self adjoint on $D(H_{01}\otimes 1)\cap D(H_{02}\otimes
1)\cap \mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$.
Let $\psi\in\mathscr{H}$, $z\in \mathds{C}$ with ${\mathrm{Im}}z\neq
0$. Suppose that $\forall \phi\in D(H_{01}\otimes 1)\cap D(1\otimes
H_{02})\cap \mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$: $$\label{eq:7}
\langle \psi , (H-z)\phi\rangle_{\mathscr{H}}=0\; .$$ Then it suffices to show that $\psi = 0$. This is done in few steps. Let $n\in \mathds{N}$ and $\phi_{n}\in D(H_{01}\otimes
1)\cap D(1\otimes H_{02})\cap \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$. For all $n\in\mathds{N}$, the space $Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$ with the scalar product: $$\label{eq:8}
\langle \,\cdot \, ,\, \cdot\,\rangle_{\mathscr{X}_{n}}= q_{H_{01}\otimes 1}(\,\cdot\, ,\,\cdot\,)+q_{1\otimes H_{02}}(\,\cdot\, ,\, \cdot\,) +(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\langle \, \cdot \, ,\, \cdot\,\rangle_{\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}}$$ is complete, and therefore a Hilbert space. We denote it by $\mathscr{X}_{n}$. Then together with Assumption \[ass:1\] imply, since $\phi_{n}\in D(H_{01}\otimes
1)\cap D(1\otimes H_{02})$: $$\label{eq:9}
\langle \psi_{n}, \phi_{n} \rangle_{\mathscr{X}_{n}}= (z+\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\langle \psi_{n} ,\phi_{n}\rangle_{\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}}-\langle \psi , H_I\phi_{n} \rangle_{\mathscr{H}}\; .$$ Use bound and then Riesz’s Lemma on $\mathscr{X}_{n}$: it follows that $\psi_{n}\in Q(H_{01}\otimes
1)\cap Q(1\otimes H_{02})\cap \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$ for any $n\in\mathds{N}$. Let $\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$. Then $\exists \{\phi^{(\alpha)}\}_{\alpha\in\mathds{N}}$ such that $\forall \alpha\in\mathds{N}$, $\phi^{(\alpha)}\in D(H_{01}\otimes 1)\cap D(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$; and $\forall n\in\mathds{N}$, $\phi_{n}^{(\alpha)}\to \phi_{n}$ in the topology induced by $\lVert \,\cdot\, \rVert_{\mathscr{X}_{n}}^{}$. Furthermore $\forall\alpha\in\mathds{N}$: $$\label{eq:11}
\langle \psi , (H-z)\phi^{(\alpha)} \rangle_{\mathscr{H}}=0\; .$$ Since $\psi_{n}\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$, we can take the limit of and obtain, $\forall \phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$: $$\label{eq:12}
q_{H_{01}\otimes 1}(\psi,\phi)+q_{1\otimes H_{02}}(\psi,\phi)+\langle \psi , H_I\phi \rangle_{\mathscr{H}}=z\langle \psi , \phi\rangle_{\mathscr{H}}\; .$$ Hence we can choose $\phi=\psi_{\leq n}:= 1\otimes \mathds{1}_{\leq
n}(\mathscr{H}_2^{(\cdot)})\psi$ in . Then, using Assumption \[ass:1\] and taking the imaginary part we obtain: $$\label{eq:15}
{\mathrm{Im}}(z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle={\mathrm{Im}}(\langle \psi -\psi_{\leq n}, H_{I}\psi_{\leq n} \rangle_{})\; .$$ Now, by Assumption \[ass:2\] (the equality holds on the suitable domain): $$H_I \bigl(1\otimes \mathds{1}_{\leq n}(\mathscr{H}_2^{(\cdot)})\bigr) =\bigl(1\otimes \mathds{1}_{\leq n+2}(\mathscr{H}_2^{(\cdot)})\bigr)H_I \bigl(1\otimes \mathds{1}_{\leq n}(\mathscr{H}_2^{(\cdot)})\bigr)\; .$$ Furthermore $1\otimes \mathds{1}_{\leq n+2}(\mathscr{H}_2^{(\cdot)})(\psi -
\psi_{\leq n})= \psi _{n+1}\oplus \psi _{n+2}$. Then Equation becomes: $${\mathrm{Im}}(z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle= \sum_{i=1}^2{\mathrm{Im}}(\langle \psi_{n+i} , H_I\psi_{\leq n}\rangle_{})\; .$$ Using the symmetry of $H_I$, and we obtain: $$\label{eq:17}
{\mathrm{Im}}(z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle={\mathrm{Im}}(\langle \psi_{n+2} , H_I\psi_{n}\rangle+\langle \psi_{n+1} , H_I\psi_{n}\rangle+\langle \psi_{n+1} , H_I\psi_{n-1}\rangle )\; .$$ Now bound using ; then we obtain $\forall n\in \mathds{N}$: $$\label{eq:18}
\begin{split}
\lvert {\mathrm{Im}}z\rvert\sum_{i=0}^n\lVert \psi_{i} \rVert_{}^2\leq C\Bigl[\lVert \psi_{n+1} \rVert\Bigl((n+1)\bigl(\lVert \psi_{n} \rVert+\lVert \psi_{n-1} \rVert \bigr)+\sqrt{n+1}\bigl(\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}+\lVert \psi_{n-1} \rVert_{\mathscr{X}_{n-1}}\bigr)\Bigr)\\+\lVert \psi_{n+2} \rVert\Bigl((n+1)\lVert \psi_{n} \rVert+\sqrt{n+1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}\Bigr)\Bigr]\\\leq 2C(n+1)\Bigl[ \sum_{i_1=0}^2\lVert \psi_{n+i_1} \rVert_{}^2+\sum_{i_2=-1}^0(n+1)^{-1}\lVert \psi_{n+i_2} \rVert_{\mathscr{X}_{n+i_2}}^2 \Bigr] \; .
\end{split}$$
For all $\alpha >0$ define: $$S:= \sum_{n=0}^{\infty}\lVert \psi_{n} \rVert_{}^2\; ; \;S_\alpha:=\sum_{n=0}^{\infty}(n+\alpha)^{-1}\lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^2\; .$$ $\psi\in\mathscr{H}$, hence $S$ is finite. We prove that also $S_{\alpha}$ is finite. Using equation with $\phi=\psi_{n}$ we obtain, for all $n\in\mathds{N}$: $$\label{eq:38}
(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2=(n+\alpha)^{-1}(z+\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \psi_{n} \rVert_{}^2-(n+\alpha)^{-1}\langle \psi , H_I\psi_{n}\rangle_{}\; .$$ Now, we can use bound on $(n+\alpha)^{-1}\lvert \langle \psi ,
H_I\psi_{n}\rangle\rvert_{}^{}$, obtaining $$\label{eq:39}
\begin{split}
(n+\alpha)^{-2}\lvert \langle \psi , H_I\psi_{n} \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2 (n+\alpha)^{-2}\sum_{i=-2}^2\lVert \psi_{n+i} \rVert^2 \Bigl[(n+1)^2\lVert \psi_{n} \rVert^2+(n+1)\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\Bigr]\\\leq C^2(\alpha)\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{}^2 \Bigl[\lVert \psi_{n} \rVert_{}^2+(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\Bigr]\; ,
\end{split}$$ for some $C(\alpha)>0$. The only terms we need to deal with are $(n+\alpha)^{-1}\lVert \psi_{n+i} \rVert_{}^2 \lVert \psi_{n}
\rVert_{\mathscr{X}_{n}}^2$. We use the fact that for any $\varepsilon,a,b>0$, $ab\leq \frac{1}{2}(\varepsilon a^2+\frac{1}{\varepsilon}b^2)$, obtaining $$\label{eq:40}
(n+\alpha)^{-1}\lVert \psi_{n+i} \rVert_{}^2 \lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^2\leq \frac{1}{2}\Bigl(\varepsilon(n+\alpha)^{-2}\lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^4+\frac{1}{\varepsilon}\lVert \psi_{n+i} \rVert_{}^4\Bigr)\; .$$ Combining with , and applying to Equation , we obtain the following bound: for all $\varepsilon,\alpha>0$, $\exists
C(\alpha,\varepsilon)>0$ such that $$\label{eq:41}
(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\leq C(\alpha,\varepsilon)\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{}^2+\varepsilon(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\; .$$ Fix $\varepsilon <1$, then for all $\alpha >0$, $\exists C(\alpha)>0$ such that $\forall \bar{n}\in
\mathds{N}$: $$\label{eq:4}
\sum_{n=0}^{\bar{n}} (n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\leq C(\alpha) S\; ;$$ uniformly in $\bar{n}$. Then we can take the limit $\bar{n}\to\infty$ and obtain $S_{\alpha}<\infty$.
The bound of Equation could seem to follow from an implicit smallness condition on the interaction $H_I$. As it will become clearer with the examples of Section \[sec:applications\], it is not the case. Roughly speaking, Assumption \[ass:2\] allows for interaction parts that are at most as singular as $(H_0+\lvert M_1\rvert_{}^{}+\lvert
M_2\rvert_{}^{})^{1/2}(N+1)^{1/2}$.
Now return to Equation . There exists $n^{*}\in\mathds{N}$ such that $\forall n\geq n^{*}$: $$\frac{1}{2}S\leq \sum_{i=0}^{n}\lVert \psi_{i} \rVert_{}^2\leq S\; .$$ Hence summing in $n^{*} \leq n \leq \bar{n}$ on both sides of we obtain for all $\bar{n}>n^{*}$: $$\begin{split}
\frac{1}{2}S\sum_{n=n^{*}}^{\bar{n}}(n+1)^{-1}\leq \sum_{n=n^{*}}^{\bar{n}}(n+1)^{-1}\sum_{i=0}^n\lVert \psi_{i} \rVert_{}^2\leq 2 \frac{C}{\lvert {\mathrm{Im}}z\rvert_{}^{}}(3S+S_1+S_2)\; .
\end{split}$$ The bound on the right hand side is uniform in $\bar{n}$: that is absurd, unless $S=S_1=S_2=0$ $\Leftrightarrow$ $\psi =0$.
Once essential self-adjointness is established, it is possible to give the following characterization of the domain of self-adjointness $D(H)$.
\[prop:1\] Assume \[ass:1\] and \[ass:2\](\[ass:3\]). If exists $K$ self-adjoint operator with domain $D(K)$ such that:
1. $D(H_{0})\cap D(K)$ is dense in $\mathscr{H}$; $\mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$ is dense in $D(K)$.
2. There exists $ 0 < \varepsilon < 1$ such that $\exists
C(\varepsilon)>0$, $\forall \phi\in D(H_{0})\cap D(K)$: $$\label{eq:20}
\lVert H_I\phi \rVert_{}^{}\leq \varepsilon\lVert H_{0}\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; .$$
Then $D(H)\cap D(K)=D(H_{0})\cap D(K)$.
Using bound , we have $\forall \phi \in D(H_{0})\cap
D(K)$: $$\label{eq:21}
\lVert H\phi \rVert_{}^{}\leq (\varepsilon+1)\lVert H_{0}\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; .$$ Then $D(H)\supseteq D(H_{0})\cap D(K)$. Now let $\phi\in D(H)\cap
D(K)$: using $$\label{eq:22}
\lVert H_{0}\phi \rVert_{}^{}\leq \varepsilon\lVert H_0\phi \rVert_{}^{} +\lVert H\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; ;$$ since $\varepsilon <1$, $D(H_0)\supseteq D(H)\cap D(K)$.
Applications {#sec:applications}
============
It is possible to apply Theorem \[thm:1\] in several situations of mathematical and physical interest. We present and discuss some of them in this section; not before a brief discussion of the “boundaries” of Theorem \[thm:1\]: it may be interesting to see how its proof fails when we consider operators that are more than quadratic in the annihilation/creation operators; and to define a quadratic operator that is not sufficiently regular for Assumption \[ass:2\](\[ass:3\]) to hold. According to this purpose, we will consider simple toy models on $\Gamma _s(\mathds{C})$. We denote by $a^{\#}$ the corresponding annihilation/creation operators.
Let’s consider a simple trilinear Hamiltonian on $\Gamma _s(\mathds{C})$: $$H_3=a^{*}a+a^{*}a^{*}a^{*}+aaa\; .$$ The free part is $H_0=a^{*}a$, and the interaction part is $H_I=a^{*}a^{*}a^{*}+aaa$. Assumption \[ass:1\] is satisfied, and Assumption \[ass:3\] is slightly modified: $i$ now ranges from $-3$ to $3$, and bounds and are replaced by the simple bound: $$\lvert \langle \psi , H_I\phi \rangle \rvert_{\mathscr{H}}\leq C(n+1)^{3/2}\lVert \phi \rVert_{\mathscr{H}^{(n)}}^{}\bigl(\lVert \psi _{n+3} \rVert_{\mathscr{H}^{(n+3)}}^{}+\lVert \psi _{n-3} \rVert_{\mathscr{H}^{(n-3)}}^{}\bigr)\; .$$ The proof of Theorem \[thm:1\] carries on, almost unchanged, up to Equation that would now read $$\lvert {\mathrm{Im}}z \rvert_{}^{}\sum_{i=0}^n \lVert \psi_i \rVert_{}^2\leq C (n+1)^{3/2}\lVert \psi _{n+3} \rVert_{}^2\; .$$ However if we now take the sum in $n$ from $n^{*}$ to $\bar{n}$ (where $n^{*}$ is such that $\frac{1}{2}\lVert \psi \rVert_{}^2\leq \sum_{i=0}^n \lVert \psi_i
\rVert_{}^2\leq \lVert \psi \rVert_{}^2$ for all $n\geq n^{*}$) we *cannot* conclude that $\lVert \psi \rVert_{}^{}$ must be zero, because the series $\sum_{n=0}^{\infty }(n+1)^{-3/2}$ converges. Hence the proof fails, and analogously would fail for any higher order polynomial of the annihilation/creation operators.
On the other hand, we introduce now a quadratic model for which Assumption \[ass:2\](\[ass:3\]) fails to hold, and thus Theorem \[thm:1\] cannot be applied. For the following operator on $L^2 (\mathds{R}^{} )\otimes \Gamma _s(\mathds{C})$ Assumption \[ass:2\] *is satisfied*: $$H_{\partial a}=-\partial _x^2+a^{*}a-i\partial _x(a^{*}+a)+a^{*}a^{*}+aa\; ,$$ where $H_0=-\partial _x^2+a^{*}a$ and $H_I=-i\partial _x(a^{*}+a)+a^{*}a^{*}+aa$. If, however, the derivative operator is coupled with the quadratic term $$H_{\partial aa}=-\partial _x^2+a^{*}a-i\partial _x(a^{*}a^{*}+aa)\; ,$$ \[ass:2\](\[ass:3\]) is no longer satisfied. The interaction in this case would be of type $H_0^{1/2}N$, and therefore too singular: Theorem \[thm:1\] *does not hold* for $H_{\partial aa}$.
Throughout the section we will adopt the following notations, in addition to the ones of Section \[sec:defin-notat\]. Let $\mathscr{K}$ a Hilbert space; we denote by $\mathcal{L}(\mathscr{K})$ the set of bounded operators on $\mathscr{K}$ and by $\lvert \,\cdot\,\rvert_{\mathcal{L}(\mathscr{K})}^{}$ the operator norm. It is also useful to define the annihilation/creation operator valued distributions $a^{\#}(x)$, $x\in\mathds{R}^d$. Let $f\in L^2 (\mathds{R}^d)$, $a^{\#}(f)$ the annihilation/creation operators on $\Gamma_s(L^2 (\mathds{R}^d))$. Then the operator valued distributions $a^{\#}(x)$ acting on $L^2 (\mathds{R}^d)$, with values on $\Gamma_s(L^2 (\mathds{R}^d))$, are defined by: $$(a^{*},f)\equiv\int_{\mathds{R}^d}a^{*}(x)f(x)dx:=a^{*}(f)\; ;\; (a,f)\equiv\int_{\mathds{R}^d}a(x)\bar{f}(x)dx:=a(f)\; .$$ They satisfy the commutation relations (inherited by the CCR) $[a(x),a^{*}(y)]=\delta(x-y)$.
Hamiltonians of identical bosons. {#sec:hamilt-many-bosons}
---------------------------------
The criterion applies to operators in the Fock space $\Gamma_s(\mathscr{K})$, for any separable Hilbert space $\mathscr{K}$. Simply choose $\mathscr{H}_1\equiv \mathds{C}$ and $\mathscr{H}_2\equiv \mathscr{K}$; then $\mathds{C}\otimes
\Gamma_s(\mathscr{K})\approx \Gamma_s(\mathscr{K})$ up to an unitary isomorphism.
An example is given by the following class of operators. Let $\mathscr{K}=L^2 (\mathds{R}^d)$; $h_0$ a positive self adjoint operator on $L^2 (\mathds{R}^{d})$ (the one-particle free Hamiltonian). Furthermore, let $V_1\in L^2 (\mathds{R}^{d})$, $V_2,
V_3\in L^2 (\mathds{R}^{2d})$, with $V_2=\overline{V}_2$, and $V_4(\cdot):\mathds{R}^d\to \mathds{R}$, such that $V_4(x)=V_4(-x)$ and $V_4\,(h_0+1)^{-1/2}\in \mathcal{L}(L^2 (\mathds{R}^d))$. Consider $$\label{eq:32}
\begin{split}
H=d\Gamma(h_0)+\int_{\mathds{R}^d}^{}\Bigl(V_1(x)a^{*}(x)+\overline{V}_1(x)a(x)\Bigr) dx+ \int_{\mathds{R}^{2d}}^{}\Bigl(V_2(x,y)a^{*}(x)a(y)+V_3(x,y)a^{*}(x)\\
a^{*}(y) +\overline{V}_3(x,y)a(x)a(y)\Bigr) dxdy+\frac{1}{2}\int_{\mathds{R}^{2d}}^{}V_4(x-y)a^{*}(x)a^{*}(y)a(x)a(y) dxdy\; .
\end{split}$$ We make the following identifications: $H_{01}\equiv 0$, $H_{02}\equiv d\Gamma(h_0)$, $H_{diag}\equiv\int
(V_4a^{*}a^{*}aa+V_2a^{*}a)$, $H_2\equiv
\int_{}^{}(V_1a^{*}+\overline{V}_1a)+\int
(V_3a^{*}a^{*}+\overline{V}_3aa)$. Assumption \[ass:1\] is trivial to verify; and Assumption \[ass:3\] follows from standard estimates on Fock space: let $\psi\in\Gamma_s(L^2 (\mathds{R}^d))$, $\phi_n\in
L^2_s(\mathds{R}^{nd})\cap Q(d\Gamma(h_0))$, $n\in\mathds{N}$, then $$\begin{gathered}
\label{eq:35}
\begin{split}
\lvert \langle \psi, H_{diag}\phi_n \rangle_{}\rvert_{}^{}\leq \Bigl(n\lVert V_2 \rVert_2^{}\lVert \phi_n \rVert_{}^{}+\lvert V_4\,(h_{0}+1)^{-1/2}\rvert_{\mathcal{L}(L^2 (\mathds{R}^d))}^{}\bigl(n^{3/2}\lVert (d\Gamma(h_0))^{1/2}\phi_n \rVert_{}^{}\\+n^2\lVert \phi_n \rVert_{}^{}\bigr)\Bigr)\lVert \psi_n \rVert_{}^{}\; ;
\end{split}\\
\label{eq:36} \lvert \langle \psi, H_2\phi_n \rangle_{}\rvert_{}^{}\leq 2\Bigl(\sqrt{n+1}\lVert V_1 \rVert_2^{}+ (n+1)\lVert V_3 \rVert_2^{}\Bigr)\lVert \phi_n \rVert_{}^{}\sum_{\substack{i=-2\\i\neq 0}}^{2}\lVert \psi_{n+i} \rVert_{}^{}\; .\end{gathered}$$ Hence we can apply Theorem \[thm:1\]; and prove essential self-adjointness of $H$ in $D(d\Gamma (h_0))\cap
f_0(L^2(\mathds{R}^{d})^{(\cdot)})$. We can also apply Proposition \[prop:1\] with $K\equiv N^{3}$, i.e. $D(H)\cap
D(N^{3})=D(d\Gamma(h_0))\cap D(N^{3})$. Observe that if $d=3$, the well-known many body Hamiltonian with Coulomb pair interaction $$H_C=d\Gamma(-\Delta)\pm\frac{1}{2}\int_{\mathds{R}^6}^{}\frac{1}{\lvert x-y\rvert_{}^{}}a^{*}(x)a^{*}(y)a(x)a(y) dxdy\; ,$$ is just the special case $h_0= -\Delta$, $V_1=V_2=V_3=0$ and $V_4=\pm \lvert x \rvert_{}^{-1}$.
Nelson-type Hamiltonians. {#sec:nels-type-hamilt}
-------------------------
We consider now the dynamics of different species of particles (or fields) interacting. A typical example is the Nelson Hamiltonian. It was introduced in a rigorous way by @nelson:1190 to describe nucleons in a meson field, and studied by several authors [e.g. @DG1; @MR1814991; @MR1809881; @GHPS].
Let $\mathscr{H}=L^2 (\mathds{R}^{pd})\otimes \Gamma_s(L^2
(\mathds{R}^d))$: the first space corresponds to $n$ non-relativistic particles; the second to a scalar relativistic field. Let $\omega$ be a positive self-adjoint operator on $L^2 (\mathds{R}^d)$ (the dispersion relation of the relativistic field), $V\in L^2_{loc}
(\mathds{R}^d, \mathds{R}_{+})$ an external potential acting on the particles. The interaction between the particles and the field is linear in the creation and annihilation operators $a^{\#}$ corresponding to the field. Let $v:\mathds{R}^{2d}\to \mathds{C}$ such that
- $(1-\Delta_x)^{-1/2}\lVert v(x,\cdot) \rVert_{L^2_{(k)}
(\mathds{R}^d)}^2(1-\Delta_x)^{-1/2}\in \mathcal{L}(L^2_{(x)}
(\mathds{R}^{d}))$;
- for all $k\in\mathds{R}^d$, $v(x,k)(1-\Delta_x)^{-1/2}\in
\mathcal{L}(L^2_{(x)} (\mathds{R}^d))$, with $\lvert
v(x,\cdot)(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)}
(\mathds{R}^{d}))}^{}\in L^2_{(k)} (\mathds{R}^d)$.
Then we define the Nelson Hamiltonian: $$\label{eq:23}
H_N=\Bigl(\sum_{i=1}^p -\Delta_{x_i}+V(x_i)\Bigr)\otimes 1+1\otimes d\Gamma(\omega)+\sum_{i=1}^pa^{*}(v(x_{i},\cdot))+a(v(x_i,\cdot))\; .$$ The function $v$ describes the coupling between the particles and the relativistic field. The assumptions above imply that it has a good behaviour both for high and small momenta; in particular in three-dimensions it acts as an UV cutoff function.
The model of @nelson:1190 was much more specific: $d=3$, $\omega(k)=\sqrt{k^2+\mu^2}$ with $\mu>0$, $V=0$ and $v(x,k)=\lambda(2\pi)^{-3/2}(2\omega(k))^{-1/2}e^{-ik\cdot
x}\mathds{1}_{\lvert \,\cdot\,\rvert \leq \sigma}(k)$ with $\lambda,\sigma >0$. With these assumptions, $v\in L^{\infty}
(\mathds{R}^3,L^2 (\mathds{R}^3))$, $\omega^{-1/2}v\in L^{\infty}
(\mathds{R}^3,L^2 (\mathds{R}^3))$; then $H_N$ (the Nelson model with UV cut off) is self-adjoint by the Kato-Rellich Theorem. However, if we consider $d=2$ and $\mu=0$ (massless relativistic field), the Kato-Rellich Theorem is not applicable because $\omega^{-1/2}v\notin L^{\infty} (\mathds{R}^2,L^2
(\mathds{R}^2))$ due to an infrared divergence. Instead assumptions \[ass:1\] and \[ass:3\] are still satisfied, thus Theorem \[thm:1\] can be used.
In order to check Assumptions \[ass:1\] and \[ass:2\] on , we make the (straightforward) identifications: $\mathscr{H}_1\equiv L^2 (\mathds{R}^{pd})$, $\mathscr{H}_2\equiv L^2
(\mathds{R}^d)$, $H_{01}\equiv\sum_i-\Delta_{x_i}+V(x_i)$, $H_{02}\equiv d\Gamma(\omega)$, $H_I\equiv
\sum_ia^{*}(v(x_i,\cdot))+a(v(x_i,\cdot))$. We do not need to introduce a decomposition of $\mathscr{H}_1$. Assumption \[ass:1\] is satisfied: for all $V\in L^2_{loc} (\mathds{R}^d,\mathds{R}_{+})$, $-\Delta+V(\cdot)$ is a positive self-adjoint operator, and the vectors with fixed number of particles are invariant for the evolution associated with the positive self-adjoint operator $d\Gamma(\omega)$. Furthermore, since $H_{01}\otimes 1$ and $1\otimes
H_{02}$ are positive self-adjoint commuting operators, $H_0$ is a positive self-adjoint operator with domain $D(H_0)= D(H_{01}\otimes
1)\cap D(1\otimes H_{02})$. Assumption \[ass:2\] is also satisfied by usual estimates: $\forall \psi \in \mathscr{H}$, $\forall\phi_n\in
L^2 (\mathds{R}^{pd})\otimes L^2_s(\mathds{R}^{nd})\cap
Q(H_{01}\otimes 1)$, $n\in\mathds{N}$, $$\label{eq:13}
\begin{split}
\lvert \langle \psi , H_I\phi_n \rangle_{}\rvert\leq \sqrt{2p}\bigl( 2 \sqrt{n} \lVert \lvert v(x,\cdot)(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)})}^{} \rVert_{L^2_{(k)}} + \lvert (1-\Delta_x)^{-1/2}\lVert v(x,\cdot) \rVert_{L^2_{(k)}}^2\\(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)})}^{1/2}\bigr)\Bigl(\Bigl\lVert \bigl(\sum_{i=1}^p-\Delta_{x_i}\bigr)^{1/2}\phi_n \Bigr\rVert+\sqrt{p} \lVert \phi_n \rVert\Bigr) \sum_{\substack{i=-1\\i\neq 0}}^{1}\lVert \psi_{n+i} \rVert\; .
\end{split}$$ Then $H_N$ is essentially self-adjoint on $D(H_0)\cap f_0(L^2
(\mathds{R}^{pd})\otimes L^2(\mathds{R}^{d})^{(\cdot)})$.
Let $H_N\rvert_s$ be the restriction of $H_N$ to $L^2_s
(\mathds{R}^{pd})\otimes \Gamma_s(L^2 (\mathds{R}^d))$. It is possible to extend $H_N\rvert_s$ to $\Gamma_s(L^2 (\mathds{R}^d))\otimes
\Gamma_s(L^2 (\mathds{R}^d))$ in the following way. Define $$\label{eq:25}
\widetilde{H}_N=d\Gamma(-\Delta+V)\otimes 1+1\otimes d\Gamma(\omega)+\int_{\mathds{R}^d}^{} \psi^{*}(x)\bigl(a^{*}(v(x,\cdot))+a(v(x,\cdot))\bigr)\psi(x)dx\; ,$$ where $\psi^{\#}$ are the creation and annihilation operators corresponding to the first Fock space. Then $H_N\rvert_s$ and $\widetilde{H}_N$ agree on the $p$-particle sector $L^2_s
(\mathds{R}^{pd})\otimes \Gamma_s(L^2 (\mathds{R}^d))$ of $\Gamma_s(L^2 (\mathds{R}^d))\otimes \Gamma_s(L^2
(\mathds{R}^d))$. The self-adjointness of $\widetilde{H}_N$ still follows from Theorem \[thm:1\] using the bound : it is sufficient to choose for $\mathscr{H}_1\equiv \Gamma_s(L^2
(\mathds{R}^d))$ the decomposition in finite particle vectors $\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}\equiv \{L^2_s (\mathds{R}^{jd})\otimes
\Gamma_s(\mathscr{H}_2) \}_{j\in\mathds{N}}$. Let $H_0\equiv d\Gamma(-\Delta+V)\otimes 1 +1\otimes d\Gamma(\omega)$, then the domain of essential self-adjointness for $\widetilde{H}_N$ is $D(H_0)\cap f_0(L^2(\mathds{R}^{d})^{(\cdot)}\otimes
L^2(\mathds{R}^{d})^{(\cdot)})$. Let $N_1$ and $N_2$ be the number operators corresponding to the first and second Fock space respectively. Then applying Proposition \[prop:1\] we also obtain $D(\widetilde{H}_N)\cap D(N_1^2+N_2^2)=D(H_0)\cap D(N_1^2+N_2^2)$.
Pauli-Fierz Hamiltonian. {#sec:pauli-fierz-hamilt}
------------------------
The last example considered is an operator describing the dynamics of rigid charges and their radiation field interacting. The model was introduced by @pauli1938theorie, and has been extensively studied by a mathematical standpoint. See @MR2097788 [and references thereof contained] for a detailed presentation.
Let $\mathscr{H}^{(spin)}=(\otimes^p
\mathds{C}^{2[\frac{d}{2}]})\otimes L^2 (\mathds{R}^{pd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$, $\mathscr{H}=
L^2 (\mathds{R}^{pd})\otimes \Gamma_s(\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))$: the first space corresponds to $p$ spin-$\frac{1}{2}$ particles, the second to spinless particles. Let $\chi\in L^2 (\mathds{R}^d)$, $V\in L^2_{loc}
(\mathds{R}^{pd},\mathds{R}_{+})$, $\omega=\lvert k\rvert $, $m_j>0$, $q_j\in \mathds{R}$ for all $j=1,\dotsc,p$. Furthermore, let $e_{\lambda}: \mathds{R}^d\to \mathds{R}^d$ such that for almost all $k\in \mathds{R}^d$, $k\cdot e_{\lambda}(k)=0$ and $e_{\lambda}(k)
\cdot e_{\lambda'}(k)=\delta_{\lambda\lambda'}$ for all $\lambda,\lambda'=1,\dotsc, d-1$. Then we define the electromagnetic vector potential in the Coulomb gauge as $$\label{eq:26}
A(x)=\sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}e_{\lambda}(k)\Bigl( a^{*}_{\lambda}(k)\chi(k)e^{ik\cdot x}+a_{\lambda}(k)\bar{\chi}(k)e^{ik\cdot x} \Bigr) dk\; ;$$ where $a_{\lambda}^{\#}$ are the creation and annihilation operators of $\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$ satisfying the canonical commutation relations $[a_{\lambda}(k),a^{*}_{\lambda'}(k')]=\delta_{\lambda\lambda'}\delta(k-k')$; the (spinless) Pauli-Fierz Hamiltonian on $\mathscr{H}$ is then $$\label{eq:27}
H_{PF}= \sum_{j=1}^p \frac{1}{2m_j}\bigl(-i\nabla_j\otimes 1+q_j A(x_j)\bigr)^2 + V(x_1,\dotsc, x_p)\otimes 1 + 1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .$$ The function $\chi$ plays the role of an ultraviolet cut off in the interaction, and is usually interpreted as the Fourier transform of the particles’ charge distribution. Let $\{\sigma^{(\mu)}\}_{\mu=1}^d$ the $2^{[\frac{d}{2}]}\times
2^{[\frac{d}{2}]}$ matrices satisfying $\sigma^{(\mu)}\sigma^{(\nu)}+\sigma^{(\nu)}\sigma^{(\mu)}=2\delta_{\mu\nu}\mathrm{Id}$. Also, denote by $\sigma_j^{(\mu)}$, $j=1,\dotsc, p$ the operator on $(\otimes^p \mathds{C}^{2[\frac{d}{2}]})$ acting as $\sigma^{(\mu)}$ on the $j$-th space of the tensor product. Then the spin-$\frac{1}{2}$ Pauli-Fierz Hamiltonian on $\mathscr{H}^{(spin)}=(\otimes^p
\mathds{C}^{2[\frac{d}{2}]})\otimes \mathscr{H}$ can be written as: $$\label{eq:28}
H_{PF}^{(spin)}=1\otimes H_{PF} +\frac{i}{2}\sum_{j=1}^pq_j\sum_{1\leq \mu<\nu\leq d}^{}\sigma_j^{(\mu)}\sigma_j^{(\nu)}\otimes \Bigl(\partial_j^{(\mu)}A^{(\nu)}(x_j)-\partial_j^{(\nu)}A^{(\mu)}(x_j)\Bigr)\; ;$$ where $A^{(\mu)}(x)$ is the $\mu$-th component of the vector $A(x)$.
The quadratic form corresponding to the Pauli-Fierz Hamiltonian is bounded from below, so it is possible to define at least one self-adjoint extension by means of the Friedrichs Extension Theorem. This type of information is not completely satisfactory, since infinitely many extensions may exist, each one dictating a different dynamics for the system. For small values of the ratios $q^2_j/m_j$ between charge and mass of the particles, and if $\chi,\chi/\sqrt{\omega}\in L^2 (\mathds{R}^d)$, a unique self-adjoint extension is given by KLMN Theorem. For arbitrary values of the ratios $q^2_j/m_j$, it is possible to prove essential self-adjointness of both $H_{PF}$ and $H_{PF}^{(spin)}$ (for the spin operator we need in addition $\omega\chi\in L^2 (\mathds{R}^d)$) by means of Theorem \[thm:1\], under the sole assumption $\chi\in L^2
(\mathds{R}^d)$. As discussed in Section \[sec:introduction\], an analogous result (on a slightly different domain) has been obtained with an argument of functional integration by @MR1891842. If the dependence on $x$ of $A(x)$ is more general, functional integration methods may not be applicable; however Theorem \[thm:1\] still holds.
In the following discussion we will focus on a simplified model, for the sake of clarity. Assumptions \[ass:1\] and \[ass:2\] are checked on $H_{PF}$ with $p=1$, $m=1/2$ and $q=-1$, i.e.: $\mathscr{H}\equiv L^2 (\mathds{R}^d)\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$ and $$\label{eq:29}
H\equiv \bigl(i\nabla_x\otimes 1+A(x)\bigr)^2 + V(x)\otimes 1 + 1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .$$
Observe that, since we are in the Coulomb gauge, $\nabla_x\cdot
A(x)=0$ hence $[-i\nabla_x\otimes 1, A(x)]=0$ on a suitable dense domain. Rewrite $H$ in the following form, to identify the free and interaction parts: $$\label{eq:30}
\begin{split}
H= \bigl(-\Delta_x+V(x)\bigr)\otimes 1+1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk +2i A(x)\cdot(\nabla_x\otimes 1)+ A^2(x)\; .
\end{split}$$ We identify $H_{01}\equiv -\Delta+V$, $H_{02}\equiv
\sum_{\lambda}\int_{\mathds{R}^d}^{}\omega a^{*}_{\lambda}a_{\lambda}$ and $H_I\equiv 2i A\cdot(\nabla\otimes 1)+
A^2$. Assumption \[ass:1\] is satisfied, as in the Nelson model above. For the interaction part, we have the following bounds: $\forall\psi\in\mathscr{H}$, $\forall\phi_n\in L^2
(\mathds{R}^d)\otimes (\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))^{\otimes_s n}\cap Q(H_{01}\otimes 1)$, $n\in\mathds{N}$, $$\label{eq:33}
\begin{aligned}
\lvert \langle \psi , &A(x)\cdot (\nabla_x\otimes 1)\phi_n\rangle_{}\rvert\leq \sqrt{2(d-1)}\lVert \chi \rVert_2^{} \sqrt{n+1}\lVert (\lvert \nabla_x\rvert_{}^{}\otimes 1) \phi_n\rVert_{}^{}\sum_{\substack{i=-1\\i\neq 0}}^1\lVert \psi_{n+i} \rVert_{}^{}\; ;\\
\lvert \langle \psi , &A^2(x)\phi_n\rangle_{}\rvert\leq 2(d-1) \lVert \chi \rVert_2^{}(n+1)\lVert \phi_n \rVert_{}^{}\sum_{i=-2}^{2}\lVert \psi_{n+i} \rVert_{}^{}\; .
\end{aligned}$$ Hence Assumption \[ass:2\] is satisfied. Then $H$ is essentially self-adjoint on $D(H_0)\cap f_0(L^2 (\mathds{R}^{d})\otimes
(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))^{(\cdot)})$.
\[rem:1\] Neither non-negativity of the Pauli-Fierz operator nor smallness of the coupling constant are necessary to prove essential self-adjointness by means of Theorem \[thm:1\]. Using operator methods (commutator estimates), self-adjointness of $H_{PF}$ with $V=0$ has been proved for general coupling constants in [@MR2436496], but the non-negativity was needed to associate a unique self-adjoint operator to the quadratic form. Theorem \[thm:1\] relies on different assumptions, and takes advantage of the fibered structure of the Fock space: boundedness from below of the operator is, in general, not necessary. In fact, the Hamiltonians considered in Sections \[sec:hamilt-many-bosons\] and \[sec:nels-type-hamilt\] are possibly unbounded from below, as well as the following extension of the Pauli-Fierz Hamiltonian to infinite degrees of freedom (for the particles). As outlined in Section \[sec:conclusive-remarks-1\], if we assume boundedness from below, Theorem \[thm:1\] can be extended to operators quartic in the creation/annihilation operators (see Assumptions \[ass:5\], \[ass:6\] and Theorem \[thm:2\]).
Let $m_j=1/2$, $q_j=-1$ and $V=\sum_{i=1}^pV_{ext}(x_i)+\sum_{i<j}V_{pair}(x_i-x_j)$ such that $V_{ext}\in L^2_{loc}(\mathds{R}^d,\mathds{R}_{+})$, $V_{pair}(x)=V_{pair}(-x)$ and $V_{pair}(1-\Delta)^{-1/2}\in\mathcal{L}(L^2 (\mathds{R}^d))$. Under these assumptions define $H_{PF}\rvert_s$ as the restriction of to $L^2_s (\mathds{R}^{pd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$. The physical interpretation is a system of $p$ identical bosonic charges subjected to an external potential, interacting via pair interaction and with their radiation field. As we did for the Nelson model in , we can extend $H_{PF}\rvert_s$ to $\Gamma_s(L^2
(\mathds{R}^d))\otimes \Gamma_s(\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))$: $$\label{eq:31}
\begin{split}
\widetilde{H}_{PF}=\int_{\mathds{R}^d}^{}\psi^{*}(x)\Bigl\{\bigl(i\nabla_x\otimes 1 + A(x)\bigr)^2+V_{ext}(x) \Bigr\}\psi(x) dx + \frac{1}{2}\int_{\mathds{R}^{2d}}^{}V_{pair}(x-y)\psi^{*}(x)\psi^{*}(y)\\\psi(x)\psi(y) dxdy +1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .
\end{split}$$ We would like to prove essential self-adjointness by means of Theorem \[thm:1\]. Identify $H_{01}\equiv \int_{}^{}\psi^{*}(-\Delta
+V_{ext})\psi $; $H_{02}\equiv
\sum_{\lambda}\int_{\mathds{R}^d}^{}\omega
a^{*}_{\lambda}a_{\lambda}$; $H_I\equiv \int \psi^{*}(2i
A\cdot(\nabla\otimes 1)+
A^2)\psi+\frac{1}{2}\int_{}^{}V_{pair}\psi^{*}\psi^{*}\psi\psi $; and $\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}\equiv \{L^2_s (\mathds{R}^{jd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))
\}_{j\in\mathds{N}}$. Then Assumptions \[ass:1\] and \[ass:2\] are satisfied using bounds analogous to and (for $V_{pair}$), for each fixed $j\in \mathds{N}$. Hence $\widetilde{H}_{PF}$ is essentially self-adjoint on $D(H_{01}\otimes
1)\cap D(1\otimes H_{02})\cap f_0(L^2
(\mathds{R}^{d})^{(\cdot)}\otimes (\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))^{(\cdot)})$.
Conclusive remarks {#sec:conclusive-remarks-1}
==================
The examples of the preceding section are not exhaustive: we focused on them because of their relevance in physical and mathematical literature. The application to operators on curved space-time, or to anti-symmetric systems may also lead to results of interest.
The Assumptions \[ass:1\], \[ass:2\] and \[ass:3\] are easy to check: in the examples above follow from basic estimates of creation and annihilation operators. The proof of Theorem \[thm:1\] itself is not complicated, and relies on the direct sum decomposition of $\Gamma_s(\mathscr{H}_2)$ and the structure of the interaction with respect to the latter. Hence this criterion gives, in our opinion, a simple yet powerful tool to prove essential self-adjointness in Fock spaces, tailored to take maximum advantage of their structure.
If we assume that $H$ is bounded from below, we can take inspiration from @masson1971 and extend our criterion to accommodate quartic operators. The modified assumptions and theorem would then read:
[B$_{H}$]{} \[ass:5\] $H$ is a densely defined symmetric operator on $\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)$ bounded from below. $H_{01}$ and $H_{02}$ are self-adjoint operators bounded from below such that $\forall t\in \mathds{R}$, $\{\mathscr{H}_2^{(n)}\}_{n\in\mathds{N}}$ is invariant for $e^{it
H_{02}}$.
[B$_I$]{} \[ass:6\] $H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in $\mathscr{H}$. Furthermore exists a complete collection $\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}$ invariant for $H_0$ and $H_I$ such that: $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$, $$\label{eq:34}
H_I\,\phi\in \bigoplus_{i=-4}^{4}\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n+i)}\; .$$ Also, $H_I$ satisfies the following bound: $\forall
j,n\in\mathds{N}$ $\exists C(j)>0$ such that $\forall\psi\in
\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$: $$\label{eq:37}
\begin{split}
\lvert \langle \psi , H_I\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(j)\sum_{i=-4}^4\lVert \psi_{j,n+i} \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}}^2 \Bigl[(n+1)^4\lVert \phi \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)^{2}\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\Bigr]\; .
\end{split}$$
\[thm:2\] Assume \[ass:5\] and \[ass:6\]. Then $H$ is essentially self adjoint on $D(H_{01}\otimes 1)\cap D(H_{02}\otimes 1)\cap
f_0(\mathscr{H}_1^{(\cdot)}\otimes \mathscr{H}_2^{(\cdot)})$.
An attempt to extend the results of [@masson1971] can be found in [@arai1991]. Theorem \[thm:2\] is a generalization of both: it can be applied to more singular situations and a more general class of spaces.
The proof of Theorem \[thm:1\] can be adapted to Theorem \[thm:2\], making use of the inferior bound for $H$. We remark that Assumption \[ass:5\], by itself, implies that $H$ has at least one self-adjoint extension: it may be tricky to prove for general operators. Theorem \[thm:2\] essentially states that for regular enough quartic interactions, existence of a particular self-adjoint extension (the Friedrichs one) is equivalent to its uniqueness. It may have interesting applications in CQFT: e.g. the $d$-dimensional (bounded from below) $Y_d$ and $(\lambda\varphi(x)^4)_d$ models with cut offs have interactions that are at most quartic and regular. It is our hope that the ideas utilized in this paper could contribute to improve the mathematical insight on interacting quantum field theories, and could be developed to study self-adjointness of more singular systems.
This work has been supported by the Centre Henri Lebesgue (programme “Investissements d’avenir” — ANR-11-LABX-0020-01). The author would like to thank Giorgio Velo, that has suggested to him the idea of a direct proof of self-adjointness on Fock spaces. Also, he would like to thank Zied Ammari and Francis Nier for precious advices and stimulating discussions during the redaction of the paper.
[27]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
Z. Ammari. Asymptotic completeness for a renormalized nonrelativistic [H]{}amiltonian in quantum field theory: the [N]{}elson model. *Math. Phys. Anal. Geom.*, 30 (3):0 217–285, 2000. ISSN 1385-0172. [doi: ]{}[10.1023/A:1011408618527]{}. URL <http://dx.doi.org/10.1023/A:1011408618527>.
Z. Ammari and M. Falconi. Wigner measures approach to the classical limit of the [N]{}elson model: Convergence of dynamics and ground state energy. *J. Stat. Phys.*, 1570 (2):0 330–362, 10 2014. [doi: ]{}[10.1007/s10955-014-1079-7]{}. URL <http://dx.doi.org/10.1007/s10955-014-1079-7>.
A. Arai. A theorem on essential self-adjointness with application to hamiltonians in nonrelativistic quantum field theory. *J. Math. Phys.*, 320 (8):0 2082–2088, 1991. URL <http://dx.doi.org/10.1063/1.529178>.
V. Bach, J. Fr[ö]{}hlich, and I. M. Sigal. Quantum electrodynamics of confined nonrelativistic particles. *Adv. Math.*, 1370 (2):0 299–395, 1998. ISSN 0001-8708. [doi: ]{}[10.1006/aima.1998.1734]{}. URL <http://dx.doi.org/10.1006/aima.1998.1734>.
V. Bach, J. Fr[ö]{}hlich, and I. M. Sigal. Spectral analysis for systems of atoms and molecules coupled to the quantized radiation field. *Comm. Math. Phys.*, 2070 (2):0 249–290, 1999. ISSN 0010-3616. [doi: ]{}[10.1007/s002200050726]{}. URL <http://dx.doi.org/10.1007/s002200050726>.
V. Bach, J. Fr[ö]{}hlich, I. M. Sigal, and A. Soffer. Positive commutators and the spectrum of [P]{}auli-[F]{}ierz [H]{}amiltonian of atoms and molecules. *Comm. Math. Phys.*, 2070 (3):0 557–587, 1999. ISSN 0010-3616. [doi: ]{}[10.1007/s002200050737]{}. URL <http://dx.doi.org/10.1007/s002200050737>.
J. M. Cook. The mathematics of second quantization. *Proc. Nat. Acad. Sci. U. S. A.*, 37:0 417–420, 1951. ISSN 0027-8424.
J. Derezi[ń]{}ski and C. G[é]{}rard. Asymptotic completeness in quantum field theory. [M]{}assive [P]{}auli-[F]{}ierz [H]{}amiltonians. *Rev. Math. Phys.*, 110 (4):0 383–450, 1999. ISSN 0129-055X. [doi: ]{}[10.1142/S0129055X99000155]{}. URL <http://dx.doi.org/10.1142/S0129055X99000155>.
J. Derezi[[ń]{}]{}ski and C. G[[é]{}]{}rard. *Mathematics of quantization and quantum fields*. Cambridge Monographs on Mathematical Physics. Cambridge University Press, Cambridge, 2013. ISBN 978-1-107-01111-3. [doi: ]{}[10.1017/CBO9780511894541]{}. URL <http://dx.doi.org/10.1017/CBO9780511894541>.
J. Derezi[[ń]{}]{}ski and V. Jak[š]{}i[[ć]{}]{}. Spectral theory of [P]{}auli-[F]{}ierz operators. *J. Funct. Anal.*, 1800 (2):0 243–327, 2001. ISSN 0022-1236. [doi: ]{}[10.1006/jfan.2000.3681]{}. URL <http://dx.doi.org/10.1006/jfan.2000.3681>.
M. Falconi. *Classical limit of the Nelson model*. PhD thesis, Dottorato di Ricerca in Matematica XXIV ciclo, Universit[à]{} di Bologna, 2012.
C. G[é]{}rard, F. Hiroshima, A. Panati, and A. Suzuki. Infrared problem for the [N]{}elson model on static space-times. *Comm. Math. Phys.*, 3080 (2):0 543–566, 2011. ISSN 0010-3616. [doi: ]{}[10.1007/s00220-011-1289-7]{}. URL <http://dx.doi.org/10.1007/s00220-011-1289-7>.
J. Ginibre and G. Velo. Renormalization of a quadratic interaction in the [H]{}amiltonian formalism. *Comm. Math. Phys.*, 18:0 65–81, 1970. ISSN 0010-3616.
J. Glimm. Yukawa coupling of quantum fields in two dimensions. [I]{}. *Comm. Math. Phys.*, 5:0 343–386, 1967. ISSN 0010-3616.
J. Glimm and A. Jaffe. *Collected papers. [V]{}ol. 2*. Birkh[ä]{}user Boston, Inc., Boston, MA, 1985. ISBN 0-8176-3272-7. Constructive quantum field theory. Selected papers, Reprint of articles published 1968–1980.
D. Hasler and I. Herbst. On the self-adjointness and domain of [P]{}auli-[F]{}ierz type [H]{}amiltonians. *Rev. Math. Phys.*, 200 (7):0 787–800, 2008. ISSN 0129-055X. [doi: ]{}[10.1142/S0129055X08003389]{}. URL <http://dx.doi.org/10.1142/S0129055X08003389>.
F. Hiroshima. Essential self-adjointness of translation-invariant quantum field models for arbitrary coupling constants. *Comm. Math. Phys.*, 2110 (3):0 585–613, 2000. ISSN 0010-3616. [doi: ]{}[10.1007/s002200050827]{}. URL <http://dx.doi.org/10.1007/s002200050827>.
F. Hiroshima. Self-adjointness of the [P]{}auli-[F]{}ierz [H]{}amiltonian for arbitrary values of coupling constants. *Ann. Henri Poincar[é]{}*, 30 (1):0 171–201, 2002. ISSN 1424-0637. [doi: ]{}[10.1007/s00023-002-8615-8]{}. URL <http://dx.doi.org/10.1007/s00023-002-8615-8>.
D. Masson and W. K. McClary. On the self-adjointness of the $(g(x)\phi^{4})_{2}$ hamiltonian. *Comm. Math. Phys.*, 210 (1):0 71–74, 1971. URL <http://projecteuclid.org/euclid.cmp/1103857260>.
E. Nelson. Interaction of nonrelativistic particles with a quantized scalar field. *J. Math. Phys.*, 50 (9):0 1190–1197, 1964. [doi: ]{}[10.1063/1.1704225]{}. URL <http://link.aip.org/link/?JMP/5/1190/1>.
E. Nelson. A quartic interaction in two dimensions. In *Mathematical [T]{}heory of [E]{}lementary [P]{}articles ([P]{}roc. [C]{}onf., [D]{}edham, [M]{}ass., 1965)*, pages 69–73. M.I.T. Press, Cambridge, Mass., 1966.
T. Okamoto and K. Yajima. Complex scaling technique in nonrelativistic massive [QED]{}. *Ann. Inst. H. Poincar[é]{} Phys. Th[é]{}or.*, 420 (3):0 311–327, 1985. ISSN 0246-0211. URL <http://www.numdam.org/item?id=AIHPB_1985__42_3_311_0>.
W. Pauli and M. Fierz. Zur theorie der emission langwelliger lichtquanten. *Il Nuovo Cimento*, 150 (3):0 167–188, 1938.
M. Reed and B. Simon. *Methods of modern mathematical physics. [II]{}. [F]{}ourier analysis, self-adjointness*. Academic Press, New York, 1975.
L. Rosen. A [$\lambda \phi ^{2n}$]{} field theory without cutoffs. *Comm. Math. Phys.*, 16:0 157–183, 1970. ISSN 0010-3616.
I. Segal. Construction of non-linear local quantum processes. [I]{}. *Ann. of Math. (2)*, 92:0 462–481, 1970. ISSN 0003-486X.
H. Spohn. *Dynamics of charged particles and their radiation field*. Cambridge University Press, Cambridge, 2004. ISBN 0-521-83697-2. [doi: ]{}[10.1017/CBO9780511535178]{}. URL <http://dx.doi.org/10.1017/CBO9780511535178>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have observed a sample of 64 small diameter sources towards the central $-$6[$^{\circ}$]{} $~<~l~<~$6[$^{\circ}$]{}, $-$2[$^{\circ}$]{} $~<~b~<~$2[$^{\circ}$]{} of the Galaxy with the aim of studying the Faraday rotation measure near the Galactic Centre (GC) region. All the sources were observed at 6 and 3.6 cm wavelengths using the ATCA and the VLA. Fifty nine of these sources are inferred to be extragalactic. The observations presented here constitute the first systematic study of the radio polarisation properties of the background sources towards this direction and increases the number of known extragalactic radio sources in this part of the sky by almost an order of magnitude. Based on the morphology, spectral indices and lack of polarised emission, we identify four Galactic HII regions in the sample.'
author:
- |
Subhashis Roy$^{1}$[^1] A. Pramesh Rao$^1$ & Ravi Subrahmanyan$^2$\
$^1$ National Centre for Radio Astrophysics (TIFR),\
Pune University Campus, Post Bag No.3, Ganeshkhind, Pune 411 007, India.\
$^2$ Australia Telescope National Facility, CSIRO, Locked bag 194, Narrabri, NSW 2390, Australia
bibliography:
- 'eg.src.bib'
title: '**Extragalactic sources towards the central region of the Galaxy**'
---
\[firstpage\]
Galaxy: center – techniques: polarimetric – radio continuum: ISM – Galaxy: HII regions
Introduction
============
Since extragalactic sources are located outside the Galaxy, the effect of ISM on the propagation properties of electromagnetic waves from these objects can be modelled without distance ambiguities as in the cases of pulsars, and thereby allowing us to observe the integrated effect of the medium along large ($\sim$20 kpc) line of sight distance. Unfortunately, only a few extragalactic sources have been identified within the central few degrees of the Galaxy ($-$6[$^{\circ}$]{} $~<~l~<~$6[$^{\circ}$]{}, $-$5[$^{\circ}$]{} $~<~b~<~$5[$^{\circ}$]{}), which limits their usefulness as probes to study the Galactic Centre (GC) ISM. High obscuration at optical wavelengths and the confusion due to the high concentration of stars at infrared wavelengths have prevented identification of extragalactic sources in this region. High angular resolution studies at centimetre wavelengths (e.g., @BECKER1994 at 5 GHz and @ZOONEMATKERMANI1990 at 1.4 GHz) have identified compact radio sources, but in the presence of a large number of Galactic sources near the Galactic Centre (GC), identifying the extragalactic sources is non-trivial, and only about half a dozen extragalactic sources in this region have been identified [@BOWER2001].
To study the Faraday rotation measure (RM) near the centre of the Galaxy ($-$6[$^{\circ}$]{} $~<~l~<~$6[$^{\circ}$]{}, $-$2[$^{\circ}$]{} $~<~b~<~$2[$^{\circ}$]{}), we have selected a sample of 64 small diameter ($< 10''$) sources (see Sect. 1.1) in the region, which we have studied with high angular resolution at 6 cm (C band) and 3.6 cm (X band) with the ATCA and the VLA. All the sources were studied for linear polarisation and the width of the frequency channels were chosen to avoid bandwidth depolarisation up to a RM of 15,000 rad m$^{-2}$. Though the NVSS [@CONDON1998] was capable of detecting polarised emission from sources, in those cases where the RM is high ($>$ 350 rad m$^{-2}$) its bandwidth of 50 MHz would cause bandwidth depolarisation. Our observations, for the first time, provide reliable measurements of the polarisation properties of the sources in the region. These observations have almost an order of magnitude higher sensitivity (in Stokes I) and up to 3 times higher resolution as compared to the previous VLA Galactic plane survey (GPS), and this high sensitivity together with higher resolution has helped to identify the Galactic sources in the initial sample.
In this paper, we provide information on the morphology, polarisation fraction, spectral indices and rotation measure of these sources, and in a companion paper (henceforth Paper II) draw inferences about the magnetic field in the GC region.
Sample selection
----------------
We surveyed the literature and formed a sample of possible extragalactic radio sources in the central $-$6[$^{\circ}$]{}$ < l < $6[$^{\circ}$]{}, $-$2[$^{\circ}$]{}$ < b < $2[$^{\circ}$]{} of the Galaxy. These sources were selected on the basis of their small scale structure ($\le$ 10$^{''}$) and non-thermal spectra ($\alpha \le -0.4$, S($\nu$) $\propto$ $\nu^ \alpha$). For the sources to have detectable linear polarisation and so be useful for the RM study, an estimate of the polarisation fraction of the sources are required. However, in the absence of any reliable information on their polarisation fraction in the literature, we assumed the unresolved sources to be polarised at the mean polarisation fraction of extragalactic small diameter sources of 2.5% [@SAIKIA1988]. Sources with measured flux density greater than 10 mJy at 5 GHz were selected. The source catalogues used for this selection were VLA images of the GC region at 327 MHz [@LAROSA2000], the VLA survey of the Galactic plane (GPS) at 1.4 GHz [@ZOONEMATKERMANI1990], [@HELFAND1992] and at 5 GHz [@BECKER1994]. Sources observed by @LAZIO1998a and the 365 MHz Texas survey [@DOUGLAS1996] were also used for this purpose. In those cases where no flux density estimates were available at 5 GHz, the flux densities at this frequency was estimated by extrapolating the 1.4 GHz flux densities using spectral indices measured between 327 MHz and 1.4 GHz. A total of 64 sources were found that satisfied all of the above criteria.
Observations and data reduction
===============================
Details of array configurations, frequencies used and the date of radio observations are in Table \[table.pol.obs\]. The ATCA observations were made using a 6 km array configuration. Twelve sources, G357.435$-$0.519, G357.865$-$0.996, G358.002$-$0.636, G358.982+0.580, G359.388+0.460, G359.568+1.146, G359.844$-$1.843, G359.911$-$1.813, G0.537+0.263, G1.028$-$1.110, G1.035+1.559, G2.143+1.772 were observed on 06 Feb 2000 during the pilot run of the project. We observed 24 more sources G353.462$-$0.691, G354.719$-$1.117, G354.740+0.138, G356.000+0.023, G356.161+1.635, G356.567+0.869, G356.719$-$1.220, G358.591+0.046, G358.917+0.072, G359.546+0.988, G359.993+1.591, G0.313+1.645, G0.846+1.173, G1.505$-$1.231, G1.954$-$1.702, G2.423-1.660, G4.005+1.403, G4.188$-$1.680, G4.256$-$0.726, G4.898+1.292, G5.260$-$0.754, G5.358+0.899, G5.511$-$1.515 and G6.183$-$1.480 on 30th September and 02nd, 06th and 08th October, 2000. To unwrap possible n$\pi$ wrap in polarisation angles measured between two frequencies, 4 more polarised sources from the pilot run G357.865$-$0.996, G359.388+0.460, G0.537+0.263, G1.028$-$1.110 were re-observed on 06th or 08th Oct 2000. For the same reason, 4 sources from the pilot run G359.388+0.460, G359.911$-$1.813, G0.537+0.263, G2.143+1.772 along with G359.993+1.591 were re-observed on 12th April 2002. All these observations were made using the multi-channel continuum observing mode, and the data were acquired in 16 independent frequency channels covering 128-MHz bands centred at the observing frequencies. Each target source was typically observed for a total of 40–50 minutes. Since we used an E-W array configuration, multiple-snapshot mode was used to get a satisfactory [*uv*]{}-coverage and each source was observed $\approx$10 times equally spaced in hour-angle. The sources 1741$-$312 and 1748$-$253 were used to calibrate the antenna based amplitudes and phases (secondary calibrators), and their flux densities were measured based on the observation of the primary flux density calibrator PKS B1934$-$638. This source (polarisation fraction $\le$ 0.2%) was also used to determine antenna based polarisation leakages. The calibration and editing of the ATCA data was performed using [MIRIAD]{}. Calibrated data was converted into Stokes I, Q, U, V, and further analysis were carried out using AIPS. Maps made at different frequencies for a particular source were convolved to the same resolution. The Polarisation angle ($\phi$) being given by $\phi$=0.5 tan$^{-1}(U/Q)$ (where, signs of Q and U are considered separately to unambiguously determine the value of $\phi$), we divided the Stokes U image by the Q image and measured the polarisation angle and its error (AIPS task COMB). Finally, the polarisation angle images at different frequency bands were fitted to the equation, $$\label{rm.equation}
\phi = RM.\lambda^2 + \phi_0 +n\pi$$ using the AIPS task RM. In this equation, $n$ is an integer, $\lambda$ is the wavelength, and $\phi_0$ denotes the intrinsic polarisation angle (i.e., when the observing frequency tends to infinity). If the rms residuals exceed four times the expected rms noise, the fitted values were rejected.
Since there were 16 frequency channels per 128 MHz band of the ATCA data, we tried to measure the RM from the ATCA observations in two ways by using the AIPS task RM. (i) Since AIPS task RM cannot fit more than four frequency channels to measure RM, in order to maximise signal to noise ratio as well as to check for high RM in the data, we divided the central 12 frequency channels of 4.8 and 5.9 GHz band into 3 equal parts. Polarisation angles were measured from first and third part of each of these bands, and these were used as input to the AIPS task RM. This allowed us to measure RM as high as 30,000 [rad m$^{-2}$]{}. (ii) If the RM measured by the previous method is not very high (i.e., $\le$ 2000 rad m$^{-2}$, and was the case for all except one source), data from each of the 4.8, 5.3, 5.9 and 8.5 GHz band were averaged and polarisation angle measured from each of these bands. These polarisation angle images and their error maps were used as the input to the AIPS task RM.
We used the VLA in its BnA configuration to observe the relatively weak sources in the sample. The default continuum mode, which provides a single frequency channel of bandwidth 50 MHz in each IF band, was used. Observations were centred at frequencies 4.63, 4.88, 8.33 and 8.68 GHz. 27 new sources, G353.410$-$0.360, G354.815+0.775, G355.424$-$0.809, G355.739+0.131, G356.905+0.082, G358.149$-$1.675, G358.156+0.028, G358.643$-$0.034, G358.605+1.440, G358.930$-$1.197, G359.392+1.272, G359.2$-$0.8 (Mouse), G359.604+0.306, G359.717$-$0.036, G359.710$-$0.904, G359.871+0.179, G0.404+1.061, G1.826+1.070, G2.922+1.028, G3.347$-$0.327, G3.748$-$1.221, G3.928+0.253, G4.005+0, G4.619+0.288, G4.752+0.255, G5.791+0.794 and G5.852+1.041 were observed on 11th and 13th February 2001 in two different bands. Moreover, the source G353.462$-$0.691 from ATCA observations was found to be almost unpolarised in the 6 cm band, and was re-observed on 11th February with the VLA. To unwrap possible n$\pi$ wrap in polarisation angles measured between two frequencies, 5 more sources from the ATCA pilot run, G358.002$-$0.636, G1.035+1.559, G359.844$-$1.843, G359.911$-$1.813 and G2.143+1.772 were re-observed on the 2nd day (13th Feb, 2001) of the VLA observations. From these sources, we re-observed G359.604+0.306 and G359.717$-$0.036 with the ATCA on 12th April 2002 to uniquely determine their RM.
The secondary calibrators used for the observation were the same as in the ATCA observations. Polarisation calibration was performed using the unpolarised source 3C84. 3C286 was used as the primary flux density calibrator and to estimate the instrumental phase difference between the two circularly polarised (RR and LL) antenna signals. Each source was observed for typically 5 minutes at 2 different hour angles. All the data were calibrated and processed using the [AIPS]{} package.
The source G356.905+0.082 has a compact core, which is weakly polarised, and an extended highly polarised halo. Due to zero spacing problem, the 8.5 GHz image of the halo could have missing flux density problems. Since the polarised intensity per beam of the core is similar to the contribution by the halo, which has small scale structures in polarised image and is not properly sampled at 8.5 GHz, we could not reliably estimate the RM from the core or the halo. Therefore, we have not provided its RM in Table \[gc.rm.list\].
-------------- ----------- --------- --------- ------------- ----------
Epoch Telescope Array Obs. Frequency No. of
config- time (GHz) sources
uration (hours) observed
06 Feb 2000 ATCA 6A 10 4.80 & 5.95 12
30 Sept 2000 ATCA 6B 12 4.80 & 5.95 12
02 Oct 2000 ATCA 6B 12 4.80 & 5.95 12
06 Oct 2000 ATCA 6B 12 5.31 & 8.51 14
08 Oct 2000 ATCA 6B 12 5.31 & 8.51 13
11 Feb 2001 VLA BnA 06 4.63 & 4.88 28
13 Feb 2001 VLA BnA 06 8.33 & 8.68 32
12 Apr 2002 ATCA 6B 11 5.06 & 5.70 08
-------------- ----------- --------- --------- ------------- ----------
: Journal of observations[]{data-label="table.pol.obs"}
Results
=======
Based on the ATCA data, 24 sources were found to have at least one polarised component. Images of these sources are shown in Fig. \[5ghz.atca.larger.maps\]. The FWHM sizes of the beams in the ATCA images are $\approx$6$^{''} \times 2^{''}$, and the typical RMS noise is 0.23 mJy/beam in Stokes I and about 0.15 mJy/beam in Stokes Q and U. These images show 7 of the 24 sources to have a single unresolved component and the remainder are either partially resolved or have multiple components.
Of the sources observed using the VLA, 21 have at least one polarised component. Images of these sources along with 2 polarised phase calibrators are shown in Fig. \[5ghz.vla\]. The FWHM sizes of the beams in the VLA images are $\sim$2$^{''} \times 1.5^{''}$ and the RMS noise is typically 75 $\mu$Jy/beam. In these images, four sources appear as single unresolved components. Since 1741$-$312 and 1748$-$253 have been observed by both ATCA and the VLA, but 1741$-$312 has a faint emission around the compact source as seen in higher resolution VLA maps, we have presented its properties in Stokes I from the VLA data.
Stokes I images of the sources in the sample that were not detected to have polarised emission are shown in Fig. \[5ghz.atca.unpol.src.maps\] and Fig. \[5ghz.vla.unpol.src.maps\] respectively.
The properties of the 24 polarised sources observed with ATCA are presented in Table \[pol.src.prop.at\] and 21 sources observed with VLA are in Table \[pol.src.prop.vla\]. In these tables, the following conventions are used. Column 1: the source name using Galactic co-ordinates ($l \pm b$). Column 2: the component designation; ‘N’ denotes Northern, ‘S’ denotes Southern, ‘E’ denotes Eastern and ‘W’ denotes Western, ‘C’ denotes central, ‘EX’ denotes highly extended and ‘R’ denotes ring type. Columns 3 and 4: Right ascension (RA) and Declination (DEC) of the radio intensity peaks of the components in J2000 co-ordinates. Column 5: the deconvolved size of the components with their major and minor axes in arc-seconds and the position angle (PA) in degrees (formatted as major axis $\times$ minor axis, PA). A few sources that are observed to have multiple resolved components in the 8.5 GHz images are, however, not well resolved in the 4.8 GHz images. For these sources, we have measured the size parameters of the components from the 8.5 GHz images and we put a ‘\*’ symbol beside these measured parameters. Columns 6 and 7: the corresponding peak and total flux density of the components at 4.8 GHz in units of mJy beam$^{-1}$ and mJy respectively. Column 8: total flux density of the component at 4.8 GHz as measured by the VLA GPS survey. Column 9: percentage polarisation of the components. Column 10: spectral indices of the components measured between 8.5 and 4.8 GHz. A few of the sources are extended over several synthesised beam-widths and for these sources we have convolved the 8.5 GHz images to the resolution at 4.8 GHz and then made spectral index images. The spectral indices of the individual components are measured from these images, we put an ‘s’ beside the spectral index for these extended sources. Columns 11 and 12: spectral indices between 4.8 and 1.4 and between 1.4 and 0.3 GHz respectively. Column 13: the source classification; ‘EG’ denotes an extragalactic source (based on the morphology), and G denotes a Galactic source. Several sources show the morphology typical of FR$-$I or FR$-$II sources and this is noted along with the extragalactic classification. Sources which appear unresolved (deconvolved source size $<<$ beam size) are denoted by U, slightly resolved (deconvolved source size $\lesssim$ beam size) by SR, double by D and T denotes a triple source consisting of a pair of lobes and a core. C+E denotes a flat spectrum core with extended emission either in the form of a lobe or jet. If there are several objects in the field which appear to be unrelated, we label the object as M.
For computing spectral indices in Table \[pol.src.prop.at\] and \[pol.src.prop.vla\], the 1.4 GHz flux densities of the sources have been taken from the VLA GPS and the NRAO VLA Sky Survey (NVSS) [@CONDON1998]. If the measured flux density of a component in the GPS differs from that in the NVSS by more than 20%, we have put a ‘$\dagger$’ mark beside the computed spectral indices (column 11). We have visually examined the NVSS images of these sources and if we find that the source is not in a confused region of the image we have used the flux density from NVSS to compute the spectral index; in such cases, we put ‘(N)’ beside the measured spectral index. For the source G359.871+0.171, the 1.4 GHz flux density has been assumed to be the same as that measured by @LAZIO1999 at 1.5 GHz and we put ‘(L)’ beside its measured spectral index between 4.8 and 1.4 GHz. For a few sources in the list, the 1.4 GHz flux density in unknown. In these cases, we put a ‘\*\*’ in column 11 and enter the spectral index between 4.8 and 0.3 GHz in column 12 with ‘(0.3/4.8)’ written below. The P-band flux densities of the sources have been taken from the Texas survey at 365 MHz [@DOUGLAS1996]. However, the Texas survey is known to have large uncertainty in flux densities for sources near the Galactic plane and which is more near the complex GC region. Therefore, if any source is detected in the GC image at 330 MHz [@LAROSA2000], we have used their flux density to compute the spectral index (column 12 in Table \[pol.src.prop.at\], \[pol.src.prop.vla\] and column 11 in Table \[unpol.src.prop.at\], \[unpol.src.prop.vla\]) and put ‘GC’ in parenthesis beside the spectral index measured. For 5 sources the flux densities at 330 MHz have been taken from @ROY2002, and we put ‘(GM)’ beside the spectral index in column 12. We have also taken the flux densities of 3 sources at 330 MHz from S. Bhatnagar (private communication) and put ‘(GM1)’ in column 12. Many of the sources resolved at frequencies of 1.4 GHz and above appear unresolved in the low frequency Texas survey. For these sources, we only compare their integrated flux densities between 1.4 and 0.3 GHz, and put ‘(i)’ beside the measured spectral index in column 12.
In Table \[unpol.src.prop.at\] & \[unpol.src.prop.vla\] we present the properties of the sources which are not detected in polarised emission. These tables are similar to Table \[pol.src.prop.at\] & \[pol.src.prop.vla\], except that we have omitted the column representing percentage polarisation (column 9 in Table \[pol.src.prop.at\] & \[pol.src.prop.vla\]). For the four Galactic HII regions we have identified, we write ‘G$-$HII’ in column 12 of this Table.
The measured RM towards 44 sources (65 components) and 2 secondary calibrators are given in Table \[gc.rm.list\], which is arranged as follows:\
Column 1: the source name in Galactic co-ordinates (G$l \pm$b). Column 2 & 3: RA (J2000) and DEC (J2000) of the source components. The co-ordinates of these components are based on the peak in the polarised intensity (if the peak in the polarised intensity do not coincide with the peak in total intensity, the component position will be slightly different than what is given in Table \[pol.src.prop.at\] & \[pol.src.prop.vla\] based on the peak in total intensity). Column 4 & 5: the measured RM (in rad m$^{-2}$) and the error in these measurements at the position of the peak in the polarised emission. Depolarisation fraction is defined as the ratio of the polarisation fraction of any component between lower to that at higher frequency, and in column 6 we provide the depolarisation fraction (D) of the source components between 4.8 and 8.5 GHz. Column 7: Percentage error in the depolarisation fraction ($\Delta$(D) in %). Assuming that the emission mechanism is synchrotron, the orientation of the electric field in the radiation has been used to infer the orientation of the magnetic field in the plasma and we provide the direction of this magnetic field ($\theta$) and the error in this measurement ($\Delta\theta$) in columns 8 & 9 respectively. Column 10: Reduced Chi-square ($\chi^2$) of the fit of Equation \[rm.equation\] to the measured polarisation angles. We show an example of bad fit ($\chi^2$=17) in Fig. \[bad.fit\], and one example of good fit ($\chi^2$=0.2) in Fig. \[good.fit\]. Column 11: Measured polarisation angles at different frequencies. The observed frequencies (in MHz) and the measured polarisation angles (in degrees) are tabulated in pairs, and each frequency, polarisation angle pair are separated by commas.
![An example of good fit of Equation \[rm.equation\] to the measured polarisation angles vs. square of wavelength plot (reduced $\chi^2$=0.2). The polarisation angles are measured towards the source G356.567+0.869.[]{data-label="good.fit"}](g358\DOT 0-1\DOT rm\DOT new){width="\textwidth"}
![An example of good fit of Equation \[rm.equation\] to the measured polarisation angles vs. square of wavelength plot (reduced $\chi^2$=0.2). The polarisation angles are measured towards the source G356.567+0.869.[]{data-label="good.fit"}](g356\DOT 5+1\DOT rm){width="\textwidth"}
The source G359.2$-$0.8 (Mouse) is a known Galactic non-thermal source. Since this source is known to be within 5 kpc from the Sun [@UCHIDA1992], the RM towards this object is expected to be quite small. To check if observations confirm its low RM, we observed this source. From Table \[gc.rm.list\], its RM is indeed measured to be very small ($-$5 $\pm$18 [rad m$^{-2}$]{}). However, our samples are selected to measure the RM introduced by the GC region, but this source only samples the local ISM, and its RM is not used in any further analysis.
Accuracy of the spectral index measurements
-------------------------------------------
The errors in the derived spectral indices depend on the accuracies of the flux densities at different frequencies. One of the drawbacks of a Fourier synthesis array is that if a source is resolved on the shortest interferometer baseline, its flux density will most likely be underestimated in the image. In the GC region, where the sky density of sources is high and emission at various size scales may co-exist, confusion could be a significant source of errors in imaging and hence in flux density estimates. Because our radio observations are performed at relatively high frequencies, the contribution from extended Galactic synchrotron background is negligible. Additionally, most of the sources in our sample have small angular sizes and, therefore, the problem of missing flux density should be minimal. The ATCA and VLA observations were both made using a single array at both the frequencies. Since problem of any missing flux density increases with increase in observing frequency, this might result in an underestimation of the source spectral index. We have detected extended emission of up to $\sim$10$'$ scale in our 4.8 GHz images of G353.410$-$0.360, G355.424$-$0.809, G355.739+0.131, G356.905+0.082, G358.149$-$1.675, G358.643$-$0.034, G359.2$-$0.8 (Mouse), G359.717$-$0.036, G0.313+1.645 and G5.791+0.794. To examine for any missing flux density at 8.5 GHz, we first restricted the Fourier Transform of the CLEAN components of the 4.8 GHz images to the [*uv*]{}-range of the 8.5 GHz visibilities. A comparison of the corresponding images with the original 4.8 GHz images (without the restriction in visibility coverage) shows that except for G353.410$-$0.360, G355.739+0.131, G356.905+0.082, G358.643$-$0.034, G358.149$-$1.675, G359.2$-$0.8 (Mouse) and G359.717$-$0.036, the missing flux density is less than 10% of the total flux density, and the corrsesponding error in their spectral indices between 8.5 and 4.8 GHz is less than 0.25. Among the above 7 sources, 4 have been identified as HII regions. The spectral indices between 8.5 and 4.8 GHz of the 2 remaining extragalactic sources G356.905+0.082 and G358.149$-$1.675 and the Galactic non-thermal source G359.2$-$0.8 (Mouse) (Table \[pol.src.prop.vla\]) have been measured only for the compact components (e.g., core and hot spots). For other sources, the error in the measured flux densities are expected to be about 5%, and the corresponding error in their spectral indices between 8.5 and 4.8 GHz is about 0.1.
Owing to extended emission and enhanced source confusion near the GC, images made from radio interferometers and with sparse spatial frequency coverage suffer from systematic errors. To estimate the errors in the source flux densities in the GPS and Texas survey, we compared the 1.4 GHz flux densities of our sample sources as measured by the GPS and the NVSS and found that the differences in flux density measurements were about 20 per cent. 20% differences in flux density contributes an error of about 0.15 in the measured spectral index between 4.8 and 1.4 GHz. We also compared the flux densities of the 12 sources which are detected in both the Texas survey and in the 330 MHz VLA GC image [@LAROSA2000] and noticed that the difference is almost a factor of two for six of these sources. The errors in the flux density measurents from the GC image are expected to be relatively smaller and, therefore, if the 0.3 GHz flux density of a source is taken from the Texas survey, an error $\sim$0.5 can occur in its spectral index measurement between 1.4 and 0.3 GHz. We note that other than the missing flux density problem, there are several other systematic errors which could lead to erroneous measurement of spectral indices. One of these errors is source variability, which we discuss below.
Effect of source variability {#variability}
----------------------------
Estimation of source spectral indices and the RM requires multi-frequency observations. Most of our multifrequency observations were separated by a time period, which varies from a few days to an year or two. Any source variability in these timescale will affect the measured spectral indices and possibly the measured RM. However, we note that the variability of the extragalactic sources at frequencies higher than a few GHz is often intrinsic in nature [@WAGNER1995] and in some cases caused by interstellar scintillation (ISS) [@LOVELL2003]. The typical variability timescale for intrinsic variability is from a few hours to years [@WAGNER1995], and from hours to months in the case of ISS [@LOVELL2003]. However, only the core dominated objects with flat spectrum ($\alpha >-0.5$) can show significant variability in timescale of days or less [@WAGNER1995]. Among the sources observed, only G356.905+0.082, G357.865$-$0.996, G358.002$-$0.636, G0.537+0.263, G3.928+0.253 falls into this category. Among them, the multifrequency observations of G356.905+0.082 and G3.928+0.253 were separated by only 2 days, and the probability of significant variability at this timescale is believed to be low. As described below, G357.865$-$0.996 is variable, G0.537+0.263 does not appear to be variable, and we do not have data to check for variability of G358.002$-$0.636. We note that variability in timescale of few hours could not be identified for any of our sources.
We identify 12 sources from Sect. 2, G357.865$-$0.996, G358.002$-$0.636, G359.388+0.460, G359.604+0.306, G359.717$-$0.036, G359.844$-$1.843, G359.911$-$1.813, G359.993+1.591, G0.537+0.263, G1.028$-$1.110, G1.035+1.559, G2.143+1.772, towards which our observations were separated upto 2 years, and their measured properties could suffer substantially from variability. Among these sources G1.028$-$1.110, G1.035+1.559, G359.604+0.306, G359.717$-$0.036 and G359.993+1.591 are extended (more than few arc seconds in angular size). Consequently, their variability timescale will be many years, and they are not considered to be variable in this paper. Among the rest of the sources, G357.865$-$0.996, G359.388+0.460, G0.537+0.263, G359.911$-$1.813, G2.143+1.772 were observed with ATCA on days, where certain frequency band was common to both the epochs. On comparing the flux densities of these sources in the same band (with appropriate correction of flux densities from their measured spectral indices due to change of frequency), from both the observing days, it was found that only the flux density of the source G357.865$-$0.996 has changed by more than 10%. Unfortunately, such a method could not be applied for the remaining 2 sources G358.002$-$0.636 and G359.844$-$1.843. Since, G359.844$-$1.843 is a steep spectrum source (Table \[pol.src.prop.at\]), the emission from it is likely to be originated from extended region (e.g., from jets), and the variability timescale is large. However, G358.002$-$0.636 is a rather flat spectrum source (core dominated), and could be variable. We note that the fit of Equation 1 to its polarisation angles is bad (Reduced $\chi^2$=17, Table \[gc.rm.list\]), and one possible reason for this is source variability. Consequently, the measured RM of G357.865$-$0.996 and G358.002$-$0.636 is excluded from the sample used to derive statistical estimates of the RM introduced by the GC region. We do not consider the effect of variability on the spectral indices of sources estimated from other catalogues (i.e., other than $\alpha_{8.5/4.8}$).
Contribution of the intrinsic Faraday rotation to the measured RM {#subsect_intrinsic}
-----------------------------------------------------------------
When the polarised emission from the source reaches the observer, rotation of the polarisation angle could occur (i) within the source, (ii) in the Inter Galactic Medium (IGM) or (iii) in the ISM of the Galaxy. Compared to the ISM of our Galaxy, the electron density of the IGM is very small, and consequently the Faraday rotation introduced by the IGM is negligible. However, if the synchrotron electrons are mixed with thermal electrons at the source, or, if there is an intervening galaxy, the ISM of which introduces RM, or if there is cluster of galaxies along the line of sight, there can be Faraday rotation introduced outside our Galaxy. As discussed by @GARDNER1966 [@KRONBERG1972; @VALLEE1980], the polarisation angle could deviate from the $\lambda^2$ law due to the following three mechanisms.\
(i) If the synchrotron optical depth becomes significant at some frequency, the polarisation direction makes a transition from parallel to perpendicular to the projected magnetic field. (ii) If there are multiple unresolved emission components with differing spectral indices and polarisation characteristics, it can cause complex wavelength dependent variations in polarisation angle. (iii) If there are significant gradients in Faraday rotation across or through the emission region of the source, then also polarisation angle can have complex dependence on the observing wavelength. The polarisation angles of the source G358.002$-$0.636 deviates significantly from the $\lambda^2$ law, and other than source variability (Sect. \[variability\]), it could have been caused by one of the above processes.
In the case of Faraday rotation outside the Galaxy, the Faraday screen is likely to be located several orders of magnitude farther away than the GC ISM. As a result, emission from different parts of the source is viewed within one synthesised beam, the linear scale of which is much larger than what is sampled in our ISM. At such length scales ($\sim$1 kpc), the magnetic field within the intervening Faraday screen is likely to be uncorrelated. As a result, there is differential Faraday rotation within the beam, which gives rise to source depolarisation. From the data, the RM towards the source G358.917+0.072 is found to be the highest with a measured value of 4768 [rad m$^{-2}$]{}, and it shows a high depolarisation fraction (0.3 between 4.8 and 8.5 GHz), which is likely to be caused by differential Faraday rotation [@KRONBERG1972].
Following the arguments given above, if the reduced $\chi^2$ of the fit is greater that 4.6, the probability of occurrence of which is less than 1% [@BEVINGTON1969] or any of the source component shows a depolarisation fraction of less than 0.6 (Table \[gc.rm.list\]) between 4.8 and 8.5 GHz, we suspect that there is significant RM introduced outside the Galaxy and the RM towards those components have not been used any further in this or in deriving the properties of the Faraday screen in Paper II. We note that the intrinsic RM of most of the extragalactic sources are quite small $\sim10$ [rad m$^{-2}$]{} [@SIMARD-NORMANDIN1980] and the RMs measured from our sample is quite high ($\sim1000$ [rad m$^{-2}$]{}), and consequently the RM introduced outside the Galaxy should have little effect on the RMs of most of the sources in our sample.
Discussion
==========
Galactic HII regions
--------------------
Extended unpolarised structures, typical of HII regions, are seen in the VLA images of the sources G353.410$-$0.360, G355.739+0.131, G358.643$-$0.034 and G359.717$-$0.036. A small diameter source is seen in the field of G353.410$-$0.360, which has a flat spectrum between 8.5 and 4.8 GHz. It is well resolved in our BnA array image and is unlikely to be the flat spectrum core of a background extragalactic source. We believe that it is a Galactic HII region.
The 8.5 GHz VLA BnA array image of G359.717$-$0.036 (Fig. \[8ghz.G359.7$-$0.SNR.candidate\]) shows a partial shell-like structure. Extended emission can also be seen around this object. Since either HII regions or shell type SNRs can have shell like morphology, we have tried to measure its spectral index accurately. We made an image of this object at 8.4 GHz using the archival VLA CD array data (Project Code AY68 made by Farhad Yusef Zadeh). We also imaged it at 1.4 GHz using the archival Galactic Centre data acquired and presented by @PEDLAR1989 using the B, C and D array of the VLA during 1981–1984. From the low resolution 8.4 and 1.4 GHz image of this source, the measured spectrum of this component is flat, indicating it to be an HII region.
The image of G358.643$-$0.034 at 8.5 GHz suffers from missing short spacings and, consequently, the measured steep spectral index between 8.5 and 4.8 GHz (Table \[unpol.src.prop.vla\]) is underestimated. However, its spectrum between 4.8 and 1.4 GHz is inverted and the source is not detected in the 330 MHz GC image. The extended structure and inverted spectrum at low frequencies suggest that it is an HII region.
G355.739+0.131 has a flat spectrum between 8.5 and 1.4 GHz and is well resolved. This source is also not detected at 0.3 GHz, which suggests that its spectrum has turned over due to significant free-free absorption at low radio frequencies: these suggest that the source is an HII region.
Are the rest of the sources extragalactic ?
-------------------------------------------
Excluding the sources described above, the rest of the sources observed are either polarised or the deconvolved diameters imply a brightness temperature greater than 10$^3$ K at 4.8 GHz. Except for a few small diameter sources, all the other sources have steep spectral indices indicating the emission from these sources is non-thermal. The small diameter objects have flat spectra ($\alpha \ge -0.3$) down to 0.3 GHz. A thermal source with brightness temperature of 10$^3$ K at 4.8 GHz should show self absorption (i.e., positive spectral index) at 0.3 GHz (otherwise, the required brightness temperature exceeds 10$^5$ K). Therefore, these objects with flat spectra at 0.3 GHz indicate that the emission from these small diameter sources are also non-thermal. Among the non-thermal sources, the object G359.2$-$0.8 (Mouse) (Table \[pol.src.prop.vla\]) [@YUSEF-ZADEH1987d; @UCHIDA1992] is known to be a Galactic non-thermal source located within 5 kpc from the Sun. We review below non-thermal emission from Galactic objects with the aim of rejecting any of the remaining sources in the sample, which might admit such a classification.
### Non thermal emission from Galactic sources
There are several types of Galactic sources that emit non-thermal emission.\
(i) Supernova Remnants (SNRs): These are the remnants of supernova explosions. The electrons in these objects are accelerated to high energies near the expanding shock front. The SNRs are usually spherical in shape, and when projected on the sky appear like a ring for the shell type SNRs, while they appear to be filled with emission in the case of plerion-type SNRs. None of the non-thermal objects in Tables \[pol.src.prop.at\], \[pol.src.prop.vla\], \[unpol.src.prop.at\] & \[unpol.src.prop.vla\] have such a morphology, indicating that there is no resolved supernova remnant among these sources. However, if an SNR is young, its angular size will be small and will appear like an unresolved source. Assuming an initial expansion velocity of 3000 [km s$^{-1}$]{}, an SNR at a distance of 10 kpc will expand to an angular size of 6$^{''}$ in about 100 years after the explosion. An object of angular size of 6$^{''}$ will be well resolved in our images and should have a ring or filled centre morphology: the absence of these rule out any SNR older than 100 years in our sample. Since the expected number of supernova explosions in our Galaxy is believed to be about 1 in 50 years, the probability of finding an SNR of age less than 100 years in the central 12[$^{\circ}$]{}$\times$ 4[$^{\circ}$]{}of the Galaxy is less than 0.1, which suggests that there is no young SNR in our sample.
\(ii) Radio Pulsars: Pulsars typically have a very steep spectrum, with spectral indices of $-1.5$ to $-4$ at cm wavelengths. They would be unresolved and almost all would be undetectable at 5 GHz owing to their steep spectrum. Since none of the unresolved sources in our sample have a very steep spectrum, we believe that there is no pulsar in our list of sources.
\(iii) Radio Stars: Most stars are weak radio emitters; however, some are detected in non-thermal emission. @BECKER1994 show that the radio flux densities from these stars are $\le$1 mJy at 5 GHz. Since the observed flux densities of our sources are much higher than the upper limit for these stars, such objects can be ruled out from our list of sources.
\(iv) Transient sources: These are highly variable and transient radio sources which include radio counterparts of X-ray sources. The sources in our sample are not only detected at 4.8 and 8.5 GHz, but also detected at 1.4 GHz in the GPS and the NVSS. These sources are also detected at 0.3 GHz, either in the Texas survey or in the 330 MHz GC observations. These observations were separated by days to years and the measured spectral indices determined between any two bands are quite close to the mean value (differences are less than 0.6). Therefore, the flux densities of these sources have not changed by more than a factor of two and this rules out the possibility that there are transient sources in our catalogue.
\(iv) Galactic Microquasar: These are stellar-mass black holes in our Galaxy that mimic, on a smaller scale, many of the phenomena seen in quasars (see @MIRABEL1999, and the references therein). For a black hole accreting at the Eddington limit, the characteristic black body temperature at the last stable orbit in the surrounding accretion disk is given by T $\sim2 \times 10^7
M^{-1/4}$ (Rees 1984). Therefore, compared to the AGNs, the emission from microquasars are shifted towards higher frequencies and the microquasars are usually identified by their X-ray properties [@MIRABEL1999].
Though many of the already known microquasars are highly variable, two of these sources, 1E1740.7$-$2942 [@MIRABEL1999] and GRS 1758$-$258 [@MARTI2002], are persistent sources of both X-rays and relativistic jets. At radio wavelengths, these two sources are morphologically similar to typical radio galaxies, which have a central compact component and two extended lobes. Therefore, based on morphology, microquasars cannot be separated from the distant radio galaxies. However, as mentioned above, microquasars are believed to have X-ray counterparts. We have searched the ROSAT PSPC all sky survey [@VOGES1999] and a catalogue of soft X-ray sources ($|l| \le$ 1.5[$^{\circ}$]{}, $|b| \le $ 2.0[$^{\circ}$]{}) in the GC region [@SIDOLI2001]. Twenty of our sources are also located within the boundary of the ASCA survey of the GC region ($|l| \le$ 2.5[$^{\circ}$]{}, $|b| \le $ 2.5[$^{\circ}$]{}) [@SAKANO2002]. However, none of the radio sources in our sample were found to have any counterpart in these catalogues. Therefore, it is unlikely that any of the sources we have observed (Tables \[pol.src.prop.at\], \[pol.src.prop.vla\], \[unpol.src.prop.at\] & \[unpol.src.prop.vla\]) is a microquasar.
Extragalactic source counts
---------------------------
The expected number of extragalactic sources (N) in 1 square arc minute of the sky at 5 GHz and with a flux density limit of S mJy is N($>$S)= $0.032 \times$ S$^{-1.13}$ [@LEDDEN1980]. Therefore, the expected number of extragalactic sources seen through the central $l \times b$ = 12[$^{\circ}$]{}$\times$ 3.6[$^{\circ}$]{} region of the Galaxy above a flux density limit of 20 mJy at 5 GHz is 168. However, as we selected sources with steep spectral indices ($\alpha <-0.4$), typical sizes $\le$10$^{''}$ and excessive confusion prevails in the region, we could identify only 59 extragalactic sources, which indicates that about two thirds of the extragalactic sources in this region are yet to be identified. The median angular size of these 59 sources is 7.6$^{''}$, and the median flux density at 1.4 GHz is 160 mJy.
Summary
=======
We have observed 64 sources towards the central $-$6[$^{\circ}$]{} $~<~l~<~$6[$^{\circ}$]{}, $-$2[$^{\circ}$]{} $~<~b~<~$2[$^{\circ}$]{} of the Galaxy using the 6 and 3.6 cm band of the ATCA and the VLA. Based on our work described herein, 59 of these sources are classified to be extragalactic. This increases the number of known extragalactic radio sources towards this unique region by almost an order of magnitude and provides the first systematic study of the polarisation properties of the background sources in the region. We provide 4.8 GHz images of all the observed sources and measure the angular sizes and the spectral indices of these sources. Based on the morphology, spectral characteristics and polarisation properties, we identify 4 Galactic HII regions in the sample.
Acknowledgements {#acknowledgements .unnumbered}
================
S.R. thanks D. J. Saikia for useful discussions. The Australia Telescope is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
\[lastpage\]
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Samuel Regandell
- Thomas Marquart
- Nikolai Piskunov
bibliography:
- 'citations.bib'
date: 'Accepted 2017-12-18'
subtitle: Putting standards into practical software
title: Inside a VAMDC data node
---
Introduction {#sec:introduction}
============
The problem of analyzing remote sensing data is encountered in a wide variety of scientific and technological fields, ranging from astronomy and climate research to computer tomography and movie production. Its solution requires detailed knowledge about absorption, scattering/reflection and emissivity of matter. This in turn builds on insight to the electronic and nuclear structure of atomic and molecular species, its dependence on the environment, as well as radiative and collisional processes leading to transitions between energy levels. Ray tracing, spectral analysis and other types of remote sensing tools are well-established in research and industry. By contrast, the collection of relevant atomic and molecular data has remained challenging. This is due to the heterogenous nature of individual data collections: Atomic and Molecular ([[A&M]{}]{}) databases are often created to serve very specific applications (e.g. modelling of stellar atmospheres) or as repositories storing the output from experimental or theoretical work by atomic physicists. Thus a particular data collection is often specialised not only in terms of content but also in terms of units, format and query language.
The Virtual Atomic and Molecular Data Centre (VAMDC, <http://vamdc.eu>) is an electronic infrastructure that solves these problems by providing standardized ways to both formulate queries and to retrieve data. All the data collections that are part of VAMDC, henceforth called **data nodes**, provide their internally heterogenous data in a commonly agreed-to, machine-redable form. VAMDC was originally created with support from an EU FP7 grant and is now maintained and further developed by a large international consortium. It offers atomic- or molecular data providers the possibility to quickly bring their results to their users. At the same time it provides the attributes of published work through including the valid references for proper citations. All major [[A&M]{}]{} databases are either integrated in VAMDC or support standards and protocols developed within this project.
![Graphical overview of the VAMDC infrastructure.[]{data-label="fig:vamdcstructure"}](figures/VAMDC_infrastructure.eps){width="40.00000%"}
To make uniform access to data possible, VAMDC maintains an infrastructure of web services. The VAMDC **Registry** lists data nodes and other services together with their respective capabilities. This allows users and programs to find resources and only select those containing the sort of data they need. Most visible from the perspective of the data-user community is the VAMDC [**portal**]{}[^1]. The portal is a web interface that helps with formulating queries and collecting data from the connected data nodes. The portal also acts as an intermediary to services that process the machine-readable results; these services can perform calculations and export to various file formats or human-readable presentations. For a full overview, see @vamdc2016.
In this paper we focus on VAMDC’s data provider infrastructure, consisting of the aforementioned data nodes. A node is hosted and maintained either by the data producer or by a group interested in collecting data from several sources. Each node may offer a wide variety of scientific data depending on its history and purpose. Regardless of its internal layout and content, a node can join VAMDC by implementing the interface defined by the VAMDC standards. This interface must:
- Understand queries formulated using the VAMDC query language.
- Produce data output in the standardized data format (XSAMS) used by VAMDC for data exchange.
- Implement the VAMDC application programming interface (API) to receive queries, to send back the results and to report additional information to clients and to the registry.
The **node software** is the software package supplied by VAMDC. It aims to help individual data nodes set up and operate. In the following we will first go into some detail on the VAMDC standards. We then describe the node software implementation and the intricacies that arise from balancing versatility, simplicity of use and performance at the same time. Finally we take an existing data node as an example to illustrate the node software in action and the process of using the co-bundled [publishing tools]{} to import and update node data content.
The VAMDC Standards for Data Nodes {#sec:vamdc_standards}
==================================
The VAMDC standards are publicly available online at and we refer the reader to that resource for all details omitted here. The purpose of this section is to summarise the parts of the standards that are particularly relevant to the node software and to highlight a few aspects that directly affect the implementation, as described in the subsequent sections.
The data format: XSAMS {#sec:output_language_xsams}
----------------------
Data returned by a node must conform to the [XML Schema for Atoms, Molecules and Solids]{} (XSAMS). XSAMS originated at the International Atomic Energy Agency (IAEA) and is now maintained and further developed by VAMDC in collaboration with NIST. XSAMS defines a strict layout for where and how data is represented in the XML document. This includes for example how to describe atomic and molecular structure, processes like collisions or radiative transitions, as well as supplemental information such as environmental properties. In addition, associated references are included and cross-referenced within the XML structure. The top levels of the XSAMS structure are shown in a simplified form in Figure \[fig:xsamstoplevels\].
![Simplified summary of XSAMS top level structure (colored boxes) with a small selection of the possible properties representing an atomic radiative transition (gray boxes). The depicted relationships (arrows) hide a lot of complexity; for example the radiative transition references the species but also individual upper/lower states. Almost all properties can also reference its own sources.[]{data-label="fig:xsamstoplevels"}](figures/XSAMSlayout.eps){width="40.00000%"}
XSAMS documents are validated against the schema. This ensures that the document is self-contained with consistent referencing (no dangling cross-references) and that no essential information is missing. The schema makes heavy use of its own types and classes to make sure that properties like data accuracies and evaluations are always represented in the same way.
VAMDC publishes its standards in an online **dictionary**[^2]. The **returnables** section of the standard defines exactly how a node returning XSAMS should name a specific data property as well as what unit and data type should be used. For example, the life time of an atomic state must be identified as *AtomStateLifeTime* and be a float given in seconds.
The query language: VSS2 {#sec:query_language_vss2}
------------------------
The query language that data nodes understand is called [VAMDC SQL-subset 2]{} (VSS2). It consists of severely restricted SQL-like `SELECT` statements. VSS2 supports querying with binary operators `<=,>,…`, `LIKE` and `IN` as well as logical relations with `AND` or `OR`. Other standard SQL features, like `BETWEEN`, are not included in the SQL-subset. Instead of database column names, SELECT statements use keywords from the VAMDC standard dictionary in a similar way as described for XSAMS in Section \[sec:output\_language\_xsams\].
Below is a full example of a query for atomic lines in FeIII:
SELECT * WHERE RadTransWavelength < 858.4
AND RadTransWavelength > 858.1
AND AtomSymbol = 'Fe'
AND AtomIonCharge = 2
Note that there is no SQL FROM-clause in the above query. This is because the internal table structure is hidden from users. The query is written as if the data existed in one large in-line table. The keys used in the WHERE-clause in place of column names are called **restrictables**. Restrictables are derived from the XSAMS itself and thus generally correspond to a certain location in the data hierarchy. While the definition of a Restrictable is fixed, it is up to each node to decide how to go about supplying it. For example, a given node may need to do convert units or perform an internal query against more than one database column or table.
It is important to note that, since XSAMS needs to be a self-contained document, one cannot expect to receive only the isolated data type specified in the SELECT-clause. That is, if the radiative transition we seek references a certain state which in turn refers to a certain ion, *both* the state and the ion must be included in the returned document for it to be valid XSAMS. The VAMDC dictionary however specifies a select number of **requestables**. One or several such requestables can replace the `*` wildcard in the query to request only a certain type of data. For instance, using the “Species” requestable in `SELECT Species WHERE...` will only return the atoms and molecules part of XSAMS. Leaving out the `WHERE` altogether will return all species that a node has data about.
The exact same query can be sent to any VAMDC data node. It is the node’s task to convert the standard keys into a query relevant to the node’s internal database. Each node also needs to specify and report just which keys are actually relevant for querying its data. This allows clients (like the VAMDC portal) to send a query only to nodes that can answer it, instead of gathering uncessary empty results.
The API: VAMDC-TAP {#sec:api}
------------------
A data node communicates with the world via HTTP, using GET/POST/HEAD requests. The definition of this web-API is known as [VAMDC-TAP]{}, where TAP stands for Table Access Protocol. This, in turn, is a subset of the International Virtual Observatory Alliance TAP standard (@2012ivoa.spec.0827D). VAMDC-TAP is a RESTful API that operates on the basis of a single request and reply. No sessions are stored or maintained. There is no connection between subsequent queries and there is no authentication.
At the specified URI endpoint data nodes accept requests (HTTP GET/POST) containing a VSS2 query as described in Section \[sec:query\_language\_vss2\]. If the request is instead made using HTTP HEAD, the node answers with statistics about the given query (in the form of HTTP headers). Such statistics include the number of species and processes that match the query as well as the expected size of the selected dataset. This allows gathering information without actually executing a potentially expensive query and subsequent big data transfer. This mechanism is used by the VAMDC Portal to gather preliminary statistics from all nodes. Nodes can choose to truncate the data if the result would be too large, reporting the truncation as part of the header return.
This is an example of a URI that uses VAMDC-TAP to query a data node:
http://<node-url>/tap/sync?
LANG=VSS2&
REQUEST=doQuery&
FORMAT=XSAMS&
QUERY=<VSS2-query>
The URI specifies the language, return format and the type of request. The QUERY parameter holds an SQL-like string as specified in Section \[sec:query\_language\_vss2\]. This example accesses the `tap/sync` endpoint to initiate a synchronous query to the node, meaning that the data is returned right away, in the response to the same HTTP request. There is currently no asynchronous alternative to avoid blocking the server in the case of very expensive queries. This is planned for the next iteration of VAMDC-TAP and node software releases.
The VAMDC-TAP definition includes the additional administrative URI endpoints `/tap/availability` and `/tap/capabilities`. The former allows the node to report planned node maintenance or downtimes. The node’s capabilities, a short XML document, include information like the URL addresses to various services, the version of the node software used as well as sample queries for testing the node in question. The capabilities document also specifies which **returnables** (see Section \[sec:output\_language\_xsams\]) and **restrictables** (see Section \[sec:query\_language\_vss2\]) this node supports. This API can be accessed by tools like the VAMDC portal as well as by custom use-case-specific applications. It is also intentionally easy to script.
VAMDC offers [processor]{} services. A processor takes an URI with a query as input, like in the example above. It then fetches the data and processes it before presenting the result to the user. A complication arises in the form of the [same-origin policy]{} that web-browsers implement for security reasons. This means that browsers don’t make cross-site requests unless the website or service explicitly annouces that it is fine with accepting them. Since VAMDC is inherently distributed and the various parts of the infrastructure are meant to interact with each other at the user’s behest, data nodes thus need to announce their willingness to take cross-site requests. This is done by answering HTTP SERVICE requests with the appropriate HTTP headers according to the cross-origin resource sharing (CORS) standard[^3].
Node software implementation {#sec:node_software_implementation}
============================
From the description of the VAMDC Standards above, it should be clear that there are obvious advantages with data nodes sharing a common software. The nodes need not only answer to the same API – incoming queries are also node-independent, as is the data output format that needs to be assembled for the reply. However, depending on the type of data they hold, nodes can be quite different internally. This concerns mainly the table layout of the underlying database and is reflected in the list of restrictables and returnables they support.
There are currently two implementations of the VAMDC standards for data nodes. One is built on the Python programming language while the other is using Java. All but two of the currently $\mathrm{\sim}$30 registered data nodes use the Python node software which is the subject of this paper. The Python node software can be obtained from its online git repository[^4]. It is released under the GPL version 3 open-source license.
Python is an open-source high-level language with a mature ecosystem of utility libraries. It is widely used within e.g. the astronomical science community and tends to be fast to learn for users with any previous programming experience. This has in our experience lowered the barrier of entry for data providers wanting to start up their own node.
Django and Data Model Abstraction
---------------------------------
The Python implementation of a VAMDC Node relies on the Django web framework[^5]. Django is widely used as the underpinnings of professional large-scale websites, is actively developed and has an active community. It features a powerful and flexible ORM (Object-Relational Mapping) which allows us to define a database schema and formulate queries to it using Python rather than SQL. This abstraction allows Django to support a wide range of different relational database engines. In effect, data providers already storing their data in an SQL database will likely find that Django can make use of it.
The database engine is not formally part of the node software, but MySQL with its MyISAM storage engine is used at many nodes. This also is the setup that has been recommended to new nodes that do not already have their data in a relational database. The recommendation still holds even though MySQL now uses InnoDB as default storage engine, because in our use-case with de-facto read-only access to the database, MyISAM has performance advantages once it has been set up.
Django’s abstraction layer between the database and the node software API implementation allows us to write code that is agnostic about a single node’s intricacies. This allows a large portion of the code to be reused by all data nodes. When a node has special needs that makes the common implementation insufficient, we add hooks where node-specific code can be plugged in.
The node software only uses a sub-set of Django’s capabilities. Still, it allows us to keep the code volume of the node software small: The core functionality is only some 1000 lines, plus another 2000 for the XML-generator. This has made maintenance and upgrades over the last seven years relatively easy, even when seen over many major Django versions. Similarly, the effort of porting the node software from Python 2.x to Python 3.x was measured in hours rather than days; both are now supported with the same code-base to ensure a smooth upgrade path for nodes.
The choice of Django has turned out to have a beneficial impact on several data nodes, beyond their involvement with VAMDC. Django’s relative ease of developing new custom interfaces for data collections has allowed nodes to significantly improve their own web presence, often heavily tailored to their pre-existing user communities. Examples of this are seen with for example CDMS[^6] (@2005JMoSt.742..215M) and HITRAN[^7] (@2013JQSRT.130....4R).
Publishing Tools {#sec:publishing_tools}
----------------
Some atomic and molecular data are produced through ab initio calculations of atomic and molecular energy structure, transition probabilities and collisional cross-sections etc. Other data come from experimental measurements of wavelengths, line strengths and lifetimes. Both approaches have advantages and problems. Occasionally, data producers merge the two e.g. by combining experimental lifetimes with theoretical branching ratios. In all cases, the new data comes in form of tables, often in peculiar units and with complex assessment of the quality.
A stand-alone part of the node software, dubbed the **publishing tools**, parses such data and populates a new relational database for use by the VAMDC node. Figure \[fig:vamdcstructure\] shows the top-level structure of the XSAMS output language the node is expected to return. For query efficiency the node should ideally try to organize its schema in a similar way. For example, the [Atom]{}, [Molecule]{} and [Transition]{} data should aim to be separate database tables.
In the simplest case, the node data provider already has an SQL database with a useful schema. Django’s native `inspectdb` mechanism can use the existing database for creating the Django data model. These auto-generated Python classes usually only need small modifications before the node software can access the database properly. In particular, the foreign key relations between tables have to be transferred to the Django model.
However, in many cases the original data is either stored using some non-SQL solution or using a schema less optimized for VAMDC queries. This is where the publishing tools come into play: they can perform various operations on the data before inserting it into the node SQL database. The goal is to do as much work as possible at database creation, when execution time does not matter, instead of repeating work every time an incoming query is processed. Possible operations include, but are not limited to, proper parsing of near-arbitrary table formats, interpretation of header information, tagging of missing data, unit conversions and pre-creation of derived quantities that are useful in the XSAMS output.
The first step for doing this data import is to define the new (empty) database schema. The new schema is formulated using Django’s model classes. Each such class represents a table in the new database, including the relationships between them. Every instance of these classes is one record in the table. This model is then used to create the actual tables in the database, again granting nodes the flexibility to choose any of the database engines that Django supports.
The mapping of data from the old data store to the new database can be the most work-intensive part of setting up a VAMDC data node, depending on the original form of the data. This is done by writing a **mapping file** that contains a Python dictionary in a form understood by the conversion tool. It must be created uniquely for each node. The components of the dictionary describe how to read and parse the records of each input file into one or more output files. Multiple inputs can for example be combined into a single output or vice versa. It is often not possible for the parser to work in parallel with all input files; for example, inter-table relations may not be possible to construct until all relevant files have been read. This requires multiple passes, something which is supported by the mapping mechanism. Since the mapping file can contain Python code, this process can be further customized with helper functions to account for any peculiarities in the input data. The publishing tool includes a set of functions for processing and handling common data formats.
Running the mapping file through the execution script produces a series of text files that exactly match the database schema – one file per database table with matching columns. These files can be efficiently read using the database’s own import functions. The syntax of this import is the only part of the node creation that depends on the particular choice of SQL database. For example, a user of MySQL would open the database client and execute:
``` {language="sql"}
mysql> LOAD DATA INFILE
filename INTO TABLE table;
```
Once the mapping file exists, it is easy to re-run and tweak the process, or to add new data to the node at a later time. The publishing tools are not part of the day-to-day operations of the VAMDC node.
With the data loaded in the new database, one can set up addtional database indices or perform other optimizations. This bit highly depends on the individual node’s database schema, size and engine.
Updates to the database schema often coincide with an update of the data itself, meaning that it can be convenient to simply use the publishing tools to recreate the entire database from an updated mapping file and input data files. If this is not desired or possible, Django comes with a very flexible system for database migrations, using small Python files that describe the changes to the database over time. These are executed in sequence by Django and, if written correctly, can even be used to revert changes and roll back to a previous version of the database in case of errors.
Receiving and executing queries {#sec:queries}
-------------------------------
One of the two main tasks of the node software is to receive queries of the form described in Section \[sec:query\_language\_vss2\] above, and execute them on the node’s database.
In order to do this, the node software needs to be told which VAMDC [Restrictables]{} correspond to what column in which table of the database. We implement this in the most straight-forward way, a Python dictionay where the key is the Restrictable and the value is the name of the Django model and the relevant data field, concatenated by an underscore. This is the Django-internal way of referring to model fields and we refer readers to the excellent Django documentation for more information. An example of this restrictables dictionary is shown in Section \[sec:vald\]. The point is that this dictionary connects the (node-independent) Restrictable keywords to the (node-specific) names of the right table and column in the database.
An incoming VSS2 query is validated as a first step. We use the third-party package `pyparsing`[^8] to formulate an SQL-dialect strictly limited to the supported SQL subset. Rejecting any SQL statements except SELECT right away also offers protection from a large set of potential malicious queries.
The parsed and validated query is then handed to what we call a node’s **query function**. This is the only piece of Python logic that is custom to each node. The query function’s goal is to turn the input query into `QuerySets`. A queryset are Django’s internal objects representing the *result* of an SQL query. A queryset is *lazily evaluated*, meaning that the query is only executed on the database once the result values are accessed. This allows the query function to quickly set up the result and pass it on to the XSAMS generator which is described in Section \[sec:xsams\_generator\] below. Only then, as we describe below, will the database deliver data that get wrapped in XML and streamed directly to the user that initiated the query.
Nodes have, by design, near-total freedom on how to write the query function. Covering all of the complexity and subtleties of [[A&M]{}]{} data collections would not be possible. Thus the only constraint the system impose is that a node’s custom query function returns the appropriate QuerySets. Documentation and examples are provided to help data providers with this step and most nodes use the method described in the following.
The node software offers a utility to convert a (parsed and validated) query into a Django `Q-object`. Most nodes use this first thing in their query function. A Q-object is Django’s way of representing a *database-agnostic* query. It is suitable for direct interaction with Django’s ORM.
The next step is to apply the Q-object to the data model. This is where nodes differ largely. For example a node that contains atomic transition line lists is likely to first restrict on the transition model. It would then use the ORM to figure out the corresponding species and states along with the references that need to be attached. Other nodes might restrict the species first or use a totally different strategy.
As mentioned above, the query function returns QuerySets that can be used to populate the output XML document. Each QuerySet needs to correspond to a top-level organizational group of XSAMS, like Atoms, Molecules, Transitions, Collisions, Sources and so on. The return value of the query function is therefore a Python dictionary where the keys identify the part of XSAMS to be used, and the values are the (yet to be evaluated) QuerySets. The names of the keys to be used are listed in the node software documentation.
The total overhead for receiving a query, parsing it and setting up the result, i.e. running a node’s query function, is typically below 0.5 sec which is negligible compared to the total response time, especially for large queries. This means that nodes typically are responsive and start returning data soon after the query is initiated.
Generating XSAMS {#sec:xsams_generator}
----------------
The second main task of the node software is to take the data that comes from the database and put it into an XSAMS document (cf. Section \[sec:output\_language\_xsams\]). In the previous section we described how nodes return QuerySets from their query function. These are passed directly into the node software’s **XSAMS generator**.
This generator is named such not only because it assembles the XSAMS data output, but also because it is a proper Python `generator` in the sense that it uses `yield` statements. Generators in Python are structures that can be looped over without being fully in memory when the loop starts. Expressions get evaluated only for the current loop element. This nicely matches the input QuerySets; only what is currently needed for the next piece of output gets fetched from the database and put into XML. Especially for large data set this is not only much more memory efficient, it also means less of a delay before the first piece of data is ready to return.
The main task of the XSAMS generator is to go through the differents parts of XSAMS in turn, passing each loop through the QuerySet it received. Each iteration of the QuerySet produces an instance of the Django data model (a row in the database) and the generator accesses its fields’ values to build up the corresponding piece of XSAMS. We do not use an XML library or a document object model for assembling the XML. For maximum control we instead use plain Python string concatenations, placeholders and `”.join()` on lists of strings.
In order to know which attribute of the current model instance needs to be mapped to which element or attribute of XSAMS, the generator needs the Returnables dictionary. This is analogous to the Restrictables dictionary from the previous section, but in the opposite direction. Again the dictionary keys are the (node-independent) keywords; these are hardcoded in the XSAMS generator to the proper XML-location. Since the list of Returnable keys is derived from XSAMS, there is a one-to-one correspondence.
The values of the dictionary tell the generator how to access the (current iteration) model instance to retrieve the wanted information. We use an internal function called `GetValue()` for this, allowing nodes to fill the Returnables dictionary values in several ways. Firstly, they can contain a static string, useful when a value is the same for all data and not even stored in the database (like units). Secondly, it can be the name of a model field in which case its value gets retrieved. Thirdly, it can be the name of a model field in a *different* model, connected by a foreign key relation; as common with Django, table traversal in models is possible by simply concatenating fields and foreign keys with periods.
The fourth possible option for filling the dictionary value is the name of a custom method on the data model. `GetValue()` calls this method and inserts whatever it returns into the current place in XSAMS. This allows for maximal flexibility, but for performance reasons nodes are encouraged to not use this if they can pre-calculate data instead. Note that while Python’s `eval()` is the easiest way to implement `GetValue()`, we went to some length to avoid it because this function gets called very often and needs to be speedy.
While the XSAMS generator is built to cover as many nodes’ requirements as possible, it is also possible to replace the common routines by custom versions at several levels of granularity. Some nodes attach a custom method to a data model that returns large pre-assembled chunks of XML, others choose to overwrite the top-level function of the generator altogether and re-use only parts of it. In practical use nodes have often found that it has been easier to create a custom routine rather than adapting the standard one to also cover a highly specialized case. By deviating from the common code base only when necessary, nodes keep the volume of custom code small.
Apart from Returnables and Restrictables, there is a third group of VAMDC standard keys, the Requestables (see Section \[sec:query\_language\_vss2\]). When one or several of these are given in the query, instead of `SELECT *`, this means that only certain parts of XSAMS are asked for. The generator in this case simply skips the execution of the other parts.
VAMDC-TAP implementation {#sec:vamdc_tap_generator}
------------------------
The previous sections describe how queries are parsed and how the reply content is assembled. What remains is the implementation of the VAMDC-TAP API. Django offers a straight-forward way to connect Python code to URL endpoints. Since the API is simple, a single `view` function per endpoint is enough. The `capabilities` and `availability` (see Section \[sec:api\]) endpoints are implemented by rendering simple templates with a few configuration variables, using Django’s own template mechanism. For example, the keys from the `Returnables` and `Restrictables` dictionaries go into the `capabilities` to tell other parts of the VAMDC infrastructure which keys are understood by the data node in question.
The `sync` endpoint receives and answers queries. All this view needs to do is to call the query validation, the node’s query function and then pass its result to the XSAMS generator. Note that due to the *lazy evaluation* of both the query and the generator, the main queries have not yet run at this point and no XML is yet held in memory. Only when the XSAMS generator is passed to Django’s `StreamingHTTPResponse` does the whole machinery start churning and gradually producing the XSAMS document for the client. XSAMS documents can be hundreds of MB in size, this method has proven superiour for our needs, compared to early attempts with Djangos templates or XML libraries.
The view-function behind `sync` also needs to distinguish between GET and HEAD requests, the latter allowing for returning statistics. In addition, incoming OPTION requests are answered according to CORS to allow cross-site requests (cf. Section \[sec:api\]). Thanks to Django’s high-level tools for handling requests, the whole API implementation is only about 150 lines of code and therefore easy to understand and maintain.
Deployment {#sec:deployment}
----------
A data node’s performance depends to a large extent on the underlying database itself. However, the assembly of large XML documents can also become computationally expensive, not the least because the XSAMS generator is generalized for use by all nodes. Instead of spending effort on optimization, it has proven beneficial for several nodes to replace the standard Python interpreter by `PyPy`[^9], an alternative interpreter employing just-in-time compilation for large gains in speed. Django and the node software run nicely with PyPy, including the recent PyPy3.
The node software must be run by an application server able to serve data over HTTP. Since data providers generally host their nodes themselves, the setup should be light weight and easy to manage. We recommend to use `gunicorn`[^10], an application server with minimal configuration. It is also written in Python and a common choice for deploying Django apps. Alternatives include Apache `mod_wsgi` and `uWSGI`. Nodes frequently choose set-ups that deviate from the recommendation, for example due to integration with other services on the same machine or local expertise with certain tools.
Furthermore, it is good practice to put a proxy between the application and the client. Its most important task is to offer `gzip` compression to clients (XML is highly compressible). The proxy can also do load balancing and provide HTTPS/TLS. The most common and recommended choice of proxy is `Nginx`[^11] but again, data providers often choose whatever webserver they have running already. `Caddy`[^12] is a relatively new webserver that can act as proxy, do compression and automatically handle free certificates for HTTPS, with minimal configuration effort.
A data node in practice: VALD {#sec:vald}
=============================
The Vienna Atomic Line Database (VALD, @vald2011) stores atomic and molecular transition data of astronomical interest[^13]. This section describes the process and experiences of using the node software to adapt VALD into a VAMDC data node.
VALD originally uses a custom binary storage method. Since the original VALD database will continue operation alongside its VAMDC counterpart, this data was extracted using VALD’s own tools into a series of text files:
- The species and their properties (charge, composition, ionization energy, isotopic fractions, InCHi keys designation etc. Each record of this file also contain a unique species number.
- A list of transitions and the corresponding energy levels, sorted by wavelength. This file has two lines per transition containing transition probability, line broadenings and the upper and lower level description (energy, Landé factor, coupling model, term designations etc).
- A list of the quantum numbers for each level in the same order as the previous file (J, L, S, parity etc).
- A list of the custom reference-ID strings used in VALD, tied to a number identifier.
- A list of all data collections included in VALD identified by an id string. Often such compilation has an associated critical evaluation paper that contains valuable quality analysis work.
- A BibTeX file linking VALD’s reference-ID strings to actual Bibtex entries.
Note that no less than three different files are involved in tracking references making the import process quite involved.
The next step was to prepare the VAMDC database. As discussed in Section \[sec:publishing\_tools\], this is done using Django **models**, each representing a database table in the database. In the VALD case, the most relevant models are:
- `Species`, containing all species data like mass number, ionization energy and identifers. This model also holds a many-to-many relationship to other species if this is a part of a molecule.
- `State` describes all atomic or molecular states and references the Species involved, the energy, Lande factor and quantum mechanical properties. Also the literature sources for the main properties are referenced.
- `Transition` contains the wavelength for each Transition alongside references to the States and Species involved. It also refers to all sources used for the data and which line list the transition data comes from.
- `Linelist` represents the line lists and sources used for every Species. It also designates if the data was obtained through empirical observations, calculations etc.
- `Reference` stores the actual bibtex references for all data in the database.
This data model was then used by Django’s `migrate`-mechanism to create the tables in an empty database. With the database schema in place we now used the publishing tools’ mapping system to map the raw data files of VALD to the VAMDC schema. Due to the layout of the raw data files the mapper was set to read several files simultaneously and collate the lines (or groups of lines) from each file to build an output record. The first pass created the records and the second established the relationships between the tables. For efficiency, a separate run created the many-to-many relationships as extra intermediary data files.
An important re-organization of data was carried out at this stage. The original data inlines upper- and lower states as part of the transition record. The XSAMS data model instead separates states and transitions and lets the latter refer to the former. We therefore needed to de-duplicate states and create several output files from the same input file. The relation between the states and transitions also had to be preserved by generating unique keys tying them together.
Since VALD is a large database with 250 million lines, this conversion process is computationally quite heavy. Originally the idea was to have the node software mapper create Django database instances directly. This turned out to be both slow and memory consuming (since a full Python object must be created for each record). Creating intermediary text files that match the database schema (cf. Section \[sec:publishing\_tools\]) turned out to be considerably faster and is now a part of the standard workflow. Furthermore we found that the import process benefited greatly from the use of the PyPy just-in-time compiling Python interpreter; we saw a speedup of up to a factor five. Consequently we kept using PyPy also for the node’s normal operation and recommend PyPy in the node software manual.
At this point the mapping tool had created a series of large text files, one for each database table. These files were read directly into the MySQL database using MySQL’s own load mechanism. This is fast as long as the database indices are not in place yet. Django can automatically create indices for certain fields but we chose to defer this until after loading the data. Then we also created additional database indices to speed up the most common types of queries.
With the database ready, we proceed with preparing the dictionaries to enable communication with VAMDC. The restrictables dictionary looks like this:
"AtomSymbol": "species__name",
"AtomMassNumber": "species__massno",
...
"Lower.StateEnergy": "lostate__energy"
"Upper.StateEnergy": "upstate__energy"
...
This allows the node software to convert incoming VSS2 request into a Django Q-object. On the left hand side are the standard VAMDC restrictable keys. The right hand side contains the corresponding model field name in Django format, using double-underscores `__` to denote membership. In the example above, a request for `AtomSymbol` data leads to querying the `name` field of the `species` table.
Next is the query function. As decscribed above, it receives the query and starts by converting the SQL into a Django Q-object. Because VALD is primarily a transition line list, our query strategy is to start from the `transitions` model and pass to it the Q-object, resulting in the (not evaluated) QuerySet for all matching transitions.
Reconstructing the QuerySets for states and references containing only the records needed for the selected transistions turned out to be slow for this particular database. So in order to prematureley evaluate the transitions-QuerySet we opted for a customized generator instead; The VALD node overrides the main function of the XML generator. Our custom version is a variant of the standard one. It uses all the sub-functions creating the XML but has one important optimization: It starts outputting the transition data right away, *before* figuring out which species, states and references will be needed to make the data complete. During the loop over transitions, we collect the ID-keys in Python `set`s. These sets are then used in simple and fast database queries to retrieve the complementary data. For queries close to the size limit this reduced the response time of the VALD node by up to a factor ten.
A drawback of this approach is that the VALD node no longer knows how much data it will include in a given request and cannot include statistics in the GET HTTP headers. To get the statistics headers from VALD an explicit HEAD request is needed. This was deemed to be an acceptable compromise for the increase in speed.
The VALD `returnables` dictionary used by the XSAMS generator to populate the XML looks like follows, making use of the full range of options described in Section \[sec:xsams\_generator\]:
...
"AtomStateID": "AtomState.id",
"AtomSymbol": "Atom.name",
...
"RadtransSpeciesRef": RadTran.get_wave_refs()",
"RadtransWavelengthUnit": "A",
"RadtransUpperStateRef": "Radtran.upstate_id",
"RadtransLowerStateRef": "Radtran.lostate_id",
...
The VAMDC keys (on the lefthand side) tell the XSAMS generator where in XSAMS to place a certain value. The dictioanry values are the constants, model fields or methods returning the content to insert. Note that we make use of Django’s features to access inter-table foreign keys through `_id` and use a custom model method for the non-trivial way that VALD handles references.
After this, the VALD node was operational and could be registered with the VAMDC network. An important aspect of the setup is that it’s easily repeatable and scriptable: When a new line list is added or a correction is done to VALD one can just re-run the process and get a VAMDC version with the latest update.
Concluding remarks {#sec:concl}
==================
We have in some detail described the VAMDC Standards that concern data nodes, and the implementation of these protocols in practice. The software package that has been the topic of this paper strikes an illustrative balance between diverging requirements from different research groups and the goal to provide common tools that make common tasks easy. It is also a case where a successful web framework like Django becomes highly useful in science.
In summary, the node software carries out the tasks of
- Helping data providers importing existing data to a standard relational database schema, starting from almost arbitrary custom storage formats.
- Making it easy to repeat the process when set up once. Such situations occur when the data is updated/expanded or a new type of data is added to a node.
- Translating the VAMDC standard query language (VSS2) to the internal names/fields/units conventions at a given node.
- Converting the data from the node’s database to the VAMDC common output format (XSAMS).
- Providing the API access needed to make a data node operate within the VAMDC infrastructure, including registry information.
Since its inception, the node software has been developed by a small team of VAMDC consortium members who each are in charge of a data node. This means that developers and users overlap significantly and that flexibility and easy modification were always prioritized over user-friendliness. One can argue that “end users” wanting to use VAMDC to retrieve data never need to use or understand the node software; it is used by data providers only. Nevertheless, significant effort has been spent into documentation, both in text and screencast video form, allowing new data nodes to be set up without significant input or help from the node software developers.
Since its initial launch, there have been several node software releases, marked with Git(Hub) tags. These follow the evolution of the VAMDC standards and have also often offered gradual but significant improvements driven by feedback from the data provider community or new requirements by an individual data node. These custom changes are merged into the main version control branch which is today considered a rolling beta release. The VAMDC standards themselves are stable and currently only evolve in a backwards-compatible manner. This means that unless a data node wants to make use of the latest features, it needs not upgrade in order to stay interoperable with the surrounding infrastructure. This fits well with the reality of data nodes; they are often run with very little manpower for oversight, maintenance and upgrades. To our knowledge this has not yet presented any problems, neither in terms of interoperability nor in security. Nevertheless we are currently preparing a release based on the latest long-term-support version of Django (1.11, April 2017). After release we will encourage all data nodes to upgrade the node software and the underlying stack to the latest version.
Acknowledgements {#sec:acknowledements}
================
The authors acknowledge the contributions, testing efforts and feedback by many VAMDC collaborators that have benefited the node software tremendously. In particular Guy Rixon, Nicolas Moreau, Christian Hill, Christan Endres, Michail Doronin, Yaye-Awa Ba, and Johannes Postler.
Appendix: XSAMS example {#appendix-xsams-example .unnumbered}
=======================
This is an annotated example of an XSAMS document (See Section \[sec:output\_language\_xsams\]) returned from a request of Ca lines between 5000.0 and 5000.1Å. The return data is from the VALD database node. The output was shortened in places for brevity.
<?xml version="1.0" encoding="UTF-8"?>
<XSAMSData xmlns="http://vamdc.org/xml/xsams/1.0"
xmlns:cml="http://www.xml-cml.org/schema"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://vamdc.org/xml/xsams/1.0">
<Processes>
<Radiative>
<RadiativeTransition id="Pvald-R571145" process="excitation">
This return concerns a radiative transition, which is identified using an ID internally consistent within this XSAMS document. Other parts of the document can then reference this as needed.
<EnergyWavelength>
<Wavelength>
<Comments>Vacuum wavelength from state energies (RITZ)</Comments>
<SourceRef>Bvald-K07</SourceRef>
<Value units="A">5000.04125034</Value>
</Wavelength>
</EnergyWavelength>
The wavelength and its unit is given. It *could* reference a `Method` to show it is a calculated result but in VALD’s case this is instead noted as a comment and the reference to the source is given (later in this document).
<UpperStateRef>Svald-26696</UpperStateRef>
<LowerStateRef>Svald-44769</LowerStateRef>
<SpeciesRef>Xvald-191</SpeciesRef>
<Probability>
<Log10WeightedOscillatorStrength>
<SourceRef>Bvald-K07</SourceRef>
<Value units="unitless">-3.607</Value>
</Log10WeightedOscillatorStrength>
</Probability>
Similarly, the transition will reference other parts of the XML document (the `AtomicState`) for the information about the upper/lower states involved in this transition. The transition probability is given with values, units and reference.
<ProcessClass />
<Broadening name="natural" envRef="Evald-natural">
<Comments>Natural Broadening</Comments>
<SourceRef>Bvald-K07</SourceRef>
<Lineshape name="lorentzian">
<LineshapeParameter name="log(gamma)">
<Value units="1/s">7.24</Value>
</LineshapeParameter>
</Lineshape>
</Broadening>
Broadening is given with value and unit as well as with references to sources and to the more detailed environment block later in the document. This exemplifies natural broadening, other broadening types (such as Van der Walls and Stark broadening) would follow (not shown here).
</RadiativeTransition>
</Radiative>
</Processes>
<Species>
<Atoms>
<Atom>
<ChemicalElement>
<NuclearCharge>20</NuclearCharge>
<ElementSymbol>Ca</ElementSymbol>
</ChemicalElement>
<Isotope>
<IsotopeParameters>
<MassNumber>40</MassNumber>
</IsotopeParameters>
<Ion speciesID="Xvald-191">
This describes the species (Isotope) that `RadiativeTransition` referenced above.
<IonCharge>0</IonCharge>
<AtomicState stateID="Svald-26696">
<Description>3p6.3d.7f 3P*</Description>
<AtomicNumericalData>
<StateEnergy>
<SourceRef>Bvald-K07</SourceRef>
<Value units="1/cm">60690.2700</Value>
</StateEnergy>
<LandeFactor>
<SourceRef>Bvald-K07</SourceRef>
<Value units="unitless">1.41</Value>
</LandeFactor>
</AtomicNumericalData>
<AtomicQuantumNumbers>
<Parity>odd</Parity>
<TotalAngularMomentum>1.0</TotalAngularMomentum>
</AtomicQuantumNumbers>
<AtomicComposition>
<Component>
<Configuration>
<AtomicCore>
<Term />
</AtomicCore>
</Configuration>
<Term>
<LS>
<L>
<Value>1</Value>
</L>
<S>1.0</S>
</LS>
</Term>
</Component>
</AtomicComposition>
</AtomicState>
The entirety of the Ion State `3p64d.7f 3P*` is described, along with its quantummechanical data. The measurements of its energy and landé factor both come with references to which sources they are coming from. Note that more States would follow after this one.
<InChI>InChI=1S/Ca</InChI>
<InChIKey>OYPRJOBELJOOCE-UHFFFAOYSA-N</InChIKey>
</Ion>
</Isotope>
</Atom>
</Atoms>
</Species>
<Sources>
<Source sourceID="Bvald-2017-02-20-14-58-15">
<Comments>
This Source is a self-reference. It represents the database
and the query that produced the xml document. The sourceID
contains a timestamp. The full URL is given in the tag
UniformResourceIdentifier but you need to unescape ampersands
and angle brackets to re-use it.
Query was: select * where
(RadTransWavelength
&gt;= 5000.0 AND RadTransWavelength
&lt;= 5000.1) AND ((AtomSymbol = 'Ca'))
</Comments>
<Year>2017</Year>
<Category>database</Category>
<UniformResourceIdentifier>
http://vald.inasan.ru/vald-node/tap/sync?LANG=VSS2&
REQUEST=doQuery&
FORMAT=XSAMS&QUERY=select+*+where+%28RadTransWavelength+
%3E%3D+5000.0+AND+RadTransWavelength+%3C%3D+5000.1%29+
AND+%28%28AtomSymbol+%3D+%27Ca%27%29%29
</UniformResourceIdentifier>
<ProductionDate>2017-02-20</ProductionDate>
<Authors>
<Author>
<Name>N.N.</Name>
</Author>
</Authors>
</Source>
The node software will always return a “self reference” as part of the XSAMS document. This details which query produced this result and is intended primarily for the end user to be able to reproduce their query. It could also potentially be used for bug reporting.
<Source sourceID="Bvald-K07">
<Authors>
<Author><Name>R. L. Kurucz</Name> </Author>
</Authors>
<Title>Robert L. Kurucz on-line database of observed and
predicted atomic transitions</Title>
<Category>private communication</Category>
<Year>2007</Year>
<SourceName>unknown</SourceName>
<Volume></Volume>
<PageBegin></PageBegin>
<PageEnd></PageEnd>
<UniformResourceIdentifier>http://kurucz.harvard.edu/atoms/1400/gfemq1400.pos <!--...--></DigitalObjectIdentifier>
<BibTeX>@misc{K07, Author = {{Kurucz}, R.~L.}, Title = {Robert L. Kurucz on-line
database of observed and predicted atomic transitions}, year = 2007,
Bdsk-Url-1 = {http://kurucz.harvard.edu/atoms/1400/gfemq1400.pos, ...}}
</BibTeX>
</Source>
The source reference contains all necessary components for properly crediting the data, including an optional BibTex field. Note that parts of this entry has been cropped.
</Sources>
<Methods>
<Method methodID="Mvald-0">
<Category>experiment</Category>
<Description>VALD exp - transition between levels with
experimentally known energies</Description>
</Method>
This is an example of the node providing more information than necessary – this method is not referenced from anywhere in the XSAMS document but is still included (other methods have been cut out to shorten the listing).
</Methods>
<Functions>
<Function functionID="Fvald-stark">
<Name>Stark Broadening</Name>
<Expression computerLanguage="Fortran">
gammawaal * (T / 10000.0) ** (1.0/6.0) * N</Expression>
<Y name="gammaL" units="1/cm3/s" />
<Arguments>
<Argument name="T" units="K">
<Description>The absolute temperature, in K</Description>
</Argument>
<Argument name="N" units="1/cm3">
<Description>Number density of neutral perturbers</Description>
</Argument>
</Arguments>
<Parameters>
<Parameter name="gammawaal" units="1/cm3/s">
<Description>Lorentzian FWHM of the line</Description>
</Parameter>
</Parameters>
<Description>This function gives the temperature dependence of
Stark broadening.</Description>
</Function>
The function block can offer functional code in computer language or other representation. Multiple other functions would follow this one.
<Environments>
<Environment envID="Evald-stark">
<Comments>A given gamma can be scaled with
gamma = gamma_given * (T / T_ref)^1/6
* number density of free electrons.</Comments>
<Temperature>
<Value units="K">1.0E4</Value>
</Temperature>
<TotalNumberDensity>
<Comments>The broadening parameters are given in
Hz per number density (i.e. 1/cm3/s), so they can simply
be scaled with the number density. Note that unless otherwise noted,
log10(gamma) is given.</Comments>
<Value units="1/cm3/s">1</Value>
</TotalNumberDensity>
</Environment>
Again, additional environment sections were cut for brevity.
</Environments>
</XSAMSData>
[^1]: <http://portal.vamdc.eu/>
[^2]: <http://dictionary.vamdc.eu/>
[^3]: https://www.w3.org/TR/cors
[^4]: <https://github.com/VAMDC/NodeSoftware>
[^5]: <https://www.djangoproject.com>
[^6]: <http://cdms.berkeley.edu>
[^7]: <https://www.cfa.harvard.edu/hitran>
[^8]: This is the node software’s only additional dependency apart from Django and the Python driver for the database engine in question.
[^9]: <http://pypy.org>
[^10]: <http://gunicorn.org>
[^11]: <http://nginx.org/en>
[^12]: <https://caddyserver.com>
[^13]: <http://vald.astro.uu.se>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report 850 $\mu$m dust polarization observations of a low-mass ($\sim$12 $M_{\odot}$) starless core in the $\rho$ Ophiuchus cloud, Ophiuchus C, made with the POL-2 instrument on the James Clerk Maxwell Telescope (JCMT) as part of the JCMT B-fields In STar-forming Region Observations (BISTRO) survey. We detect an ordered magnetic field projected on the plane of sky in the starless core. The magnetic field across the $\sim$0.1 pc core shows a predominant northeast-southwest orientation centering between $\sim$40$\degr$ to $\sim$100$\degr$, indicating that the field in the core is well aligned with the magnetic field in lower-density regions of the cloud probed by near-infrared observations and also the cloud-scale magnetic field traced by Planck observations. The polarization percentage ($P$) decreases with an increasing total intensity ($I$) with a power-law index of $-$1.03 $\pm$ 0.05. We estimate the plane-of-sky field strength ($B_{\mathrm{pos}}$) using modified Davis-Chandrasekhar-Fermi (DCF) methods based on structure function (SF), auto-correlation (ACF), and unsharp masking (UM) analyses. We find that the estimates from the SF, ACF, and UM methods yield strengths of 103 $\pm$ 46 $\mu$G, 136 $\pm$ 69 $\mu$G, and 213 $\pm$ 115 $\mu$G, respectively. Our calculations suggest that the Ophiuchus C core is near magnetically critical or slightly magnetically supercritical (i.e. unstable to collapse). The total magnetic energy calculated from the SF method is comparable to the turbulent energy in Ophiuchus C, while the ACF method and the UM method only set upper limits for the total magnetic energy because of large uncertainties.'
author:
- Junhao Liu
- Keping Qiu
- David Berry
- James Di Francesco
- Pierre Bastien
- 'Patrick M. Koch'
- 'Ray S. Furuya'
- 'Kee-Tae Kim'
- Simon Coudé
- Chang Won Lee
- Archana Soam
- Chakali Eswaraiah
- Di Li
- Jihye Hwang
- 'A-Ran Lyo'
- Kate Pattle
- Tetsuo Hasegawa
- Woojin Kwon
- 'Shih-Ping Lai'
- 'Derek Ward-Thompson'
- 'Tao-Chung Ching'
- Zhiwei Chen
- Qilao Gu
- Dalei Li
- 'Hua-bai Li'
- 'Hong-Li Liu'
- Lei Qian
- Hongchi Wang
- Jinghua Yuan
- 'Chuan-Peng Zhang'
- Guoyin Zhang
- 'Ya-Peng Zhang'
- Jianjun Zhou
- Lei Zhu
- Philippe André
- Doris Arzoumanian
- Yusuke Aso
- 'Do-Young Byun'
- 'Michael Chun-Yuan Chen'
- 'Huei-Ru Vivien Chen'
- Wen Ping Chen
- Jungyeon Cho
- Minho Choi
- Antonio Chrysostomou
- Eun Jung Chung
- Yasuo Doi
- 'Emily Drabek-Maunder'
- 'C. Darren Dowell'
- 'Stewart P. S. Eyres'
- Sam Falle
- Lapo Fanciullo
- Jason Fiege
- Erica Franzmann
- Per Friberg
- 'Rachel K. Friesen'
- Gary Fuller
- Tim Gledhill
- 'Sarah F. Graves'
- 'Jane S. Greaves'
- 'Matt J. Griffin'
- Ilseung Han
- Jennifer Hatchell
- 'Saeko S. Hayashi'
- Thiem Hoang
- Wayne Holland
- Martin Houde
- Tsuyoshi Inoue
- 'Shu-ichiro Inutsuka'
- Kazunari Iwasaki
- 'Il-Gyo Jeong'
- Doug Johnstone
- Yoshihiro Kanamori
- 'Ji-hyun Kang'
- Miju Kang
- 'Sung-ju Kang'
- Akimasa Kataoka
- 'Koji S. Kawabata'
- Francisca Kemper
- Gwanjeong Kim
- Jongsoo Kim
- Kyoung Hee Kim
- 'Mi-Ryang Kim'
- Shinyoung Kim
- 'Jason M. Kirk'
- 'Masato I. N. Kobayashi'
- Takayoshi Kusune
- Jungmi Kwon
- 'Kevin M. Lacaille'
- 'Chin-Fei Lee'
- 'Jeong-Eun Lee'
- Hyeseung Lee
- 'Sang-Sung Lee'
- 'Sheng-Yuan Liu'
- Tie Liu
- Sven van Loo
- Steve Mairs
- Masafumi Matsumura
- 'Brenda C. Matthews'
- 'Gerald H. Moriarty-Schieven'
- Tetsuya Nagata
- Fumitaka Nakamura
- Hiroyuki Nakanishi
- Nagayoshi Ohashi
- Takashi Onaka
- Josh Parker
- Harriet Parsons
- Enzo Pascale
- Nicolas Peretto
- Andy Pon
- 'Tae-Soo Pyo'
- Ramprasad Rao
- 'Mark G. Rawlings'
- Brendan Retter
- John Richer
- Andrew Rigby
- 'Jean-François Robitaille'
- Sarah Sadavoy
- Hiro Saito
- Giorgio Savini
- 'Anna M. M. Scaife'
- Masumichi Seta
- Hiroko Shinnaga
- Motohide Tamura
- 'Ya-Wen Tang'
- Kohji Tomisaka
- Yusuke Tsukamoto
- 'Jia-Wei Wang'
- 'Anthony P. Whitworth'
- 'Hsi-Wei Yen'
- Hyunju Yoo
- Tetsuya Zenko
title: 'The JCMT BISTRO Survey: The Magnetic Field In The Starless Core $\rho$ Ophiuchus C'
---
introduction
============
The role of magnetic fields (B-fields) has long been a hot topic under debates in the star formation studies . There are two major classes of star-formation theories that significantly differ in the role played by magnetic fields. “Strong magnetic field models” suggest that molecular clouds are supported by magnetic fields, which quasi-statically dissipate via ambipolar diffusion. Eventually self-gravity overcomes the magnetic force, inducing the collapse of molecular cloud cores and the formation of stars [@2006ApJ...646.1043M]. In contrast, “weak field models” suggest that turbulent flows, instead of magnetic fields, dominate the evolution of molecular clouds and create overdense regions where stars form [@2004RvMP...76..125M]. Recently, results from simulations indicate that magnetic field and turbulence are both essential to provide support against gravitational collapse [@2014prpl.conf...77P and references therein]. Observational studies of magnetic fields in star-forming regions can directly test these theoretical models, providing deep insights into the relative importance of magnetic fields and gravity/turbulence in cloud evolution and star formation.
Observing the polarized emission of dust grains and the polarization of background stars is one of powerful ways to investigate the plane-of-sky magnetic field structure in star-forming regions [@1988QJRAS..29..327H]. The starlight polarization was first discovered by @1949Sci...109..165H and @1949Sci...109..166H. Later on, the observed polarization of starlight was explained by the partial extinction of starlight by magnetically aligned dust grains , where the short axes of spinning dust grains align with magnetic field lines. This explaination is widely accepted. There are many theories trying to explain why dust grains are aligned with magnetic fields. Among them, the Radiative Alighment Torque (RAT) theory is most accepted [@2007JQSRT.106..225L]. Although the detail of the the gain alignment mechanism is still unclear, the plane-of-sky magnetic field structure in star formation regions has been successfully traced using polarization observations . Polarization observations at near-infrared (NIR) wavelengths, which are expected to trace polarization produced by dust extinction of background starlight, are often used to investigate the magnetic filed structure in dense molecular regions [@2014ApJ...783....1S; @2015ApJS..220...17K]. However, NIR polarization observations are not sufficient to trace the magnetic field in regions with high extinction or associated with few background stars. Polarization observations at sub-millimeter (sub-mm) wavelengths, which trace dust thermal emission, are essential to overcome the drawback of NIR polarization observations and to probe the magnetic field structure in denser enviroments such as filaments and dense cores.
Among an increasing number of polarization observations toward low-mass star formation regions, studies of the protostellar phase of Young Stellar Objects (YSO) have attracted most interests. One important approach of these studies is to find hourglass-shaped magnetic fields. As predicted by the theoretical model and simulations [@1993ApJ...417..220G; @1993ApJ...417..243G], the magnetic field of magnetically dominated dense regions is expected to show an hourglass shape in the collapse phase. At 0.001-0.01 pc scales, dust polarization observations towards low-mass protostellar systems have revealed the expected hourglass-shaped field morphologies [@2006Sci...313..812G; @2009ApJ...707..921R; @2013ApJ...769L..15S]. More chaotic field morphologies, which are expected in weakly magnetic environment or probably affected by stellar feedback or complex geometry, are also reported at this scale [@2014ApJS..213...13H; @2017ApJ...842L...9H]. At larger scales (0.01-0.1 pc), hints of the hourglass shape are less obvious [@2009ApJS..182..143M; @2010ApJS..186..406D; @2014ApJS..213...13H] and the role of magnetic field at this scale is comparatively less understood.
Since the magnetic field in protostellar cores can suffer from feedback by star-forming activities, polarization studies of starless cores are essential to help us understand the role of magnetic field in the early stages of star formation. The relatively weak polarized dust emission in starless cores, however, is far more difficult to detect and, as a result, there are only a handful of dust polarization observations toward starless cores . The role of magnetic fields in the initial phase of star formation remains an open question.
The Ophiuchus molecular cloud is a low-mass star-forming region located at a distance of $\sim$137 pc [@2017ApJ...834..141O]. It is one of the nearest star formation regions and has been widely studied [@2008hsf2.book..351W and references therein]. Star formation in this cloud is heavily influenced by compression of expanding shock shells from the nearby Sco-Cen OB association [@1977AJ.....82..198V]. A detailed DCO$^+$ emission study has identified several dense cores, Ophiuchus A to Ophiuchus F (hereafter Oph-A to Oph-F), in the main body of Ophiuchus [@1990ApJ...365..269L]. Among these dense cores, our target, Oph-C, which harbors no embedded protostars [@2009ApJ...692..973E] and is not associated with Herschel 70 $\mu$m emission [@2015MNRAS.450.1094P], appears to be the least evolved and is extremely quiescent. The 850 $\mu$m continuum of Oph-C was observed as part of the James Clerk Maxwell Telescope (JCMT) Gould Belt Survey (GBS) [@2007PASP..119..855W; @2015MNRAS.450.1094P]. @2015MNRAS.450.1094P identified a few low-mass, pressure-confined, virially bound, and $\sim$0.01 pc-scale ($\sim$ 3000 AU) sub-cores in Oph-C based on the GBS data. The 850 $\mu$m polarization data of the Ophiuchus cloud obtained using SCUPOL, the previous JCMT polarimeter, was catalogued by @2009ApJS..182..143M. More recently, the large-scale plane-of-sky magnetic field map of the Ophiuchus cloud was presented by at 5$\arcmin$ resolution as part of the Planck project. @2015ApJS..220...17K conducted NIR polarimetry of Ophiuchus cloud and suggested the magnetic field structures in the cloud may have been influenced by the nearby Sco-Cen OB association.
Here we present 850 $\mu$m dust polarization observations using the POL-2 polarimeter in combination with the Summillimetre Common-User Bolometer Array 2 (SCUBA-2) on the JCMT toward the Oph-C region as part of the B-fields In STar-forming Region Observations (BISTRO) survey [@2017ApJ...842...66W]. The BISTRO survey is aimed at using POL-2 to map the polarized dust emission in the densest parts of all of the Gould Belt star-forming regions including Orion A [@2017ApJ...846..122P], Oph-A [@2018ApJ...859....4K], M16 [@2018ApJ...860L...6P], Oph-B [@2018ApJ...861...65S], and several other regions (papers in preparation). With the unique resolution offered by JCMT, which can resolve the magnetic field structures down to scales of $\sim$1000 AU in nearby star formation regions, these POL-2 observations are crucial to test theoretical models of star formation at an intermediate scale and to generate a large sample of polarization maps of dense cores obtained in a uniform and consistent way for statistical studies [@2017ApJ...842...66W]. The B-field structures traced by POL-2 agree well with those traced by the previous SCUPOL observations, but the POL-2 maps are more sensitive than the previous SCUPOL data and trace larger areas [e.g., @2018ApJ...859....4K; @2018ApJ...861...65S].
This paper is organized as follows: in section 2, we describe the observations and data reduction; in section 3, we present the results of the observations and derive the B-field strength; in section 4, we discuss our results; and section 5 is given for a summary of this paper..
Observations
============
The polarized emission of Oph-C was observed at 850 $\mu$m with SCUBA-2 [@2013MNRAS.430.2513H] along with POL-2 [@2016SPIE.9914E..03F, Bastien et al. in prep.] between 2016 May 22 and September 10. The region was observed 20 times, among which 19 datasets had an average integration time of 42 minutes and 1 bad dataset was excluded. The observations were made with the POL-2 DAISY mode, which produces a map with high signal-to-noise ratio at the central region of 3$\arcmin$ radius and with increasing noise to the edge. The effective beam size of JCMT is 14.1$\arcsec$ ($\sim$9 mpc at 137 pc) at 850 $\mu$m.
The data were reduced using the SMURF [@2013ascl.soft10007J] package in Starlink [@2014ASPC..485..391C]. Firstly, the *calcqu* command is used to convert the raw bolometer timestreams into separate Stokes $I$, $Q$, and $U$ timestreams. Then, the *makemap* routine in the *pol2map* script creates individual $I$ maps from the $I$ timestreams of each observation, and coadds them to produce an initial reference $I$ map. Secondly, the *pol2map* is re-run with the initial $I$ map to generate an ASTMASK, which is used to define the signal-to-noise-based background regions that are set to zero until the last iteration, and a PCAMASK, which defines the source regions that are excluded when creating the background models within *makemap*. With the ASTMASK and the PCAMASK, *pol2map* is again re-run to reduce the previously created $I$ timestreams of each observation, creating improved $I$ maps, These individual improved $I$ maps are then coadded to produce a final improved $I$ map. Finally, with the same masks, *pol2map* creates the $Q$ and $U$ maps, along with their variance maps, and the debiased polarization catalogue, from the $Q$ and $U$ timestreams. The final improved $I$ map is used for instrumental polarization correction. The final $I$, $Q$, $U$ maps and the polarization catalogue are gridded to 7$\arcsec$ pixels for a Nyquist sampling.
The absolute calibration is performed by applying a flux conversion factor (FCF) of 725 Jy beam$^{-1}$ pW$^{-1}$ to the output $I$, $Q$, and $U$ maps, converting the units of these maps from pW to Jy beam$^{-1}$. Due to the additional losses from POL-2, this FCF is 1.35 times larger than the standard SCUBA-2 FCF of 537 Jy beam$^{-1}$ pW$^{-1}$ [@2013MNRAS.430.2534D]. The uncertainty on the flux calibration is 5% [@2013MNRAS.430.2534D]. Figure \[fig:figiqu\] shows the $Q$, $U$, and $I$ maps of our POL-2 data. The rms noises of the background regions in the $Q$ or $U$ maps are $\sim$3.5 mJy beam$^{-1}$. From the corresponding variance maps, the average $Q$ or $U$ variances are $\sim$2 mJy beam$^{-1}$, reaching the target sensitivity value for the BISTRO survey.
Figure \[fig:figdiff\] shows the total intensity map toward the same region made with SCUBA-2 as part of the Gould Belt Survey (GBS) project [@2015MNRAS.450.1094P] and the difference between the SCUBA-2 $I$ map and the POL-2 $I$ map. Because of the difference in the data reduction procedures of the POL-2 data and the SCUBA-2 data and the slower scanning speed of the POL-2 observation, the large-scale structures seen in the SCUBA-2 $I$ map are suppressed in the POL-2 $I$ map. So the BISTRO $I$ map is much fainter than the GBS $I$ map.
Because of the uncertainties in the $Q$ $\&$ $U$ values and that the polarized intensity and polarized percentage are defined as positive values, the measured polarized intensities are biased toward larger values [@2006PASP..118.1340V]. The debiased polarized intensity and its corresponding uncertainty are calculated as: $$PI = \sqrt{Q^2 + U^2 - 0.5(\delta Q^2 + \delta U^2)},$$ and $$\delta PI = \sqrt{\frac{(Q^2 \delta Q^2 + U^2 \delta U^2)}{(Q^2 + U^2)}},$$ where $PI$ is the polarized intensity, $\delta Q$ the uncertainty of $Q$, and $\delta U$ is the uncertainty of $U$. The debiased polarization percentage, $P$ and its uncertainty $\delta P$ are therefore derived by: $$P = \frac{PI}{I},$$ and $$\delta P = \sqrt{(\frac{\delta PI^2}{I^2} + \frac{\delta I^2 (Q^2 + U^2)}{I^4})},$$ where $\delta I$ is the uncertainty of the total intensity.
Finally, the polarization position angle $\theta$ and its uncertainty $\delta \theta$ are estimated to be: $$\theta = \frac{1}{2} \tan ^{-1} (\frac{U}{Q}),$$ and $$\delta \theta = \frac{1}{2} \sqrt{\frac{(Q^2 \delta U^2 + U^2 \delta Q^2)}{(Q^2 + U^2)^2}},$$
Results
=======
The magnetic field morphology in the Oph-C region
-------------------------------------------------
Assuming that the shortest axis of dust grains is perfectly aligned with the magnetic field, we can derive the orientation of the magnetic field projected on the plane of the sky by rotating the observed polarization vectors by 90$\degr$. Figure \[fig:figmap\] shows the B-field segments of our POL-2 observations. These segments have lengths proportional to the polarization degrees and orientations along the local B-field. Note that our POL-2 segments are Nyquist sampled with a pixel size of $7''$. With these criteria, the vectors in the Oph-C region are well separated from those in other dense regions of the Ophiuchus cloud. The magnetic field orientations do not appear random, and have a predominant northeast-southwest orientation.
Figure \[fig:figcom\] compares the magnetic field orientations from our POL-2 data with previous observations with the older JCMT polarimeter [SCUPOL, @2009ApJS..182..143M]. We use criteria of $P/\delta P>2$ and $\delta P<4$% for both the POL-2 data and the SCUPOL data. Compared to the previous SCUPOL observations, our POL-2 observations show significant improvements by detecting dust polarization over a much larger area and toward the center of the core.
![Position angle (after 90$\degr$ rotation) histogram for $B$-field vectors with $P/\delta P>2$ and $\delta P<4$%. The bin size is 10$\degr$. The POL-2 vectors are shown in black. The SCUPOL vectors are shown in red. Angles are measured east of north. \[fig:figang\]](./ang.pdf)
In Figure \[fig:figang\], histograms of the position angles of the B-field segments from the POL-2 data and the SCUPOL data are shown. The POL-2 histogram has a broad peak between $\sim$40$\degr$ to $\sim$100$\degr$. The standard deviation of the position angles of these POL-2 vectors is $\sim$33$\degr$. The SCUPOL vectors are randomly distributed, which is inconsistent with the POL-2 vectors.
![Comparision of position angles (after 90$\degr$ rotation) for overlaping SCUPOL and POL-2 vectors with $P/\delta P>2$ and $\delta P<4$%. Data points correspond to POL-2 data with $P/\delta P>3$ are marked in blue. Angles are measured east of north. \[fig:figangang\]](./ang_ang.pdf)
To further compare the SCUPOL data and the POL-2 data, we resampled the POL-2 data to the same pixel size as that of the SCUPOL data (10$\arcsec$) and aligned the World Co-ordinate System of the two data sets. We found 31 pairs of spatially overlapping vectors between the two data sets with vector selection criteria of $P/\delta P>2$ and $\delta P<4$%. Figure \[fig:figangang\] shows the comparision of position angles (after 90$\degr$ rotation) for these overlapping vectors. Large angular differences in the position angles of overlapping vectors can be seen in this figure. The average angular difference of overlapping vectors is estimated to be $\sim$39$\degr$. We further computed the Kolmogorov-Smirnov (K-S) statistic on the POL-2 and SCUPOL position angles and found a probability of 0.06, which suggests the inconsistency in position angles between the two samples. Such a difference, along with the aforementioned inconsistency in histograms of the position angles, can be explained by the lower signal-to-noise ratio of the SCUPOL data.
The strength of the magnetic field {#sec:strength}
----------------------------------
@1951PhRv...81..890D and @1953ApJ...118..113C proposed that the strength of the B-field could be estimated by interpreting the observed deviation of polarization angles from the mean polarization angle orientation as being due to Alfvén waves induced by turbulent perturbations. This interpretation implies that ${\delta}B/B_0\sim\sigma_v/V_{\rm A},$ where ${\delta}B$ is the magnitude of a turbulent component of the B-field, $B_0$ is the strength of the large-scale B-field, $\sigma_v$ is the one-dimensional nonthermal velocity dispersion, and $V_{\rm A}=B_0/\sqrt{4\pi\rho}$ is the Alfvén speed for a gas with a mass density of $\rho$ [also see @2009ApJ...696..567H]. Such a method (the Davis-Chandrasekhar-Fermi method, DCF method hearafter), in its modified form, has been widely used in estimating the plane-of-sky magnetic field strength $B_{\rm pos}$ from a polarization map by implicitly assuming that ${\delta}B/B_{\rm pos}\sim\sigma_\theta$, where $\sigma_\theta$ is the measured dispersion of polarization angles about a mean or modeled B-field.
Recently, progress has been made toward more accurately quantifying ${\delta}B/B_{\rm pos}$ from a statistical analysis of polarization angles. In this context, there are different methods based on the “structure function” (SF) of polarization angles [@2009ApJ...696..567H] or the “auto-correlation function" (ACF) of polarization angles [@2009ApJ...706.1504H]. Yet another approach is to measure the polarization dispersion with a method analogous to “unsharp masking” (UM) [@2017ApJ...846..122P]. Here we use these methods to estimate $B_{\rm pos}$ in Oph C and compare the results. For the analyses, we use our vector selection criteria of $P/\delta P>3$ and $\delta P<5$%.
In the original version of DCF’s field model [@1951PhRv...81..890D; @1953ApJ...118..113C], the effects of signal integration along line-of-sight and across the beam (hereafter beam-integration effect) were not taken into account. Results from theoretical and numerical works have shown that the beam-integration effect can cause the angular dispersion in polarization maps to be underestimated, therefore overestimating the magnetic field strength [@2001ApJ...561..800H; @2001ApJ...546..980O; @2001ApJ...559.1005P; @2008ApJ...679..537F; @2009ApJ...706.1504H; @2016ApJ...821...21C]. To account for this effect, we take a conventional correction factor (we use $Q_c$ to represent this factor throughout this paper) value of 0.5 [@2001ApJ...546..980O] to correct the measured angular dispersions and the corresponding magnetic field strength in the SF and UM analyses. The correction parameter $Q_c$ is further discussed in Section \[methods\].
Assuming the optically thin dust emission, a dust-to-gas ratio $\Lambda$ of 1:100 [@1991ApJ...381..250B], and an opacity index $\beta$ of 2 [@1983QJRAS..24..267H], we calculate gas column density $N(H_2)$ following: $$N(H_2) = \frac{I_{\nu}}{\mu m_{\mathrm{H}} \kappa_{\nu} B_{\nu} (T)},$$ where $I_{\nu}$ is the continuum intensity at frequency $\nu$, $\mu = 2.86$ is the mean molecular weight [@2013MNRAS.432.1424K; @2015MNRAS.450.1094P], $m_{\mathrm{H}}$ is the atomic mass of hydrogen, $\kappa_{\nu} = 0.1 (\nu / 1 \mathrm{THz})^{\beta}$ is the dust opacity [@1983QJRAS..24..267H] in cm$^2$ g$^{-1}$, and $B_{\nu} (T)$ is the Planck function at temperature $T$. In our analyses, we adopted a dust temperature of 10 $\pm$ 3 K [@2007MNRAS.379.1390S]. The uncertainty on the estimation of column density mainly comes from the uncertainty of $\kappa_{\nu}$ . Conservatively, we adopt a fractional uncertainty of 50% for $\kappa_{\nu}$. In our calculations, we ignore the uncertainties on $\Lambda$ and $\mu$. The column density was estimated over the area with Stokes $I>250$ mJy beam$^{-1}$ (see Figure \[fig:figmap\]). The measured area $A$ is 14544 arcsec$^2$ (0.0053 pc$^2$). Taking into account the uncertainties on $\kappa_{\nu}$, temperature, and flux calibration, the fractional uncertainty on the estimated column density is 59%. Therefore, the mean column density in the concerned region is estimated to be (1.05 $\pm$ 0.62) $\times$ 10$^{23}$ cm$^{-2}$. Since the Oph-C core is highly ‘centrally condensed’ , we adopt a spherical geometry and a core volume ($V$) of 4/3($A^3$/$\pi$)$^{1/2}$. Again, we ignore the uncertainty on the geometry assumption. The average volume density $n_{\mathrm{H_2}}$ is estimated to be (6.4 $\pm$ 3.7) $\times$ 10$^{5}$ cm$^{-3}$. The total mass in our measured volume, $M = \mu m_\mathrm{H} N (\mathrm{H_2}) A$, is 12 $\pm$ 7 solar masses.
To calculate the plane-of-sky magnetic field strength, we need information about the velocity dispersion of the gas. Assuming isotropic velocity perturbations, we adopt the line-of-sight velocity dispersion estimated by . In their work, they carried out N$_2$H$^+$ (1–0) observations toward the Ophiuchus main cloud with a 26$\arcsec$ beam using the IRAM 30m telescope, and found that the average line-of-sight nonthermal velocity dispersion, $\sigma_v$, of the dense structures in Oph-C is 0.13 $\pm$ 0.02 km s$^{-1}$. Their N$_2$H$^+$ data are appropriate to trace the velocity dispersion in Oph-C because of many reasons. At 10 K, the critical density of N$_2$H$^+$ (1–0) is 6.1 $\times$ 10$^4$ cm$^{-3}$ [@2015PASP..127..299S], which is sufficient to probe the dense materials in Oph-C. Also the masses of dense structures in Oph-C traced by the N$_2$H$^+$ (1–0) data and the SCUBA-2 850 $\mu$m continuum data are in good agreement [@2015MNRAS.450.1094P] and the N$_2$H$^+$ (1–0) in Oph-C is optically thin , indicating the N$_2$H$^+$ data and our SCUBA-2/POL-2 data generally trace the same material. Although the beam size of the N$_2$H$^+$ observation is nearly twice the beam size of our POL-2 observation, the spatial resolution of the N$_2$H$^+$ observation, 26$\arcsec$, is still sufficient to resolve the Oph-C core that has a diameter of $\sim$2$\arcmin$. Thus, it could be concluded that the average line-of-sight nonthermal velocity dispersion of the dense structures in Oph-C traced by N$_2$H$^+$ (1–0) is well suited to represent the average gas motions in our concerned region.
Parameter Description SF ACF UM
--------------------------------------------------------- -------------------------------------------------- ----------------- ----------------- -------------------
$\Delta \theta$ (degrees) Angular dispersion 22 $\pm$ 1 21 $\pm$ 8 11 $\pm$ 1
$\langle \delta B^2 \rangle $/$ \langle B_0^2 \rangle$ Turbulent-to-ordered magnetic field energy ratio 0.15 $\pm$ 0.01 0.13 $\pm$ 0.10 0.035 $\pm$ 0.004
$B_{\rm pos}$ ($\mu$G) Plane-of-sky magnetic field strength 206 $\pm$ 68 223 $\pm$ 113 426 $\pm$ 141
Parameter Description SF ACF UM
--------------------------------------------------------- -------------------------------------------------- ----------------- ----------------- -----------------
$\Delta \theta$ (degrees) Angular dispersion 45 $\pm$ 14 34 $\pm$ 13 21 $\pm$ 7
$\langle \delta B^2 \rangle $/$ \langle B_0^2 \rangle$ Turbulent-to-ordered magnetic field energy ratio 0.61 $\pm$ 0.37 0.35 $\pm$ 0.27 0.14 $\pm$ 0.09
$B_{\rm pos}$ ($\mu$G) Plane-of-sky magnetic field strength 103 $\pm$ 46 136 $\pm$ 69 213 $\pm$ 115
$\lambda$ Observed magnetic stability critical parameter 7.8 $\pm$ 5.7 5.9 $\pm$ 4.6 3.8 $\pm$ 3.0
$\lambda_c$ Corrected magnetic stability critical parameter 2.6 $\pm$ 1.9 1.9 $\pm$ 1.5 1.3 $\pm$ 1.0
$E_{\mathrm{B}}$ ($10^{35}$ J) Total magnetic energy 5.4 $\pm$ 4.8 9.5 $\pm$ 9.7 23.2 $\pm$ 25.0
### Structure function analysis
In the SF method [@2009ApJ...696..567H], the magnetic field is assumed to consist of a large-scale structured field, $B_0$, and a turbulent component, ${\delta}B$. The structure function infers the behavior of position angle dispersion as a function of vector separation $l$. At some scale larger than the turbulent scale $\delta$, ${\delta}B$ should reach its maximum value. At scales smaller than a scale $d$, the higher-order terms of the Taylor expansion of $B_0$ can be cancelled out. When $\delta < l \ll d$, the angular dispersion function follows the form: $$\langle \Delta \Phi ^2 (l)\rangle_{\mathrm{tot}} \simeq b^2 + m^2l^2+\sigma_M^2(l).$$ In this equation, $\langle \Delta \Phi ^2 (l)\rangle_{\mathrm{tot}}$, the square of the total measured dispersion function, consists of $b^2$, a constant turbulent contribution, $m^2l^2$, the contribution from the large-scale structured field, and $\sigma_M^2(l)$, the contribution of the measurement uncertainty. The ratio of the turbulent component and the large-scale component of the magnetic field is given by: $$\frac{\langle {\delta}B^2 \rangle^{1/2}}{B_0} = \frac{b}{\sqrt{2-b^2}}.$$ And $B_0$ is estimated according to: $$B_0 \simeq \sqrt{(2-b^2)4\pi \mu m_{\mathrm{H}} n_{\mathrm{H_2}}} \frac{\sigma_v}{b}.$$ Then the estimated plane-of-sky magnetic field strength is corrected by $Q_c$: $$B_{\mathrm{pos}} = Q_cB_0.$$
![Left-hand axis: Angular dispersion function for Oph-C. The angle dispersion segments are shown in black solid circles with error bars. The measurement uncertainties were removed. The best fit is shown by the blue dashed line. Right-hand axis: the number of independent vectors (green dashed line) included in each data bin. The bin size is 7$\arcsec$. \[fig:figstru\]](./structure.pdf)
Figure \[fig:figstru\] shows the angular dispersion corrected by uncertainty ($\langle \Delta \Phi ^2 (l)\rangle_{\mathrm{tot}} - \sigma_{M}^2(l)$) as a function of distance measured from the polarization map. Following @2009ApJ...696..567H, the data are divided into separate distance bins with separations corresponding to the pixel size. At scales of 0-25$\arcsec$, the angular dispersion function increases steeply with the segment distance, most possibly due to the contribution of the turbulent field. At scales larger than 25$\arcsec$, the function continues increasing with a shallower slope, which we may attribute to the large-scale ordered magnetic field structure, and reaches its maximum at $\sim$100$\arcsec$. The maximum of the angular dispersion function is lower than the value expected for a random field [52$\degr$, @2010ApJ...716..893P]. The angular dispersion function presents wave-like “jitter” features at $l >$25$\arcsec$. have attributed the jitter features to the sparse sampling of the vectors in the observed region, which means the independent vectors involved in each distance bin are not enough to achieve statistical significance. We performed simple Monte Carlo simulations (see Appendix \[modeling\]) and found that the uncertainty from sparse sampling is $\sim$1.5$\degr$ in the structure function for models with SFs similar to that of our data in amounts of large-scale spatial correlation and random angular dispersions. We fit the structure function over $25\arcsec<l<100\arcsec$. During the fitting, both the uncertainties from the sparse sampling and from simply propagating the measurement uncertainties of the observed position angles have been taken into account. The reduced chi-squared ($\chi_{red}^2$) of the fitting is 1.2. The calculated values of parameters are given in Table \[tab:para\_int\] (without correction for the beam-integration effect) and Table \[tab:para\] (with correction for the beam-integration effect).
### Auto-correlation function analysis
The ACF method [@2009ApJ...706.1504H] expands the SF method by including the effect of signal integration along the line of sight and within the beam in the analysis. @2009ApJ...706.1504H write the angular dispersion function in the form: $$1 - \langle \cos \lbrack \Delta \Phi (l)\rbrack \rangle \simeq \frac{1}{N} \frac{\langle {\delta}B^2 \rangle}{\langle B_0^2 \rangle} \times \lbrack 1 - e^{-l^2/2(\delta^2+2W^2)}\rbrack + a_2\arcmin l^2,$$ where $\Delta \Phi (l)$ is the difference in position angles of two vectors seperated by a distance $l$, $W$ the beam radius (6.0$\arcsec$ for JCMT, i.e., the FWHM beam divided by $\sqrt{8 \ln{2}}$), $a_2\arcmin$ is the slope of the second-order term of the Taylor expansion, and $\delta$ is the turbulent correlation length mentioned before. $N$ is the number of turbulent cells probed by the telescope beam and is given by: $$N = \frac{(\delta^2 + 2W^2)\Delta\arcmin}{\sqrt{2\pi}\delta^3},$$ where $\Delta\arcmin$ is the effective thickness of the cloud. The ordered magnetic field strength can be derived by: $$B_0 \simeq \sqrt{4\pi \mu m_{\mathrm{H}} n_{\mathrm{H_2}}} \sigma_v \left[ \frac{\langle {\delta}B^2 \rangle}{\langle B_0^2 \rangle} \right]^{-1/2}.$$
Figure \[fig:figcorre\](a) shows the angular dispersion function of the polarization segments in the Oph-C region. Figure \[fig:figcorre\](b) shows the correlated component of the dispersion function. The uncertainty from sparse sampling is $\sim$0.015 in the auto-correlation function (see Appendix \[modeling\]). We fit the function at $l<100\arcsec$. Again, both the uncertainties from the sparse sampling and from the measurements have been taken into account. In our fitting, $\Delta \arcmin$ is set to 20$\arcsec$, which is roughly the FWHM of the starless sub-core identified by @2015MNRAS.450.1094P. The $\chi_{red}^2$ of the fitting is 1.1. The turbulent correlation length $\delta$ is found to be 7.0$\arcsec$ $\pm$ 2.7$\arcsec$ (4.3 $\pm$ 1.6 mpc). The number of turbulent cells is derived to be 2.5 $\pm$ 0.5. The calculated values of other parameters are given in Tables \[tab:para\_int\] and \[tab:para\].
### Unsharp masking analysis
In this section, we followed @2004ApJ...600..279C to derive the plane-of-sky magnetic field strength with the expression: $$\label{eq:eqcf}
B_{\mathrm{pos}} = Q_c \sqrt{4\pi \mu m_{\mathrm{H}} n_{\mathrm{H_2}}}\frac{\sigma_\mathrm{v}}{\sigma_{\mathrm{\theta}}}.$$ The dispersion of the magnetic field angle, $\sigma_{\mathrm{\theta}}$, is measured following the unsharp masking method developed by @2017ApJ...846..122P. Firstly, a 3 $\times$ 3 pixel boxcar average is applied to the measured angles to show the local mean field orientation. With a 3 $\times$ 3 pixel boxcar, the effect of the curvature of the large-scale ordered field on the smoothing is minimized. Then the deviation in field angle from the mean field orientation is derived by subtracting the smoothed map from the observed magnetic field map. Finally, the standard deviation of the residual angles is measured to represent the angular dispersion of the magnetic field angle.
We applied the UM method on our data and restricted the analysis to pixels where the maximum angle difference within the boxcar is $< 90 \degr$. Figure \[fig:figunsharp\] shows the observed position angles $\theta_{\rm obs}$, the position angles $<\theta>$ of a mean B-field derived by smoothing the observed position angles with a $3\times3$ pixel boxcar filter, and the residual values $\theta_{\rm obs}-<\theta>$. We then calculated the standard deviation of magnetic field angles ($\sigma_{\theta}$) as a cumulative function of the maximum permitted angle uncertainty ($\delta \theta_{\mathrm{max}}$) in the 3 $\times$ 3 pixel smoothing box (see Figure \[fig:figcumu\]). With Monte Carlo simulations, @2017ApJ...846..122P found that $\sigma_{\theta}$ can well represent the true angular dispersion when $\delta \theta_{\mathrm{max}}$ is small, while $\sigma_{\theta}$ tend to increase with $\delta \theta_{\mathrm{max}}$ when $\delta \theta_{\mathrm{max}}$ is large. In our case, we restrict our analysis to 12$\degr < \delta \theta_{\mathrm{max}} <$ 47$\degr$, where $\sigma_{\theta}$ remains relatively constant within this $\delta \theta_{\mathrm{max}}$ range. The average standard deviation is measured to be 10.7$\degr \pm$0.6$\degr$ (see Figure \[fig:figcumu\]). This value is introduced in Equation \[eq:eqcf\] as $\sigma_{\mathrm{\theta}}$. The calculated values of other parameters are given in Tables \[tab:para\_int\] and \[tab:para\].
Discussion
==========
Structure and orientation of the magnetic field
-----------------------------------------------
Oph-C is unique in the Ophiuchus cloud as it is fully a starless core. Investigating the magnetic field structure in starless cores is essential for us to explore the initial conditions of star formation. Previous polarization observations toward cores in the starless phase have shown relatively smooth and uniform magnetic field structures [@2000ApJ...537L.135W; @2004ApJ...600..279C; @2009MNRAS.398..394W]. Recently, @2017ApJ...845...32K presented the first detection of an hourglass-shaped magnetic field in a starless core with NIR polarization observations toward FeSt 1–457 (also known as Pipe-109), suggesting that the magnetic field lines can be distorted by mass condensation in the starless phase. However, the NIR polarization observations cannot trace the densest materials in the core and the hourglass morphology was not found in sub-mm polarization observations toward the same source . Our observations toward Oph-C, which present the most sensitive sub-mm polarization observation in a low-mass starless core to date, reveal an relatively ordered B-field with a prevailing northeast-southwest orientation (see Figure \[fig:figmap\]). However, the B-field structure in Oph-C shows no evidences of an hourglass morphology, which is consistent with previous observations that an hourglass morphology is not generally found in other cold dense cores from sub-mm polarization observations at scales $>$ 0.01 pc. This suggests that mass condensation does not significantly distort the local B-field structure at scales $>$ 0.01 pc in the densest materials of dense cores at both the starless phase and prestellar phase.
The role of magnetic field in dense cores may vary with the evolution of the core. As part of the BISTRO survey [@2017ApJ...842...66W], polarization observations towards two protostellar cores with similar masses as that of Oph-C in the Ophiuchus cloud, Oph-A and Oph-B, have been made and are ready to be compared with our data. Our observations of Oph-C show that the overall magnetic field geometry in Oph-C is ordered and the polarization position angles show large angular dispersions. This behaviour is similar with that in Oph-B, which is relatively a quiescent core in Ophiuchus but is more evolved than Oph-C, while the B-field in Oph-A, which is the warmest and the only core with substantially gravitationally bound sub-cores found in Ophiuchus, is mostly well organized and with small angular dispersions [@2009ApJ...692..973E; @2015MNRAS.450.1094P; @2018ApJ...859....4K; @2018ApJ...861...65S]. In addition, the angular dispersions in Oph-C (($\sim$11$\degr$ to $\sim$22 $\degr$)) and Oph-B ($\sim$15$\degr$) is larger than that in Oph-A ($\sim$2 to $\sim$6$\degr$). These indicate that the star formation process may possibly reduce angular dispersions in the magnetic field in the late stages of star formation.
Our observations reveal that there is a prevailing orientation in the B-field in Oph-C centering at $\sim$40$\degr$ to $\sim$100$\degr$ (see Figure \[fig:figang\]). This orientation agrees with the B-field orientations in Oph-A, where the B-field components center at $\sim$40$\degr$ to $\sim$100$\degr$ [@2018ApJ...859....4K], and Oph-B, where the position angle of B-field peaks at $\sim$50$\degr$ to $\sim$80$\degr$ [@2018ApJ...861...65S]. The B-field position angle distribution in Oph-C is consistent with the $\sim$50$\degr$ B-field component in lower-density regions of the Ophiuchus cloud traced by the NIR polarization map of @2015ApJS..220...17K, and is also aligned with the cloud-scale B-field orientation probed by Planck . The consistence of B-field orientation from cloud to core scales indicates that the large-scale magnetic field plays a dominant role in the formation of dense cores in the Ophiuchus cloud.
Depolarization effect
---------------------
A clear trend of decreasing polarization percentage with increasing dust emission intensity is seen in Figure \[fig:figmap\]. Such an effect is more evident in Figure \[fig:figpi\](a), in which the $P$-$I$ relation suggests depolarization toward high density. Considering that the overall field does not change orientation while threading the core, the depolarization in Oph-C seems unlikely to be a by-product of field tangling of complex small-scale field lines within the JCMT beam. In addition, the subsonic nonthermal gas motions of Oph-C indicate that the polarization percentage has not been significantly affected by the number of turbulent cells along line of sight. The RAT mechanism [@2007JQSRT.106..225L], which suggests inefficient grain alignment toward high density regions, also cannot fully explain the depolarization effect because of the lack of an internal or external radiation field in the Oph-C region. Alternatively, grain characteristics such as size, shape, and composition, which are related to grain alignment mechanism, may explain the depolarization effect. The turbulent structure of the magnetic field, which can induce the field’s tangling and therefore reduce $P$, also provides a plausible explanation to the decreasing of the polarization percentage towards higher intensities .
We fitted the $P$-$I$ relation with a power-law slope, and found the slope index is $-$1.03 $\pm$ 0.05, which indicates that the polarized intensity is almost constant in Oph-C. The nearly constant polarized intensity is more clearly shown in Figure \[fig:figpi\](b). The slope index for Oph-C is slightly lower than the index of $-$0.92 $\pm$ 0.05 for the entire Ophiuchus cloud . For other dense cores in Ophiuchus, a slope index of -0.7 $\sim$ -0.8 was found in Oph-A [@2018ApJ...859....4K], and an index of around -0.9 was found in Oph-B [@2018ApJ...861...65S]. Considering that Oph-A is the warmest and most evolved among the Oph cores, Oph-B is more quiescent than Oph-A, and Oph-C is the most quiescent region in Ophiuchus [@2015MNRAS.450.1094P], it appears that the power-law slope of the $P$-$I$ relation is shallower in more evolved dense cores in the Ophiuchus cloud. This trend could be explained by the improved alignment efficiency resulting from the additional internal radiation [predicted by the RAT theory @2007JQSRT.106..225L] in more evolved regions. Alternatively, if the depolarization is caused by turbulence , the stronger turbulence in more evolved dense cores may also be a possible reason for the variation in the slope index. More detailed analysis of the depolarization effect in the Ophiuchus Cloud will be presented in a separate publication by the BISTRO team.
Magnetic field strength
-----------------------
### Comparison of three modified DCF methods {#methods}
While the morphologies of magnetic fields can help us to qualitatively understand its role in the star formation process, the magnetic field strength is important in quantitatively assessing the significance of the magnetic field compared to gravity based on the mass-to-flux ratio, and compared to turbulence based on the ratio of random-to-ordered components in polarization angle statistics. The strengths of magnetic fields, however, cannot be measured directly from polarization observations. In this paper, we estimated the average magnetic field strength in Oph-C from different modified DCF methods. Results of these methods are shown in Table \[tab:para\_int\] and Table \[tab:para\]. From the statistical analyses of the dispersion of dust polarization angles, the beam-integrated angular dispersions derived from the SF and ACF methods are consistent with each other ($\sim$21$\degr$ to $\sim$22$\degr$), and are larger than that derived from the UM method ($\sim$11$\degr$), indicating the magnetic field strength estimated from the UM method could be systematically larger than that derived from the SF method and the ACF method. Similar behavior was found when applying these methods on the polarization map of OMC-1 [@2009ApJ...696..567H; @2009ApJ...706.1504H; @2017ApJ...846..122P], a region that has a relatively stronger magnetic field (the $B_{\rm pos}$ is $\sim$13.2 mG estimated from the UM method and $\sim$3.5 to $\sim$3.8 mG estimated from the SF/ACF method without correction for beam integration) than that in Oph-C. The estimated $B_{\rm pos}$ in Oph-C ($\sim$0.1 to $\sim$0.2 mG) is lower than $B_{\rm pos}$ in Oph-A ($\sim$0.2 to $\sim$5 mG) and Oph-B ($\sim$0.6 mG).
Because the results of the dispersion function analysis could be affected by the bin size [@2010ApJ...721..815K], we have redone the SF and ACF analyses to find the dependence on the bin size. We found that oversampling (with bin size $<$7$\arcsec$) would inject additional noises into the dispersion functions (both SF and ACF), thus leading to overestimation of the angular dispersion and underestimation of the B-field strength. The origin of the additional noise is possibly related to the wrongly generated masks due to small pixel size in the POL-2 data reduction process, and needs to be further investigated. Increasing the bin size, on the other hand, shows little effects on the SF method, and leads to larger values of turbulent scale and B-field strength for the results of the ACF method. We also found that, by undersampling, the turbulent scale estimated from the ACF method is always approximately equal to the bin size. The effects on the ACF method can be simply explained by a loss of information on small scales due to undersampling. @2010ApJ...721..815K has also investigated the dependence of the SF method on the bin size, but got different results from ours: oversampling shows little effect on the SF, while undersampling biases the analysis toward larger dispersion values. This indicates that the dependence of dispersion function on the bin size is not simple. Considering these factors and that the derived turbulent scale ($\sim$ 7.0$\arcsec$) is approximately equal to the Nyquist sampling interval of our data, we note that the turbulent scale along with the B-field strength derived from the ACF method could be overestimated.
Increasing the box size for smoothing would significantly overestimate the angular dispersion derived from the UM method because of field curvature, while in a zero-curvature case, decreasing the box size would slightly underestimate the angular dispersion @2017ApJ...846..122P. Since the B-field in Oph-C does not show well-defined shapes, it is unclear whether the angular dispersion in Oph-C is underestimated or overestimated by the UM method. We checked the dependence on the box size in our UM analysis by re-applying the UM method to our data with a 5 $\times$ 5 pixel smoothing box and a 7 $\times$ 7 pixel smoothing box, and derived angular dispersions of $\sim$11$\degr$ to $\sim$12$\degr$, indicating that larger smoothing box would not significantly change the results of our UM analysis.
Systematic uncertainties of the DCF method may arise from the beam-integration effect. For the UM method and the SF method, we use a correction factor $Q_c$ to account for the averaging effect of turbulent cells along the line of sight. @2001ApJ...546..980O found that $Q_c$ is in the range of 0.46-0.51 for angular dispersions less than 25$\degr$. In our case, the Gaussian fitting of the position angles of polarization segments in Oph-C gives a standard deviation of angle of 33$\degr$, which is larger than the angular dispersion limit of @2001ApJ...546..980O. However, this standard deviation includes the contribution from the curvature of the large-scale field. Excluding the angular variations of the large-scale field, we got standard deviations $<$ 25$\degr$ from the modified DCF methods. So we adopted a conventional $Q_c$ value of 0.5. The uncertainty of the $Q_c$ value is $\sim$30 % [@2004ApJ...600..279C]. On the other hand, the ACF method takes into account the beam-integration effect by directly fitting the angular dispersion function. The number of turbulent cells of $\sim$2.5 derived from the ACF method is equivalent to a $Q_c$ of $\sim$0.63, which is slightly larger than the correlation factor adopted by the SF analysis and the UM analysis.
As mentioned by , even applying the most complicated modified DCF method on the highest quality data would lead to a $B_{\rm pos}$ value with an uncertainty varying by a factor of two or more due to various reasons. It is essential to assess the accuracy of these methods by comparing the results of these methods on polarization maps from simulations. Although the magnetic field strengths estimated from the three modified DCF methods may have systematic differences, they are consistent with each other within the uncertainties, indicating that these results are robust to some extent, and that we can still compare the relative importance of magnetic field with gravity and turbulence with these results.
### Magnetic field vs. gravity
To find out whether or not the magnetic field can support Oph-C against gravity, we compared the mass-to-magnetic-flux ratio with the critical ratio using the local magnetic stability critical parameter $\lambda$ [@2004ApJ...600..279C]: $$\lambda = \frac{(M/\Phi)_{observed}}{(M/\Phi)_{critical}},$$ where $(M/\Phi)_{observed}$ is the observed mass-to-magnetic-flux ratio: $$(\frac{M}{\Phi})_{observed} = \frac{\mu m_{\mathrm{H}}N(H_2)}{B},$$ and $(M/\Phi)_{critical}$ is the critical mass-to-magnetic-flux ratio: $$(\frac{M}{\Phi})_{critical} = \frac{1}{2\pi\sqrt{G}}.$$ We estimated $\lambda$ using the relation in @2004ApJ...600..279C: $$\lambda = 7.6 \times 10^{-21} \frac{N(H_2)}{B_{\rm pos}}$$
The observed critical parameters derived from the SF, ACF, and UM methods are 7.8 $\pm$ 5.7, 5.9 $\pm$ 4.6, and 3.8 $\pm$ 3.0 (see Table \[tab:para\]), respectively. @2004ApJ...600..279C proposed that the observed $M/\phi$ along with $\lambda$ are overestimated because of geometrical effects. @1993ApJ...407..175C found a average line-of-sight B-field strength ($B_{\mathrm{los}}$) of $+$ 6.8 $\pm$ 2.5 $\mu$G in the Ophiuchus cloud based on OH Zeeman observations with a 18$\arcmin$ beam. Their estimated $B_{\mathrm{los}}$ is much smaller than the B$_{\mathrm{pos}}$ infered by our analyses, indicating that the B-field in Oph-C is possibly lying near the plane of sky. However, since quasi-thermal OH emissions cannot trace high density materials with $n$(H$_2$) $> 10^4 $ cm$^{-3}$ and the beam of the OH Zeeman observation is much larger than that of our polarization maps, it is more likely that the line-of-sight B-field strength in the Oph-C is underestimated by @1993ApJ...407..175C. As the degree of the underestimation of $B_{\mathrm{los}}$ is unknown, the correction factor for the geometrical bias cannot be derived by simply comparing $B_{\mathrm{los}}$ and $B_{\mathrm{pos}}$. Alternatively, we adopted a statistical correction factor of 3 [@2004ApJ...600..279C]. By applying this correction, we obtain corrected critical parameters ($\lambda_c$) of 2.6 $\pm$ 1.9, 1.9 $\pm$ 1.5, and 1.3 $\pm$ 1.0 (see Table \[tab:para\]) for the SF, ACF, and UM methods, respectively. These values indicate that the Oph-C region is near magnetically critical or slightly magnetically supercritical (i.e. unstable to collapse).
### Magnetic field vs. turbulence
To compare the relative importance of the magnetic field and turbulence in Oph-C, we calculated the magnetic field energy and the internal nonthermal kinetic energy. The total magnetic field energy is given by: $$E_B = \frac{B^2 V}{2 \mu_0}$$ in SI units, where $\mu_0$ is the permeability of vacuum and $B = \frac{4}{\pi}B_{\rm pos}$ [@2004ApJ...600..279C] is the total magnetic field strength. And the internal nonthermal kinetic energy is derived by: $$E_{K,NT} = \frac{3 M \sigma_v^2}{2}.$$ For the estimated volume (see Section \[sec:strength\]) of Oph-C, the internal nonthermal kinetic energy is (6.1 $\pm$ 2.0) $\times$ 10$^{35}$ J. The total magnetic field energy measured from the SF, ACF, and UM methods are (5.4 $\pm$ 4.8) $\times$ 10$^{35}$ J, (9.5 $\pm$ 9.7) $\times$ 10$^{35}$ J, and (2.3 $\pm$ 2.5) $\times$ 10$^{36}$ J (see Table \[tab:para\]), respectively. The $E_B$ calculated from the SF method is comparable to $E_{K,NT}$, while the values of $E_B$ estimated from the ACF and UM methods are greater than $E_{K,NT}$. However, the uncertainty is more than 100% for the values of $E_B$ calculated from the ACF and UM methods. Thus, we can only set upper limits for the total magnetic field energy in Oph-C from these two methods.
Summary
=======
As part of the BISTRO survey, we have presented the 850 $\mu$m polarization observations toward the Oph-C region with the POL-2 instrument at the JCMT. The main conclusions of this work are as follows:
1. Our POL-2 observations are much more sensitive and trace a larger area than previous SCUPOL observations. Unlike the randomly distributed magnetic field orientations traced by the SCUPOL observations, the magnetic field traced by our POL-2 observations show an ordered field geometry with a predominant orientation of northeast-southwest. We found the average angular difference of spatially overlapping vectors between the two data sets to be $\sim$39$\degr$. We performed a K-S test on the position angles, and found that the POL-2 data and the SCUPOL data have low probability (0.06) to be drawn from the same distribution. The inconsistency between the POL-2 and the SCUPOL data may be explained by the low signal-to-noise ratio of the SCUPOL data.
2. The B-field orientation in Oph-C is consistent with the B-field orientations in Oph-A and Oph-B. The orientation also agrees with the B-field component in lower-density regions traced by NIR observations, and is aligned with the cloud-scale B-field orientation revealed by Planck.
3. We detect a decreasing polarization percentage as a function of increasing total intensity in the Oph-C region. The power-law slope index is found to be $-$1.03 $\pm$ 0.05, suggesting that the polarized intensity is almost constant in Oph-C.
4. We compare the plane-of-sky magnetic field strength in Oph-C calculated from different modified DCF methods. The $B_{pos}$ calculated by the SF method, the ACF method, and the UM method are 103 $\pm$ 46 $\mu$G, 136 $\pm$ 69 $\mu$G, and 213 $\pm$ 115 $\mu$G, respectively.
5. The mass-to-magnetic-flux ratio of Oph-C is found to be comparable to or slightly higher than its critical value, suggesting that the Oph-C region is near magnetically critical magnetically supercritical (i.e., unstable to collapse).
6. In Oph-C, the total magnetic energy calculated from the SF method is comparable to the turbulent energy. Due to large uncertainties, the ACF method and the UM method only set upper limits for the total magnetic energy.
7. We compared our work with studies of two other dense cores in the Ophiuchus cloud. We find the B-fields in Oph-C and Oph-B have larger angular dispersions than Oph-A. We also find a possible trend of shallower $P$-$I$ relationship with evolution in the three dense cores in the Ophiuchus region. In addition, the $B_{pos}$ in Oph-C is lower than $B_{pos}$ of more evolved regions (e.g., Oph-A and Oph-B) in Ophiuchus.
We thank Dr. Qizhou Zhang for helpful discussions on dispersion function analysis. J.L., K.Q., D.L., and L.Q. are supported by National Key R&D Program of China No. 2017YFA0402600. J.L. and K.Q. acknowledge the support from National Natural Science Foundation of China (NSFC) through grants U1731237, 11590781, and 11629302. J.L. acknowledges the support from the program of China Scholarship Council (No. 201806190134) and from the Smithsonian Astrophysical Observatory pre-doctoral fellowship. K.P. was an International Research Fellow of the Japan Society for the Promotion of Science. D.W.T. and K.P. acknowledge Science and Technology Facilities Council (STFC) support under grant numbers ST/K002023/1 and ST/M000877/1. J.K. was supported by MEXT KAKENHI grant number 16H07479 and the Astrobiology Center of NINS. M.T. was supported by MEXT KAKENHI grant number 22000005. D.W.T. and K.P. acknowledge Science and Technology Facilities Council (STFC) support under grant numbers ST/K002023/1 and ST/M000877/1. C.W.L. and M.K. were supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education, Science and Technology (CWL: NRF-2016R1A2B4012593) and the Ministry of Science, ICT and Future Planning (MK: NRF-2015R1C1A1A01052160). K.P. and S.P.L. acknowledge the support of the Ministry of Science and Technology of Taiwan (Grant No. 106-2119-M-007-021-MY3). W.K. was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF-2016R1C1B2013642). J.E.L. is supported by the Basic Science Research Program through the National Research Foundation of Korea (grant No. NRF-2018R1A2B6003423) and the Korea Astronomy and Space Science Institute under the R&D program supervised by the Ministry of Science, ICT and Future Planning. A.S. acknowledge the support from KASI for postdoctoral fellowship. T.L. is supported by a KASI fellowship and an EACOA fellowship. The James Clerk Maxwell Telescope is operated by the East Asian Observatory on behalf of the National Astronomical Observatory of Japan, the Academia Sinica Institute of Astronomy and Astrophysics, the Korea Astronomy and Space Science Institute, the National Astronomical Observatories of China, and the Chinese Academy of Sciences (Grant No. XDB09000000), with additional funding support from the Science and Technology Facilities Council of the United Kingdom and participating universities in the United Kingdom and Canada. The James Clerk Maxwell Telescope has historically been operated by the Joint Astronomy Centre on behalf of the Science and Technology Facilities Council of the United Kingdom, the National Research Council of Canada, and the Netherlands Organisation for Scientific Research. Additional funds for the construction of SCUBA-2 and POL-2 were provided by the Canada Foundation for Innovation. The data used in this paper were taken under project code M16AL004. This research made use of APLpy, an open-source plotting package for Python [@2012ascl.soft08017R], Astropy, a community-developed core Python package for Astronomy , and Matplotlib, a Python 2D plotting library for Python [@2007CSE.....9...90H].
Alves, F. O., Frau, P., Girart, J. M., et al. 2014, , 569, L1 Andr[é]{}, P., Belloche, A., Motte, F., & Peretto, N. 2007, , 472, 519 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, , 558, A33 Beckwith, S. V. W., & Sargent, A. I. 1991, , 381, 250 Berry, D. S., Gledhill, T. M., Greaves, J. S., & Jenness, T. 2005, Astronomical Polarimetry: Current Status and Future Directions, 343, 71 Chapin, E., Gibb, A. G., Jenness, T., et al. 2013, Starlink User Note, 258 Crutcher, R. M. 2012, , 50, 29 Crutcher, R. M., Nutter, D. J., Ward-Thompson, D., & Kirk, J. M. 2004, , 600, 279 Crutcher, R. M., Troland, T. H., Goodman, A. A., et al. 1993, , 407, 175 Chandrasekhar, S., & Fermi, E. 1953, , 118, 113 Cho, J., & Yoo, H. 2016, , 821, 21 Currie, M. J., Berry, D. S., Jenness, T., et al. 2014, Astronomical Data Analysis Software and Systems XXIII, 485, 391 Dotson, J. L., Vaillancourt, J. E., Kirby, L., et al. 2010, , 186, 406 Davis, L. 1951, Physical Review, 81, 890 Dempsey, J. T., Friberg, P., Jenness, T., et al. 2013, , 430, 2534 Enoch, M. L., Evans, N. J., II, Sargent, A. I., & Glenn, J. 2009, , 692, 973-997 Falceta-Gon[ç]{}alves, D., Lazarian, A., & Kowal, G. 2008, , 679, 537 Friberg, P., Bastien, P., Berry, D., et al. 2016, , 9914, 991403 Galli, D., & Shu, F. H. 1993, , 417, 220 Galli, D., & Shu, F. H. 1993, , 417, 243 Girart, J. M., Crutcher, R. M., & Rao, R. 1999, , 525, L109 Girart, J. M., Rao, R., & Marrone, D. P. 2006, Science, 313, 812 Girart, J. M., Beltr[á]{}n, M. T., Zhang, Q., Rao, R., & Estalella, R. 2009, Science, 324, 1408 Hall, J. S. 1949, Science, 109, 166 Heitsch, F., Zweibel, E. G., Mac Low, M.-M., Li, P., & Norman, M. L. 2001, , 561, 800 Henning, T., Michel, B., & Stognienko, R. 1995, , 43, 1333 Hildebrand, R. H., Kirby, L., Dotson, J. L., Houde, M., & Vaillancourt, J. E. 2009, , 696, 567 Hildebrand, R. H. 1988, , 29, 327 Hildebrand, R. H. 1988, Astrophysical Letters and Communications, 26, 263 Hildebrand, R. H. 1983, , 24, 267 Hiltner, W. A. 1949, Science, 109, 165 Holland, W. S., Bintley, D., Chapin, E. L., et al. 2013, , 430, 2513 Houde, M., Vaillancourt, J. E., Hildebrand, R. H., Chitsazzadeh, S., & Kirby, L. 2009, , 706, 1504 Hull, C. L. H., Mocz, P., Burkhart, B., et al. 2017, , 842, L9 Hull, C. L. H., Plambeck, R. L., Kwon, W., et al. 2014, , 213, 13 Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90 Kandori, R., Tamura, M., Kusakabe, N., et al. 2017, , 845, 32 Kirk, J. M., Ward-Thompson, D., Palmeirim, P., et al. 2013, , 432, 1424 Koch, P. M., Tang, Y.-W., & Ho, P. T. P. 2010, , 721, 815 Kwon, J., Tamura, M., Hough, J. H., et al. 2015, , 220, 17 Kwon, J., Doi, Y., Tamura, M., et al. 2018, , 859, 4 Jenness, T., Chapin, E. L., Berry, D. S., et al. 2013, Astrophysics Source Code Library, ascl:1310.007 Lazarian, A. 2007, , 106, 225 Loren, R. B., Wootten, A., & Wilking, B. A. 1990, , 365, 269 Mac Low, M.-M., & Klessen, R. S. 2004, Reviews of Modern Physics, 76, 125 Matthews, B. C., McPhee, C. A., Fissel, L. M., & Curran, R. L. 2009, , 182, 143 Motte, F., Andre, P., & Neri, R. 1998, , 336, 150 Mouschovias, T. C., Tassis, K., & Kunz, M. W. 2006, ApJ, 646, 1043 Naghizadeh-Khouei, J., & Clarke, D. 1993, , 274, 968 Ortiz-Le[ó]{}n, G. N., Loinard, L., Kounkel, M. A., et al. 2017, , 834, 141 Ossenkopf, V. & Henning, T. 1994, , 291, 943 Ostriker, E. C., Stone, J. M., & Gammie, C. F. 2001, , 546, 980 Padoan, P., Federrath, C., Chabrier, G., et al. 2014, Protostars and Planets VI, 77 Padoan, P., Goodman, A., Draine, B. T., et al. 2001, , 559, 1005 Pattle, K., Ward-Thompson, D., Kirk, J. M., et al. 2015, , 450, 1094 Pattle, K., Ward-Thompson, D., Berry, D., et al. 2017, , 846, 122 Pattle, K., Ward-Thompson, D., Hasegawa, T., et al. 2018, , 860, L6 Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2018, arXiv:1807.06212 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2016, , 586, A138 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2015, , 576, A105 Poidevin, F., Bastien, P., & Matthews, B. C. 2010, , 716, 893 Qiu, K., Zhang, Q., Menten, K. M., et al. 2014, , 794, L18 Rao, R., Girart, J. M., Marrone, D. P., Lai, S.-P., & Schnee, S. 2009, , 707, 921 Robitaille, T., & Bressert, E. 2012, APLpy: Astronomical Plotting Library in Python, Astrophysics Source Code Library, ascl:1208.017 Roy, A., Andr[é]{}, P., Palmeirim, P., et al. 2014, , 562, A138 Santos, F. P., Franco, G. A. P., Roman-Lopes, A., Reis, W., & Rom[á]{}n-Z[ú]{}[ñ]{}iga, C. G. 2014, , 783, 1 Shirley, Y. L. 2015, , 127, 299 Soam, A., Pattle, K., Ward-Thompson, D., et al. 2018, , 861, 65 Soler, J. D., Alves, F., Boulanger, F., et al. 2016, , 596, A93 Soler, J. D., Hennebelle, P., Martin, P. G., et al. 2013, , 774, 128 Stamatellos, D., Whitworth, A. P., & Ward-Thompson, D. 2007, , 379, 1390 Stephens, I. W., Looney, L. W., Kwon, W., et al. 2013, , 769, L15 Vaillancourt, J. E. 2006, , 118, 1340 Vrba, F. J. 1977, , 82, 198 Ward-Thompson, D., Pattle, K., Bastien, P., et al. 2017, , 842, 66 Ward-Thompson, D., Sen, A. K., Kirk, J. M., & Nutter, D. 2009, , 398, 394 Ward-Thompson, D., Di Francesco, J., Hatchell, J., et al. 2007, , 119, 855 Ward-Thompson, D., Kirk, J. M., Crutcher, R. M., et al. 2000, , 537, L135 Wilking, B. A., Gagn[é]{}, M., & Allen, L. E. 2008, Handbook of Star Forming Regions, Volume II, 5, 351
Uncertainty from sparse sampling {#modeling}
================================
Here we derive the uncertainty in the dispersion function caused by the lack of vector samples (sparse sampling). We perform simple Monte Carlo simulations of modeled structured fields with randomly generated Gaussian dispersions to roughly estimate the uncertainty of sparse sampling in the dispersion function of our data. It should be noted that because simulating the beam-integration effect is extremely time-consuming and only affects the first two or three data points of the dispersion function, the beam-integration effect is not taken into account in our toy models.
We start with generating the underlying field model. We note that since the uncertainty in the angular dispersion function due to sparse sampling is only related to the amount of spatial correlation of field orientations across the sky and the amount of angular dispersion relative to the structured field , the choice of the underlying field model is arbitrary. We build a set of underlyling parabola models [e.g., @2006Sci...313..812G; @2009ApJ...707..921R; @2014ApJ...794L..18Q] with the form: $$y = g + gCx^2,$$ where x is the offsets in pixels along the field axis from the center of symmetry. In Figure \[fig:figmod\] (a) and (b), a parabola field model with $C = 0.13$ is shown in magenta curves as an example.
We then derive the orientation of the modeled sparsely sampled B-vectors by applying a Gaussian angular dispersion of 22$\degr$ (to match the angular dispersion of $\sim$21$\degr$ to $\sim$22$\degr$ derived from the SF and ACF methods) to the underlying B-vectors with the same spatial distributions as those of the observed B-vectors in Oph-C with $P/\delta P>3$ and $\delta P<5$%, while the offsets and angle of the modeled B-vectors with respect to the center of symmetry of the underlying modeled field is random. An example of the modeled sparsely sampled B-vectors is shown in Figure \[fig:figmod\] (a).
In a similar way, we also derived the orientation of “unbiased” samples of B-vectors with a Gaussian angular dispersion of 22$\degr$ and spatial separation of 1 pixel (7$\arcsec$) for comparison. There are enough vectors in the “unbiased” sample to achieve statistical significance. An example of the modeled “unbiased” B-vectors is shown in Figure \[fig:figmod\] (b).
We calculate the SF and ACF (see Figure \[fig:figmod\] (c) and (d) for examples of the SF and ACF) from the sparse samples and “unbiased” samples of modeled vectors, and find that the average deviation of the SFs and ACFs between the two sets of samples are $\sim$1.5$\degr$ and $\sim$0.015 over $25\arcsec<l<100\arcsec$, relatively for SFs and ACFs with similar amounts of large-scale spatial correlation and random anglular dispersions (e.g., similar SF and ACF shapes over $25\arcsec<l<100\arcsec$) to the dispersion functions calculated from the observed data. These average deviations, which are larger than the statistical uncertainties ( $\sim$0.6$\degr$ for the SF and $\sim$0.007 for the ACF over $25\arcsec<l<100\arcsec$ in average) propagated from the measurement uncertainty, are introduced in our analyses as the uncertainties due to sparse sampling.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A regular polygon circumscribing another regular polygon (with a different side number) may be tightened to minimize the difference of both areas. The manuscripts computes the optimum result under the restriction that both polygons are concentric, and obtains limits if the process is repeated in a two-dimensional Babuschka-doll fashion with side numbers increasing or decreasing by one or stepping through the prime numbers. The new aspect compared to the circumscription discussed in the literature so far is that further squeezing of the outer polygon is possible as we drop the requirement of drawing intermediate spacing circles between the polygon pairs.'
author:
- 'Richard J. Mathar'
bibliography:
- 'all.bib'
title: Tightly Circumscribed Regular Polygons
---
Introduction
============
Notation
--------
A regular $n$-gon is drawn with $n$ edges of some common side length $s_n$. The perimeter is $ns_n$. The incircle with inradius $r_n^{(i)}$ touches each edge at the mid point. Each edge covers an angle of $$\phi_n=\frac{2\pi}{n}$$ if viewed from the incircle center. The area of the polygon is comprised of $n$ rotated copies of an isosceles triangle in which the short edge has length $s_n$, facing the angle $\phi_n$, and the two other edges have length $r_n^{(o)}$, which also is the radius of the circumcircle. This isosceles triangle might be sliced into two symmetric rectilinear triangles by drawing a line (apothem) from its base center to the midpoint of the polygon’s incircle; the definition of the tangent and sine functions in these yield $$\tan\frac{\phi_n}{2} = \frac{s_n/2}{r_n^{(i)}},$$ $$\sin\frac{\phi_n}{2} = \frac{s_n/2}{r_n^{(o)}},$$ and therefore $$r_n^{(i)} = \frac{s_n}{2\tan\frac{\phi_n}{2}},
\label{eq.rni}$$ $$r_n^{(o)} = \frac{s_n}{2\sin\frac{\phi_n}{2}}.
\label{eq.rno}$$ As a bridge between two-dimensional geometry and numerical algebra, we define the *standard* position of the polygon in the Cartesian $(x,y)$ plane by mapping $x$ and $y$ to the real and imaginary part of the complex plane, placing the vertices labeled $j=0,1,\ldots n-1$ counter-clock-wise at the coordinates $$x+iy = r_n^{(o)} e^{2\pi j i/n},
\label{eq.std}$$ where $i\equiv \sqrt{-1}$ is the imaginary unit. Edges/sides are also enumerated from $0$ to $n-1$ by calling the smaller of the two vertex labels that are joined.
Tight Circumscription {#sec.hist}
---------------------
The circumscription of a regular $n$-gon by a regular $m$-gon has been constructed earlier by an elementary step drawing the $n$-gon, its circumcircle, declaring this circle to be also the incircle of the $m$-gon with $r_m^{(i)}=r_n^{(o)}$, and drawing the $m$-gon around its incircle [@Finch p. 428][@WeissteinCRC p. 2300][@EIS A051762].
The theme of this manuscript is to drop the requirement of equating the two circles and to search for smaller circumscribing regular $m$-gons in the extended range $r_n^{(i)}\le r_m^{(i)}\le r_n^{(o)}$.
This requires positions where some edges of the $m$-gon cut through the circumcircle of the $n$-gon. The tighter solution, however, may exist only within a restricted range of (relative) orientations of the polygons. The manuscript works out a full representation of the smallest circumscribing polygons, using the ratio $r_m^{(o)}/r_n^{(o)}$ as a figure of merit.
In overview, the achievable size ratios are calculated in Section \[sec.std\] for concentric polygon pairs aligned such that the common center, a vertex of the inner polygon and a vertex of the outer polygon are collinear (standard positions). In Section \[sec.rot\] further size reductions of the outer polygon are found if the side number $n$ of the inner polygon is even and the outer polygon is turned around the common center. A cursory outlook in Section \[sec.trans\] shows that shifting the center of the outer polygon away from the center of the inner polygon may define even smaller circumscribing $m$-gons.
Concentric Standard Placements {#sec.std}
==============================
General side numbers
--------------------
The definition of circumscription implies that the outer $m$-gon must stay further away from the origin than the inner $n$-gon at all viewing directions; at one or more points of contact, both polygons have the same distance $r_n^{(o)}$ to the origin.
![ Circumscribing a 3-gon by a 7-gon. []{data-label="fig.Rpol37"}](Rpol37)
Finding the smallest $m$-gon for small side numbers like Fig. \[fig.Rpol37\] works as follows. Any point on the side of the outer polygon of unit radius has a position $\exp(2\pi i j_o/m)+t[
\exp(2\pi i (j_o+1)/m)
-\exp(2\pi i j_o/m) ]$ with parameter $0\le t\le 1$. Once the vertex number $j_i$ of the inner polygon and the side number $j_o$ of the outer polygon which it touches are known, the point of contact between both polygons in the complex plane solves $$r_{m}^{(o)}
\{e^{2\pi i j_o/m}+t[
e^{2\pi i (j_o+1)/m}
-e^{2\pi i j_o/m) }]\}
=
r_{n}^{(o)}e^{2\pi i j_i/n}
.
\label{eq.shrink}$$ and after division through $r_m^{(0)}e^{2\pi i j_0/m}$ $$1+t[
e^{2\pi i /m}
-1]
=
\frac{r_{n}^{(o)}}{r_{m}^{(o)}}e^{2\pi i (j_i/n-j_o/m)}
.
\label{eq.jijo}$$ Real and imaginary part of this equation establish an inhomogeneous $2\times 2$ linear system of equations for the unknown $t$ and $r_n^{(o)}/r_m^{(o)}$: $$\begin{aligned}
1+t[
\cos(2\pi /m)
-1]
&=&
\frac{r_{n}^{(o)}}{r_{m}^{(o)}}\cos[2\pi (j_i/n-j_o/m)],
\\
t
\sin(2\pi /m)
&=&
\frac{r_{n}^{(o)}}{r_{m}^{(o)}}\sin[2\pi (j_i/n-j_o/m)]
.\end{aligned}$$ $$\begin{aligned}
\left(
\begin{array}{cc}
\cos[2\pi (j_i/n-j_o/m)] & 1-\cos(2\pi/m) \\
\sin[2\pi (j_i/n-j_o/m)] & -\sin(2\pi/m)
\end{array}
\right)
&&
\nonumber \\
\cdot
\left(
\begin{array}{c}
r_{n}^{(o)}/r_{m}^{(o)} \\
t
\end{array}
\right)
&=&
\left(
\begin{array}{c}
1\\
0
\end{array}
\right)
.\end{aligned}$$ The solution is obtained with Cramer’s rule. The inverse ratio is [@AS 4.3.16,4,3,35] $$\begin{aligned}
r_{m}^{(o)}/r_{n}^{(o)}
&=&
\frac
{
\sin(j_i\phi_n-j_o\phi_m)
-\sin[j_i\phi_n-(j_o+1)\phi_m]
}
{\sin \phi_m}
\nonumber \\
&=&
\frac
{
\cos[j_i\phi_n-(j_o+1/2)\phi_m]
}
{\cos( \phi_m/2)}
\nonumber \\
&=&
\frac
{
\cos[\frac{\pi}{nm}\{2j_im-(2j_o+1)n\}]
}
{\cos( \pi/m)}
\label{eq.rrat}
.\end{aligned}$$ Let $d=m-n$ be the difference in the side numbers; then $$2j_im-(2j_o+1)n = 2j_id-(2j_o-2j_i+1)n
\label{eq.mism}$$ in the argument of the cosine is a “mismatch” value in the angular directions.
The geometric interpretation of this equation: The minimum radius of the outer polygon is determined by the edge $j_o$ that first hits a nearby vertex $j_i$ while shrinking. The relevant index pair is the one that maximizes the cosine in the numerator, so the phase angle is steered towards zero or $2\pi$, equivalent to $j_i\phi_n\approx (j_0+1/2)\phi_m$. This means the relevant phase angle in the complex plane and viewing direction is where the vertex $j_i$ points near the middle of edge $j_o$.
Examples:
- In Fig. \[fig.Rpol37\] we have set $n=3$, $m=7$, $r_n^{(o)}=1$ and observe that $j_o=2$, $j_i=1$. Eq. (\[eq.rrat\]) obtains $r_m^{(o)}/r_n^{(o)}=\cos(\pi/21)/\cos(\pi/7) \approx 1.097519$.
- In the inner pair of Fig. \[fig.Rpol3to16\] we have set $n=3$, $m=4$, $r_n^{(o)}=1$ and observe that $j_o=1$, $j_i=1$. Eq. (\[eq.rrat\]) obtains $r_m^{(o)}/r_n^{(o)}= \sqrt{2}\cos(\pi/12)=(1+\sqrt{3})/2
\approx 1.366025$ [@Conwayarxiv98; @GirstmairAA81].
- In the inner pair of Fig. \[fig.Rpol3toPri\] we have set $n=3$, $m=5$, $r_n^{(o)}=1$ and observe that $j_o=1$, $j_i=1$. The equation yields $r_m^{(o)}/r_n^{(o)}=\cos(\pi/15)/\cos(\pi/5)=\frac{3-\sqrt{5}}{4}+\sqrt{3}\sqrt{\frac{5-\surd 5}{8}}
\approx 1.2090569$.
Numerical examples of the size ratios of are gathered in Table \[tab.reg\]. \[In solutions with interlaced circles—summarized in Section \[sec.hist\]—the ratio $r_m^{(o)}/r_n^{(o)}$ always equals $1/\cos(\phi_m/2)$ derived with Eqs. (\[eq.rni\]) and (\[eq.rno\]). This restricted search space would have put constant values down each column.\]
Where $m$ is a multiple of $n$, the table entries equal one. In these cases one can re-use $n$ vertices of the inner polygon as vertices of the outer polygon, and obtains its remaining $m-n$ vertices by regular subdivision of the angle, $
\phi(m) = \frac{\phi(n)}{m/n}.
$ The circumradii are the same, $r_m^{(o)}=r_n^{(o)}$, and their ratio equals one.
3 4 5 6 7 8 9 10
---- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
3 1.00000000 1.36602540 1.20905693 1.00000000 1.09751942 1.07313218 1.00000000 1.04570220
4 2.00000000 1.00000000 1.23606798 1.15470054 1.10991626 1.00000000 1.06417777 1.05146222
5 1.95629520 1.39680225 1.00000000 1.14837497 1.10544807 1.07905555 1.06158549 1.00000000
6 2.00000000 1.36602540 1.23606798 1.00000000 1.10991626 1.07313218 1.06417777 1.04570220
7 1.97766165 1.40532128 1.23109193 1.15147176 1.00000000 1.08068940 1.06285492 1.05040347
8 2.00000000 1.41421356 1.23606798 1.15470054 1.10991626 1.00000000 1.06417777 1.05146222
9 1.87938524 1.40883205 1.23305698 1.13715804 1.10853655 1.08136200 1.00000000 1.05082170
10 2.00000000 1.39680225 1.23606798 1.14837497 1.10991626 1.07905555 1.06417777 1.00000000
Consecutive side numbers
------------------------
![ Circumscribing a 3-gon by a 4-gon by a 5-gon …up to a 16-gon, all at concentric standard positions. []{data-label="fig.Rpol3to16"}](Rpol3to16)
The radial growth of a repeated, possibly infinite nesting is illustrated in Figure \[fig.Rpol3to16\]. The circumradius is the partial products of the first upper sub-diagonal of Table \[tab.reg\].
The heuristics is here with $m=n+1$ that
- for odd $n$, the edge $j_o=m/2-1$ hits the vertex $j_i=m/2-1$ with residual mismatch (\[eq.mism\]) equal to $-1$. \[This is the best possible absolute value because $2j_im$ is even and $(2j_o+1)n$ is odd then.\] So (\[eq.rrat\]) is $$r_{n+1}^{(o)}/r_{n}^{(o)}
=
\frac
{
\cos\frac{\pi}{n(n+1)}
}
{\cos \frac{\pi}{n+1}},\quad n\,\mathrm{odd},$$
- For even $n$ and $m=n+1$, the edge $j_o=n/2$ hits the vertex $j_i=n/2$ on the negative real axis with residual mismatch (\[eq.mism\]) equal to zero. So (\[eq.rrat\]) yields $$r_{n+1}^{(o)}/r_{n}^{(o)}
=
\frac
{
1
}
{\cos \frac{\pi}{n+1}}, n\,\mathrm{even}.$$
These two equations constitute the first upper diagonal of Table \[tab.reg\]. Fencing the polygons up to infinity defines the limiting radius as an alternating product of these two factors, $$\begin{aligned}
\frac{r_\infty^{(o)}}{r_3^{(o)}}
&=&
\frac{r_4^{(o)}}{r_3^{(o)}}
\times
\frac{r_5^{(o)}}{r_4^{(o)}}
\times
\frac{r_6^{(o)}}{r_5^{(o)}}
\times
\cdots
\nonumber\\
&=&
\prod_{n=3,5,7,\ldots} \frac{ \cos\frac{\pi}{n(n+1)} }{\cos\frac{\pi}{n+1}}
\prod_{n=4,6,8,\ldots} \frac{1}{\cos\frac{\pi}{n+1}}
\nonumber \\
&=&
\frac{\prod_{n=3,5,7,\ldots} \cos\frac{\pi}{n(n+1)}}
{\prod_{n=3}^\infty \cos\frac{\pi}{n+1}}
\nonumber \\
&=&
\frac{1}{2K'}
\prod_{n=3,5,7,\ldots} \cos\frac{\pi}{n(n+1)}
\nonumber \\
&\approx& 4.16674437148793
\label{eq.Couse}\end{aligned}$$ where we have inserted (\[eq.Co\]) and [@Finch][@EIS A085365] $$K'\equiv \prod_{n=3}^\infty \cos\frac{\pi}{n}
\approx
0.1149420448532962007.
\label{eq.Kprime}$$
In Figure \[fig.Rpol16to3\] the edge count of the circumscribed polygon is *decreased* from 16 to 3. The ratio of the circumradii of the 3-gon and the 16-gon is $\approx 6.2$ in the image. The observation is that here for $m=n-1$ and
- $n$ even, the outer edge $j_o=n/2-1$ and $j_i=n/2$. The value of (\[eq.rrat\]) is $$r_{n-1}^{(o)}/r_{n}^{(o)}
=
\frac
{
1
}
{\cos\frac{\pi}{n-1}}
,\quad n\, \mathrm{even}.$$
- whereas for $n$ odd $j_o=(n-3)/2$ and $j_i=(n-1)/2$. The value of (\[eq.rrat\]) is $$r_{n-1}^{(o)}/r_{n}^{(o)}
=
\frac
{
\cos\frac{\pi}{n(n-1)}
}
{\cos\frac{\pi}{n-1}}
,\quad n\,\mathrm{odd}$$
The alternating infinite product of these terms, the finite radius of the free inner region in Figure \[fig.Rpol16to3\] if inscribing indefinitely: $$\begin{aligned}
&& r_3^{(o)}/r_{\infty}^{(o)}
=
\frac{r_3^{(o)}}{r_4^{(o)}}
\times
\frac{r_4^{(o)}}{r_5^{(o)}}
\times
\frac{r_5^{(o)}}{r_6^{(o)}}
\times\cdots
\nonumber \\
&&=
\prod_{n=4,6,8,\ldots}
\frac{r_{n-1}^{(o)}}{r_n^{(o)}}
\times
\frac{r_{n}^{(o)}}{r_{n+1}^{(o)}}
\nonumber \\
&&=
\prod_{n=4,6,8\ldots}
\frac{1}{\cos\frac{\pi}{n-1}}
\times
\frac
{
\cos\frac{\pi}{(n+1)n}
}
{\cos\frac{\pi}{n}}
\nonumber \\
&& \approx 8.5526818319553
\label{eq.Kstd}
.\end{aligned}$$ This is slightly smaller than the equivalent polygon circumscribing constant $1/K' \approx 8.7000366\ldots$ [@EIS A051762][@WeissteinCRC p. 2300][@Finch; @GrimstoneMG64 p. 428] by the factor $$\prod_{n=4,6,8,\ldots} \cos\frac{\pi}{n(n+1)}\approx 0.98306273874458351\ldots
,
\label{eq.cosCorr}$$ based on (\[eq.Kprime\]). The areas have been smaller relative to the published construction with interspersed circles. The logarithm of the new constant is evaluated in Appendix \[sec.logcos\].
![ Circumscribing a 16-gon by a 15-gon by a 14-gon …down to a 3-gon, all with the same center. []{data-label="fig.Rpol16to3"}](Rpol16to3)
Prime side numbers
------------------
If $m$ is the next prime after $n$, the indices of the vertex of $n$ and edge of $m$ that describes the contact is irregular, see Fig. \[fig.Rpol3toPri\] and Table \[tab.rrpr\]. It is given by the pair $(j_o,j_i)$ which maximizes the value of (\[eq.rrat\]). The mismatch of (\[eq.mism\]), $2j_im-(2j_o+1)n$ cannot be nulled for odd primes $m$ and $n$ because $2j_im$ is even and $(2j_o+1)n$ is odd. But the value can apparently be forced to $\pm 1$ (where the sign is not important because this is an argument to the even function of the cosine), as demonstrated in Table \[tab.rrpr\]. \[The value of Eq. (\[eq.mism\]) is either $+1$ or $-1$ depending on whether the odd number $(2j_o-2j_i+1)n$ is to be incremented or decremented to reach a multiple of 4, since the prime gaps $d$ are even and the values of $2j_id$ are multiples of 4.\]
Assuming this $\pm 1$ heuristics if always correct, the circumcircle radius in Figure \[fig.Rpol3toPri\] grows to $$\begin{aligned}
\frac{r_\infty^{(o)}}{r_3^{(o)}}
&=&
\frac{\cos\frac{\pi}{nm}}{\cos\frac{\pi}{m}}
=
\frac{\cos\frac{\pi}{3\cdot 5}}{\cos\frac{\pi}{5}}
\times
\frac{\cos\frac{\pi}{5\cdot 7}}{\cos\frac{\pi}{7}}
\times
\frac{\cos\frac{\pi}{7\cdot 11}}{\cos\frac{\pi}{11}}
\times
\cdots
\nonumber \\
&=&
\frac{1}{2}
\frac{1}{\cos\frac{\pi}{3}}
\frac{\cos\frac{\pi}{3\cdot 5}}{\cos\frac{\pi}{5}}
\times
\frac{\cos\frac{\pi}{5\cdot 7}}{\cos\frac{\pi}{7}}
\times
\frac{\cos\frac{\pi}{7\cdot 11}}{\cos\frac{\pi}{11}}
\times
\cdots
\nonumber \\
&=&
\frac{1}{2K_p'}
\prod_{p_j\ge 3} \cos\frac{\pi}{p_jp_{j+1}}
\approx 1.5550895739...
\label{eq.ratPri}\end{aligned}$$ where $$K_p' = \prod_{p=3,5,7,11\ldots} \cos\frac{\pi}{p} \approx 0.3128329$$ is Kitson’s product over odd primes $p$ [@KitsonArxiv06; @KitsonMG92][@EIS A131671]. The infinite product of cosines in (\[eq.ratPri\]) is evaluated in (\[eq.Cprim\]) and smaller than unity, so the equation says that our construction squeezes the circumradius of the casting prime-sided regular polygons by more than a factor two compared to Kitson’s variant of construction.
Fig. \[fig.Rpol3toPri\] illustrates why: A non-zero mismatch angle reflects that no vertex of the inner polygon touches a midpoint of a side of the outer polygon; in consequence the circumradius of the inner polygon is larger than the inradius of the outer polygon for each individual pair of polygons.
$n$ $m$ $j_i$ $j_o$ $2j_im-(2j_o+1)n$
----- ----- ------- ------- -------------------
3 5 1 1 1
5 7 1 1 -1
7 11 1 1 1
11 13 3 3 1
13 17 5 6 1
17 19 4 4 -1
19 23 7 8 -1
23 29 2 2 1
29 31 7 7 -1
31 37 13 15 1
37 41 14 15 1
41 43 10 10 -1
43 47 16 17 -1
47 53 4 4 1
53 59 22 24 -1
59 61 15 15 1
61 67 5 5 -1
67 71 25 26 -1
71 73 18 18 1
: Vertex and edge indices $j_i$ and $j_o$ that maximize (\[eq.rrat\]) for adjacent primes $n$ and $m$, describing the concentric polygons in Fig. \[fig.Rpol3toPri\]. []{data-label="tab.rrpr"}
![ Circumscribing a 3-gon by a 5-gon by a 7-gon by a 11-gon etc up to a 541-gon, all with the same center, using all odd primes as edge numbers. []{data-label="fig.Rpol3toPri"}](Rpol3toPri)
If polygons with sides of odd prime numbers are stacked in *reverse* order, inscribing a 5-gon in a 3-gon, a 7-gon in the 5-gon, a 11-gon in the 7-gon etc., there is no substantial modification to the calculation, because interchanging the values of $n$ and $m$ in Table \[tab.rrpr\] appears to lead again to a list of $\pm 1$ in the mismatches. Now the ratio of the circumradius of the triangle divided by the radius of the circular inner hole is $$\begin{aligned}
\frac{r_3^{(o)}}{r_{\infty}^{(o)}}
=
\cdots \times
\frac{r_7^{(o)}}{r_{11}^{(o)}}
\times
\frac{r_5^{(o)}}{r_7^{(o)}}
\times
\frac{r_3^{(o)}}{r_5^{(o)}}
\nonumber \\
=
\frac{\prod_{p_j\ge 3}\cos\frac{\pi}{p_jp_{j+1}}}
{\prod_{p_j\ge 3} \cos\frac{\pi}{p_j}}
=
\frac{1}{K_p'}
\prod_{p_j\ge 3}\cos\frac{\pi}{p_jp_{j+1}},\end{aligned}$$ so there is a straight factor of 2 relative to the value in (\[eq.ratPri\]).
Concentric, Rotations allowed {#sec.rot}
=============================
If the outer polygon is rotated by an angle $\alpha_m$ relative to the standard position (\[eq.std\]), the vertices move to $$x+iy = r_m^{(o)} e^{2\pi j i/m+\alpha_m},\quad 0\le j< m.
\label{eq.rot}$$
In consequence, all three factors on the left hand side of (\[eq.shrink\]) are multiplied by $e^{i\alpha_m}$, Eq. (\[eq.jijo\]) obtains an additional factor $e^{-i\alpha_m}$ on the right hand side, and the phase shift finally enters Eq. (\[eq.rrat\]): $$\frac{r_m^{(o)}}{r_n^{(o)}}
= \frac{\cos[j_i\phi_n-(j_o+1/2)\phi_m-\alpha_m]}{\cos(\phi_m/2)}
.
\label{eq.ratWalpha}$$
The tightest solution for fixed $\alpha_m$ is represented by the pair $(j_i,j_0)$ which maximizes the value of $r_m^{(o)}/r_n^{(o)}$ and maximizes the value of the cosine in the numerator (because $m$ and the denominator are fixed). Shifts of $\alpha_m$ induce reduction of some peaks and rises of others in the bi-periodic domain spanned by the $j_i$ and $j_o$. The best solution is obtained where the value of the cosine becomes degenerate with highest multiplicity on the grid of the ($j_i$, $j_o$). In geometrical terms, rotated solutions seek to maximize the number of contact points between the two polygons, illustrated in Fig. \[fig.Rpol45a\] and \[fig.Rpol46a\].
![Tight 5-gon around the 4-gon in the standard placement, where the angle $\alpha_5$ in (\[eq.rot\]) is kept at zero, and another—tighter—solution where $\alpha_5$ is set to $-\pi/20=-9^\circ$ to yield a smaller 5-gon. The standard solution generates $r_5^{(o)}/r_4^{(o)}\approx 1.236$—see Table \[tab.reg\]— whereas the solution allowing rotation yields $r_5^{(o)}/r_4^{(o)}\approx 1.222$—see Table \[tab.rot\]. []{data-label="fig.Rpol45a"}](Rpol45a)
![Tight 6-gon around the 4-gon in the standard placement \[where the angle $\alpha_6$ in (\[eq.rot\]) is kept at zero\], and another tighter solution where $\alpha_6$ is set to $\pi/12=15^\circ$ to yield a smaller 6-gon. The standard placement achieves $r_6^{(o)}/r_4^{(o)}\approx 1.154$ according to Table \[tab.reg\], and the version allowing rotation achieves $r_6^{(o)}/r_4^{(o)}\approx 1.115$ reported in Table \[tab.rot\]. []{data-label="fig.Rpol46a"}](Rpol46a)
The interesting range is $0\le \alpha_m \le \min(\phi_m,\phi_n)$, because rotation of the inner polygon by integer multiples of $\phi_n$ or rotation of the outer polygon by integer multiples of $\phi_m$ leaves the graph invariant. \[Or, formally speaking, changes of $\alpha_m$ modulo $\phi_n$ or modulo $\phi_m$ can be absorbed into resetting the integers $j_i$ or $j_o$ in the numerator.\] The ratio $r_m^{(o)}/r_n^{(o)}$ exercises $m$ and also $n$ periods if $\alpha$ is turned through a full angle of $2\pi$, and contains therefore $\operatorname{lcm}(m,n)$ periods. ($\operatorname{lcm}$ is the least common multiple of both.)
Because $j_i\phi_n$ and $(j_o+1/2)\phi_m$ have an integer representation if measured in units of $\pi/(mn)$, because the cosine is a smooth function of its argument, because the periodicity with respect to $\alpha_m$ means its extremal values can only occur at multiples of half the period, and because $nm=\operatorname{lcm}(n,m)\gcd(n,m)$, we may encode all relevant angles as $\alpha_m \equiv s_{n,m}\pi/(mn)$ with integer-valued $s_{n,m}$. The phase angle in the numerator of (\[eq.ratWalpha\]) becomes $$\frac{\pi}{nm}[2j_im-(2j_0+1)n-s_{n,m}]
.
\label{eq.phass}$$
Investigation all possible pairs of polygons up to the $88$-gon leads to the following heuristics:
- If $n$ is odd, then $s_{n,m}=0$. \[Interpretation: the mismatch (\[eq.mism\]) is odd; no vertex points exactly to the center of an edge. This establishes the following stability/frustration argument: By the up-down symmetry of the graph, infinitesimal rotation of the inner polygon requires pushing at least that edge of the outer polygon outwards, which necessarily growth in size instead of shrinking as requested.\] This implies that neither the polygon pair in Figure \[fig.Rpol37\] nor the cascaded stack with the primal edges numbers in Figure \[fig.Rpol3toPri\] can be compressed by adding rotations.
- Periodicity: $s_{n,m}=s_{n,m+n}$. This seems to be a consequence of the modular property mentioned above; a change of $m$ by a multiple of $n$ is absorbed by modifying $j_i$ or $j_o$ by integer units. The $\gcd(n,m)$ (the period length of the cosine) is also preserved. Both aspects combined seem to freeze the number of the contacts between the two polygons.
- $s_{n,n}=0$. If the edge numbers are equal, the circumscribed polygon is a copy of the inscribed polygon.
- If $n$ is even,
- $s_{n,n/2}=n/2$. This says that an outer polygon with half as many vertices as the inner polygon may be constructed by outwards extension of one over the other edge of the inner polygon (which requires a rotation by half of the angle $\phi_m$ relative to the standard position). This achieves $r_m^{(i)}=r_n^{(i)}$.
- A half period exists with palindromic symmetry: $s(n,n/2+k)=s(n,n/2-k)$. Reason: The mirror symmetry of the standard placement leads to equivalent solutions if the outer polygon is rotated either clockwise or counter-clockwise. Solutions are *even* functions of $\alpha_m$, so sign flips of $s_{n,m}$ are irrelevant. The half period then results from a general property of (Fourier series of) periodic even functions.
Consuming these rules, we need to tabulate the $s_{n,m}$ only in the triangle of even $n$ with $0\le m\le n/2$ for a full coverage. Then
- If $m\le n/2$ is odd and $n$ is even, then $s(n,m)=\gcd(n/2,m)$.
- If $m< n/2$ is even, and
- $n$ is two times an odd number, $s_{n,m}=0$.
- $n$ is two times an even number, $s_{n,m}=2\gcd(n/2,m/2)$. This selection is apparently aligning edge 0 of the circumscribing polygon parallel to edge 0 of the inscribed polygon with the aim to increase the number of contacts to a multiple of four, similar to what is observed in Fig. \[fig.Rpol46a\].
As an application, the cumulative wrench angle of the vertex direction of the outermost regular polygon in Figure \[fig.Rpol3to16a\] relative to its position in Figure \[fig.Rpol3to16\] is calculated as $$\begin{aligned}
\sum_{n\ge 3,m=n+1} \pi\frac{s_{n,m}}{nm}
=
\pi \sum_{n=2,3,4,\ldots} \frac{s_{2n,2n+1}}{2n(2n+1)}
\nonumber \\
=
\pi \sum_{n=2,3,4,\ldots} \frac{s_{2n,1}}{2n(2n+1)}
=
\pi \sum_{n=2,3,4,\ldots} \frac{1}{2n(2n+1)}
\nonumber \\
=
\pi[\frac{5}{6}-\log(2)]\approx 25.23^\circ.\end{aligned}$$
![ Concentric encircling the 3-gon by a 4-gon by a 5-gon and so on as in Figure \[fig.Rpol3to16\], but minimizing the areas from the 4-gon upwards by rotating these polygons by variation of $\alpha_m$. []{data-label="fig.Rpol3to16a"}](Rpol3toa)
Another heuristic observation with this rule is that the absolute value of the mismatch (\[eq.mism\]) for $m=n+1$ is kept at 1 if $n$ is odd and at 2 if $n$ is even. The growth of the radius in Figure \[fig.Rpol3to16a\] is limited to the infinite product of terms of the form (\[eq.ratWalpha\]), $$\begin{aligned}
\frac{r_{\infty}^{(o)}}
{r_3^{(o)}}
=
\frac{\cos\frac{2\pi}{3\cdot 4}}{\cos\frac{\pi}{4}}
\times \frac{\cos\frac{\pi}{4\cdot 5}}{\cos\frac{\pi}{5}}
\times \frac{\cos\frac{2\pi}{5\cdot 6}}{\cos\frac{\pi}{6}}
\times \frac{\cos\frac{\pi}{6\cdot 7}}{\cos\frac{\pi}{7}}
\times\cdots
\nonumber \\
=
\frac{1}{2K'}
\prod_{k=3,5,7,\ldots}\cos\frac{2\pi}{k(k+1)}
\prod_{k=4,6,8,\ldots}\cos\frac{\pi}{k(k+1)}
\nonumber \\
=
\frac{1}{2K'\cos\frac{\pi}{2\cdot 3}}
C_e
\prod_{k=3,5,7,\ldots}\cos\frac{2\pi}{k(k+1)}
\nonumber \\
=
\frac{1}{\sqrt{3}K'}
C_e
\prod_{k=3,5,7}\cos\frac{2\pi}{k(k+1)}
\approx
3.5809046865583,
\label{eq.roteoInf}\end{aligned}$$ where $K'$, $C_e$ and the infinite product are taken from (\[eq.Kprime\]), (\[eq.Ce\]) and (\[eq.cos2pi\]). This circumradius including rotations is considerably smaller than the circumradius (\[eq.Couse\]) in the standard positions.
3 4 5 6 7 8 9 10
---- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
3 1.00000000 1.36602540 1.20905693 1.00000000 1.09751942 1.07313218 1.00000000 1.04570220
4 1.93716632 1.00000000 1.22204076 1.11535507 1.10348396 1.00000000 1.06044555 1.03851698
5 1.95629520 1.39680225 1.00000000 1.14837497 1.10544807 1.07905555 1.06158549 1.00000000
6 1.73205081 1.36602540 1.22929667 1.00000000 1.10681271 1.07313218 1.04801052 1.04570220
7 1.97766165 1.40532128 1.23109193 1.15147176 1.00000000 1.08068940 1.06285492 1.05040347
8 1.98422940 1.30656296 1.23255619 1.14559538 1.10830702 1.00000000 1.06324431 1.04847492
9 1.87938524 1.40883205 1.23305698 1.13715804 1.10853655 1.08136200 1.00000000 1.05082170
10 1.98904379 1.39680225 1.17557050 1.14837497 1.10879865 1.07905555 1.06352950 1.00000000
Translated Centers {#sec.trans}
==================
A glance at Figure \[fig.Rpol3to16\] or \[fig.Rpol45a\] for example reveals that further compression of the outer polygon would be possible if either one is shifted sideways, giving up the requirement that the two polygons be concentric.
In the complex plane this adds a displacement $z_m^{(o)}$ of the outer polygon as a new parameter to Eq. (\[eq.shrink\]): $$\begin{aligned}
&&z_m^{(o)}
+
r_{m}^{(o)}
\{e^{2\pi i j_o/m}+t[
e^{2\pi i (j_o+1)/m}
-e^{2\pi i j_o/m) }]\}
\nonumber \\
&&=
r_{n}^{(o)}e^{2\pi i j_i/n}
.\end{aligned}$$ Assuming that $z_m^{(o)}$ is real-valued (that center shifts are sideways), this can be written as $$\frac{z_m^{(o)}}{r_m^{(o)}}\cos(j_o\phi_m)
+
1+t[
\cos(\phi_m)
-1]
=
\frac{r_{n}^{(o)}}{r_m^{(o)}}\cos(j_i\phi_n-j_o\phi_m)
.$$ $$-\frac{z_m^{(o)}}{r_m^{(o)}}\sin(j_o\phi_m)
+ t \sin(\phi_m)
=
\frac{r_{n}^{(o)}}{r_m^{(o)}}\sin(j_i\phi_n-j_o\phi_m)
.$$ We do not discuss this parameter space systematically or solutions obtained by combined translations and rotations, but merely illustrate this aspect by the simplest examples:
- The triangle $n=3$ could touch the quadrangle $m=4$ in Figure \[fig.Rpol3to16\] at its right vertex on the horizontal axis, at $j_o=j_i=t=0$, which gives $$\frac{z_m^{(o)}}{r_m^{(o)}}
+ 1
=
\frac{r_{n}^{(o)}}{r_m^{(o)}}
\label{eq.trStd}$$ and $0=0$ for the imaginary part. (This equation and overlap of the two vertices is possible whenever $m>n$.) The other two vertices of the triangle would stay glued to the quadrangle’s sides, one point at $j_i=1=j_o$, which is $$1-t
=
\frac{r_{n}^{(o)}}{r_m^{(o)}}\cos(2\pi/12)
.$$ $$-\frac{z_m^{(o)}}{r_m^{(o)}}
+ t
=
\frac{r_{n}^{(o)}}{r_m^{(o)}}\sin(2\pi/12)
.$$ The solution to these three linear equations for the three unknown $t$, $z_m^{(o)}/r_m^{(o)}$ and $r_n^{(o)}/r_m^{(o)}$ is $$\begin{aligned}
\frac{z_m^{(o)}}{r_m^{(o)}} = 1-\frac{2}{\surd 3}\approx -0.1547005,
\\
\frac{r_n^{(o)}}{r_m^{(o)}}=2(1-\frac{1}{\sqrt{3}})\approx 0.84529946.\end{aligned}$$ Therefore $r_m^{(o)}/r_n^{(o)}\approx 1.1830127$ which is indeed smaller than $r_4^{(o)}/r_3^{(o)}$ in Tables \[tab.reg\] and \[tab.rot\].
- For $n=4$, $m=3$, the 3-gon circumscribing the 4-gon in Figure \[fig.Rpol16to3\], the shift leads for the contact on the negative real line where $t=1/2$, $j_o=(m-1)/2$, $j_i=n/2$ to $$-\frac{1}{2}\frac{z_m^{(o)}}{r_m^{(o)}}
+\frac{1}{4}=
\frac{1}{2}\frac{r_n^{(o)}}{r_m^{(o)}}
.$$
\[The generic equation if $n$ is even, $m$ is odd and $n>m$ is $$-\frac{z_m^{(o)}}{r_m^{(o)}}\sin(\frac{\phi_m}{2})
+\frac{1}{2}\sin(\phi_m)=
\frac{r_n^{(o)}}{r_m^{(o)}}\sin(\frac{\phi_m}{2})
.$$ \] Two more equations are established by $j_i=1$, $j_o=0$ if the upper and lower vertex of the quadrangle meets the other two edges of the triangle, namely $$\begin{aligned}
\frac{z_m^{(o)}}{r_m^{(o)}} +1-\frac{3}{2}t=0;
\\
\frac{\sqrt{3}}{2}t=
\frac{r_n^{(o)}}{r_m^{(o)}}\end{aligned}$$ such that $$\frac{r_n^{(o)}}{r_m^{(o)}}=\frac{3}{4}(\sqrt{3}-1)
\approx 1/1.821367$$ with center displacement $$\frac{z_m^{(o)}}{r_m^{(o)}}
=\frac{1}{4}(5-3^{3/2}).$$
- For the 5-gon circumscribing the 3-gon in Figure \[fig.Rpol3toPri\] the shift leads for the contact on the positive real axis again to (\[eq.trStd\]) plus two equations established by $j_i=j_o=1$, such that $$\begin{aligned}
\frac{r_n^{(o)}}{r_m^{(o)}}=
-\frac{2}{3}-\frac{4}{3}\cos\frac{2\pi}{5}
+\frac{2}{3}\cos\frac{2\pi}{15}
+2\cos\frac{4\pi}{15}
\nonumber \\
\approx 1/1.1512750\end{aligned}$$ with center displacement $$\frac{z_m^{(o)}}{r_m^{(o)}}
=
-\frac{5}{3}-\frac{4}{3}\cos\frac{2\pi}{5}
+\frac{2}{3}\cos\frac{2\pi}{15}
+2\cos\frac{4\pi}{15}
.$$
Summary
=======
We have defined and computed the smallest ratio of the circumradii of a pair of non-overlapping concentric regular polygons, and have pointed at infinite products of cosines that arise if some infinite sets of regular polygons are nested defined by simple strides in the sets of side numbers.
Quenching Factor of the Kepler-Bouwkamp Constant {#sec.logcos}
================================================
Even lower term in the product
------------------------------
The constant (\[eq.cosCorr\]) is approached by calculating its logarithm (and including one more term to put the result into a more general perspective) [@StephensMG79], $$\begin{aligned}
\log \prod_{n=2,4,6,8,\ldots} \cos\frac{\pi}{n(n+1)}
\nonumber\\
=
\sum_{n=2,4,6,8,\ldots} \log\cos\frac{\pi}{n(n+1)}
\nonumber\\
\approx -0.160923373349205036366901529
\label{eq.numlog}\end{aligned}$$ via the associated Taylor series [@EIS A046991][@GR 1.518] $$\log\cos\epsilon = -\frac{\epsilon^2}{2}-\frac{\epsilon^4}{12}-\frac{\epsilon^6}{45}-\frac{17\epsilon^8}{2520}
-\frac{31\epsilon^{10}}{14175}
-\frac{691\epsilon^{12}}{935550}
-\cdots
\label{eq.taylogcos}$$ as follows: $$\begin{aligned}
&&-\log \prod_{n=2,4,6,8,\ldots} \cos\frac{\pi}{n(n+1)}
\nonumber\\
=
\sum_{k=1}^\infty
&&
[
\frac{\pi^2}{2(2k)^2(2k+1)^2}
+\frac{\pi^4}{12(2k)^4(2k+1)^4}
\nonumber\\
&&
+\frac{\pi^6}{45(2k)^6(2k+1)^6}
+\frac{17\pi^8}{2520(2k)^8(2k+1)^8}+\cdots
]
\label{eq.logcost}
.\end{aligned}$$ Partial fraction decompositions of the individual terms have the following format [@GR 2.102] [@MahoneyJCAM9; @VellemanAMM109; @Xinarxiv04; @EusticeAMM86] $$\frac{1}{n^2(n+1)^2}
=\frac{1}{n^2}+\frac{1}{(n+1)^2}-\frac{2}{n(n+1)},$$ $$\begin{aligned}
\frac{1}{n^4(n+1)^4}
=
\frac{1}{n^4}+\frac{1}{(n+1)^4}
-\frac{4}{n^3}+\frac{4}{(n+1)^3}
\nonumber\\
+\frac{10}{n^2}+\frac{10}{(n+1)^2}
-\frac{20}{n(n+1)},\end{aligned}$$ $$\frac{1}{n^{2s}(n+1)^{2s}}
=
\sum_{t=1}^{2s}\
\binom{4s-t-1}{2s-1}
\left[\frac{(-)^t}{n^t}+\frac{1}{(n+1)^t}\right].$$
Sums of reciprocal powers of the even or odd integers are in terms of Riemann’s $\zeta$-function [@GR 0.233][@Jolley (335)] $$\sum_{k=1}^\infty \frac{1}{(2k)^t}
=
\frac{1}{2^t}\zeta(t),$$ and $$\sum_{k=1}^\infty \frac{1}{(2k+1)^t}
=
[1-\frac{1}{2^t}]\zeta(t) -1.$$ Combining the previous three equations generates (with a little extra care at $t=1$ [@GR 0.234]) $$\begin{aligned}
&&T_e(2s)\equiv \sum_{k=1}^\infty \frac{1}{(2k)^{2s}(2k+1)^{2s}}
\nonumber\\
&&=
\sum_{k=1}^\infty \sum_{t=1}^{2s}
\binom{4s-t-1}{2s-1}
\left[
\frac{(-)^t}{(2k)^t}+\frac{1}{(2k+1)^t}
\right]
\nonumber\\
&&=
\sum_{k=1}^\infty \bigg\{
-\frac{\binom{4s-2}{2s-1}}{(2k)(2k+1)}
\nonumber\\
&&
\qquad +\sum_{t=2}^{2s}
\binom{4s-t-1}{2s-1}
\left[
\frac{(-)^t}{(2k)^t}+\frac{1}{(2k+1)^t}
\right]
\bigg\}
\nonumber\\
&&=
-\binom{4s-2}{2s-1}[1-\ln 2]
\nonumber\\
&&
\qquad
+\sum_{k=1}^\infty \sum_{t=2}^{2s}
\binom{4s-t-1}{2s-1}
\left[
\frac{(-)^t}{(2k)^t}+\frac{1}{(2k+1)^t}
\right]
\nonumber\\
&&=
\sum_{t=1}^{2s}
\binom{4s-t-1}{2s-1}
\left\{
-1 +[1-\frac{1-(-)^t}{2^t}]\zeta(t)
\right\}
.
\label{eq.ksum}\end{aligned}$$ For odd $t$, $[1-\frac{1-(-)^t}{2^t}]\zeta(t)$ equals Dirichlet’s $\eta$-function, in particular $\eta(1)=\log 2$ at the pole of $\zeta(1)$ [@AS Tab. 23.3].
The three base examples of this format are: $$\begin{aligned}
&&\sum_{k=1}^\infty
\frac{1}{(2k)^2(2k+1)^2}
=-3+\frac{\pi^2}{6}+2\log 2
\nonumber\\
&&\approx 0.03122842796811705530687941 ;\end{aligned}$$ $$\begin{aligned}
&&\sum_{k=1}^\infty
\frac{1}{(2k)^4(2k+1)^4}
=-35+\frac{\pi^4}{90}+3\zeta(3)+\frac{5\pi^2}{3}+20\log 2
\nonumber\\
&&\approx 0.00077822287109160078401223 ;\end{aligned}$$ $$\begin{aligned}
&&\sum_{k=1}^\infty
\frac{1}{(2k)^6(2k+1)^6}
=-462+21\pi^2 +42\zeta(3)
\nonumber\\
&&
+252 \log 2
+\frac{7\pi^4}{30}+\frac{\pi^6}{945}
+\frac{45}{8}\zeta(5)
\nonumber\\
&&\approx 0.0000214492855159526203348.\end{aligned}$$ Interchange of the summation over the Taylor orders and over the $k$ and insertion of (\[eq.ksum\]) into (\[eq.logcost\]) leads to the value (\[eq.numlog\]). Exponentiation gives $$\begin{aligned}
C_e \equiv
\prod_{n=2,4,6,\ldots}^\infty \cos\frac{\pi}{n(n+1)}
\nonumber\\
\approx 0.85135730526671405636170.
\label{eq.Ce}\end{aligned}$$
Odd lower term in the product
-----------------------------
In (\[eq.numlog\]), the smaller factor in the product $n(n+1)$ was always even. With exactly the same technique we obtain a “complete” version of (\[eq.ksum\]) where the smaller term steps through all positive integers: $$\begin{aligned}
&&T(2s)\equiv \sum_{n=1}^\infty \frac{1}{n^{2s}(n+1)^{2s}}
\nonumber\\
&&=
\sum_{t=1}^{2s}\binom{4s-t-1}{2s-1}
\left\{
[1+(-)^t]\zeta(t)-1
\right\}.\end{aligned}$$ Here $[1+(-1)^t]\zeta(t)$ is to be interpreted as $0$ if $t=1$, ignoring the pole of $\zeta$. The three basic examples are $$\begin{aligned}
&&\sum_{n=1}^\infty \frac{1}{n^2(n+1)^2} = \frac{\pi^2}{3}-3
\nonumber\\
&&\approx 0.2898681336964528729448303333 ;\end{aligned}$$ $$\begin{aligned}
&&\sum_{n=1}^\infty \frac{1}{n^4(n+1)^4} = -35+\frac{10\pi^2}{3}+\frac{\pi^4}{45}
\nonumber\\
&&\approx 0.0633278043868051124803107260 ;\end{aligned}$$ $$\begin{aligned}
&&\sum_{n=1}^\infty \frac{1}{n^6(n+1)^6} = -462+42\pi^2+\frac{7\pi^4}{15}+\frac{2\pi^6}{945}
\nonumber\\
&&\approx 0.0156467855897643141498131091.\end{aligned}$$ The complete version of (\[eq.numlog\]) does not exist because the term at $n=1$ contributes $\log \cos (\pi/2) = -\infty$. We drop this term at $n=1$ and use $T(2s)-2^{-2s}$ with (\[eq.taylogcos\]) to compute $$\log \prod_{n=2}^\infty \cos \frac{\pi}{n(n+1)}
\approx
-0.2039684236116246918364049,$$ and its exponential value $$C\equiv \prod_{n=2}^\infty \cos \frac{\pi}{n(n+1)}
\approx
0.815488120950370848344387.
\label{eq.C}$$
If the smaller factor in $n(n+1)$ is odd, the difference is involved: $$T_o(2s) \equiv \sum_{n=1,3,5,\ldots}^\infty \frac{1}{n^{2s}(n+1)^{2s}}
=T(2s)-T_e(2s).$$ Division of (\[eq.C\]) through (\[eq.Ce\]) yields the complement $$\begin{aligned}
&& C_o \equiv \prod_{n=3,5,7\ldots}^\infty \cos \frac{\pi}{n(n+1)}
=\frac{C}{C_e}
\nonumber \\
&& \approx
0.95786823687957188013580826171688
\label{eq.Co}\end{aligned}$$ for use in (\[eq.Couse\]).
Numerator $2\pi$ with odd lower term
------------------------------------
The factor $2\pi$ in the numerator of (\[eq.roteoInf\]), $$\prod_{k=3,5,7\ldots}\cos\frac{2\pi}{k(k+1)}
=
\prod_{n=1}^\infty \cos\frac{\pi}{(2n+1)(n+1)}$$ causes slow convergence of the methods shown above. An acceleration method with deferred summation is available [@SebahGourdon]: The partial product up to some $n\le M$ is calculated explicitly, and the logarithm of the remaining infinite product is expanded in a Taylor series in $1/n$: $$\begin{aligned}
&& \sum_{n>M} \log \cos\frac{\pi}{(2n+1)(n+1)}
\nonumber\\
&&
=\sum_{n>M} [
-\frac{\pi^2}{8}\frac{1}{n^4}
+\frac{3\pi^2}{8}\frac{1}{n^5}
-\frac{23\pi^2}{32}\frac{1}{n^6}
+\frac{9\pi^2}{8}\frac{1}{n^7}+\cdots
]
\label{eq.cos2pT}\end{aligned}$$ Each term on the right hand side is then replaced by an incomplete $\zeta$-function, $$\sum_{n>M} \frac{1}{n^s} = \zeta(s)-\sum_{n=1}^M\frac{1}{n^s}.$$ With $M=10$ and (\[eq.cos2pT\]) expanded up to $O(n^{-30})$ we obtain for example $$\prod_{k=3,5,7\ldots}\cos\frac{2\pi}{k(k+1)}
\approx
0.8373758680415481080004775.
\label{eq.cos2pi}$$
Quenching Factor of Kitson’s Constant {#sec.logp}
=====================================
The logarithm of the product in (\[eq.ratPri\]) is the a sum over all odd primes $p_j\ge 3$: $$\log \prod_{p_j\ge 3} \cos\frac{\pi}{p_jp_{j+1}}
=
\sum_{p_j\ge 3} \log \cos\frac{\pi}{p_jp_{j+1}}$$ Again with (\[eq.taylogcos\]) we evaluate $$T(2s)\equiv \sum_{p_j} \frac{1}{p_j^{2s}p_{j+1}^{2s}}$$ for integer $s$. $$\begin{aligned}
T(2)&\approx & 0.005519522774559;
\\
T(4)&\approx & 0.0000204508599535;
\\
T(6)&\approx & 0.000000088340410739027;
\\
T(8)&\approx & 0.000000000390629312549651477,\end{aligned}$$ so $$\sum_{p_j\ge 3} \log\cos\frac{\pi}{p_jp_{j+1}} \approx -0.02740567$$ and after exponentiation $$\begin{aligned}
\prod_{p_j\ge 3} \cos\frac{\pi}{p_jp_{j+1}}
\approx 0.9729664541346255360938192\ldots
\label{eq.Cprim}\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Coherent scattering of solar, atmospheric and diffuse supernovae neutrinos creates an irreducible background for direct dark matter experiments with sensitivities to WIMP-nucleon spin-independent scattering cross sections of $10^{-46}$-$10^{-48}$ cm$^2$, depending on the WIMP mass. Even if one could eliminate all other backgrounds, this “neutrino floor” will limit future experiments with projected sensitivities to cross sections as small as $10^{-48}$ cm$^2$. Direction-sensitive detectors have the potential to study dark matter beyond the neutrino bound by fitting event distributions in multiple dimensions: recoil kinetic energy, recoil track angle with respect to the sun, and event time. This work quantitatively explores the impact of direction sensitivity on the neutrino bound in dark matter direct detection.'
author:
- Philipp Grothaus
- Malcolm Fairbairn
- Jocelyn Monroe
bibliography:
- 'nulimit.bib'
title: Directional Dark Matter Detection Beyond the Neutrino Bound
---
\[sec:introduction\]Introduction
================================
Dark matter comprises approximately 25% of the energy density of the Universe [@spergel2003; @Ade2013], yet its particle properties are unknown. There are a large class of dark matter candidates with masses and interaction energy scales expected to appear just beyond the electroweak scale [@snowmass2013]. Such models are interesting since they can lead naturally to an abundance of dark matter in agreement with cosmology. They also suggest cross sections for scattering off of nuclei which are within reach of current and near-future direct detection experiments.
The precise predictions for scattering cross sections with nuclei in these models vary a great deal and even within a given model the scattering strongly depends upon the values of the input parameters, such that the range for the WIMP-nucleon interaction cross section $\sigma_p$ spans many orders of magnitudes. Very small WIMP-nucleon cross sections can arise in models in which the dark matter candidate is a mixed state such that small mixing angles may suppress the interaction, or when the self annihilation cross section of dark matter in the Early Universe (which determines relic density today) is enhanced via kinematics rather than couplings. Another possibility is that different contributions to the interaction of the WIMP with nucleons cancel each other for specific choices of input parameters.
These possibilities require direct dark matter searches that are sensitive to very small WIMP-nucleon cross sections where, as we will see, coherent neutrino-nucleus scattering will become a problem [@drukier1986; @Cabrera1984; @Monroe2007; @Strigari2009].
Direct detection experiments search for dark matter particles using the coherent elastic scattering process. Neutrinos also interact coherently with atomic nuclei, causing the nucleus to recoil with energies up to tens of keV. Such recoils would be indistinguishable from dark matter interactions individually. The scale of the ambient neutrino flux in this energy range is $10^6$ cm$^{-2}$s$^{-1}$, and the coherent neutrino-nucleus cross section is of order 10$^{-39}$ cm$^2$. Background interactions due to these neutrinos represent a lower “neutrino bound” on the achievable sensitivity of dark matter direct detection experiments [@Gutlein2010; @Billard2013].
Discovering dark matter with a cross section close to or below the neutrino limit will be difficult. Equivalently, if we discover dark matter relatively soon with cross sections far above the neutrino limit we will still want to build larger detectors to study the dark matter in more detail, in which case the precision of such measurements will be limited by background neutrinos.
In this paper, we estimate the impact of backgrounds in dark matter detectors caused by coherent neutrino-nucleus elastic scattering of ambient solar, atmospheric and supernovae neutrinos, taking into account recoil energy, direction and time modulation sensitivity. We calculate probability distribution functions for the dark matter signal and the neutrino background in the dimensions of recoil energy, recoil direction, and event time. We find that direction sensitivity adds approximately an order of magnitude sensitivity beyond nondirectional searches for light dark matter, and depending on the target species and energy threshold, this sensitivity can leap far beyond the solar neutrino bound.
\[sec:detectors\]Dark Matter Detection
======================================
Direct detection experiments seek to look for the signal of dark matter particles via their elastic scattering interactions with detector nuclei [@gaitskell2004].
Recent experiments have limited the magnitude of the scattering cross section to be less than approximately $10^{-45}$ cm$^2$ [@xenon100; @LUX]. This corresponds roughly to one event per 100 kilograms of detector fiducial mass per day of detector live time. The next generation of ton-scale plus experiments are expected to increase this sensitivity by two or three orders of magnitude.
Directional detection experiments measure both the energy and track direction of the recoil nuclei. To measure the direction of such low-energy tracks, gas targets are used at pressures of 0.05-0.1 atmospheres, with high-density readout of charge and optical signals [@Ahlen2009]. Current directional detectors are at the research and development stage, and in small prototypes have demonstrated energy thresholds of a few keV, and at higher thresholds (50-100 keV) have demonstrated angular resolution of 30-55 degrees [@Ahlen2009; @Ahlen:2010ub; @Santos:2011xk]. For the studies here we use the event angle $\theta_{\rm sun}$, which is defined as the angle between the recoiling nucleus track and the Earth-Sun direction.
Earlier work on directional dark matter detection has shown how dark matter properties can be constrained [@Billard2010], how exclusion limits with directional detectors may be set [@Billard2010b], how well a dark matter signal may be distinguished from an isotropic background [@Morgan2004], or how the dark matter velocity distribution may be tested [@Lee2012]. We will investigate here the implications of directional dark matter detectors for dark matter searches in the presence of neutrino background.
\[sec:signal\_xsec\] Dark Matter Scattering Cross-Section
----------------------------------------------------------
The current approach in direct dark matter detection to test the wide range of theoretical models for dark matter is to measure the event rate of dark matter particles scattering off of target nuclei [@Goodman1984].
The zero momentum WIMP-nucleus cross section is given by $$\sigma_0=\frac{4\mu_T^2}{\pi}\left( Z f_p + (A-Z) f_n \right)^2~,$$ where $f_p$ and $f_n$ are the couplings of the dark matter particle to the proton and neutron, respectively, $\mu_T$ is the dark matter-nucleus reduced mass, $A$ the atomic number and $Z$ the number of protons of the target nucleus. We assume $f_p$ and $f_n$ to be approximately equal such that the estimation $\sigma_0 \propto A^2 f_p^2 \mu_T^2$ holds. Then, we can cast $\sigma_0$ into the WIMP-proton cross section $\sigma_p$ via $\sigma_0 = \sigma_p \left( \mu_T / \mu_p \right)^2 A^2$ and use the event rate to constrain $\sigma_p$.
\[sec:phase\_dist\] Dark Matter Velocity Distribution
-----------------------------------------------------
We assume a local density for the dark matter of 0.3 GeV cm$^{-3}$ which is in agreement with current astrophysical values [@fairbairn2013].
We assume that the WIMPs have a Maxwellian distribution $f(\vec{v})$ with a cut off at the halo escape velocity $v_{\rm esc} = 544$ km/s. It is well known that limits for light dark matter depend strongly on astrophysical uncertainties, but considering these is not part of this paper. If $|\vec{v}|<v_{\rm esc}$, the distribution in the halo rest frame is $${\label{eq:vel_dis}}
f(\vec{v})_{halo} = \frac{1}{N_{\rm esc}}\left(\frac{3}{2\pi\sigma^2_v}\right)^{3/2}\exp\left[-\frac{3\left(\vec{v}\right)^2}{2\sigma^2_v}\right] ~,$$ with $N_{esc}=\rm{erf}(z)-2z~\rm{exp}(-z^2)/\sqrt{\pi}$ accounting for the truncation. We have $z=v_{\rm esc}/\bar{v}$ and $\bar{v}=220$ km/s as the most probable WIMP velocity, which is related to the width of the distribution via $\sigma_v=\sqrt{3/2}\bar{v}$. For $|\vec{v}|>v_{\rm esc}$ we assume that $f(\vec{v})$ vanishes.
In the lab frame, we need to take into account the Earth’s overall velocity vector which has contributions from the Sun’s movement around the Galactic center, the peculiar movement of the Sun relative to the local standard of rest and the Earth’s velocity vector relative to the sun which changes throughout the year. The velocity distribution is therefore time dependent. A detailed description of the Earth’s overall velocity vector that has been used in this work can be found in [@McCabe2013; @Lee2013]. We integrate this time dependence over the exposure time from $t_0$ to $t_1$ to account for the annual modulation in the event rate.
\[sec:signal\_dist\] Dark Matter Signal Distribution
----------------------------------------------------
The stronger the scattering cross section, the larger the event rate in an experiment. The differential rate is $$\frac{dR_{\rm DM}}{dE_r} = M_{\rm det} ~\frac{\rho_0 \sigma_0}{2m_{\rm DM}\mu^2_T}F^2(E_r) \int_{t_0}^{t_1} \int_{v_{\rm min}}^{\infty}\frac{f(\vec{v},t)}{v}d^3v~dt~.$$ Here, $M_{\rm det}$ is the detector mass, $\rho_0$ the local dark matter density, $m_{\rm DM}$ the dark matter mass, $f(\vec{v},t)$ the dark matter velocity distribution in Earth’s frame of reference and $F(Q)$ the form factor which describes the distribution of weak hypercharge within the nucleus. The form factor depends on the momentum transfer squared, $Q^2=2 m_T E_{\rm r}$. In this work we use Helm form factors, see e.g. [@lewin1995] or [@horowitz1981]. To obtain the differential rate, an integral over this velocity distribution must be performed from a minimum velocity $v_{\rm min}=\sqrt{2 E_{\rm min}/m_{\rm DM}}$ that depends on the recoil energy through the minimal WIMP energy $E_{\rm min}=E_{\rm r} (m_{\rm DM}+m_T)^2/(4 m_{\rm DM} m_T)$ necessary to obtain such a recoil energy $E_{\rm r}$.
The number of dark matter events is calculated as an integral over the differential rate and the energy-dependent detection efficiency $\epsilon(E_{\rm r})$ of an experiment: $${\label{eq:rate_dm}}
s=\int_{E_{\rm thr}}^{E_{\rm up}} \epsilon(E_{\rm r}) \frac{dR_{\rm DM}}{dE_{\rm r}} dE_{\rm r} ~.$$ If a dark matter particle with kinetic energy $E_{\rm DM}$ scatters off a target nucleus with scattering angle $\theta$ with respect to its incoming direction, the resulting recoil energy is $${\label{eq:dm_Erec}}
E_{\rm r} = E_{\rm DM} r (1-\cos \theta)/2 ~,$$ with $r=4m_{\rm DM} m_T /(m_{\rm DM} + m_T)^2$. In this work we assume isotropic scattering in $\cos \theta$. The scattering angle of the recoiling nucleus with respect to the incoming dark matter velocity is then given by $$\tan \theta'= \frac{p' \sin \theta}{\sqrt{2 m_{\rm DM} E_{\rm DM}}-p' \cos \theta} ~,$$ with $p'=\sqrt{2m_{\rm DM} E_{\rm DM} - 2 m_T E_{\rm r}}$.
![\[fig:dm\_pdf\] Two dimensional dark matter probability distribution $\rho$ of recoil energy and event angle for a 6 GeV dark matter particle in a CF$_4$ detector with 5 keV threshold in September.](plots/pdf_dm.pdf){height="7.cm"}
Figure \[fig:dm\_pdf\] shows the two dimensional probability distribution of event angle and recoil energy in a tetrafluoromethane, CF$_4$, detector with 5 keV energy threshold for a 6 GeV dark matter particle. Two distinct features should be noted. First, the event angles of dark matter scattering events preferably lie at large $\cos~\theta_{\rm sun}$ (small angles) because there is more solid angle (on the sphere) there. Second, the probability distribution drops to zero above the largest possible recoil energy for the given dark matter mass and escape velocity. The power of directionality is that dark matter masses that create an energy spectrum very similar to the neutrino background can easily be distinguished when the event angle is taken into account. As we will see, for light dark matter a strong gain in sensitivity compared to nondirectional detectors is therefore expected.
A third feature that is not directly visible in figure \[fig:dm\_pdf\], but is important nonetheless, is a variation of the peak of the dark matter probability distribution in time. The direction of the Earth’s overall velocity vector will point approximately towards the radio galaxy Cygnus A [^1], such that the incoming dark matter particles in the lab frame will have a preferred direction coming from Cygnus A. The relative angle between the Sun and Cygnus A changes over the year, such that the peak in the dark matter probability distribution will follow a similar pattern.
The annual modulation in the event rate of light dark matter has a maximum in June because at this time the velocity vector of the Earth and the Sun are parallel to each other [@freeseannualmod]. Both vectors approximately point into the direction of Cygnus A. In December, these two vectors are antiparallel resulting in a minimum of the event rate. The angle between the Earth-Sun direction and the Earth-Cygnus A direction, $\theta_{\rm sun-CygnA}$, is expected to be the same in June and December, because the Earth has simply moved to the other side of the Sun. However, in September the Earth is between the Sun and Cygnus A, such that $\theta_{\rm sun-CygnA}$ is at its largest value. The two objects appear on opposite directions in the sky. Analogously, in March when the Earth is behind the Sun relative to Cygnus, $\theta_{\rm sun-CygnA}$ is at its smallest value. These situations were studied to test the coordinate system of our simulations.
![\[fig:dm\_cygn\_sun\] Distribution of the angle between the incoming dark matter velocity and the Earth-Sun direction over the year for events above a 5 keV threshold in a CF$_4$ detector. For each month $1\times10^4$ dark matter events have been simulated. The maximum of the distribution follows the expected pattern as described in the text.](plots/theta_DM_sun_time.pdf){height="7.cm"}
The time evolution of the peak in the two dimensional dark matter probability distribution arises because of this modulation in the relative angle between the incoming dark matter velocity vector and the Earth-Sun direction, $\theta_{\rm DM-sun}$. Since in September the Sun and Cygnus A appear in different directions on the sky, the velocities of the incoming WIMPs that can produce an event above a detector’s fixed energy threshold therefore preferentially point along the Earth-Sun direction. In March, however, the incoming dark matter velocities will point away from the Sun, resulting in a large $\theta_{\rm DM-sun}$. When simulating light dark matter events for each month of the year and producing a histogram for $\theta_{\rm DM-sun}$, we expect the peaks of these histograms to show a modulation that follows exactly this pattern. In figure \[fig:dm\_cygn\_sun\] we color code the number of events in each angular bin. It is visible that the distribution in $\theta_{\rm DM-sun}$ follows the expected pattern with a maximum in March and a minimum in September.
Having presented the dark matter event rate as a function of energy, time and direction, we now turn to the neutrinos.
\[sec:bgnd\]Neutrino Backgrounds
================================
Dark matter experiments are potentially sensitive to two separate types of neutrino interactions: the first is $\nu$-$e^-$ neutral current elastic scattering, where the neutrino interacts with the atomic electrons, and the second is $\nu$-$A$ neutral current coherent elastic scattering, where the neutrino interacts with the target nucleus. The fact that the former process can lead to events in a dark matter experiment has long been realized and has led to it being suggested as a method for solar neutrino detection [@bahcall1995]. The maximum recoil electron kinetic energy from $\nu$-$e^-$ events can be as large as a few hundred keV, and the cross sections are of order $10^{-44}$ cm$^2$. The latter process has never been observed since the maximum nuclear recoil kinetic energy is only a few tens of keV, however, the cross section is relatively large, approximately $10^{-39}$ cm$^2$. This work focuses exclusively on coherent $\nu$-$A$ scattering.
Although coherent $\nu$-$A$ scattering has never been observed, the process is theoretically well understood. The calculated Standard Model cross section is relatively large, of order $10^{-39}$ cm$^2$ [@freedman1977; @drukier1984]. There has been interest in using this process to make precision weak interaction measurements at the SNS [@scholberg2006], to search for supernova neutrinos [@horowitz2003] and to measure neutrinos produced in the Sun [@Cabrera1984]. Even before direct dark matter detection experiments existed, this process was anticipated as a background [@drukier1986]. On the other hand, one could also take the neutrino events as a signal and test neutrino physics using dark matter detectors, see e.g. [@Harnik2012].
Here we calculate the background rates caused by $\nu$-$A$ coherent scattering in target materials relevant to current dark matter searches. We consider the recently measured solar, e.g. [@Cleveland1998; @Ahmed2003], the atmospheric, e.g. [@Fukuda1998; @Adamson2005; @Abbasi2010], and the predicted diffuse neutrino flux from supernovae throughout the Universe and include the nuclear form factors in the coherent cross section calculation. We include the direction dependence of the recoil signal, and its sidereal and annual modulation.
\[sec:xsec\]Neutrino Scattering Cross Sections
----------------------------------------------
The maximum recoil kinetic energy in $\nu$-$A$ coherent scattering is $$E_{\rm r,max} \ = \ \frac{2 E_{\nu}^2}{m_T + 2 E_{\nu}} ~,$$ where $E_{\nu}$ is the incident neutrino energy, and $m_T$ is the mass of the target nucleus. The four-momentum exchange is related to the recoil energy by $Q^2$ = 2$m_T E_{\rm r}$, and the three-momentum exchange $q$ is approximately equal to $\sqrt{2 m_T E_{\rm r}}$. For neutrino energies below 20 MeV and nuclear targets from $^{12}$C to $^{132}$Xe, the maximum recoil kinetic energy ranges between 50 and 5 keV, meaning that the maximum possible $q$ is quite small, $<$1 fm$^{-1}$. Typical nuclear radii, $R$, are 3-5 fm, and therefore the product $q R < 1$. In this regime, the neutrino scatters coherently off the weak charge of the entire nucleus, which is given by $$Q_W \ = \ N - (1 - 4 \sin^2 \theta_W) Z ~,$$ where $N$ and $Z$ are the number of target neutrons and protons respectively, and $\theta_W$ is the weak mixing angle. Through the dependence on $Q_W$, coherence enhances the scattering cross section with respect to the single nucleon cross section by approximately a factor of $N^2$.
The $\nu$-$A$ coherent scattering cross section is given by [@Freedman1973; @freedman1977; @drukier1984] $$\frac{d\sigma}{d(\cos \theta)} \ = \ \frac{G_F^2}{8\pi} \ Q_W^2 \ E_{\nu}^2 \ (1 + \cos \theta) \ F(Q^2)^2 ~,$$ where $G_F$ is the Fermi coupling constant, $Q_W$ is the weak charge of the target nucleus, $E_{\nu}$ is the projectile neutrino energy, $\cos \theta$ is the scattering angle in the lab frame of the outgoing neutrino direction with respect to the incoming neutrino direction, and $F(Q^2)$ is again the nuclear form factor. The suppression of the cross section by the nuclear form factor depends on the target material and grows with the momentum transfer in a collision.
The dependence of the cross section on scattering angle means that solar neutrino elastic scattering events will, in principle, point back to the sun. However, the majority of dark matter detectors do not have directional sensitivity, and so we calculate event rates here as a function of recoil nucleus kinetic energy as well. The scattering angle and the recoil kinetic energy are related via 2-body kinematics and the cross section can be expressed in terms of the kinetic energy, $E_{\rm rec}$, of the recoiling nucleus as $$\frac{d\sigma}{dE_{\rm r}} \ = \ \frac{G_F^2}{4\pi} \ Q_W^2 \ M^2 \ (1 - \frac{m_T E_{\rm r}}{2 E_{\nu}^2}) \ F(Q^2)^2.$$ The theoretical uncertainty on the coherent $\nu$-$A$ scattering cross section comes from uncertainty in the form factor; for neutrino energies of 10 MeV the uncertainty is expected to be less than 10% [@horowitz2003].
\[sec:flux\]Neutrino Fluxes
---------------------------
![\[fig:flux\] The neutrino fluxes considered in this work. The grey colored fluxes will not give events above thresholds considered in this paper. Important fluxes for coherent neutrino nucleon scattering originate from solar neutrinos (red), atmospheric neutrinos (blue) and diffuse supernovae neutrinos (green).](plots/flux.pdf){height="7.cm"}
There are many sources that contribute to the large flux of ambient neutrinos and antineutrinos. The main sources are fusion reactions in the Sun, radioactive decays in the Earth’s mantle and core, decay products of cosmic ray collisions with the atmosphere, relic supernovae neutrinos and neutrinos from fission processes at nuclear reactors. We show the approximate energy ranges and fluxes of neutrinos in table \[tab:flux\_table\]. For this work, we consider the fluxes of solar, atmospheric and supernovae neutrinos. In reference [@Monroe2007] it has been shown that the contribution of geoneutrinos to the background of dark matter searches can be neglected. In figure \[fig:flux\] we show the energy dependent fluxes used here.
The largest contribution from solar neutrinos is the $^8$B neutrino flux, which is well understood. The predicted flux normalization, shown in table \[tab:flux\_table\], agrees with the measured flux at the 2% level [@sno2007]. The uncertainty of the measured flux that includes neutrino oscillations is only 3.5% [@sk2005], even though the predicted flux normalization has an uncertainty of 16% [@bahcall2004]. For the predicted atmospheric neutrino flux the estimated normalization uncertainty is 10% for neutrino energies below 100 MeV [@superk2005] which agrees well with measurements by a number of experiments. We note that the normalization of the low energy component of the atmospheric neutrino flux is strongly dependent on the latitude. This is due to the geomagnetic cut off, e.g. the flux at Super-Kamiokande [@honda2001] is approximately half of the flux at the SNO experiment.
The background of diffuse supernovae neutrinos is the integrated flux from all supernovae that occurred in the Universe. The neutrino energy spectrum of a single supernovae is assumed to be similar to a Fermi-Dirac spectrum with temperatures of 3 MeV for electron neutrinos, 5 MeV for electron antineutrino and 8 MeV for the other four flavors. For more details on diffuse supernovae neutrinos see for example [@Beacom2010]. We assume an uncertainty of 10% on the supernovae neutrino flux.
Source Predicted flux Energy (MeV)
--------------------------------------------------- ------------------------ --------------
[@bahcall2004] Solar $\nu$ pp 5.99$\times10^{10}$ $<$0.4
[@bahcall2004] Solar $\nu$ CNO 5.46$\times10^{8}$ $<$2
[@bahcall2004] Solar $\nu$ $^7$Be 4.84$\times10^{9}$ 03, 0.8
[@bahcall2004] Solar $\nu$ $^8$B 5.69$\times10^{6}$ $<$12
[@bahcall2004] Solar $\nu$ h.e.p. 7.93$\times10^{3}$ $<$18
[@gaisser2002] Atmospheric $\nu$+$\overline{\nu}$ O(1/E(GeV)$^{2.7})$) 0-10$^3$
[@Beacom2010]Diffuse Supernovae $T_{\nu} \approx8$ MeV $0-10^2 $
: \[tab:flux\_table\]Ambient sources of neutrinos. Fluxes are given in number per cm$^2$ per second.
The calculations here use the predicted neutrino fluxes without including neutrino oscillations. The coherent scattering process is neutrino-flavor independent to leading order, and we assume no sterile neutrino participation in oscillations, thus the oscillated and unoscillated predicted neutrino fluxes are, in practice, equivalent for our calculation.
\[sec:signal\] Neutrino Signal Distribution
-------------------------------------------
The neutrino event rate is, similarly to the dark matter event rate, given by an integral over the differential recoil rate and the energy efficiency, $${\label{eq:rate_nu}}
b=\int_{E_{\rm thr}}^{E_{\rm up}} \epsilon(E_{\rm r})\frac{dR_{\nu}}{dE_{\rm r}} dE_{\rm r}~.$$ The differential rate is $${\label{eq:diffrate_nu}}
\frac{dR_{\nu}}{dE_r} = n_T ~ \int_{t_0}^{t_1} \int_{E^{\rm min}_\nu}^{\infty} \frac{dN(t)}{dE_\nu} ~ \frac{d\sigma(E_\nu, E_{\rm r})}{dE_{\rm r}}~ dE_\nu~dt ~,$$
![\[fig:cf4\_rate\] Neutrino event rate in a CF$_4$ detector. For this plot a perfect energy efficiency and an upper threshold of 100 keV were considered. For the rest of the paper we assume a more realistic energy efficiency function lowering the total event rate.](plots/rate_CF4.pdf){height="7.cm"}
with $n_T$ the number of target nuclei in the detector, the flux $\frac{dN(t)}{dE_\nu}$ and the differential cross section $\frac{d\sigma(E_\nu, E_r)}{dE_r}$. The dependence on time, $t$, in the flux is due to the change in distance between the Sun and the Earth over the year. We integrate the time dependence over the exposure time of a given experiment to calculate rates. Note that the only thing that changes with time is the normalization in the solar neutrino flux, not the shape of the spectrum. As a first approximation, we take the flux of atmospheric and supernovae neutrinos to be time independent, although there is a time variation in the atmospheric flux due to temperature changes in the Earth’s atmosphere [@Tilav2010]. This change in the event rate is, however, smaller than the annual modulation of the dark matter rate or the modulation of the solar neutrino rate. As we found that both of these are not contributing significantly to the sensitivity of the simulated detectors we neglect the variation of the atmospheric neutrino flux here.
The integral over the neutrino energy starts at the minimal neutrino energy $E_{\rm \nu}$ necessary to get a recoil event over threshold and is given by $E_{\rm \nu}^{\rm min}=\sqrt{m_TE_{\rm r}/2}$. In figure \[fig:cf4\_rate\] we show the event rate for a CF$_4$ detector. For the threshold that we will consider in this work (5 keV), only ${}^8B$ and hep neutrinos from the Sun as well as all atmospheric and supernovae neutrinos are important.
The scattering angle of the nucleus with respect to the incoming neutrino direction can then be found from scattering kinematics to be$${\label{eq:scat_angle_nu}}
\cos \theta'= \frac{E_{\nu}+m_T}{E_{\nu}}\sqrt{\frac{E_{\rm r}}{2m_T}} ~.$$ Figure \[fig:neutrino\_pdf\] shows the two dimensional probability distribution of recoil energy and event angle for neutrinos in a CF$_4$ detector with a 5 keV energy threshold. The significant difference to the dark matter probability distribution is the clear peak at $\cos \theta_{\rm sun} = -1$ and small recoil energies due to the solar neutrino events. Atmospheric and supernovae neutrinos contribute as a smooth, isotropic background. For a 5 keV CF$_4$ detector we can see in figure \[fig:cf4\_rate\] that the nonsolar neutrinos have only a small contribution such that in this example the probability distribution function falls off steeply away from the solar peak. The ratio of the solar peak to the smooth background of nonsolar neutrinos depends on the target material and the recoil energy threshold. In different detector configurations the dominance of the solar peak over the nonsolar background is not necessarily this significant.
![\[fig:neutrino\_pdf\] The two dimensional probability distribution $\rho$ of recoil energy and event angle of neutrinos in a CF$_4$ detector with 5 keV threshold.](plots/pdf_solar.pdf){height="7.cm"}
\[sec:analysis\]Dark Matter Searches in the Presence of Neutrino Backgrounds
============================================================================
Having obtained detailed spectra for dark matter and neutrino events as a function of energy, direction and time, we need a statistic to test these signal and background distributions in a given experiment. In order to do this, we perform a CLs test [@Read2000] to distinguish between background and signal + background hypotheses, in which the background comes from solar, atmospheric and diffuse supernovae neutrino coherent elastic scattering. We consider a range of targets and moderately optimistic energy thresholds, as well as energy and angular resolutions, which should be realistically achievable by the next-generation experiments.
\[sec:stats\] Statistical Test
------------------------------
The presence of backgrounds in direct searches of any kind implies that a given set of observed events is either pure background or contains background plus signal. One way to distinguish between these two cases statistically is to perform a hypothesis test. Such a test can be carried out by looking at the ratio between the probability densities of the measured data $\vec{X}$ being either signal plus background or background only, $\widetilde{Q}=\frac{\mathcal{L}(\vec{X}, S+B)}{\mathcal{L}(\vec{X},B)}$ [@Read2000]. We take this as the definition of our test statistic: $$\label{eqn:Q}
\widetilde{Q}=\frac{p_{b+s}(n)}{p_b(n)} \frac{\prod_{j=1}^n\frac{sS_t(t_j)+bB_t(t_j)}{s+b}~\frac{sS_{\theta,E}^{(t)}(\theta_j,E_j)+bB_{\theta,E}(\theta_j,E_j)}{s+b}}{\prod_{j=1}^n B_t(t_j)~B_{\theta,E}(\theta_j,E_j)} ~.$$ Throughout this work, we use the notation $p(x)=dP(x)/dx$ as the probability distribution function of the variable $x$ where $P(x)$ is therefore the cumulative probability of this quantity at $x$. In equation \[eqn:Q\], $s$ is the number of expected dark matter events given by equation \[eq:rate\_dm\], $b$ the number of expected neutrino events given by equation \[eq:rate\_nu\] and $n$ the total number of observed events in an experiment. The functions with capital letters $B$ or $S$ denote different normalised probability distribution functions for the neutrino and dark matter events, respectively. $p_{\lambda}(n)$ is the Poisson distribution centered at $\lambda$, where $\lambda$ is either $b$ or $b+s$ (we discuss in section \[sec:uncertainties\] how to exactly obtain $s$, $b$ and $n$). The variables $t_j, E_j, \theta_j$ denote the time, recoil energy and event angle of the j-th event. We define the event angle as the angle between the track of the recoiling nucleus and the Earth-Sun direction.
$B_t$ describes the annual modulation of the neutrino event rate. It has a maximum in January, when the neutrino flux and hence the neutrino event rate is largest, and a minimum in July, when the distance between the Earth and the Sun is at its maximum. $S_t$ encodes the information of the annual modulation of the dark matter event rate and depends on the dark matter mass. For light dark matter this function has a maximum in June and a minimum in December.
$B_{\theta,E}$ is the two dimensional probability distribution of the recoil energy and the event angle for neutrino events and $S_{\theta,E}^{(t)}$ the corresponding one for dark matter events. Visualized examples of these distributions are the figures \[fig:neutrino\_pdf\] and \[fig:dm\_pdf\], respectively. The dark matter distribution carries an additional index for time because of its variation over the year as described in section \[sec:signal\_dist\]. To include this time variation, we choose ten equally distributed days over one year and create one probability distribution function for each of these days. A given event will then interpolate linearly between the two probability distribution functions closest to the signal event time. Equation \[eqn:Q\] can be simplified to: $$\label{eq:simpleQ}
\widetilde{Q}=e^{-s}\left(\frac{b}{s+b}\right)^n \, \prod_{j=1}^n \left(1+\frac{sS_t(t_j)}{bB_t(t_j)} \right) \left(1+\frac{sS_{\theta,E}^{(t)}(\theta_j,E_j)}{bB_{\theta,E}(\theta_j,E_j)}\right) ~.$$ In the following we will discuss the log-ratio $Q=-2\log\widetilde{Q}$.
An advantage of this procedure is that experimental uncertainties can easily be incorporated by smearing the probability distributions. Dark matter searches have to deal with imperfect energy and angular resolution (in the case of directional experiments), as we discuss in section \[sec:detassumptions\], leading to a smearing of $B_{\theta, E}$ and $S_{\theta,E}^{(t)}$. The background of nonsolar neutrinos ensures a non-zero value for $B_{\theta, E}$ for all values of $\theta$ and $E$ such that $Q$ is well behaved. See section \[sec:ev\_sim\] for more details on how $B_{\theta, E}$ and $S_{\theta,E}^{(t)}$ are created.
For every dark matter mass and cross section we want to find out wether a fixed detector setup (target material, energy threshold, exposure, energy and angular resolution) is capable of distinguishing whether the observed events are pure background or contain a dark matter signal. To do so, Q has to be evaluated twice: First we simulate pseudoexperiments with only neutrino events and obtain a distribution $p_B(Q_B)$ for the background only hypothesis using equation \[eq:simpleQ\]. As, in this case, the pseudodata is more consistent with the background expectations, $p_B(Q_B)$ will peak at positive Q values.
We then repeat the exercise and simulate pseudoexperiments with dark matter and neutrino events to get a distribution $p_{SB}(Q_{SB})$. This distribution will, in contrast, peak at negative Q values.
We can then decide whether the detector setup is sensitive to a given $m_{\rm DM}$ and $\sigma_{\rm p}$, if we look at the separation between these two distributions. The clearer this separation, the easier it is for the chosen detector configuration to distinguish between these two hypotheses and the more sensitive the detector. The lower the signal rate and the more similar the signal expectations are to the background expectations, the closer the distributions will be until they start to overlap. If the overlap becomes too large, the experiment will lose its sensitivity completely. See figure \[fig:Q\_dis\] for a visualization.
![\[fig:Q\_dis\] The normalised background only distribution $p_B(Q_B)$ (blue) and signal plus background distribution $p_{SB}(Q_{SB})$ (red) including angular information (top) and excluding angular information (bottom) for s=10 and b=500 for a 6 GeV dark matter particle in a CF$_4$ detector. The gain in sensitivity when using directionality is clearly visible in the separation of the two distribution in the upper plot.](plots/Q_dis.pdf){height="7.cm"}
To quantify the sensitivity of a given dark matter experiment for a specific dark matter mass and cross section we calculate the overlap of these two distributions as follows. We integrate both, $$\beta_{SB}=\int_{-\infty}^{q} p_{SB}(Q_{SB})~dQ_{SB}~,$$ $$\beta_{B}=\int_{-\infty}^{q} p_{B}(Q_{B})~dQ_{B}~,$$ up to a $q$ value for which $$1-\beta_{SB}=\beta_B\equiv \alpha ~.$$ We take the confidence level at which the signal plus background hypothesis can be distinguished from the background-only hypothesis to be $(1-\alpha)$. In this work we are interested in separations of both hypotheses at $90\%$ confidence level, corresponding to $\alpha$ equal to 0.1, and in 3$\sigma$ separations ($\alpha=0.00135$).
The statistical approach has uncertainties due to a finite sample size of pseudoexperiments, a finite number of events to create the two-dimensional probability distributions, as well as a finite bin width when creating the histograms of the test statistics. We estimate this numerical error to be 5% in the overlap and add it to the error due to the systematic uncertainties.
\[sec:detassumptions\]Detector Performance Assumptions
------------------------------------------------------
In this work we will estimate future sensitivities of dark matter detectors including the irreducible neutrino nucleus coherent scattering as a background. To see how the mass of the target material influences the sensitivities, we look at tetrafluoromethane, CF$_4$, as a light and Xenon, Xe, as a heavy target material.
Interesting directional technologies are already in existence for experiments based on CF$_4$ [@Ahlen2009; @Monroe2011] and we show here that scaling these detectors to large masses can test cross sections beyond the neutrino background. For Xenon, on the other hand, there are at the moment no directional techniques demonstrated and our directional sensitivities are in this sense futuristic. However, we think it is still interesting to see how directional information would help if a heavy target material was used.
For the CF$_4$ detectors, we model the energy efficiency of the detectors with $$\epsilon(E_{\rm rec}) = c_1\left(1 + {\rm erf}\left[ \frac{(E_{\rm rec} - c_2)}{c_3} \right] \right) ~,$$ and choose $c_1=0.5$, $c_2$ as the energy threshold $E_{\rm thr}$ and $c_3=15$ keV. These values for the efficiency asymptote at 50% and are consistent with current directional searches [@Ahlen2009]. The 5 keV energy threshold we assume is optimistic relative to current searches (although a number of CF$_4$ directional detectors use 5.9 keV $^{55}$Fe sources for calibration, and track images with directionality at this energy have been measured in small prototypes [@Santos:2011xk]). Large direct dark matter searches based on Xenon have been carried out in the past already [@LUX; @xenon100], so we use the efficiency curve published by the LUX experiment and shift it to smaller energy thresholds, such as 2 keV. For the Xenon detectors we assume an upper energy window cutoff of 40 keV, for the CF$_4$ detectors we take 100 keV.
The energy resolution is modeled as $$\sigma_E = 0.1~\sqrt{E/{\rm keV}}~,$$ and the angular resolution as $$\sigma_\theta = \frac{30^{\circ}}{\sqrt{E/{\rm keV}}} ~.$$ For the angular efficiency we assume $100\%$.
\[sec:ev\_sim\] Event Simulation
--------------------------------
To simulate dark matter events we use the rest frame of the static, spherically symmetric dark matter halo and draw a random velocity magnitude $v$ from the velocity distribution according to equation \[eq:vel\_dis\] and, to fix the dark matter direction, additionally two angles ($\theta$,$\varphi$) in a spherically symmetric way. We then calculate the cartesian coordinates of the dark matter velocity vector $\vec{v}$ in galactic coordinates. Drawing a random event time $t$ from a uniform distribution between $t_0$ and $t_1$ gives us the Earth’s overall velocity vector in galactic coordinates from reference [@McCabe2013]. After a coordinate transformation into the rest frame of the Earth, we have the incoming WIMP velocity vector.
As we assume isotropic scattering, we draw a uniform scattering angle and obtain the recoil energy $E_{\rm r}$ of the event from equation \[eq:dm\_Erec\]. In this way, the dark matter direction and annual modulation are both included in the event simulation as we use the full information of the Earth’s velocity vector and start from a spherically symmetric halo. We calculate the event angle $\theta_{\rm sun}$ by projecting the track of the recoiling nucleus onto the Earth-Sun direction and perform the energy and angular Gaussian smearing. To take the energy efficiency into account, we only accept the corresponding fraction of events at each recoil energy and apply the energy thresholds as hard cutoffs. For each of the ten dark matter probability distributions of each dark matter mass we simulate $10^6$ events and bin the data into 30 energy and 15 angular bins.
To create the two dimensional probability distribution function for neutrinos, we perform the event rate calculation, equation \[eq:rate\_nu\], for each neutrino type separately to know exactly how many events of each type can be expected in a given detector configuration.
To simulate neutrino events, we draw a random neutrino energy according to the energy dependent flux. For a solar neutrino the direction is known. When simulating the atmospheric and supernovae neutrinos we assume an isotropic incoming neutrino direction. We use the differential cross section and its dependence on the neutrino energy to create a probability distribution for a given neutrino energy $E_{\nu}$ to give an event of recoil energy energy $E_{\rm r}$. From this we draw a random $E_{\rm r}$ and obtain the scattering angle via equation \[eq:scat\_angle\_nu\].
These real event values are smeared according to the detector resolutions as explained in section \[sec:detassumptions\] and the energy efficiency and thresholds are applied. The event time is drawn uniformly for the atmospheric and supernovae neutrinos, but from a non uniform distribution for the solar neutrinos that follows the annual modulation of the event rate. For the neutrino probability distribution we simulate $1.5\times 10^6$ neutrino events and bin them into 30 energy and 15 angular bins.
![\[fig:combo\_pdf\] The combined two dimensional probability distribution $\rho$ of the recoil energy and event angle for a 6 GeV dark matter particle and neutrinos in a CF$_4$ detector. The expected signal rate is fixed to s=10 and the expected background rate to b=500.](plots/pdf_combined.pdf){height="7.cm"}
In figure \[fig:combo\_pdf\] we present the combined two dimensional signal plus background probability distribution of event angle and recoil energy for a 6 GeV dark matter particle in a CF$_4$ detector. As can be seen in equation \[eqn:Q\], the constituent probability distribution functions of signal and background are weighted according to the expected number of background and signal events. We present the case for b=500 and s=10. Even for such low count rates, a significant excess at large $\cos \theta_{\rm sun}$ is visible, compared to figure \[fig:neutrino\_pdf\].
Given the differences between the energy, angle, and time distributions of dark matter signal and neutrino background events, a directional experiment could in principle fit simultaneously for the normalizations of both fluxes and include the systematic uncertainties as nuisance parameters. This technique has been employed by a number of experiments to constrain systematic uncertainties on both background and signal distributions, for example the current most precise measurement of the solar $^8$B flux [@Aharmim2009]. However, the degree to which this approach is successful depends strongly on the number of events, the separation of the signal and background distributions (which in the energy dimension depends on the dark matter mass), and the degree of correlation of the systematic uncertainties. Therefore we assume for this work that the neutrino flux uncertainty is externally determined, and we make a semioptimistic assumption about its magnitude in the future.
\[sec:uncertainties\]Systematic Uncertainties
---------------------------------------------
We have seen that the neutrino background is made up of three distinct populations - solar neutrinos, diffuse neutrino background from supernova explosions and neutrinos from cosmic rays hitting the atmosphere. The amount by which the uncertainties on these three fluxes will reduce in future years is uncertain and can only be estimated roughly. The solar flux in our energy range of interest is dominated by $^8$B and hep neutrinos, and the flux of neutrinos due to both of these emission mechanisms is very sensitive to the iron abundance in the Sun which affects the opacity in the core [@PenaGaray2008]. Understanding the iron opacity in the Sun is challenging - currently the solar composition as observed at the surface of the Sun [@Asplund2009] is not in good agreement with that deduced from helioseismology [@2014ApJ787]. There are at least two future experiments that will help reduce the uncertainty on the solar flux. SNO+ will make good measurements of both the Boron-8 and the Beryllium-7 neutrinos both of which depend sensitively upon the iron abundance in the core, which will indirectly constrain the hep fluxes [@Maneira2013]. If approved, Hyper-Kamiokande should be able to detect several hundred $^8$B neutrinos per day collecting such high statistics that it will look for time variation in the $^8$B flux [@Abe2011]. With these observations, it seems not impossible that the uncertainties in the solar flux may drop by a factor of several in a few decades, if not orders of magnitude. See the recent work [@Billard2014] for how data from future dark matter detectors could help to test solar models.
The diffuse supernova background (DSNB) neutrinos are more complicated since while we are able to increase our understanding of the historic stellar evolution history using astronomical observations, we are not so certain of the spectrum of neutrinos emitted from a single supernova. It seems that there are good prospects for Super-Kamiokande to improve its sensitivity, in particular by using dissolved gadolinium to improve neutron tagging, which could significantly enhance its sensitivity to the DNSB neutrinos ultimately leading to a discovery in a decade or so of running which could constrain the magnitude of the spectrum between 10-20 MeV [@Beacom2013]. Again, Hyper-Kamiokande would do much better, allowing one to measure the spectrum in great detail. It is not unreasonable to suggest a large drop in the uncertainties in this flux.
The same enhancement of Super-Kamiokande could detect the atmospheric neutrino flux. This flux is perhaps more elusive than the other two since factors which may affect it include the primary cosmic ray flux (although this will be constrained by AMS02), the geomagnetic field, the solar wind and nuclear propagation models. The theoretical uncertainties in the development of the shower have been studied by observing the interaction between protons and thin targets of $O_2$ and $N_2$ at the HARP experiment in CERN as well as observing the muon flux in the atmosphere using balloon experiments. Models like DPMJET-III and JAM then aim to constrain the resulting neutrino flux [@Honda2011]. Because of this, the uncertainties are likely to fall more slowly, although Hyper-Kamiokande and an upgraded Super-Kamiokande will probably be able to detect the flux therefore constraining it more tightly.
These considerations lead us to the semioptimistic approach to take half the current flux uncertainties as a basis for the simulated detectors in our analysis. This translates into an uncertainty of 8% for the solar neutrino fluxes (note that only $^8$B and hep neutrinos can give events above threshold in this work), and 10% for the atmospheric and supernovae neutrino fluxes.
To include these neutrino flux uncertainties we first obtain a central value result. This means that we assume the incoming fluxes to have their nominal measured values resulting in a background rate $b_0$. The number of observed events $n$ in a pseudo experiment is drawn from a Poisson distribution centered at a value $\lambda$ which is either equal to $b_0$ for the background only or $b_0+s$ for the signal plus background simulation. For each pseudo experiment we simulate these $n$ events as we discussed in section \[sec:ev\_sim\].
To account for the unknown real flux value when performing the experiment we vary the expectation of each pseudo experiment, that is $b$ in equation \[eq:simpleQ\]. Hence, for each pseudo experiment we draw a random flux value for each neutrino flux type from a Gaussian with 1$\sigma$ corresponding to the uncertainties. This results in a different expected background rate $b$ for each pseudoexperiment via equation \[eq:diffrate\_nu\] and widens the $Q$ distributions. We then repeat the procedure shifting $b_0$ up and down by one sigma to obtain a 1 sigma band for the estimated exclusion limits.
Results
=======
\[sec:results\] Estimation of Detector Sensitivities
----------------------------------------------------
In order to see directly the gain in sensitivity when directional information is used, we evaluate the sensitivity that we obtain from our statistical approach for both cases, excluding (red bands) and including directional information (green bands). To compare the results to the WIMP discovery limit that was presented in [@Billard2013], we show this limit as a light grey line. Note here that the limits from [@Billard2013] are discovery limits at the $3\sigma$ level and based on a profile likelihood approach, whereas we perform a hypotheses test. Therefore, any direct comparison should be taken with care. A strict discovery limit exists for dark matter masses that match the energy spectrum of the neutrino background perfectly, see [@Billard2013]. This is, for example, the case for a 6 GeV dark matter particle and the background of $^8$B neutrinos in a Xenon detector. We reproduce this limit and the discovery limits for heavy dark matter from [@Billard2013] with very good accuracy; see also section \[sec:max\_sensitivity\]. In the dark matter mass region around 10 GeV where a steep increase in sensitivity towards smaller cross sections is observed, however, we find slightly less constraining discovery limits, as will become clear when we discuss the Xenon detector.
In this section we will look at sensitivity limits at the 90% C.L. and 3$\sigma$ level for experiments with different target materials and energy thresholds. To compare the different simulations, the detector exposure is scaled such that the simulated experiment will observe 500 neutrino events, i.e. the background contribution is sizable. As an example for a dark matter detector with directionality, we estimated the sensitivity of tetrafluoromethane CF$_4$ as target material. As a light target CF$_4$ is promising to distinguish solar neutrinos from light dark matter. We set the energy thresholds in our run to 5 keV.
![\[fig:cf4\_run1\] Estimated sensitivity limits at 3$\sigma$ level for a nondirectional (red band) and directional (green band) CF$_4$ detector with 36 t-yrs exposure and 5 keV energy threshold resulting in 500 expected neutrino events. The fainter bands indicate corresponding sensitivity limits at 90% CL, the grey curve is the neutrino bound. ](plots/run_CF5keV.pdf){height="7.cm"}
Figure \[fig:cf4\_run1\] shows the obtained sensitivity bands for a 36.6 ton-year CF$_4$ experiment with a 5 keV energy threshold. The 500 neutrino events consist of 499.8 expected solar and 0.2 expected nonsolar neutrinos. The green and red bands represent limits that can be obtained with directional and nondirectional detectors at a 3$\sigma$ level, respectively, and the grey curve is the neutrino bound. The fainter colors show corresponding limits at 90% C.L. The separation of the green band from the red band clearly shows the impact of directional information. A strong increase in sensitvity for directional detectors towards smaller cross sections is observed which is larger the smaller the dark matter mass. This is easily understood when considering the clear seperation of the neutrino and dark matter peak in the two dimensional probability distribution functions. The lighter the dark matter particle is, the more significant this separation. For a light dark matter event to be above threshold, the track of the recoiling nucleus has to lie closer along the incoming dark matter direction in order to produce a large enough recoil. Hence, the dark matter signal also has a strong directional character, as discussed in section \[sec:detectors\]. Since the event angle distribution is different for the neutrinos, directional information has a large impact.
We find that cross sections below the solar neutrino bound can be tested at the 3$\sigma$ level when directional information is taken into account.
Towards heavier dark matter masses, we see that the sensitivity curves approach each other and directionality loses some impact. For heavy dark matter, the distinction of signal and solar background is already easy when the energy spectrum is considered on its own, because the recoil energies of solar neutrinos are much smaller compared to heavy dark matter. Besides, the dark matter events lose their directional character more and more: Light dark matter can only give recoil energies above threshold for the largest dark matter velocities in the halo, such that only those particles coming from Cygnus A can give a recoil event in the detector. The kinetic energy of heavy dark matter particles is, in contrast, also large for small dark matter velocities. Hence, the incoming direction of dark matter particles that give a signal event in the detector becomes unconstrained and more and more isotropic. A competing effect is that the track resolution for small recoil energies is worse, but improves for larger recoil energies and thus for heavier dark matter. Overall, we see that directional information is also useful for heavier dark matter. This is mainly because when heavy dark matter particles give recoil energies comparable to the recoil energies of solar neutrinos, the dark matter events can be distinguished using directional information, which would not be possible otherwise.
![\[fig:Xe\_run1\] Estimated sensitivity limits at 3$\sigma$ level for a nondirectional (red band) and directional (green bands) Xenon detector with 367 t-yrs exposure and 2 keV energy threshold resulting in 500 expected neutrino events. The fainter bands indicate corresponding sensitivity limits at 90% CL, the grey curve is the neutrino bound. ](plots/run_Xe2keV.pdf){height="7.cm"}
At the moment, the strongest constraints on the WIMP-nucleon cross section are set by experiments that use Xenon as a target material. These detectors have no directional information and no technology exists up to now that could achieve this. However, it is still interesting to ask which cross section experiments with heavy target materials would be able to probe if they could use directional information. There has been recent interest in developing a direction-sensitive Xenon detector technology based on recombination dependence on the recoil angle relative to the detector $\vec{E}$ field [@Nygren:2013nda], so perhaps this will be a possibility for the future.
Therefore, we additionally choose Xenon as a target material and perform the same tests. Estimated sensitivity curves for a hypothetical experiment with 367.7 ton-year exposure using a 2 keV threshold can be seen in figure \[fig:Xe\_run1\]. The 500 neutrino background events consist of 485.8 expected solar and 14.2 expected nonsolar neutrinos.
Our statistical test finds that even without directional information cross sections below the discovery limit from [@Billard2013] (grey curve in the plots) can be tested at 3$\sigma$ level. For example, an 8 GeV WIMP with a cross section of $2.3\times10^{-46} \rm{cm}^2$ would give about 470 dark matter events. We note here, that we assumed half the flux uncertainties and took a different statistical approach than reference [@Billard2013]. The nondirectional 3$\sigma$ limit should hence be seen as a WIMP-discovery limit obtained from our approach rather than testing cross sections beyond the discovery limit. Again, we see that directional detectors can go beyond and probe smaller cross sections compared to nondirectional detectors. The same trend that directional and nondirectional detectors give similar sensitivities for heavy dark matter particles is visible; the limits are basically identical for the Xenon detector.
Compared to the light target material CF$_4$ we find that the impact of directional information is less significant in this Xenon detector configuration when searching for heavy dark matter. With Xenon as a heavy target material solar neutrinos can give recoil energies only up to approximately 5 keV. Hence, the range of recoil energies for which directionality is the only indicator to distinguish the signal from the solar neutrino background is small. For the light target material CF$_4$ this range is larger: solar neutrinos can recoil up to approximately 30 keV, see figure \[fig:cf4\_rate\]. We can therefore conclude that the larger the range of possible recoil energies of solar neutrinos is compared to the total energy range of the detector, the larger the gain in sensitivity from directional information. On the other hand, for the same number of background events Xenon can probe smaller cross sections.
It is not clear how a Xenon detector might be made directional. However, an additional motivation to pursue directionality is that ultimately very large Xenon detectors would be limited by the background of solar neutrino-electron elastic scattering events. It is important to note that we do not take this background source into account in our simulations, although it is expected to become significant at the $10^{-48}$ cm$^2$ level [@Baudis:2013qla]. Additionally, the electron discrimination in Xenon detectors is less efficient than in other detectors, for example liquid argon. Directionality could aid in that discrimination considerably, because these events have a direction that points back to the Sun, allowing deeper cross sections to be probed. Of course it is important to note that such large detectors would cost a great deal of money since Xenon is $\sim$\$1000 per kg at today’s prices and only 25 tons are obtained from the air annually.
In this section we have presented results from detector simulations for which we fixed the number of expected neutrino events to 500 in order to estimate possible sensitivity limits in the presence of neutrino backgrounds. It was shown that cross sections beyond the discovery limit can be probed when directional information is taken into account. Directional detectors have significantly larger sensitivities for light dark matter masses. For a light target material this is also true for heavy dark matter. We will now move on and discuss how these limits behave as a function of exposure.
\[sec:max\_sensitivity\] Projected Sensitivity
----------------------------------------------
For both detector configurations that were presented in section \[sec:results\], we choose three dark matter masses and for different exposures find the minimal cross section that can be tested at a 3$\sigma$ level with and without directional information. In this way we can find the exposure necessary to go beyond the discovery limit. As the dark matter masses we choose 6 GeV in order to see how directionality helps when the energy spectrum of solar $^8B$ neutrinos and dark matter are identical, 1000 GeV as a heavy dark matter mass and 30 GeV as a mass for which nondirectional experiments have close to maximal sensitivity. We expect the projected nondirectional sensitivity limits for the Xenon detector to flatten out below the limits from reference [@Billard2013] by a factor of 2 because we assume half the neutrino flux uncertainties. For each dark matter mass and cross section we simulate $5 \times 10^3$ pseudoexperiments.
We present the results for the Xenon detector in figure \[fig:sens\_xenon\] and for the CF$_4$ detector in figure \[fig:sens\_cf4\]. The projected sensitivity limits for directional detectors are presented as solid lines, nondirectional detectors are shown as dashed-dotted lines. The discovery limits of [@Billard2013] are indicated as horizontal dashed lines. We color code the three different masses in blue (6 GeV), red (30 GeV) and black (1000 GeV).
![\[fig:sens\_xenon\] Estimated sensitivity limits for a directional Xenon detector with a 2 keV energy threshold for a 6 GeV (blue), 30 GeV (red) and 1000 GeV (black) dark matter particle. The solid lines show directional detectors, the dash-dotted lines show nondirectional detectors. The horizontal dashed lines indicate the discovery limit of [@Billard2013] for each dark matter mass. The vertical grey line shows the simulated detector of section \[sec:results\]](plots/sensXe.pdf){height="7.cm"}
As expected, for $m_{\rm DM} = 6$ GeV the sensitivity of nondirectional Xenon detector flattens out just below the discovery limit. This shows our agreement with earlier work [@Billard2013] and shows directly the impact of improved knowledge on neutrino fluxes by a factor of two. For the nondirectional CF$_4$ detector (figure \[fig:sens\_cf4\]) we see that the curve becomes flat already above the discovery limit. This is simply because the discovery limits were calculated for Xenon and we consider a different target material here. The solid blue line indicates possible limits if directional detectors were constructed. It is visible that the solar neutrino floor disappears for both target materials once there is a clear way to distinguish signal from background using directionality.
![\[fig:sens\_cf4\] Estimated sensitivity limits for a directional CF$_4$ detector with a 5 keV energy threshold for a 6 GeV (blue), 30 GeV (red) and 1000 GeV (black) dark matter particle. The solid lines show directional detectors, the dash-dotted lines show nondirectional detectors. The horizontal dashed lines indicate the discovery limit of [@Billard2013] for each dark matter mass. The vertical grey line shows the simulated detector of section \[sec:results\].](plots/sensCF4.pdf){height="7.cm"}
For the larger dark matter masses the directional and nondirectional limits are basically the same when Xenon is used as a target material, as we discussed in section \[sec:results\]. We see that the projected sensitivities reach a boundary close to the discovery limit of [@Billard2013]. For the 1000 GeV WIMP this is slightly below the boundary as we expect from the reduced flux uncertainties. The 30 GeV sensitivity line stays just above the neutrino bound. We note our that the discovery limits in [@Billard2013] were obtained with a 4 keV energy threshold, compared to a 2 keV threshold here and the different statistical approaches that were taken. Sensitivities for dark matter masses in this region show a dependence on the threshold energy. Besides, this change in threshold energy affects the neutrino energy spectrum significantly, which influences sensitivities of medium dark matter more than of heavy dark matter.
Because the directional and nondirectional projected sensitivities are close to identical, going beyond discovery limits for these masses with a heavy target material is only possible if the uncertainties of the neutrino fluxes are reduced and extremely large exposures become possible. Note here that our analysis does not include the $pp$ solar neutrino-electron scattering background, which is relevant for cross sections below $10^{-48}$ cm$^2$.
In figure \[fig:sens\_cf4\] we show the sensitivity limits for a CF$_4$ detector. Similar trends are visible. Directional information allows to go beyond the discovery limits for light WIMPs even with imperfect flux knowledge. Directionality contributes significantly for the complete dark matter mass range and can help more than an order of magnitude in the cross section. For the light target material, we see a significant contribution from directionality even to the sensitivity of heavy WIMPs, but their discovery limits cannot be reached with the exposures we consider.
In the presence of backgrounds, the sensitivity is not expected to scale linearly with exposure. In a Poisson dominated regime, a scaling behavior as the square root of the number of background events is expected. Since we are using additional information here, coming from the energy spectrum, which might be identical for background and signal, and the directionality of the events, a better scaling behavior is possible. For the two detectors we simulated, the isotropic background of nonsolar neutrinos is smaller for the CF$_4$ compared to the Xenon case, resulting in a better scaling behavior. This, however, depends on the energy threshold.
In figures \[fig:sens\_xenon\] and \[fig:sens\_cf4\] the improvement in sensitivity is visible when directional information is included in addition to event time and recoil energy. In our analysis we find that time information adds only little on top of the sensitivity of a pure spectral analysis. For a 6 GeV dark matter particle in a CF$_4$ detector, e.g., we can find an improving effect of more than 10% on the sensitivity when measuring annual modulation only if there are about $10^3$ background events. We see that annual modulation becomes important only for large background rates, and that the impact of directional information is much larger.
The estimated sensitivities of directional detectors also depend on the chosen angular and energy resolutions. We find the angular resolution to be the more important one. For a 6 GeV dark matter particle decreasing the energy resolution by a factor 2 would leave the sensitivity unchanged up to a few percent. The same change in the angular resolution, however, would reduce the sensitivity by a factor $\sim 3$ for 500 background events or even a factor $\sim 5$ if there are 5000 background events.
These plots show that in principle there is no solar discovery limit for direct dark matter searches if directional detectors are constructed. Going beyond the discovery limit for a 6 GeV dark matter particle is possible for an exposure of approximately 5 ton-years for a directional CF$_4$ and around 10 ton-years for a directional Xenon detector. For this dark matter mass, a directional experiment can reach the discovery limit with an exposure which is smaller by about an order of magnitude compared to the nondirectional case. Directionality has more impact the lighter the dark matter particle. For events to be above the energy threshold, the incoming light dark matter particles need a large velocity and, hence, have a clear arrival direction from Cygnus A. For a light target material directionality also adds to the sensitivity of heavy dark matter candidates.
\[sec:end\] Discussion and Conclusions
======================================
In this work we looked at future sensitivities of direct dark matter searches when irreducible neutrino backgrounds from coherent neutrino-nucleus elastic scattering is present. We investigated how time, recoil energy and directional information can help one distinguish signal from background. To do so, we performed a hypotheses test, as explained in section \[sec:stats\], and demanded separations of the two hypotheses at 90% confidence and three sigma level.
For the simulated detectors we assumed moderately optimistic energy thresholds and energy efficiency behavior as well as realistic smearing in the energy and angular resolution. In order to see how the target mass influences the searches, we looked at tetrafluoromethane, CF$_4$, as a light and Xenon as a heavy target material. For CF$_4$ there are detector technologies that measure the recoil track of the nucleus, whereas they have not yet been developed for Xenon.
In section \[sec:detectors\] we presented two dimensional probability distributions in recoil energy and event angle, $\theta_{\rm sun}$, for neutrino and dark matter events. In figure \[fig:dm\_pdf\] we showed the distribution for light dark matter and pointed out that it peaks at large values of $\cos \theta_{\rm sun}$. We discussed how the position of this peak evolves over the year due to the motion of the Earth around the Sun. We remarked that the lightest dark matter particles that a detector is sensitive to, need to have a large incoming velocity such that their arrival direction points back to Cygnus A.
In the same section we presented in figure \[fig:neutrino\_pdf\] the corresponding distribution for the neutrino events. The way we defined the event angle removes any time dependence of this distribution. Compared to dark matter, we noted that the solar neutrinos peak at small recoil energies and at $\cos \theta_{\rm sun} \approx -1$. The nonsolar neutrinos were assumed to have an isotropic distribution in the detector frame and act as a smooth background for the distribution.
In section \[sec:results\] we simulated one detector for each target material and fixed the exposure such that there are 500 expected neutrino events. We found that with directional information cross sections beyond the neutrino discovery limit may be probed. For the light target material we see that directional information is helpful for the complete dark matter mass range, whereas for the heavy target nuclei, directional and non-directonal detectors will give the same limits for heavy dark matter. In both cases, directional detectors can test more than an order of magnitude smaller cross sections compared to nondirectional detectors for some light dark matter masses.
We projected possible sensitivities as a function of exposure in section \[sec:max\_sensitivity\]. In figure \[fig:sens\_xenon\] and \[fig:sens\_cf4\] we saw that directional information removes the solar neutrino discovery limits and is especially useful for light dark matter. For a 6 GeV dark matter particle an exposure of approximately 5 ton-years for a directional CF$_4$ and around 10 ton-years for a directional Xenon detector is sufficient to go beyond the discovery limit of nondirectional detectors. The limit is reached with an exposure that is about an order of magnitude smaller compared to the nondirectional case. If a light target material is used, a gain in sensitivity exists also for heavy dark matter candidates, but very large exposures are needed to reach their discovery limits. We noted that perfect flux knowledge would also remove any discovery limit and conclude that there is no neutrino bound for directional dark matter searches.
This work was supported by the ERC and the STFC. We are grateful to Peter Fisher for collaboration in the early stages of this project and commenting on the manuscript. Also, we wish to thank Tevong You and Thomas Richardson for helpful discussions, Julien Billard for valuable comments on the draft and Nigel Arnot for help with software, hardware and network problems. MF is grateful for the hospitality of NORDITA during the "What is dark matter” program where many relevant interesting conversations took place.
[^1]: As the direction of Cygnus A we take a right ascention of $19^{\rm h} 59^{\rm min}28.4^{\rm s}$ and a declination of $40^{\circ} 44' 1.0''$
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Nandinii Barbosa–Cendejas'
- 'Roberto Cartas–Fuentevilla'
- 'Alfredo Herrera–Aguilar'
- 'Refugio Rigel Mora–Luna'
- Roldão da Rocha
title: A de Sitter tachyonic braneworld revisited
---
Introduction
============
In the last two decades, the paradigm underlying the observable Universe as a 3–brane, embedded in a higher-dimensional spacetime, has become a fruitful scenario for addressing several questions in physics, that cover a comprehensive plethora of phenomena, ranging from low energy physics [@lep1; @lep2; @lep3; @lep4], gravity [@gravity1; @gravity2; @gravity3; @Bazeia:2004dh; @Bazeia:2008zx], astrophysics [@astrophysics1; @astrophysics2; @astrophysics3; @astrophysics4], and high energy physics [@hep1; @hep2; @hep3; @hep4; @hep5] to cosmology [@cosmology1; @cosmology2; @cosmology3; @cosmology4; @cosmology5]. Regarding some prominent cases, the braneworld paradigm leads to plausible reformulations or even complete solutions of these problems (see Refs. [@Liu:2017gcn; @Dutra:2014xla; @thbrs; @Maartens:2010ar] for complete reviews). A striking property of these models consists of pointing out that a higher dimensional world could encode our universe, without any conflict with 4D current experiments and recent observational data.
In the context of the braneworld physics, there is a branch of models in which the fifth dimension is modeled by bulk scalar fields, extending the idea of thin branes to thick brane configurations. Several different thick brane configurations, whose dynamics is generated by a 5D gravitational action with a bulk scalar field and a self-interacting potential have been proposed. The nature of the scalar field setup leads to different specific scenarios (see the reviews [@Liu:2017gcn; @Dutra:2014xla; @thbrs]).
Despite the usefulness of the braneworld paradigm for approaching the above mentioned problems, the quest for a more realistic approach imposes a series of important tests and physical constrains. Namely, the braneworld models, whose signatures must be reflected as small corrections to the 4D physical laws according to experimental and observational data, must be able to recover the 4D physics of our Universe for precise physical limits. Hence, it is essential to define a robust physical notion of 4D gravity and matter fields localization on the brane within these models with the less possible amount of parameters.
In order to get a consistent physical model the braneworld configuration must be [*stable*]{} under small fluctuations of all the gauge, matter, and gravity background fields involved in the setup [@KKS]. Checking for this kind of stability is a highly non-trivial task, strongly dependent on the field content of the model. Another important aspect is the localization of the Standard Model matter (gauge, scalar and fermion fields) onto the brane [@fieldslocalization1; @fieldslocalization2; @fieldslocalization3; @fieldslocalization4; @fieldslocalization5; @fieldslocalization6; @fieldslocalization7; @alencar1]. So far, different scenarios of thick branes have been studied in the literature, revealing that the localization mechanism of matter fields directly depends on the warp factor into the 5D metric [@thickbraneworlds].
Within this braneworld framework, a tachyonic scalar field has been used to generate such models with recent applications to cosmology [@tachyoncosmology; @mazumdar], localization of 4D gravity on expanding 3–branes [@GADRR; @tachyonbrane1; @tachyonbrane2] and localization of the Standard Model matter fields [@coulombcorrec; @ahagauge; @AARRH; @fieldslocalizationtachyon].
The original form of the tachyonic effective action was proposed along a series of articles [@Sen0; @garousi; @bergshoeff; @kluson] as a supersymmetric generalization of the Dirac–Born–Infeld action that describes the dynamics of light modes (massless and tachyonic) on the world-volume of a non-BPS D-brane within the context of type II string theory in Minkowski spacetime [@senWSL1; @senWSL2]. This tachyonic effective action has found several applications within string cosmology [@stringcosmo1; @stringcosmo2; @stringcosmo3; @stringcosmo4; @stringcosmo5; @stringcosmo6] and supergravity [@bazeia]. Moreover, solar system constraints were imposed on its parameters by considering this action as a scalar-tensor model [@devi].
From the cosmological point of view, the early Universe and the observed accelerating expansion can be qualitatively described within this de Sitter braneworld model. Thus, these attempts to construct a consistent inflationary braneworld take into account the fact that the cosmic inflationary theory is in good agreement with temperature fluctuation properties of the Cosmic Microwave Background Radiation, and that the inflationary epoch likely took place at very high temperatures [@tachyoniflation].
Thus, the construction of a tachyonic braneworld model, that can fulfill the full set of aforementioned tests and physical constraints, is not an easy task. In fact, the corresponding Einstein and field equations are extremely nonlinear, due to the Lagrangian associated with the tachyon field. Despite these difficulties, a great effort has been made in [@GADRR] to provide an interesting tachyonic thick braneworld supported by a tachyon scalar field coupled to gravity with a bulk cosmological constant that yields certain phenomenological aspects of our 4D Universe. This model possesses several appealing properties: a) it contains an expanding metric induced on the 3–brane which arises as an exact solution to the 5D field equations; b) the field configuration is completely [*regular and stable*]{} under small perturbations [@German:2015cna]; c) The de Sitter 3–brane describes the inflationary epochs of our universe; d) the braneworld is [*asymptotically flat*]{}, despite the presence of a negative 5D cosmological constant (usually braneworlds are asymptotically dS or AdS); e) it contains a graviton spectrum with a single massless bound state that accounts for 4D gravity localized on the brane; f) it has a mass gap that makes the 5D corrections to Newton’s and Coulomb’s laws decay exponentially; g) finally, gauge, scalar and fermion fields were shown to be localized on this braneworld.
Thus, in this work we show that this tachyonic braneworld model allows for a [*more general*]{} nontrivial solution with no bulk cosmological constant, that preserves all the above mentioned remarkable properties with a less amount of parameters, constituting an important contribution to the construction of a realistic cosmological braneworld model.
The paper is organized as follows: In Sect. 2 the tachyonic thick braneworld model with the 5D cosmological constant is briefly reviewed. In Sect. 3, we obtain a second exact solution [*without*]{} the bulk cosmological constant, also analyzing its underlying parameters and some relevant derived quantities (the tachyonic self–interaction potential, the effective 4D Planck mass, and the curvature scalar), hence comparing them with their expressions when the bulk cosmological constant was present. Finally, a brief discussion of the new solution for our tachyonic braneworld model is presented and elucidated as a set of conclusions in Sect. 4.
Review of the tachyonic de Sitter thick braneworld
===================================================
The 5D action for a thick braneworld model that is produced by gravity and a bulk tachyonic scalar field in the presence of a 5D cosmological constant reads [@GADRR] $$S = \int d^5 x \sqrt{-g} \left(\frac{1}{2\kappa_5^2} R - \Lambda_5 -
V(T)\sqrt{1+g^{AB}\partial_{A} T\partial_{B} T} \right),
\label{action}$$ where $R$ is the 5D scalar curvature, $\Lambda_{5}$ is the bulk cosmological constant, the tachyon field $T$ represents the matter in the 5D bulk, $V(T)$ denotes its self–interaction potential, $\kappa_5^2=8\pi G_5$ with $G_5$ being the 5D Newton constant, and $A, B =0,1,2,3,5.$ Besides, a 5D metric ansatz with an induced 3–brane of FLRW background type is be taken to be $$\begin{aligned}
\label{metricw}
ds^2 = e^{2f(w)} \left[- d t^2 + a^2(t) \left(d x^2 + d y^2 + d
z^2 \right)+ d w^2 \right],\end{aligned}$$ where $\text{e}^{2f(w)}$ and $a(t)$ are the warp factor and the scale factor of the brane, respectively, and $w$ represents the extra–dimensional coordinate. The corresponding Einstein’s equations for this model read $$\begin{aligned}
\label{EinsteinEq_5d}
G_{AB} = - \kappa_5^2 ~\Lambda_5 g_{AB} + \kappa_5^2 ~T_{AB},\end{aligned}$$ where $T_{AB}$ is the energy–momentum tensor for the bulk tachyonic scalar field, described by $$T_{AB} = \left[ - g_{AB} \, V(T) \sqrt{1 +
(\nabla T)^2} + \frac {V(T)}{\sqrt{1+ (\nabla T)^2}} \,
\partial_{A} T \, \partial_{B} T \right].$$ The matter field equation corresponding to the action (\[action\]) is expressed in the following form: $$\Box T-\frac{\nabla_C\nabla_D T\,\, \nabla^{C} T\,\, \nabla^{D}
T}{1+(\nabla T)^2}= \frac{1}{V} \frac{\partial V(T) }{\partial T}.
\label{fieldequ1}$$
By using the metric ansatz (\[metricw\]) and the fact that the tachyon field depends only on the extra coordinate in (\[EinsteinEq\_5d\]) and (\[fieldequ1\]) (i. e. $T=T(w)$), it is straightforward to write a set of nonlinear coupled field equations, which have the following complete solution: $$a(t)=e^{H\,t}, \qquad \qquad f(w)=\frac{1}{2}\ln\left[s\,{\rm
sech}\left(\,H\,(2w+c)\right)\right], \label{scalewarpfactors}$$ $$\begin{aligned}
T(w) &=& \pm\sqrt{\frac{-3}{2\,\kappa_5^2\,\Lambda_5}}\
\mbox{arctanh}\left(\frac{\sinh\left[\frac{H\,\left(2w+c\right)}{2}\right]}
{\sqrt{\cosh\left[\,H\,(2w+c)\right]}}\right), \label{Tw}\end{aligned}$$ $$\begin{aligned}
V(T) &=& - \Lambda_5\
\mbox{sech}\left(\sqrt{-\frac{2}{3}\kappa_5^2\,\Lambda_5}\ T\right)
\sqrt{6\ \mbox{sech}^2\left(\sqrt{-\frac{2}{3}\kappa_5^2\,\Lambda_5}\ T\right)-1}\nonumber \\
&=& - \Lambda_5\ \sqrt{\left(1 +
\mbox{sech}\left[\,H\,(2w+c)\right]\right) \left(1+\frac{3}{2}\
\mbox{sech}\left[\,H\,(2w+c)\right]\right)}, \label{VT1}\end{aligned}$$ where $H$, $c$ and $s>0$ are arbitrary constants. By consistency of the above presented field equations, the constant $s$ must be $$\label{spar}
s=-\frac{6H^2}{\kappa_5^2\,\Lambda_5},$$ where the arbitrary 5D cosmological constant is negative, $\Lambda_5<0$, to ensure the real nature of the tachyonic field and its potential. It is important to note that we have an explicit expression for the self-interaction potential in terms of the tachyon field. This potential has a maximum at the position of the brane. Since the tachyonic field has a bounded domain in order to be real, it then leads to a real and bounded potential as well.
By looking at the structure of Eq. (\[spar\]) one can see that the field configuration has a limiting case when the bulk cosmological constant vanishes $\Lambda_{5} \rightarrow 0$ with the same rapidity as $H^2 \rightarrow 0$ in such a way that $s$ remains finite. In this two-fold limit, the tachyon field becomes linear $T =\pm \frac{\sqrt{s}}{4}(2w+c)$, the self-interaction potential vanishes $V(T) = 0$, whereas the scale and warp factor become a constant that can be further ignored through a rescaling of the metric coordinates, leading to a flat 5D spacetime [@GADRR].
The corresponding 5D curvature scalar reads $$R=-\frac{14}{3}\kappa_5^2\,\Lambda_5\,\mbox{sech}\left[\,H\,(2w+c)\right],
\label{R}$$ which is positive definite and asymptotically flat along the fifth dimension. This assertion is easy to see by observing the action (\[action\]) along with the asymptotic behavior of the potential (\[VT1\]) which cancels the value of the 5D cosmological constant, while $T$ remains constant at infinity.
An important point that should be stressed is that this 5D spacetime is completely free of naked singularities, that usually arise when the mass spectrum of Kaluza-Klein tensorial fluctuations display a mass gap, giving rise to an important feature of the graviton mass spectrum from the phenomenological point of view (see Refs. [@gravity1; @gravity2; @gravity3] for further details).
A detailed analysis of the stability of the tensorial metric fluctuations of the tachyonic braneworld was performed in [@GADRR], showing that 4D gravity is consistently localized on it. Moreover, the corrections to the Newton’s law, due to the 5D massive fluctuations, were computed in the thin brane limit and shown to be exponentially suppressed. On the other hand, the stability of the model under small fluctuations of the tachyon field and the scalar metric modes was successfully achieved with the aid of an auxiliary Sturm–Liouville eigenvalue problem for any value of the bulk cosmological constant in [@German:2015cna].
The thick brane with $\Lambda_5=0 $ model and its solution {#cap:New solution}
==========================================================
The action (\[action\]) that describes our tachyonic braneworld model presents a second exact solution that can be obtained from the Einstein’s equations when the bulk cosmological constant vanishes $$G_{AB} = \kappa_5^2 ~T_{AB}.
\label{einequ}$$ By making use of the metric ansatz (\[metricw\]), the Einstein tensor components read $$\begin{aligned}
G_{00} &=& 3\, \frac{\dot a^2}{a^2} - 3\, \left( f^{''} + f^{'2}
\right),
\label{eqeintach1}\nonumber \\
G_{\alpha\alpha} &=& - 2\, \ddot aa - \dot a^2 + 3\, a^2 \left(
f^{''} + f^{'2} \right) \label{eqeintach2},
\nonumber\\
G_{5 5} &=& -3\,\left( \frac{\ddot a}{a} + \frac{\dot a^2}{a^2}
\right) + 6f^{'2}, \label{eqeintach3}\end{aligned}$$ where ${``\prime"}$ and ${``\cdotp"}$ are derivatives with respect to the extra dimension and time, respectively, whereas the index $\alpha$ labels the spatial dimensions $x,$ $y$ and $z$.
The main goal of this work is to construct a new relevant solution in the context of the above mentioned tachyonic thick braneworld with a vanishing 5D cosmological constant and analyze whether the appealing properties of the brane field configuration (\[scalewarpfactors\])–(\[VT1\]) are still present in the new solution.
By taking into account the fact that the tachyon field depends only on the fifth dimension, $T(w)$, Eq. (\[fieldequ1\]) adopts the form $$T^{''}-f^{'}T^{'}+4f^{'}T^{'}(1+e^{-2 f}T^{'2})=(e^{2
f}+T^{'2})\frac{\partial_{T}V(T) }{V(T)}.
\label{fieldequ}$$ At the same time the Einstein’s equations (\[einequ\]) can be rewritten as $$\begin{aligned}
f^{''}- f^{'2} + \frac{\ddot a}{a} &=& -
\kappa_5^2\frac{V(T)T^{'2}}{3\sqrt{1 + e^{-2 f}\,T^{'2}}} ,
\label{einsteinequ} \\
f^{'2} - \frac{1}{2} \left(\frac{\ddot a}{a}+\frac{\dot a^2}{a^2}
\right)&=& - \kappa_5^2\frac{e^{2 f}\,V(T)}{6\sqrt{1 + e^{-2
f}\,T^{'2}}} .
\label{restriccion}\end{aligned}$$ By demanding consistency of this system of equations, the 3–brane solution corresponds to a 4D de Sitter cosmological background defined by $$a(t)= e^{H\,(t-t_0)}, \label{scalefactor}$$ however, the constant factor $a_0=1/e^{H\,t_{0}}$ can be absorbed by a rescaling of the spatial coordinates. Thus, this result in a braneworld model in which the induced metric on the 3–brane is described by a $dS_4$ geometry.
By further taking into account the form of the scale factor $a(t)= e^{H\,t}$, we can recast (\[einsteinequ\]) and (\[restriccion\]) as follows $$\begin{aligned}
f^{''} - f^{'2} + H^2 &=& - \kappa_5^2\frac{V(T)T^{'2}}{3\sqrt{1 +
e^{-2 f}T^{'2}}},
\label{Ee1z} \\
f^{'2} - H^2 &=& - \kappa_5^2\frac{e^{2 f}V(T)}{6\sqrt{1 + e^{-2
f}T^{'2}}}. \label{Ee2z}\end{aligned}$$ Now it is straightforward to obtain separate equations for the derivative of the scalar field $T$ and the self-interaction potential $V(T)$ from Eqs. (\[Ee1z\]) and (\[Ee2z\]), leading to: $$\begin{aligned}
T'&=&\pm e^f\sqrt{\frac{f''-f'^2+H^2}{2\left(f'^2-H^2\right)}},
\label{Tprime}\\
V(T)&=& \pm \frac{3}{\kappa_5^2}\,e^{-2f}\,\sqrt{2\left(f''+
f'^2-H^2\right)\left(f'^2 -H^2\right)}. \label{V}\end{aligned}$$ Therefore, by determining the nature of the scaling and warp factors, the self-interaction potential $ V(T) $ and the derivative of the tachyon field $T'$ are completely determined. However, care must be taken, since the tachyon scalar field must be real and have a form that ensures the localization of 4D gravity, whereas the self-interaction potential must be real and well defined in the sense that ensures the stability of the whole braneworld field configuration. These restrictions are very demanding, several warp factors with “convenient" localizing behavior lead to a complex tachyon field $T$. On the other hand, ensuring a real and stable tachyon potential often yields either a complex tachyon field $T$ or a warp factor that does not enable the localization of 4D gravity. Hence, we propose the most general form for the warp factor as follows $$\label{fwg}
f(w)=-n\ln\left[\frac{\cosh\left[\,H\,(\frac{w}{n}+c)\right]}{s}\right],$$ where $H$, $c$, $s$, and $n$ are constants. Here we should demand $s>0$ and $n>0$ in order to have a warp factor that ensures the localization of 4D gravity and other matter fields. It is straightforward to realize that this warp factor is a solution of the Einstein and field equations if the tachyon scalar field adopts the [*general*]{} form $$\label{Twg}
T(w) = \pm i\frac{\sqrt{\frac{n}{2 (1-n)}}\,\, s^n \,\,
_2F_1\left(\frac{1}{2},\frac{1-n}{2};\frac{3-n}{2};\cosh ^2\left[H
\left(\frac{w}{n}+c\right)\right]\right)}{H \cosh ^{n-1}\left[H
\left(\frac{w}{n}+c\right)\right]} + k, \quad n\neq 1,$$ where $k$ denotes a pure imaginary number. The tachyon scalar field is purely real within the domain $0<n<1$ andm under a suitable choice of the parameter $k$ (for non–integer values $n>1$ the tachyon field is purely imaginary), and the tachyon potential is given by the following expression $$\label{VTg}
V(w)=\frac{3 H^2 \sqrt{\frac{2 (n+1)}{n}} }{k_5^2 s^{2
n}}\text{sech}^{2(1-n)}\left[H \left(\frac{w}{n}+c\right)\right].$$ It is convenient to have at hand a real tachyon scalar field that allow us to explicitly write the self-interaction potential $V(w)$ in terms of the scalar field $T$. In order to satisfy both conditions we consider the particular case when $n=\frac{1}{2}$. For this specific value, the warp factor, the tachyon scalar field and the self-interaction potential read $$\begin{aligned}
f(w)
&=&-\frac{1}{2}\ln\left[\frac{\cosh\left[\,H\,(2w+c)\right]}{s}\right],
\label{fw}\\
T(w) &=& \pm \frac{1}{\sqrt{2}b}\,i\,\text{EllipticF}\left( i H
\left(
w+\frac{c}{2}\right),2\right), \label{Tw2}\\
V(T) &=& \frac{3 \sqrt{6}\, b^2}{k_5^2 }\,\text{sec}\left[2\text{JacobiAmplitude}\left(i \sqrt{2} b T,2\right)\right]\nonumber \\
&=& \frac{3 \sqrt{6}\, b^2}{k_5^2 }\,\text{sech} \left(H \left(2
w+c\right)\right), \label{VT}\end{aligned}$$ where the $\text{JacobiAmplitude}(u,m)$ gives the amplitude $\text{am}(u | m)$ for Jacobi elliptic functions, i. e., it is the inverse of the elliptic integral of the first kind. In the last two equations we set $$s=\frac{H^2}{b^2}.$$ where $b$ is an arbitrary constant and the scale factor remains the same.
![The profile of the tachyonic scalar field $T$. The thin line represents the positive branch of the field and the thick line is associated to the negative branch of the tachyon. Here we have set $n=1/2$, $c=0,$ $H=1$, $2\kappa_{5}^2=1$ and $s=1$ for simplicity.[]{data-label="fig_T"}](tkink.eps){width="7cm"}
-5mm
![The shape of the self–interaction potential of the tachyonic scalar field $V(T)$. We set $n=1/2$, $c=0,$ $H=1$, $2\kappa_{5}^2=1$ and $s=1$ for simplicity.[]{data-label="fig_VT"}](vt.eps){width="7cm"}
-5mm
Therefore, the set of equations (\[scalefactor\]), (\[fw\]), (\[Tw2\]), and (\[VT\]) provide a new relevant tachyonic thick braneworld configuration. By comparing with the previous solution (\[scalewarpfactors\]), (\[Tw\]) and (\[VT1\]), the new warp factor (\[fwg\]) also has a decaying and vanishing asymptotic behavior for $0<n<1$ and coincides with (\[scalewarpfactors\]) when $n=1/2$. Similarly, the new tachyon scalar field (\[Tw2\]) is real and possesses a kink– or antikink–like profile as shown in Fig. 1, but with the positive and negative branches opposite to those presented by the field (\[Tw\]). On the other hand, the potential (\[VT\]) also has a maximum at the position of the brane, it is positive definite and bounded (from below and above), but unlike (\[VT1\]), it vanishes as $w$ tends to infinity or, equivalently, as $T$ approaches it asymptotic value (see Fig. 2), emphasizing the fact that the braneworld is asymptotically flat.
It is important to recall that the new solution presented in this work cannot be recovered from the solution (\[scalewarpfactors\])-(\[VT1\]) with the aid of a two-fold limit. On the other hand, it can be noticed as well that when $\Lambda_5=0$, the tachyon scalar field (\[Tw\]) diverges, while the potential (\[VT1\]) vanishes, indicating that this is not a physical limit.
By substituting Eqs. (\[Twg\]) and (\[VTg\]) into Eqs. (\[Ee1z\]) and (\[Ee2z\]), we can see that there is no restriction on the $s$ parameter coming from the field equations of the system. This contrasts with the first solution generated by the warp factor (\[scalewarpfactors\]) with $\Lambda_5 \neq 0$, where the field equations demand the relation $s=-\frac{6H^2}{\kappa_5^2\,\Lambda_5} \label{s}$.
However, when we compute the effective 4D Planck mass $M_{pl}$ for the special value $n=\frac{1}{2}$ with no bulk cosmological constant $\Lambda_{5}=0$, then we obtain $s=H^2/b^2$, where $b$ is an arbitrary constant, yielding $$\label{Jm1}
M_{\text{Pl}}^{2} \sim M_*^3 \frac{\sqrt{\frac{2}{\pi }} \Gamma
\left(\frac{3}{4}\right)^2 H^2}{b^3}.$$ This expression shows the relationship between the Planck mass in 5D and 4D, adjusted with the $b$ parameter. Comparing to the braneworld model presented in [@GADRR], the effective Planck mass in 4D is related to the 5D one by $$\label{Jm2}
M_{\text{Pl}}^{2} \sim - M_*^3 \frac{\sqrt{\frac{2}{\pi }}\Gamma
\left(\frac{3}{4}\right)^2 H^2}{\Lambda_5^3},$$ which diverges when taking the single limit $\Lambda_5 \rightarrow 0$. However, one can still perform a two-fold limit, leading to a finite value of the 4D Planck mass $M_{\text{Pl}}^{2}$.
Finally, computing the 5D curvature scalar for our solution (\[fwg\]) yields $$R=\frac{4 H^2 (3 n+2)}{n\,s^{2 n} } \text{sech} ^{2 (1-n)}\left[H
\,\left(\frac{w}{n}+c\right)\right],
\label{R5n}$$ in agreement with the asymptotic form dictated by the self-interaction potential (\[VT\]). For the special value $n=1/2$, it reads $$R=28\,b^2\,\mbox{sech}\left[\,H\,(2w+c)\right]. \label{R5}$$
This 5D invariant is positive definite and asymptotically vanishes, yielding an asymptotically 5D Minkowski spacetime as the solution (\[scalewarpfactors\])-(\[VT1\]) does. However, unlike the curvature scalar corresponding to the braneworld [@GADRR], this 5D invariant is obtained within a simpler theory, since the original action (\[action\]) has one less input parameter ($\Lambda_5 =0$).
Discussion and conclusions
==========================
We presented a new complete solution for the braneworld model (\[action\]), generated by a tachyonic scalar field minimally coupled to gravity, with no bulk cosmological constant, and a 5D warped metric ansatz (\[metricw\]), with respect to which the 3–brane metric defines a FLRW geometry. The resulting spacetime is regular and stable along the whole fifth dimension for certain values of the $n$ parameter. Analytic expressions were derived for the physically meaningful warp factor (\[fwg\]), the tachyon scalar field (\[Twg\]), the self-interaction potential (\[VTg\]) and the curvature scalar (\[R5n\]), for the significant values of the $n$ parameter (for $0<n<1$).
On the one hand, the profile of the warp factor (\[fwg\]) and the 5D Ricci scalar (\[R5n\]) show a regular structure within the domain $n\in(0,1)$. These geometrical quantities show that 4D gravity, as well as other Standard Model matter fields, can be localized in the tachyonic de Sitter braneworld model, whenever $0<n<1$. On the other hand, the tachyon scalar field (\[Twg\]) and its self-interaction potential (\[VTg\]) are real only for the aforementioned domain $0<n<1$. Moreover, the potential (\[VTg\]) leads to a completely [*stable*]{} braneworld configuration, in full compliance to the treatment presented in [@German:2015cna]. In fact, it has a maximum for precisely this range of the $n$ parameter.
It is worth noticing as well a physically consistent picture at the level of the Einstein’s equations: the geometrical left hand side localizes the known 4D interactions of our world and renders a regular, asymptotically flat braneworld for $0<n<1$, whereas the matter right hand side is real and makes the whole field configuration stable, for the same domain $0<n<1$ within our setup.
It is important to note that the 5D spacetime, for the special value $n=\frac{1}{2}$, supports the same geometrical configuration as the solution presented in [@GADRR]. However, it has one parameter less involved in the initial braneworld action, since $\Lambda_5=0$. This fact yields the localization properties of the model, with $n=\frac{1}{2}$ remaining the same, since the form of the warp factor determines the character of the 5D graviton spectrum of KK massive modes (which can be gapless or possess a mass gap, for instance) and its localizing properties.
Thus, the braneworld solution (\[fw\]) presented in Section \[cap:New solution\] preserves all the aforementioned appealing properties that we look for in such a model for being realistic. In fact, it is completely regular and stable under small metric and tachyon field perturbations, it localizes 4D gravity and from the low-energy point of view recovers the Newton’s law [@GADRR] and the Coulomb’s law [@coulombcorrec] in the thin brane limit with small, exponentially suppressed corrections that come from the extra dimension. Moreover, the structure of the warp factor given in Eq. (\[fw\]) for this model makes possible the localization of massive (massless) scalar fields with spin-$0$ [@AARRH], gauge boson fields with spin–[1]{} [@ahagauge], and spin–$1/2$ fermions [@coulombcorrec].
It seems that the localization of these matter fields can be achieved for any value of the $n$ parameter within the interval $0<n<1$. However, it is not straightforward to invert the relation we obtain upon integration of (\[Tprime\]) in order to get an explicit expression for the tachyonic scalar field in terms of the fifth dimension and an analytic expression for the self-interaction potential in terms of the tachyon field for a value $n\ne1/2$, representing a major technical difficulty that we are currently trying to address.
Finally, one might use the information entropy on thick branes [@Correa:2016pgr; @Correa:2015vka], to determine a more precise bound on the $n$ parameter and to derive refinements of the model here presented.
Acknowledgments {#acknowledgments .unnumbered}
===============
NBC is grateful to FIE-UMSNH for the designation of a working space and to Juan Herrera for technical support. The work of A.H.-A. was completed at the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1607611 and a Simons Foundation grant as well. He expresses his gratitude to the ACP for providing an inspiring and encouraging atmosphere for conducting part of this research. RCF and AHA acknowledge a VIEP-BUAP grant. RCF, NBC, AHA, and RRML thank SNI for support. RdR is grateful to CNPq (Grant No. 303293/2015-2), and to FAPESP (Grants No. 2015/10270-0 and No. 2017/18897-8), for partial financial support.
[999]{}
R. Linares, H. A. Morales-Técotl and O. Pedraza, *Gravitational potential of a point mass in a brane world,* Phys. Rev. D [**86**]{} (2014) 066002 \[[arXiv:1312.7060 \[hep-ph\]]{}\].
R. Linares, H. A. Morales-Técotl and O. Pedraza, *Casimir force in brane worlds: coinciding results from Green’s function and Zeta function approaches*, Phys. Rev. D [**81**]{} (2010) 126013 \[[arXiv:1003.4286 \[hep-th\]]{}\].
R. Linares, H. A. Morales-Técotl and O. Pedraza, *Casimir force for a scalar field in warped brane worlds*, Phys. Rev. D [**77**]{} (2008) 066012. \[[arXiv:0712.3963 \[hep-ph\]]{}\]. H. A. Morales-Técotl, O. Pedraza and L. O. Pimentel, *Low-energy effects in brane worlds: Liennard-Wiechert potentials and Hydrogen Lamb shift*, *Gen. Rel. Grav.* [**39**]{} (2007) 1185 \[[arXiv: physics/0611241]{}\].
C. Csaki, J. Erlich, T. Hollowood and Y. Shirman, *Universal Aspects of gravity localized on thick branes*, Nucl. Phys. B[**581**]{} (2000) 309 \[[hep-th/0001033]{}\]. M. Gremm, *Four–dimensional gravity on a thick domain wall*, Phys. Lett. B[**478**]{} (2000) 434 \[[hep-th/9912060]{}\].
M. Gremm, *Thick domain walls and singular spaces*, Phys. Rev. D[**62**]{} (2000) 044017 \[[hep-th/0002040]{}\]. D. Bazeia and A. R. Gomes, *Bloch brane*, JHEP [**0405**]{} (2004) 012 \[[hep-th/0403141]{}\]. D. Bazeia, A. R. Gomes, L. Losano and R. Menezes, *Braneworld Models of Scalar Fields with Generalized Dynamics*, Phys. Lett. B [**671**]{} (2009) 402 \[[arXiv:0808.1815 \[hep-th\]]{}\]. F. X. Linares, M. A. García-Aspeitia and L. A. Ureña-López, *Stellar models in Brane Worlds*, Phys. Rev. D[**92**]{} (2015) no.2, 024037 \[[arXiv:1501.04869 \[gr-qc\]]{}\]. M. A. García-Aspeitia, *Branes constrictions with White Dwarfs*, Eur. Phys. J. C[**75**]{} (2015) no.11, 53 \[[arXiv:1510.06814 \[gr-qc\]]{}\]. J. E. Lidsey, T. Matos and L. A. Ureña- López, *The inflaton field as selfinteracting dark matter in the brane world scenario*, Phys. Rev. D[**66**]{} (2002) 023514 \[[astro-ph/0111292]{}\]. J. Garriga and M. Sasaki, *Brane world creation and black holes*, Phys. Rev. D[**62**]{} (2000) 043523 \[[hep-th/9912118]{}\].
M. Becker and H. Päs, *Leptonic Flavor Structure in the Brane Shifted Extra Dimensional Seesaw Mechanism*, \[[arXiv:1707.02882 \[hep-ph\]]{}\]. G. Itsios, N. Jokela, and A. V. Ramallo, *Collective excitations of massive flavor branes*, Nucl. Phys. B[**909**]{} (2016) 677 \[[arXiv:1602.06106 \[hep-th\]]{}\]. J. A. Cabrer, G. von Gersdorff, and M. Quiros, *Flavor Phenomenology in General 5D Warped Spaces*, JHEP [**1201**]{} (2012) 033 \[[arXiv:1110.3324 \[hep-ph]{}\]\]. J. A. Cabrer, G. von Gersdorff, and M. Quiros, *Improving Naturalness in Warped Models with a Heavy Bulk Higgs Boson*, Phys. Rev. D[**84**]{} (2011) 035024 \[[arXiv:1104.3149 \[hep-ph\]]{}\]. J. A. Cabrer, G. von Gersdorff, and M. Quiros, *Suppressing Electroweak Precision Observables in 5D Warped Models*, JHEP [**1105**]{} (2011) 083 \[[arXiv:1103.1388 \[hep-ph\]]{}\].
N. Bilic, S. Domazet and G. S. Djordjevic, *Particle creation and reheating in a braneworld inflationary scenario*, \[[arXiv:1707.06023 \[hep-th\]]{}\]. N. Bilic, S. Domazet and G. S. Djordjevic, *Tachyon with an inverse power-law potential in a braneworld cosmology*, Class. Quant. Grav. [**34**]{} (2017) no.16, 165006 \[[arXiv:1704.01072 \[gr-qc\]]{}\]. S. Bag and V. Sahni, *Constraining the Cosmology of the Phantom Brane using Distance Measures*, Phys. Rev. D [**95**]{} (2017) no.2, 023524 \[[arXiv:1605.04707 \[astro-ph.CO\]]{}\]. N. Bilic, S. Domazet and G. S. Djordjevic, K. Koyama, *Cosmological Tests of Modified Gravity*, Rept. Prog. Phys. [**79**]{} (2016) no.4, 046902 \[[arXiv:1504.04623 \[astro-ph.CO\]]{}\]. I. P. Neupane, *Natural Braneworld Inflation in Light of Recent Results from Planck and BICEP2*, Phys. Rev. D [**90**]{} (2014) no.12, 123502 \[[arXiv:1409.8647 \[astro-ph.CO\]]{}\].
Y. X. Liu, *Introduction to Extra Dimensions and Thick Braneworlds*, \[[arXiv:1707.08541 \[hep-th\]]{}\].
A. de Souza Dutra, G. P. de Brito and J. M. Hoff da Silva, *Method for obtaining thick brane models,* Phys. Rev. D [**91**]{} (2015) no.8, 086016 \[[arXiv:1412.5543 \[hep-th\]]{}\].
V. Dzhunushaliev, V. Folomeev and M. Minamitsuji, *Thick brane solutions*, Rept. Prog. Phys. [**73**]{} (2010) 066901 \[[arXiv:0904.1775 \[gr-qc\]]{}\].
R. Maartens and K. Koyama, *Brane-World Gravity*, Living Rev. Rel. [**13**]{} (2010) 5 \[[arXiv:1004.3962 \[hep-th\]]{}\].
S. Kobayashi, K. Koyama and J. Soda, *Thick brane worlds and their stability*, Phys. Rev. D[**65**]{} (2002) 064014 \[[hep-th/0107025]{}\].
A. E. Bernardini and R. da Rocha, *Matter localization on brane-worlds generated by deformed defects*, Adv. High Energy Phys. [**2016**]{} (2016) 3650632 \[[arXiv:1606.05921 \[hep-th\]]{}\]. Z. H. Zhao, Q. Y. Xie and Y. Zhong, *New localization method of $U(1)$ gauge vector field on flat branes in (asymptotic) $AdS_{5}$ spacetime*, Class. Quant. Grav. [**32**]{} (2015) no.3, 035020 \[[arXiv:1406.3098 \[hep-th\]]{}\]. A. E. R. Chumbes, J. M. Hoff da Silva and M. B. Hott, *A model to localize gauge and tensor fields on thick branes*, Phys. Rev. D [**85**]{} (2012) 085003 \[[arXiv:1108.3821 \[hep-th\]]{}\].
K. P. S. de Brito and R. da Rocha, *New fermions in the bulk*, J. Phys. A [**49**]{} (2016) no.41, 415403 \[[arXiv:1609.06495 \[hep-th\]]{}\]. N. Barbosa-Cendejas, D. Malagón-Morejón and R. R. Mora-Luna, *Universal spin-1/2 fermion field localization on a 5D braneworld*, Gen. Rel. Grav. [**47**]{} (2015) no.7, 77 \[[arXiv:1503.07900 \[hep-th\]]{}\]. A. Melfo, N. Pantoja and J. D. Tempo, *Fermion localization on thick branes*, Phys. Rev. D [**73**]{} (2006) 044033 \[[hep-th/0601161]{}\]. O. DeWolfe, D. Z. Freedman, S. S. Gubser and A. Karch, *Modeling the fifth dimension with scalars and gravity*, Phys. Rev. D[**62**]{} (2000) 046008 \[[hep-th/9909134]{}\]. G. Alencar, R. R. Landim, M. O. Tahim and R. N. Costa Filho, *Gauge Field Localization on the Brane Through Geometrical Coupling*, Phys. Lett. B [**739**]{} (2014) 125 \[arXiv:1409.4396 \[hep-th\]\]. N. Barbosa-Cendejas, A. Herrera-Aguilar, U. Nucamendi, I. Quiros and K. Kanakoglou, *Mass hierarchy, mass gap and corrections to Newton’s law on thick branes with Poincaré symmetry*, Gen. Rel. Grav. [**46**]{} (2014) 1631 \[[arXiv0712.3098 \[hep-th\]]{}\].
V. Kamali and E. N. Nik, *Tachyon logamediate inflation on the brane*, Eur. Phys. J. C [**77**]{} (2017) no.7, 449 \[[arXiv:1707.02773 \[gr-qc\]]{}\].
N. Barbosa-Cendejas, J. De-Santiago, G. Germán, J. C. Hidalgo and R. R. Mora-Luna, *Tachyon inflation in the large-$N$ formalism*, JCAP [**1511**]{} (2015) 020 \[[arXiv:1506.09172 \[astro-ph.CO\]]{}\].
A. Mazumdar, S. Panda, and A. Pérez-Lorenzana, *Assisted inflation via tachyon condensation*, Nucl. Phys. B[**614**]{} (2001) 101 \[[arXiv: hep-ph/0107058]{}\].
G. Germán, A. Herrera-Aguilar, D. Malagón-Morejón, R. R. Mora-Luna and R. da Rocha, *A de Sitter tachyon thick braneworld and gravity localization*, JCAP [**1302**]{} (2013) 035 \[[arXiv:1210.0721 \[hep-th\]]{}\].
R. Koley and S. Kar, *A Novel braneworld model with a bulk scalar field*, Phys. Lett. B[**623**]{} (2005) 244 [*Erratum–ibid.*]{} B[**631**]{} (2005) 199 \[[hep-th/0507277]{}\]. S. Pal and S. Kar, de Sitter branes with a bulk scalar, Gen. Rel. Grav. [**41**]{} (2009) 1165 \[[hep-th/0701266]{}\].
R. Cartas-Fuentevilla, A. Escalante, G. Germán, A. Herrera-Aguilar and R. R. Mora-Luna, *Coulomb’s law corrections and fermion field localization in a tachyonic de Sitter thick braneworld*, JCAP [**1605**]{} (2016) no.05, 026 \[[arXiv:1412.8710 \[hep-th\]]{}\].
A. Herrera-Aguilar, A. D. Rojas, E. Santos-Rodríguez,*Localization of gauge fields in a tachyonic de Sitter thick braneworld*, Eur. Phys. J. C[**74**]{} (2014) 2770 \[[arXiv:1401.0999 \[hep-th\]]{}\].
A. Díaz-Furlong, A. Herrera-Aguilar, R. Linares, R. R. Mora-Luna and H. A. Morales-Técotl, *On localization of universal scalar fields in a tachyonic de Sitter thick braneworld*, Gen. Rel. Grav. [**46**]{} (2014) no.12, 1815 \[[arXiv:1407.0131 \[hep-th\]]{}\].
X.-H. Zhang, Y.-X. Liu, and Y.-S. Duan, *Localization of Fermionic Fields on Braneworlds with Bulk Tachyon Matter*, Mod. Phys. Lett. A[**23**]{} (2008) 2093 \[[arXiv: 0709.1888 \[hep-th\]]{}\].
A. Sen, *Supersymmetric world volume action for nonBPS D-branes*, *JHEP* [**9910**]{} (1999) 008 \[[hep-th/9909062]{}\].
M. R. Garousi, *Tachyon couplings on non-BPS D-branes and Dirac-Born-Infeld action*, *Nucl. Phys. B* [**584**]{} (2000) 284 \[[hep-th/0003122]{}\].
E. A. Bergshoeff, M. de Roo, T. C. de Wit, E. Eyras, S. Panda, *T duality and actions for non-BPS D-branes*, *JHEP* [**0005**]{} (2000) 009 \[[hep-th/0003221]{}\].
J. Kluson, *Proposal for non-BPS D-brane action*, *Phys. Rev. D* [**62**]{} (2000) 126003 \[[hep-th/0004106]{}\].
A. Sen, *Tachyon Condensates on the Brane-Antibrane System*, *JHEP* [**9808**]{} (1998) 012 \[[hep-th/9805170]{}\].
A. Sen,*Non-BPS States and Branes in String Theory*, in *Advanced School on Supersymmetry in the Theories of Fields, Strings and Branes*, Eds. J. L. F. Barbon and J. M. F. Labastida (Singapore, World Scientific, 2001) 307 p. \[[hep-th/9904207]{}\].
A. Sen, *Time and tachyon*, *Int. J. Mod. Phys. A* [**18**]{} (2003) 4869 \[[hep-th/0209122]{}\]. A. Sen, *Rolling tachyon*, *JHEP* [**0204**]{} (2002) 048 \[[hep-th/0203211]{}\]. D. Choudhury, D. Ghoshal, D. P. Jatkar, S. Panda, *Hybrid inflation and brane – anti-brane system*, *JCAP* [**0307**]{} (2003) 009 \[[hep-th/0305104]{}\]. D. Choudhury, D. Ghoshal, D. P. Jatkar, S. Panda, *On the cosmological relevance of the tachyon*, *Phys. Lett. B* [**544**]{} (2002) 231 \[[hep-th/0204204]{}\]. G. W. Gibbons, [*Cosmological evolution of the rolling tachyon.* ]{}, *Phys. Lett. B* [**537**]{} (2002) 1 \[[hep-th/0204008]{}\]. T. Padmanabhan, *Accelerated expansion of the universe driven by tachyonic matter*, *Phys. Rev. D* [**66**]{} (2002) 021301 \[[hep-th/0204150]{}\].
D. Bazeia, F. A. Brito and J. R. Nascimento, *Supergravity brane worlds and tachyon potentials*, Phys. Rev. D [**68**]{} (2003) 085007 \[[hep-th/0306284]{}\].
N. C. Devi, S. Panda, A. A. Sen, *Solar System Constraints on Scalar Tensor Theories with Non-Standard Action*, *Phys. Rev. D* [**84**]{} (2011) 063521 \[[arXiv: 1104.0152 \[gr-qc\]]{}\].
G. Germán, A. Herrera-Aguilar, A. M. Kuerten, D. Malagón-Morejón and R. da Rocha, *Stability of a tachyon braneworld*, JCAP [**1601**]{} (2016) no.01, 047 \[[arXiv:1508.03867 \[hep-th\]]{}\].
R. A. C. Correa, D. M. Dantas, C. A. S. Almeida and R. da Rocha, *Bounds on topological Abelian string-vortex and string-cigar from information-entropic measure*, Phys. Lett. B [**755**]{} (2016) 358 \[[arXiv:1601.00076 \[hep-th\]]{}\]. R. A. C. Correa and R. da Rocha, *Configurational entropy in brane-world models*, Eur. Phys. J. C [**75**]{} (2015) no.11, 522 \[[arXiv:1502.02283 \[hep-th\]]{}\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define higher order infinitesimal noncommutative probability space and infinitesimal non-crossing cumulant functionals. In this framework, we generalize to higher order the notion of infinitesimal freeness, via a vanishing of mixed cumulants condition. We also introduce and study some non-crossing partitions related to this notion. Finally, as an application, we show how to compute the successive derivatives of the free convolution of two time-indexed families of distributions from their individual derivatives.'
author:
- 'M. Février[^1]'
title: ' Higher order infinitesimal freeness[^2]'
---
[**1. Introduction**]{}
Free probability theory was introduced by Voiculescu in the eighties with motivations from operator algebras [@vdn], but many connections to other fields of mathematics like random matrices (see [@voi]) or combinatorics appeared. The combinatorial side of free probability, as noticed by Speicher [@spe], is linked to the convolution on the lattices of non-crossing partitions, which have been first studied by Kreweras [@kre]. Biane proved in [@bia] that these lattices of non-crossing partitions can be embedded into the Cayley graphs of the symmetric groups, also known as the type A in the classification of finite reflection groups. Reiner introduced non-crossing partitions related to other types in this classification [@rei].\
In [@bgn], the authors showed that it is possible to build a free probability theory of type B, by replacing the occurences of the symmetric groups and the non-crossing partitions of type A by their type B analogues, namely the hyperoctaedral groups and the non-crossing partitions of type B. In their work, a central role is played by the boxed convolution which is a combinatorial operation having a natural type B analogue and describing the multiplication of two freely independent noncommutative random variables. The specificity of the boxed convolution of type B led the authors to define a noncommutative probability space of type B as a system $(\mathcal{A},\varphi ,\mathcal{V},f,\Phi )$, where $(\mathcal{A},\varphi )$ is a noncommutative probability space, $\mathcal{V}$ is a complex vector space, $f : \mathcal{V} \longrightarrow {\mathbb{C}}$ is a linear functional, $\Phi : \mathcal{A} \times \mathcal{V} \times \mathcal{A} \longrightarrow \mathcal{V}$ is a two-sided action of $\mathcal{A}$ on $\mathcal{V}$. A type B noncommutative random variable is therefore a couple $(a,\xi )\in \mathcal{A} \times \mathcal{V}$, its distribution is ${\mathbb{C}}^2$-valued and the non-crossing cumulant functionals of type B introduced in [@bgn] are also with values in ${\mathbb{C}}^2$. An important remark is that the first component of a non-crossing cumulant of type B in $(\mathcal{A},\varphi ,\mathcal{V},f,\Phi )$ is simply a non-crossing cumulant of type A in $(\mathcal{A},\varphi )$. It follows that the notion of freeness of type B for $(\mathcal{A}_1, \mathcal{V}_1) , \ldots , (\mathcal{A}_m, \mathcal{V}_m)$ in $(\mathcal{A},\varphi ,\mathcal{V},f,\Phi )$, defined in [@bgn] in terms of moments to ensure that the vanishing of mixed cumulants of type B holds, implies the freeness of $\mathcal{A}_1, \ldots , \mathcal{A}_m$ in $(\mathcal{A},\varphi )$. The free additive convolution of type B, denoted by $\boxplus ^{(B)}$, which describes the distribution of the sum of two type B noncommutative random variables that are free of type B, is an operation on the set of couples $(\mu ,\mu ')$ of linear functionals on ${\mathbb{C}}[X]$ satisfying $\mu (1)=1$ and $\mu '(1)=0$. Later, Popa stated in [@pop] type B versions of usual limit theorems and defined a S-transform for noncommutative random variables of type B.\
Recently, the analytic aspects of free probability theory of type B were investigated in [@bs] ; in particular, the authors outlined an interesting application of the free probability of type B that they called infinitesimal freeness : defining (when they exist) the zeroth and first derivatives at $0$ of a time-indexed family of distributions $(\mu _t)_{t>0}$ as the couple of distributions $(\mu ^{(0)}, \mu ^{(1)})$ defined by $$\mu ^{(0)}=\lim_{t\rightarrow 0} \mu _t$$ and $$\mu ^{(1)}=\frac{d}{dt}_{|t=0}\mu _t=\lim_{t\rightarrow 0} \frac{1}{t}(\mu _t-\mu ^{(0)}),$$ they prove that, given two such time-indexed families of distributions $(\mu _t)_{t>0}$ and $(\nu _t)_{t>0}$, the zeroth and first derivatives at $0$ of $\mu _t \boxplus \nu _t$, denoted by $((\mu \boxplus \nu )^{(0)} , (\mu \boxplus \nu )^{(1)})$, satisfy : $$((\mu \boxplus \nu )^{(0)} , (\mu \boxplus \nu )^{(1)})=
(\mu ^{(0)}, \mu ^{(1)})\boxplus ^{(B)}(\nu ^{(0)}, \nu ^{(1)}).$$\
Following this insight, a new approach of free probability of type B was developed in [@fn], named infinitesimal freeness. The equivalent structures considered there, simplifying and generalizing the noncommutative probability space of type B from [@bgn], are the infinitesimal noncommutative probability space $(\mathcal {A},\varphi ,\varphi ')$ consisting in a noncommutative probability space $(\mathcal {A},\varphi )$ to which has been added another linear functional $\varphi '$ on $\mathcal {A}$ satisfying $\varphi '(1_\mathcal {A})=0$, and the scarce $\mathbb{G}$-noncommutative probability space $(\mathcal {A},\tilde \varphi )$, where $\tilde \varphi $ is a linear map which consolidates the two functionals $\varphi ,\varphi '$ in a single one from $\mathcal {A}$ into a two-dimensional Grassman algebra $\mathbb{G}$ generated by an element $\varepsilon $ which satisfies $\varepsilon ^2 =0$. A scarce $\mathbb{G}$-noncommutative probability space appears in the framework of a noncommutative probability space of type B $(\mathcal{A},\varphi ,\mathcal{V},f,\Phi )$ when one considers the link-algebra $\mathcal{A}\times \mathcal{V}$ together with the map $(\varphi , f)$. Infinitesimal freeness of unital subalgebras $\mathcal{A}_1, \ldots , \mathcal{A}_m$ of an infinitesimal noncommutative probability space $(\mathcal {A},\varphi ,\varphi ')$ is defined as the rewriting of the condition defining freeness of type B in a more general context. More precisely, $\mathcal{A}_1, \ldots , \mathcal{A}_m$ are infinitesimally free if whenever $i_1, \ldots , i_n \in \{ 1, \ldots , k \}$ are such that $i_1 \neq i_2, i_2 \neq i_3 , \ldots , i_{n-1} \neq i_n$, and $a_1 \in \mathcal{A}_{i_1}, \ldots , a_n \in \mathcal{A}_{i_n}$ are such that $\varphi (a_1) = \cdots = \varphi (a_n) = 0$, then $\varphi ( a_1 \cdots a_n ) = 0$ and $$\varphi '( a_1 \cdots a_n ) =
\left\{ \begin{array}{l}
\varphi (a_1 \, a_n) \varphi (a_2 \, a_{n-1}) \cdots
\varphi ( a_{(n-1)/2} \, a_{(n+3)/2} ) \cdot
\varphi '( a_{(n+1)/2} ), \\
\mbox{$\ \ $} \ \ \mbox{ if $n$ is odd and $i_1 = i_n ,
i_2 = i_{n-1}, \ldots , i_{(n-1)/2} = i_{(n+3)/2}$, } \\
0, \ \mbox{ otherwise.}
\end{array} \right.$$ It is clear from this definition that infinitesimally free unital subalgebras of an infinitesimal noncommutative probability space $(\mathcal {A},\varphi ,\varphi ')$ are in particular free in $(\mathcal {A},\varphi )$. A converse is proved in [@fn] : given free unital subalgebras $\mathcal{A}_1, \ldots , \mathcal{A}_m$ of a noncommutative probability space $(\mathcal {A},\varphi )$, $\mathcal{A}_1, \ldots , \mathcal{A}_m$ are infinitesimally free in the infinitesimal noncommutative probability space $(\mathcal {A},\varphi ,\varphi ')$, for instance when we set $\varphi ':=\varphi \circ D$, where $D : \mathcal {A}\longrightarrow \mathcal {A}$ is a derivation such that $\forall 1\leq i\leq m, D(\mathcal{A}_i)\subseteq \mathcal{A}_i$. Moreover, a method is presented to obtain analogues in the framework of interest of an infinitesimal noncommutative probability space $(\mathcal {A},\varphi ,\varphi ')$ for results already established in usual free probability. This method is roughly to work in $(\mathcal {A},\tilde \varphi )$, where the computations are easy in the sense that the combinatorics is exactly the same as in a usual noncommutative probability space, and to take advantage of the equivalence between the structures $(\mathcal {A},\tilde \varphi )$ and $(\mathcal {A},\varphi ,\varphi ')$. This method is applied in [@fn] to find the right notion of infinitesimal non-crossing cumulant functional, and to compute the formulas for alternating products of infinitesimally free noncommutative random variables. These formulas make the non-crossing partitions of type B appear, as a reminder of the type B origin of infinitesimal freeness. The present work is in the lineage of [@fn].\
With the motivation to obtain higher order derivatives at $0$ of $\mu _t \boxplus \nu _t$ from those of $\mu _t$ and $\nu _t$, we generalize indeed to higher order the notion of infinitesimal noncommutative probability space from [@fn], by adding to the noncommutative probability space $(\mathcal {A},\varphi ^{(0)})$ a certain number $k$ of other linear functionals $(\varphi ^{(i)})_{1\leq i\leq k}$ on $\mathcal {A}$ satisfying $\varphi ^{(i)}(1_\mathcal {A})=0$. Following the same idea as [@fn], some formulas, the infinitesimal analogue of the free moment-cumulant formula for instance, will be simplified in the equivalent scarce $\mathcal{C}_k$ structure $(\mathcal {A},\tilde \varphi )$, where the $k+1$ linear functionals $(\varphi ^{(i)})_{0\leq i\leq k}$ are consolidated in a unique linear map $\tilde \varphi $, but with values in a certain $(k+1)$-dimensional algebra $\mathcal{C}_k$. The main benefit coming from this trick is that the formulas in $(\mathcal {A},\tilde \varphi )$ are the same as in usual free probability, with the only difference that they take place in the $(k+1)$-dimensional algebra $\mathcal{C}_k$ instead of the field of complex numbers. In what follows, we will continuously switch from the infinitesimal framework $(\mathcal{A} , (\varphi ^{(i)})_{0\leq i\leq k})$ which is the one of interest to the scarce $\mathcal{C}_k$-structure $(\mathcal{A} , \tilde \varphi )$ which is handy because the computations are easier in it.\
As noticed above, in infinitesimal freeness from [@fn], some formulas involving $\varphi '$ also involve the lattices of non-crossing partitions of type B, due to the link of infinitesimal freeness with free probability of type B pointed out in [@fn]. In higher order infinitesimal freeness, new non-crossing partitions appear in the formulas involving $\varphi ^{(k)}$. These so-called non-crossing partitions of type $k$, generalizing both non-crossing partitions of type A (corresponding to the case $k=0$) and type B (corresponding to the case $k=1$), are introduced and studied in Section 6.\
Our approach is in a sense the opposite of the approach in [@bgn]. Indeed, in [@bgn], the authors substitute the symmetric groups by the hyperoctahedral groups, and by the way non-crossing partitions of type A by their type B analogues and thus they obtain the noncommutative probability space of type B. In the present work, we directly substitute the noncommutative probability space by the $k$-th order infinitesimal noncommutative probability space, and we look for the non-crossing partitions of type $k$ appearing this way.\
Following this introduction, the paper is divided in seven other sections. In Section 2, we introduce the two equivalent notions of infinitesimal noncommutative probability space of order $k$ and of scarce $\mathcal{C}_k$-noncommutative probability space and discuss their relations with other structures. In Section 3, we introduce infinitesimal non-crossing cumulant functionals of order $k$, and define infinitesimal freeness of order $k$ by a condition of vanishing mixed cumulants. Section 4 is devoted to the addition and multiplication of infinitesimally free variables. The formula expressing the infinitesimal cumulants of the product of two infinitesimally free noncommutative random variables may be written as a sum on certain non-crossing partitions generalizing the non-crossing partitions of type B reviewed in Section 5. These special non-crossing partitions, called non-crossing partitions of type $k$, and the boxed convolution operation associated to them are introduced and studied in Sections 6 and 7. Finally, we give in Section 8 an important application of higher order infitesimal freeness : a recipe for computing the higher order derivatives of the free convolutions of two distributions.
[**2. Infinitesimal noncommutative probability space of order $k$**]{}
Throughout the paper, the integer $k\in {\mathbb{N}}$ is fixed. In this section, we introduce the two equivalent structures of infinitesimal noncommutative probability space and of scarce $\mathcal{C}_k$ noncommutative probability space and we discuss their relations to previously defined structures.
[**2.1 Infinitesimal noncommutative probability space of order $k$**]{}
The object of this subsection is to introduce the structure which is the framework for our notion of infinitesimal freeness of order $k$, namely the [*infinitesimal noncommutative probability space of order $k$*]{}.
\[incps\] We call infinitesimal noncommutative probability space of order $k$ a structure $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ where $\mathcal{A}$ is a unital algebra over ${\mathbb{C}}$, $\varphi ^{(0)}:\mathcal{A}\longrightarrow {\mathbb{C}}$ is a linear map with $\varphi ^{(0)}(1_\mathcal{A})=1$, and $\varphi ^{(i)}:\mathcal{A}\longrightarrow {\mathbb{C}}$, $1\leq i\leq k$, are linear maps with $\varphi ^{(i)}(1_\mathcal{A})=0$.
The notion of infinitesimal noncommutative probability space of order $1$ coincides with the notion of infinitesimal noncommutative probability space introduced in [@fn]. The structure defined above is therefore a generalization of this latter object, and the use of the adjective infinitesimal is justified.
An element $a\in (\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ of an infinitesimal noncommutative probability space of order $k$ is called an [*infinitesimal noncommutative random variable of order $k$*]{}. The [*infinitesimal distribution of order $k$*]{} of a $n$-tuple $(a_1, \ldots ,a_n)\in \mathcal{A} ^n$ of infinitesimal noncommutative random variables of order $k$ is the $(k+1)$-tuple $(\mu ^{(i)})_{0\leq i\leq k}$ of linear functionals on ${\mathbb{C}}\langle X_1,\ldots ,X_n \rangle$ defined by : $$\mu ^{(i)}(P(X_1, \ldots ,X_n)) : =\varphi ^{(i)}(P(a_1, \ldots ,a_n)).$$ The range of infinitesimal distributions is the set of [*infinitesimal laws of order $k$*]{}, introduced below.
An infinitesimal law (of order $k$) on $n$ variables is a $(k+1)$-tuple of linear functionals $(\mu ^{(i)})_{0\leq i\leq k}$, where $\mu ^{(i)} : {\mathbb{C}}\langle X_1,\ldots ,X_n \rangle \rightarrow {\mathbb{C}}$ is defined on the algebra of noncommutative polynomials and satisfies $\mu ^{(i)}(1)=\delta _i^0$.
For some purposes, it is handy to consider, instead of $k+1$ linear functionals as in Definition \[incps\], an equivalent unique linear map with values in a $(k+1)$-dimensional algebra. The relevant algebra, denoted by $\mathcal{C}_k$, is described below.
[**2.2 The algebra $\mathcal{C}_k$**]{}
In [@fn], the two linear maps $\varphi $ and $\varphi '$ of an infinitesimal noncommutative probability space $(\mathcal{A}, \varphi , \varphi ')$ are consolidated in a single linear map $\tilde \varphi $ on $\mathcal{A}$ with values in the two-dimensional Grassman algebra $\mathbb{G}$ generated by an element $\varepsilon $ which satisfies $\varepsilon ^2 =0$ : $$\mathbb{G} = \{ \alpha + \varepsilon \beta \mid \alpha , \beta \in {\mathbb{C}}\}.$$ This algebra has a quite natural $(k+1)$-dimensional generalization introduced below.
Let $\mathcal{C}_k$ denote the $(k+1)$-dimensional complex algebra ${\mathbb{C}}^{k+1}$ with usual vector space structure and multiplication given by the following rule: if $\alpha = (\alpha ^{(0)},\ldots ,\alpha ^{(k)})\in \mathcal{C}_k$ and $\beta = (\beta ^{(0)},\ldots ,\beta ^{(k)})\in \mathcal{C}_k$, then $$\alpha \cdot \beta =(\gamma ^{(0)},\ldots ,\gamma ^{(k)})$$ is defined by $$\label{prod}
\gamma ^{(i)} : =\sum_{j=0}^i C_i^j \alpha ^{(j)}\beta ^{(i-j)}.$$
The algebra $\mathcal{C}_k$ is a unital complex commutative algebra. Its unit is $1_{\mathcal{C}_k}=(1,0,\ldots ,0)$. An element is invertible in the algebra $\mathcal{C}_k$ if and only if its first coordinate is non-zero.\
The analogy between formula defining the product in $\mathcal{C}_k$ and the well-known Leibniz rule giving the recipe for computing the derivatives of the product of two smooth functions makes it easy to establish the formula for the product $\beta =\alpha _1 \cdots \alpha _n$ of $n$ elements $\alpha _1,\ldots ,\alpha _n\in \mathcal{C}_k$. Precisely, if $\alpha _j= (\alpha _j^{(0)},\ldots ,\alpha _j^{(k)})$ and $\beta = (\beta ^{(0)},\ldots ,\beta ^{(k)})$, one has : $$\beta ^{(i)}=\sum_{\lambda \in \Lambda _{n,i}} C_i^{\lambda _1,\ldots ,\lambda _n}
\prod_{j=1}^{n} \alpha _j^{(\lambda _j)},$$ where $$C_i^{\lambda _1,\ldots ,\lambda _n}=\frac{i!}{\lambda _1!\cdots \lambda _n!}$$ and $$\label{lambdaset}
\Lambda _{n,i}:=\{\lambda =(\lambda _1,\ldots ,\lambda _n)\in {\mathbb{N}}^{n}\mid \sum_{j=1}^{n}\lambda _j=i\}.$$\
There is an alternative description of the algebra $\mathcal{C}_k$ : it may be identified with the algebra of $(k+1)$-by-$(k+1)$ upper triangular Toeplitz matrices (with usual matricial operations) as follows : $$(\alpha^{(0)},\ldots ,\alpha ^{(k)})\simeq \left(\begin{array}{ccccc}\alpha ^{(0)}&\alpha ^{(1)}&\ldots &\frac{\alpha ^{(k-1)}}{(k-1)!}&\frac{\alpha ^{(k)}}{k!}\\
0&\alpha ^{(0)}&\ldots &\ldots &\frac{\alpha ^{(k-1)}}{(k-1)!}\\
\ldots &\ldots &\ldots &\ldots &\ldots \\
\ldots &\ldots &\ldots &\alpha ^{(0)}&\alpha ^{(1)}\\
0&0&\ldots &\ldots &\alpha ^{(0)}\\
\end{array} \right).$$ Consider $$\varepsilon :=\left(\begin{array}{ccccc}0&1&\ldots &0&0\\
0&0&\ldots &\ldots &0\\
\ldots &\ldots &\ldots &\ldots &\ldots \\
\ldots &\ldots &\ldots &0&1\\
0&0&\ldots &\ldots &0\\
\end{array} \right).$$ It is easy to compute the values of $\varepsilon ^i$ for $0\leq i\leq k+1$ ; in particular $\varepsilon ^{k+1}=0_{\mathcal{C}_k}$ and any element $\alpha = (\alpha ^{(0)},\ldots ,\alpha ^{(k)})\in \mathcal{C}_k$ may be uniquely decomposed $$\alpha = \sum_{i=0}^{k} \alpha ^{(i)} \frac{\epsilon ^i}{i!}.$$ The family $(\frac{\varepsilon ^i}{i!} , 0\leq i\leq k)$ is thus a linear basis of $\mathcal{C}_k$, to which we will refer as the canonical basis of $\mathcal{C}_k$. In particular, $\mathcal{C}_k\simeq {\mathbb{C}}[\varepsilon ]={\mathbb{C}}_k[\varepsilon ]\simeq {\mathbb{C}}[X]/(X^{k+1})$.\
In the definition of a usual noncommutative probability space, if one asks for the state to be $\mathcal{C}_k$-valued, one obtains a slightly different structure, introduced in the next section.
[**2.3 Scarce $\mathcal{C}_k$-noncommutative probability space**]{}
By [*scarce $\mathcal{C}_k$-noncommutative probability space*]{}, we mean a couple $( \mathcal{A} , \tilde \varphi )$, where $\mathcal{A}$ is a unital algebra over ${\mathbb{C}}$ and $\tilde \varphi : \mathcal{A} \to \mathcal{C}_k$ is a linear map satisfying $\tilde \varphi (1_{\mathcal{A}}) = 1_{\mathcal{C}_k}$.
The notion of scarce noncommutative probability space was introduced in [@oan], but only the particular case of scarce $\mathbb{G}$-noncommutative probability space was considered there. This same structure has been studied later in [@fn] in connection with infinitesimal noncommutative probability space and free probability of type B.
\[association\] To any infinitesimal noncommutative probability space of order $k$ $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$, we may associate a natural scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$, by putting $$\label{decomp}
\tilde \varphi :=\sum_{i=0}^k \varphi ^{(i)} \frac{\varepsilon ^i}{i!}$$ Reciprocally, given a scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$, the linear decomposition of $\tilde \varphi $ in the canonical basis of $\mathcal{C}_k$ (see equation ) gives rise to $k+1$ linear functionals $(\varphi ^{(i)})_{0\leq i\leq k}$, and consequently to an infinitesimal noncommutative probability space of order $k$ : $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$.\
The equivalence between the infinitesimal noncommutative probability space of order $k$ $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ and its associated scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$ is fundamental in what follows. Indeed, we will continuously switch from one structure to the other, according to the principle that our interest is in the infinitesimal structure whereas the computations are easier in the scarce $\mathcal{C}_k$ structure, in the sense that they mimetize those from usual free probability.
An element $a$ of a scarce $\mathcal{C}_k$-noncommutative probability space $(\mathcal{A}, \tilde \varphi )$ is called a [*$\mathcal{C}_k$-noncommutative random variable*]{}. We associate to such an $a\in \mathcal{A}$ the sequence of its [*$\mathcal{C}_k$-valued moments*]{} $(\tilde \varphi (a^n))_{n\in {\mathbb{N}}^\ast }$. We call [*$\mathcal{C}_k$-valued distribution*]{} of $a$ the whole sequence of its moments, or equivalently, the linear map from ${\mathbb{C}}[X]$ into $\mathcal{C}_k$ which maps any polynomial $P$ to $\tilde \varphi (P(a))$. One may find easier to collect all the $\mathcal{C}_k$-valued moments in a formal power series, as follows :
Let $\mathcal{C}$ be a unital commutative algebra over ${\mathbb{C}}$. We denote by $\Theta _{\mathcal{C}}^{(A)}$ the set of power series of the form $$f(z)=\sum_{n=1}^\infty \alpha _nz^n,$$ where the $\alpha _n$’s are elements of $\mathcal{C}$.
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. The [*$\mathcal{C}_k$-valued moment series*]{} of $a\in \mathcal{A} $ is the power series $\tilde M_a\in \Theta _{\mathcal{C}_k}^{(A)}$ defined as follows: $$\tilde M_a(z):=\sum_{n=1}^{\infty } \tilde \varphi (a^n) z^n.$$
The notion of $\mathcal{C}_k$-valued distribution is easily generalized to $n$-tuples of variables :
The [*$\mathcal{C}_k$-valued distribution*]{} of a $n$-tuple $(a_1,\ldots ,a_n)\in \mathcal{A} ^n$ of $\mathcal{C}_k$-noncommutative random variables in a scarce $\mathcal{C}_k$-noncommutative probability space $(\mathcal{A}, \tilde \varphi )$ is the linear map $\tilde \mu _{(a_1,\ldots ,a_n)} : {\mathbb{C}}\langle X_1,\ldots ,X_n \rangle \rightarrow \mathcal{C}_k$ defined by $$\tilde \mu _{(a_1,\ldots ,a_n)}(P(X_1,\ldots ,X_n)):=\tilde \varphi (P(a_1,\ldots ,a_n)).$$
As mentioned in [@fn], scarce $\mathbb{G}$-noncommutative probability space and infinitesimal noncommutative probability space provide a nice framework to do free probability of type B. The equivalent structures defined above are therefore the natural setting for generalizing free probability of type B. There is another structure linked to free probability of type B that one may find interesting to generalize here : the noncommutative probability space of type B, proposed in the original work on free probability of type B [@bgn]. Its natural generalization is the [*noncommutative probability space of type $k$*]{} :
By a noncommutative probability space of type $k$ we understand a system $(\mathcal{V}^{(0)},f^{(0)},\ldots ,\mathcal{V}^{(k)},f^{(k)},(\Phi _{i,j})_{0\leq i,j\leq k})$, where $(\mathcal{V}^{(0)},f^{(0)})$ is a noncommutative probability space of type A, $\mathcal{V}^{(i)}$, $1\leq i\leq k$, are complex vector spaces, $f^{(i)}:\mathcal{V}^{(i)}\longrightarrow {\mathbb{C}}$, $1\leq i\leq k$, are linear maps, $\Phi _{i,j}:\mathcal{V}^{(i)}\times \mathcal{V}^{(j)}\longrightarrow \mathcal{V}^{(i+j)}$,$0\leq i,j\leq k$, are bilinear maps satisfying $$\Phi _{h+i,j}(\Phi _{h,i}(x,y),z)=\Phi _{h,i+j}(x,\Phi _{i,j}(y,z)),$$ $\forall h,i,j\in {\mathbb{N}}, \forall x\in \mathcal{V}^{(h)}, \forall y\in \mathcal{V}^{(i)}, \forall z\in \mathcal{V}^{(j)}$.
To make the preceding definition work, we put $\mathcal{V}^{(i)}=\{0\}$, when $i\geq k+1$. The following fact noticed in [@fn] still holds : noncommutative probability spaces of type $k$ are particular cases of scarce $\mathcal{C}_k$-noncommutative probability spaces. Indeed, given a noncommutative probability space of type $k$ $(\mathcal{V}^{(0)},f^{(0)},\ldots ,\mathcal{V}^{(k)},f^{(k)},(\Phi _{i,j})_{0\leq i,j\leq k})$, the direct product $\prod_{i=0}^{k} \mathcal{V}^{(i)}$ can be endowed with a complex unital algebra structure, via the maps $(\Phi _{i,j})_{0\leq i,j\leq k}$. This algebra, together with the linear map $\tilde \varphi (x_0,\ldots ,x_{k}):=(f^{(0)}(x_0),\ldots ,f^{(k)}(x_{k}))$, forms a scarce $\mathcal{C}_k$-noncommutative probability space.\
There are natural equivalent notions of freeness on the structures introduced above, generalizing both infinitesimal freeness from [@bs] and [@fn] and freeness of type B from [@bgn]. In [@fn], infinitesimal freeness in $(\mathcal{A},\varphi ,\varphi ')$ is defined by two conditions on the linear functionals $\varphi ,\varphi '$ ; its generalization to an infinitesimal noncommutative probability space of order $k$ denoted by $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ would require $k+1$ conditions on the linear functionals $(\varphi ^{(i)})_{0\leq i\leq k}$. Infinitesimal freeness from [@fn] being also equivalent to the vanishing of the infinitesimal non-crossing cumulants, we adopt this approach and define the infinitesimal freeness of order $k$ by the vanishing of some multilinear functionals, called infinitesimal non-crossing cumulant functionals of order $k$ and introduced in the next section.
[**3. Infinitesimal non-crossing cumulants of order $k$**]{}
We begin this section by reviewing some background on non-crossing partitions.
[**3.1 Miscellaneous facts on non-crossing partitions of type A**]{}
A partition $p$ of a finite set $X$ is a family of disjoint non-empty subsets of $X$, called the blocks of $p$, whose reunion is $X$. The set of blocks of a partition $p$ of $X$ will be denoted throughout these notes by $\mbox{bl}(p)$ ; its cardinal by $| p |$.\
For $a$ and $b$ in $X$, we write $a\sim _p b$ and say that $a$ and $b$ are linked (in $p$) to denote that $a$ and $b$ are in the same block of the partition $p$ of $X$. The set of partitions of a finite set $X$ together with the reverse refinement order ($p\preceq q$ if every block of $p$ is contained in a block of $q$) is a lattice.\
Now suppose $(X,\leq )$ is a totally ordered set.\
A partition $p$ of $X$ is called non-crossing if, whenever you have $a<b<c<d$ in $X$ such that $a\sim _p c$ and $b\sim _p d$, then $a\sim _p b$.\
The set $(NC^{(A)}(X),\preceq )$ of non-crossing partitions of $X$ together with the reverse refinement order is itself a lattice. Its maximal element $1_X$ has $X$ as its only block ; its minimal element $0_X$ has every singleton as a block.\
When $X=[m]:=\{1<\ldots <m\}$, we write $NC^{(A)}(m)$ instead of $NC^{(A)}([m])$.\
A nice way to represent a non-crossing partition $p\in NC^{(A)}(m)$ is to view $1,\ldots ,m$ as equidistant clockwisely ordered points on a circle, and to draw for each block of $p$ the convex polygone whose vertices are the elements of this block. It is a necessary and sufficient condition for a partition to be non-crossing that the polygones built this way do not intersect.\
Biane found in [@bia] a bijection between the set of non-crossing partitions of $[m]$ and the set of points lying on a geodesic in the Cayley graph of the symmetric group $S_m$ with generators the set of all transpositions. This bijection $t$ associates to any non-crossing partition $p\in NC^{(A)}(m)$ the permutation $t(p)\in S_m$ whose restriction to each block $V$ of $p$ is the trace of the cycle $(1,\ldots ,m)\in S_m$ on $V$. For $a\in [ m ] $, $t(p)(a)$ is called the neighbour of $a$ in $p$. Geometrically, it is the first point linked to $a$ that one meets when one goes clockwisely around the circle, starting from $a$.\
Let us recall that the Kreweras complementation map, denoted by $\mbox{Kr}$, is the anti-isomorphism of the lattice $NC^{(A)}(m)$ of non-crossing partitions of $[m]$ introduced by Kreweras in [@kre] and defined in the following way : consider a copy $$\overline {[m]}:=\{\overline{1}<\ldots <\overline{m}\}$$ of $[m]$ and order the elements of $[m]\cup \overline{[m]}$ as follows : $$\{1<\overline{1}<\ldots <m<\overline{m}\}.$$ Given $p$ a non-crossing partition of $[m]$, $\mbox{Kr}(p)$ is the biggest (for the reverse refinement order) partition of $\overline{[m]}$ such that $p \cup \mbox{Kr}(p)$ is a non-crossing partition of $[m]\cup \overline{[m]}$. See [@ns97] for a nice geometric construction of the Kreweras complement.
On $[m]\cup \overline{[m]}$, we could have considered the alternative order $$\{\overline 1<1<\overline 2<\ldots <\overline m<m\}.$$ This would have led to another anti-isomorphism of $NC^{(A)}(m)$, also called Kreweras complementation map and denoted $\mbox{Kr}'$, which turns out to be the inverse of $\mbox{Kr}$.
There is an important equality (see [@kre]) verified by the number of blocks of the Kreweras complement of a non-crossing partition:
$$\label{nbblocks}
| p |+| \mbox{Kr}(p) |=m+1, \forall p\in NC^{(A)}(m).$$
Notice that, for $p\in NC^{(A)}(m)$, $\mbox{Kr}^2(p)$ can be easily described in the geometric representation given above : $\mbox{Kr}^2(p)$ is the anti-clockwise rotation of $p$ with angle $\frac{2\pi}{m}$.
We conclude this subsection by the introduction of a total order on the blocks of a fixed non-crossing partition $p$ of $[m]$.
Let $p \in NC^{(A)}(m)$, and $V,W\in \mbox{bl}(p)$.\
$1^o$ $V$ is said to be nested in $W$ if $\min W<\min V\leq \max V<\max W$.\
$2^o$ $V$ is said to be on the left of $W$ if $\max V<\min W$.\
$3^o$ $V\sqsubset W \Leftrightarrow V$ is nested in $W$ or $V$ is on the left of $W$.
The proof of the next proposition is trivial and left to the reader.
$\sqsubset $ is a total order on $\mbox{bl}(p)$.
If $p\in NC^{(A)}(m)$, we have seen that $p\cup \mbox{Kr}(p)$ is a non-crossing partition of $[m]\cup \overline{[m]}$ in $m+1$ blocks. These blocks will be listed in two different ways.\
The first way is to list them all together in the increasing order $\sqsubset $ : we will write $\mbox{Mix}(p,i)$ for the $i$-th block of $p\cup \mbox{Kr}(p)$ in the increasing order $\sqsubset $, for $1\leq i\leq m+1$.\
For some purposes, it is nice to list separately the blocks of $p$ and of $\mbox{Kr}(p)$, and we will write $\mbox{Sep}(p,i)$ to denote the $i$-th block of $p$ in the increasing order $\sqsubset $ if $1\leq i\leq | p |$ and to denote the $(i-| p |)$-th block of $\mbox{Kr}(p)$ in the increasing order $\sqsubset $ if $| p |+1\leq i\leq m+1$.\
It is interesting to look at the first blocks in the two resulting lists : $\mbox{Mix}(p,1)$ is a singleton in $[m]\cup \overline{[m]}$, $\mbox{Sep}(p,1)$ is an interval in $[m]$. In particular, we can deduce the well-known fact that a non-crossing partition always owns an interval-block.
[**3.2 $\mathcal{C}_k$-non-crossing cumulant functionals**]{}
In this subsection, we define non-crossing cumulant functionals in the framework of a scarce $\mathcal{C}_k$-noncommutative probability space by the free moment-cumulant formula from usual free probability, with the only difference that the computations take place in the algebra $\mathcal{C}_k$ instead of the field of complex numbers ${\mathbb{C}}$. The following notations are commonly used in the combinatorial theory of free probability.
Let $(a_1,\ldots ,a_n)\in \mathcal{A}^n$, and let $V=\{v_1<\ldots <v_m\}\subseteq [n]$, then we denote $$(a_1,\ldots ,a_n) \mid V:=(a_{v_1},\ldots ,a_{v_m})\in \mathcal{A}^m.$$ For a family of multilinear maps $(r_n: \mathcal{A}^n \rightarrow \mathcal{C}_k)_{n=1}^\infty $, we define for any $n\in {\mathbb{N}}$ and any $\pi \in NC^{(A)}(n)$ the $n$-linear functional $r_\pi : \mathcal{A}^n \rightarrow \mathcal{C}_k$ by $$r_\pi (a_1,\ldots ,a_n):=\prod _{V\in \pi } r_{|V|}((a_1,\ldots ,a_n) \mid V).$$
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. The [*$\mathcal{C}_k$-non-crossing cumulant functionals*]{} are a family of multilinear maps $(\tilde \kappa _n: \mathcal{A}^n \rightarrow \mathcal{C}_k)_{n=1}^\infty $, uniquely determined by the following equation : for every $n\geq 1$ and every $a_1,\ldots ,a_n\in \mathcal{A}$, $$\label{fmc}
\sum_{p\in NC^{(A)}(n)}\tilde \kappa _p(a_1,\ldots ,a_n)=\tilde \varphi (a_1\cdots a_n).$$
In free probability of type A, the formula above is known as the free moment-cumulant formula [@hp]. The only difference is that computations here take place in the unital commutative complex algebra $\mathcal{C}_k$ instead of ${\mathbb{C}}$. However, the proofs (see [@nsbook]) of the following classical results remain valid in this setting. That is why we record them without proof.\
For every $n\geq 1$ and every $a_1,\ldots ,a_n\in \mathcal{A}$ we have that: $$\label{invfmc}
\tilde \kappa _n(a_1,\ldots ,a_n)=\sum_{p\in NC^{(A)}(n)} \mbox{M\"ob}(p,1_n) \tilde \varphi _p(a_1,\ldots ,a_n),$$ where is the Möbius function of the lattice of non-crossing partitions. Obviously, the multilinear maps $(\tilde \varphi _n: \mathcal{A}^n \rightarrow \mathcal{C}_k)_{n=1}^\infty $ implicitely used in formula are defined by $\tilde \varphi _n(a_1,\ldots ,a_n)=\tilde \varphi (a_1\cdots a_n)$.
\[cwsae\] One has that $\tilde \kappa _n (a_1, \ldots , a_n) = 0$ whenever $n \geq 2$, $a_1, \ldots , a_n \in \mathcal{A}$, and there exists $1 \leq i \leq n$ such that $a_i \in {\mathbb{C}}1_{\mathcal{A}}$.
\[cwpae\] Let $x_1, \ldots , x_s$ be in $\mathcal{A}$ and consider some products of the form $$a_1 = x_1 \cdots x_{s_1}, \ a_2 = x_{s_1 +1} \cdots x_{s_2},
\ \ldots , \ a_n = x_{s_{n-1} +1} \cdots x_{s_n},$$ where $1 \leq s_1 < s_2 < \cdots < s_n = s$. Then $$\tilde \kappa _n (a_1, \ldots , a_n) =
\sum_{ \begin{array}{c}
{\scriptstyle \pi \in NC(s) \ such} \\
{\scriptstyle that \ \pi \vee \theta = 1_s}
\end{array} } \ \ \tilde \kappa _{\pi} (x_1, \ldots , x_s),$$ where $\theta \in NC(s)$ is the partition : $$\theta = \{ \{ 1, \ldots , s_1 \}, \{ s_1 + 1, \ldots , s_2 \} , \ldots ,
\{ s_{n-1} + 1, \ldots , s_n \} \}.$$
Given a $\mathcal{C}_k$-noncommutative random variable $a\in ( \mathcal{A} , \tilde \varphi )$, the quantities $\tilde \kappa _n(a,\ldots ,a)$ are called its [*$\mathcal{C}_k$-valued cumulants*]{}, and they are collected in a power series :
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. The [*$\mathcal{C}_k$-valued R-transform*]{} of $a\in \mathcal{A} $ is the power series $\tilde R_a\in \Theta _{\mathcal{C}_k}^{(A)}$ defined as follows : $$\tilde R_a(z):=\sum_{n=1}^{\infty } \tilde \kappa _n(a,\ldots ,a) z^n.$$
Following the well-known result of [@spe] stating roughly that, in a usual noncommutative probability space, subsets are free if and only if they satisfy the vanishing of mixed cumulants condition, we generalize this condition to our setting :
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space and $\mathcal{M}_1,\ldots ,\mathcal{M}_n$ be subsets of $\mathcal{A} $. We say that $\mathcal{M}_1,\ldots ,\mathcal{M}_n$ have [*vanishing mixed $\mathcal{C}_k$-cumulants*]{} if $$\tilde \kappa _m(a_1,\ldots ,a_m)=0$$ whenever $a_1\in \mathcal{M}_{i_1},\ldots ,a_m\in \mathcal{M}_{i_m}$ and $\exists 1\leq s<t\leq m$, such that, $i_s\not=i_t$.
As announced, infinitesimal freeness of order $k$ is defined by the vanishing of mixed $\mathcal{C}_k$-cumulants condition. More precisely :
\[f\] We will say that subsets $\mathcal{M}_1,\ldots ,\mathcal{M}_n\subseteq \mathcal{A} $ of a scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$ are [*infinitesimally free of order $k$*]{} if they have vanishing mixed $\mathcal{C}_k$-cumulants.
Using a classical argument in free probability, one can prove that, if $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ are unital subalgebras which are infinitesimally free of order $k$ in a scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$, then one has : $$\tilde \varphi (a_1 \cdots a_m)=0$$ whenever $a_1\in \mathcal{A}_{i_1},\ldots ,a_m\in \mathcal{A}_{i_m}$ with $i_1\not=\ldots \not=i_m$ satisfy $\tilde \varphi (a_1)=\ldots =\tilde \varphi (a_m)=0$.\
The converse in our $\mathcal{C}_k$-valued situation is not true, because one cannot use the nice “centering trick”, as noticed in $\cite{fn}$ Remark 4.9.
In the next subsection, we switch to the infinitesimal framework, and define infinitesimal non-crossing cumulant functionals, with the intuition that they should appear as the coefficients in the decomposition of the $\mathcal{C}_k$-non-crossing cumulant functionals in the canonical basis of $\mathcal{C}_k$.
[**3.3 Infinitesimal non-crossing cumulant functionals**]{}
In this short subsection, we focus on an infinitesimal noncommutative probability space of order $k$ structure $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$. The aim is to define cumulants and freeness in this setting, in a consistent way with the last subsection. For convenience, we will use the following notation :
For a family of multilinear maps $(r_n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $, we define for any $n\in {\mathbb{N}}$, any $\pi = \{V_1\sqsubset \cdots \sqsubset V_h\} \in NC^{(A)}(n)$ and any $\lambda \in \Lambda _{n,h}$ (defined by ) the $n$-linear functional $r_\pi ^{(\lambda )} : \mathcal{A}^n \rightarrow {\mathbb{C}}$ by $$r_\pi ^{(\lambda )}(a_1,\ldots ,a_n):=\prod _{i=1}^h r_{|V_i|}^{(\lambda _i)}((a_1,\ldots ,a_n) \mid V_i).$$
The underlying idea is to consider the $\mathcal{C}_k$-non-crossing cumulant functionals $(\tilde \kappa _n: \mathcal{A}^n \rightarrow \mathcal{C}_k)_{n=1}^\infty $ in the associated scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$ (see Remark \[association\]), and then to define the required $n$-th infinitesimal non-crossing cumulant functionals as the $n$-linear forms appearing as coefficients in the linear decomposition of $\tilde \kappa _n: \mathcal{A}^n \rightarrow \mathcal{C}_k$ in the canonical basis of $\mathcal{C}_k$. This leads to the following definition:
Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$. The [*infinitesimal non-crossing cumulant functionals of order $k$*]{} are a family of multilinear maps $(\kappa _n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $, uniquely determined by the following equation : for every $n\geq 1$, every $0\leq i\leq k$ and every $a_1,\ldots ,a_n\in \mathcal{A}$ we have that: $$\label{inffmc}
\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }} \sum_{\lambda \in \Lambda _{h,i}}
C_i^{\lambda _1,\ldots ,\lambda _h} \kappa _{p}^{(\lambda )}(a_1,\ldots ,a_n)=
\varphi ^{(i)}(a_1\cdots a_n).$$
Infinitesimal freeness in the framework of an infinitesimal noncommutative probability space of order $k$ is obviously defined by the vanishing of mixed infinitesimal cumulants.
\[inff\] We will say that subsets $\mathcal{M}_1,\ldots ,\mathcal{M}_n$ of an infinitesimal noncommutative probability space of order $k$ are [*infinitesimally free of order $k$*]{} if they have [*vanishing mixed infinitesimal cumulants*]{}, which means that, for each $0\leq i\leq k$, $$\kappa _m^{(i)}(a_1,\ldots ,a_m)=0$$ whenever $a_1\in \mathcal{M}_{i_1},\ldots ,a_m\in \mathcal{M}_{i_m}$ and $\exists 1\leq s<t\leq m$, such that, $i_s\not=i_t$.
It is straightforward to check, using formula , that the infinitesimal non-crossing cumulant functionals of an infinitesimal noncommutative probability space of order $k$ are indeed linked to the $\mathcal{C}_k$-non-crossing cumulant functionals of the associated scarce $\mathcal{C}_k$-noncommutative probability space by : $$\label{decompcum}
\tilde \kappa _n=\sum_{i=0}^k \kappa _n^{(i)} \frac{\varepsilon ^i}{i!}.$$ A consequence of formulas and is the validity of the following inverse formula: $$\label{infinvfmc}
\kappa _n^{(i)}(a_1,\ldots ,a_n)=
\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }}\sum_{\lambda \in \Lambda _{h,i}}\mbox{M\"ob}(p,1_n)
C_i^{\lambda _1,\ldots ,\lambda _h} \varphi _{p}^{(\lambda )}(a_1,\ldots ,a_n) ,$$ and of the following proposition :
One has that $\kappa _n^{(i)} (a_1, \ldots , a_n) = 0$ whenever $0\leq i\leq k$, $n \geq 2$, $a_1, \ldots , a_n \in \mathcal{A}$, and there exists $1 \leq j \leq n$ such that $a_j \in {\mathbb{C}}1_{\mathcal{A}}$.
Another consequence of relation is that subsets $\mathcal{M}_1,\ldots ,\mathcal{M}_n$ of an infinitesimal noncommutative probability space of order $k$ are infinitesimally free of order $k$ if and only if they are infinitesimally free of order $k$ in the associated scarce $\mathcal{C}_k$-noncommutative probability space.
\[inffimpliesf\] Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$, and consider its infinitesimal non-crossing cumulant functionals $(\kappa _n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $. It is interesting to notice that the multilinear maps $(\kappa _n^{(0)}: \mathcal{A}^n \rightarrow {\mathbb{C}})_{n=1}^\infty $ and $(\kappa _n^{(1)}: \mathcal{A}^n \rightarrow {\mathbb{C}})_{n=1}^\infty $ are respectively the usual non-crossing cumulant functionals in the noncommutative probability space $(\mathcal{A}, \varphi ^{(0)})$ and the infinitesimal non-crossing cumulant functionals of [@fn] in the infinitesimal noncommutative probability space $(\mathcal{A}, \varphi ^{(0)}, \varphi ^{(1)})$. This implies that subsets that are infinitesimally free of order $k$ are in particular free in $(\mathcal{A}, \varphi ^{(0)})$ and infinitesimally free in $(\mathcal{A}, \varphi ^{(0)}, \varphi ^{(1)})$ in the sense of [@fn].\
Infinitesimal freeness of unital subalgebras in [@fn], as well as freeness of type B in [@bgn], is defined in terms of moments. Section 8 will provide such a characterization of the infinitesimal freeness of order $k$ of unital subalgebras of an infinitesimal noncommutative probability space of order $k$ in terms of moments.
As stated in Remark \[inffimpliesf\], infinitesimal freeness of order $k$ of unital subalgebras $\mathcal{A}_1,\ldots ,\mathcal{A}_n\subseteq (\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ of an infinitesimal noncommutative probability space of order $k$ implies freeness of $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ in the noncommutative probability space $(\mathcal{A}, \varphi ^{(0)})$. Conversely, is it possible to “upgrade” freeness of given unital subalgebras of a noncommutative probability space to infinitesimal freeness of order $k$ ? This question is discussed in the next subsection.
[**3.4 Upgrading freeness to infinitesimal freeness of order $k$**]{}
Given a noncommutative probability space $(\mathcal{A},\varphi )$ and free unital subalgebras $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ of $\mathcal{A}$, the question of how to build a linear form $\varphi '$ on $\mathcal{A}$ such that $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ are infinitesimally free in the infinitesimal noncommutative probability space $(\mathcal{A},\varphi ,\varphi ')$ is adressed in [@fn]. Among the answers given there, there is the idea to define $\varphi ':=\varphi \circ D$, where $D$ is a derivation of $\mathcal{A}$ (a linear map $D : \mathcal{A}\longrightarrow \mathcal{A}$ satisfying $\forall a,b\in \mathcal{A}, D(a\cdot b)=D(a)\cdot b+a\cdot D(b)$) such that $D(\mathcal{A}_j)\subseteq \mathcal{A}_j$ for each $1\leq j\leq n$. We examine the question of how to build linear forms $\varphi ^{(1)},\ldots ,\varphi ^{(k)}$ on $\mathcal{A}$ such that $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ are infinitesimally free of order $k$ in the infinitesimal noncommutative probability space $(\mathcal{A},\varphi ,\varphi ^{(1)},\ldots ,\varphi ^{(k)})$. The natural idea consisting in defining $\varphi ^{(i)}:=\varphi \circ D^{i}$ where $D$ is a derivation of $\mathcal{A}$ such that $D(\mathcal{A}_j)\subseteq \mathcal{A}_j$ for each $1\leq j\leq n$ is a possible answer, as proved below :
\[upgrade\] Let $(\mathcal{A},\varphi )$ be a noncommutative probability space and let $D : \mathcal{A}\rightarrow \mathcal{A}$ be a derivation. Define $\varphi ^{(i)}:=\varphi \circ D^{i}$. Let the infinitesimal non-crossing cumulant functionals associated to $(\mathcal{A},\varphi ,\varphi ^{(1)},\ldots ,\varphi ^{(k)})$ be denoted by $(\kappa _n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $. Then, for every $n\geq 1, 0\leq i\leq k$ and $a_1,\ldots ,a_n\in \mathcal{A}$ one has $$\kappa _n^{(i)}(a_1,\ldots ,a_n)=\sum _{\lambda \in \Lambda _{n,i}} C_i^{\lambda _1,\ldots ,\lambda _n}
\kappa _n(D^{\lambda _1}(a_1),\ldots ,D^{\lambda _n}(a_n)).$$
Define the family of multilinear functionals $(\eta _n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $ by the following formulas : for every $n\geq 1, 0\leq i\leq k$ and $b_1,\ldots ,b_n\in \mathcal{A}$ $$\eta _n^{(i)}(b_1,\ldots ,b_n)=\sum _{\lambda \in \Lambda _{n,i}} C_i^{\lambda _1,\ldots ,\lambda _n}
\kappa _n(D^{\lambda _1}(b_1),\ldots ,D^{\lambda _n}(b_n)).$$ Our aim is then to prove that, for every $n\geq 1, 0\leq i\leq k$, $\eta _n^{(i)}=\kappa _n^{(i)}$. We verify that the functionals $(\eta _n^{(i)}, 0\leq i\leq k)_{n=1}^\infty $ satisfy the equations defining the infinitesimal non-crossing cumulant functionals. The left-hand side of this formula writes : $$\label{sum}
\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }}\sum_{\lambda \in \Lambda _{h,i}}
C_i^{\lambda _1,\ldots ,\lambda _h}\eta _{p}^{(\lambda )}(a_1,\ldots ,a_n).$$ Each $\eta _{|V_j|}^{(\lambda _j)}((a_1,\ldots ,a_n) \mid V_j)$ in the latter is a sum indexed by $\Lambda _{|V_j|,\lambda _j}$, involving variables $a_i, i\in V_j$. Given $p:=\{V_1,\ldots ,V_h \}\in NC^{(A)}(n)$, there is a very natural bijection between $\{(\lambda ,(\lambda ^1,\ldots ,\lambda ^h))\in \Lambda _{h,i}\times \Lambda _{n,i}\mid \lambda ^j\in \Lambda _{|V_j|,\lambda _j}\}$ and the set $\Lambda _{n,i}$. Thus, the quantity rewrites : $$\sum_{p\in NC^{(A)}(n)}\sum_{\lambda \in \Lambda _{n,i}}
C_i^{\lambda _1,\ldots ,\lambda _n}\kappa _{p}(D^{\lambda _1}(a_1),\ldots ,D^{\lambda _n}(a_n)).$$ By exchanging the summation signs, the usual free moment-cumulant formula appears, and one obtains : $$\label{leib}
\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }}\sum_{\lambda \in \Lambda _{h,i}}
C_i^{\lambda _1,\ldots ,\lambda _h} \eta _{p}^{(\lambda )}(a_1,\ldots ,a_n)=
\sum_{\lambda \in \Lambda _{n,i}} C_i^{\lambda _1,\ldots ,\lambda _n}
\varphi (D^{\lambda _1}(a_1)\cdots D^{\lambda _n}(a_n)).$$ Using Leibniz rule in the right-hand side of , one may conclude : $$\begin{aligned}
\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }}\sum_{\lambda \in \Lambda _{h,i}}
C_i^{\lambda _1,\ldots ,\lambda _h} \eta _{p}^{(\lambda )}(a_1,\ldots ,a_n)\\
=\varphi (\sum_{\lambda \in \Lambda _{n,i}} C_i^{\lambda _1,\ldots ,\lambda _n}
D^{\lambda _1}(a_1)\cdots D^{\lambda _n}(a_n))\\
=\varphi (D^i(a_1\cdots a_n))\\
=\varphi ^{(i)}(a_1\cdots a_n).\end{aligned}$$
In the notations of Proposition \[upgrade\], let $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ be unital subalgebras of $\mathcal{A}$ which are free in $(\mathcal{A}, \varphi )$, and such that $D(\mathcal{A}_j)\subseteq \mathcal{A}_j$ for $1\leq j\leq n$. Then $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ are infinitesimally free of order $k$ in $(\mathcal{A},\varphi ,\varphi ^{(1)},\ldots ,\varphi ^{(k)})$.
[**4. Addition and multiplication of infinitesimally free random variables**]{}
In this section, we consider $n$-tuples of infinitesimal noncommutative random variables $(a_1,\ldots ,a_n),(b_1,\ldots ,b_n)\in \mathcal{A}^n$ (where $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ is an infinitesimal noncommutative probability space of order $k$), with respective infinitesimal distributions $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$. We assume that the sets $\{a_1,\ldots ,a_n\}$ and $\{b_1,\ldots ,b_n\}$ are infinitesimally free of order $k$ and we are interested in the distributions of the sum $(a_1,\ldots ,a_n)+(b_1,\ldots ,b_n)$ and of the product $(a_1b_1,\ldots ,a_nb_n)$.
[**4.1. Addition of infinitesimally free random variables**]{}
We do not provide a proof of the following result, which is a straightforward calculation using multilinearity of the infinitesimal cumulant functionals and definition of infinitesimal freeness.
\[propinfadd\] Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$. Consider subsets $\mathcal{M}_1,\mathcal{M}_2$ of $\mathcal{A}$ that are infinitesimally free of order $k$. Then, one has, for each $n\geq 1$, each $n$-tuples $(a_1,\ldots ,a_n)\in \mathcal{M}_1^n,(b_1,\ldots ,b_n)\in \mathcal{M}_2^n$ and each $0\leq i\leq k$ : $$\label{infadd}
\kappa _n^{(i)}(a_1+b_1,\ldots ,a_n+b_n)=\kappa _n^{(i)}(a_1,\ldots ,a_n)+\kappa _n^{(i)}(b_1,\ldots ,b_n).$$
Using formulas and , the quantities $\kappa _m^{(i)}(a_{i_1},\ldots ,a_{i_m})$, $\kappa _m^{(i)}(b_{j_1},\ldots ,b_{j_m})$ for each $0\leq i\leq k$, each $m\geq 1$ and each subsets $\{i_1,\ldots ,i_m\} , \{j_1,\ldots ,j_m\}\subseteq [n]$ called respectively infinitesimal cumulants of $(a_1,\ldots ,a_n)$ and $(b_1,\ldots ,b_n)$ completely determine and are completely determined by the infinitesimal distributions of $(a_1,\ldots ,a_n)$ and $(b_1,\ldots ,b_n)$. Proposition \[propinfadd\] thus has the following corollary.
Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$, and $(a_1,\ldots ,a_n),(b_1,\ldots ,b_n)\in \mathcal{A}^n$ with respective infinitesimal distributions $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$. If the sets $\{a_1,\ldots ,a_n\}$ and $\{b_1,\ldots ,b_n\}$ are infinitesimally free of order $k$, then the infinitesimal distribution of $(a_1,\ldots ,a_n)+(b_1,\ldots ,b_n)$ only depends on $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$. It is called the infinitesimal free additive convolution of order $k$ of $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$ and denoted by $(\mu ^{(i)})_{0\leq i\leq k}\boxplus ^{(k)} (\nu ^{(i)})_{0\leq i\leq k}$.
The corollary above means that the infinitesimal free additive convolution of order $k$ defines an operation on infinitesimal laws. The practical way to compute the infinitesimal free additive convolution of order $k$ of two infinitesimal laws is to use consecutively the inverse of the infinitesimal version of the free moment-cumulant formula (formula ), the additivity of infinitesimal cumulants (formula ), and finally the infinitesimal version of the free moment-cumulant formula (formula ). One may find easier to make the computations in a scarce $\mathcal{C}_k$-noncommutative probability space.\
Taking into account the link between infinitesimal cumulant functionals and $\mathcal{C}_k$-non-crossing cumulant functionals, Proposition \[propinfadd\] admits the following corollaries :
\[propadd\] Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. Consider subsets $\mathcal{M}_1,\mathcal{M}_2$ of $\mathcal{A}$ that are infinitesimally free of order $k$. Then, one has, for each $n\geq 1$ and each $n$-tuples $(a_1,\ldots ,a_n)\in \mathcal{M}_1^n,(b_1,\ldots ,b_n)\in \mathcal{M}_2^n$ $$\tilde \kappa _n(a_1+b_1,\ldots ,a_n+b_n)=\tilde \kappa _n(a_1,\ldots ,a_n)+\tilde \kappa _n(b_1,\ldots ,b_n).$$
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. Consider $a,b\in \mathcal{A} $ that are infinitesimally free of order $k$, then $$\tilde R_{a+b}=\tilde R_a+\tilde R_b.$$
Using Corollary \[propadd\], it is possible to state and prove $\mathcal{C}_k$-valued versions of some famous limit theorems of free probability. We discuss this without going into the details ; for a more complete discussion of limit theorems in free probability of type B, we refer to [@pop] and [@bs]. In a scarce $\mathcal{C}_k$-noncommutative probability space $( \mathcal{A} , \tilde \varphi )$, consider a sequence $(a_n)_{n\in {\mathbb{N}}}\in \mathcal{A} ^{\mathbb{N}}$ of centered infinitesimally free identically distributed $\mathcal{C}_k$-valued noncommutative random variables. Then the moments of the (rescaled by a $\frac{1}{\sqrt{N}}$ factor) sum $\frac{1}{\sqrt{N}} \sum _{n=1}^N a_n$ converge to a $\mathcal{C}_k$-valued distribution characterized by the vanishing of all of its cumulants except the second one : this is the $\mathcal{C}_k$-valued version of the free central limit theorem. The distributions that appear as limits in the preceding result deserve to be named $\mathcal{C}_k$-valued semicircular elements. Their moments may be computed using the $\mathcal{C}_k$-valued free moment-cumulant formula. Paralelly, a $\mathcal{C}_k$-valued version of the free Poisson theorem may also be stated and proved, and thus a $\mathcal{C}_k$-valued Poisson distribution may be defined.
[**4.2 Multiplication of infinitesimally free random variables**]{}
We now investigate the distribution of the product of $n$-tuples of noncommutative random variables that are infinitesimally free of order $k$. We first focus on a $\mathcal{C}_k$-noncommutative probability space because, the combinatorics being the same in this setting as in usual free probability, the proofs and results will be straightforward adaptations of the usual ones, which can be found in [@nsbook] for instance.
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. Consider subsets $\mathcal{M}_1,\mathcal{M}_2$ of $\mathcal{A}$ that are infinitesimally free of order $k$. Then, one has, for each $n\geq 1$ and each $n$-tuples $(a_1,\ldots ,a_n)\in \mathcal{M}_1^n,(b_1,\ldots ,b_n)\in \mathcal{M}_2^n$ $$\label{altprod}
\tilde \kappa _n(a_1b_1,\ldots ,a_nb_n)=\sum_{p\in NC^{(A)}(n)}
\tilde \kappa _p(a_1,\ldots ,a_n) \tilde \kappa _{\mbox{Kr}(p)}(b_1,\ldots ,b_n).$$
Using Proposition \[cwpae\], the left-hand side of is equal to $$\sum_{ \begin{array}{c}
{\scriptstyle \pi \in NC(2n) \ such} \\
{\scriptstyle that \ \pi \vee \theta = 1_s}
\end{array} }\tilde \kappa _\pi (a_1,b_1,a_2,\ldots ,b_{n-1},a_n,b_n),$$ where $\theta $ is the partition $\{\{1,2\},\ldots ,\{2n-1,2n\}\}$.\
By the vanishing of mixed cumulants condition, the only contributing terms are those indexed by non-crossing partitions $\pi $ which are reunion of a non-crossing partition $p$ of $\{1,3,\ldots ,2n-1\}$ and a non-crossing partition $q$ of $\{2,4,\ldots ,2n\}$. The condition $\pi \vee \theta = 1_s$ for such a partition $\pi $ may be reinterpreted as $q=\mbox{Kr}(p)$ (up to the identifications $\{1,3,\ldots ,2n-1\}\leftrightarrow [n]$ and $\{2,4,\ldots ,2n\}\leftrightarrow \overline{[n]}$).
Switching to the infinitesimal framework, one can state the following result.
Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$, and $(a_1,\ldots ,a_n),(b_1,\ldots ,b_n)\in \mathcal{A}^n$ be $n$-tuples with respective infinitesimal distributions $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$.\
If the sets $\{a_1,\ldots ,a_n\}$ and $\{b_1,\ldots ,b_n\}$ are infinitesimally free of order $k$, then the infinitesimal distribution of $(a_1b_1,\ldots ,a_nb_n)$ only depends on $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$. It is denoted by $(\mu ^{(i)})_{0\leq i\leq k}\boxtimes ^{(k)} (\nu ^{(i)})_{0\leq i\leq k}$ and called the infinitesimal free multiplicative convolution of order $k$ of $(\mu ^{(i)})_{0\leq i\leq k}$ and $(\nu ^{(i)})_{0\leq i\leq k}$.
If $a,b\in \mathcal{A}$ are $\mathcal{C}_k$-noncommutative random variables that are infinitesimally free of order $k$ in a scarce $\mathcal{C}_k$-noncommutative probability space, the $\mathcal{C}_k$-valued R-transform of $a\cdot b$ is $\tilde R_{a\cdot b}=\tilde R_{a}$$_{\mathcal{C}_k}\tilde R_{b}$, where $_{\mathcal{C}_k}$ is the version of the boxed convolution operation introduced in [@ns96] with scalars in $\mathcal{C}_k$. We recall in the next subsection the definition and main properties of this operation.
[**4.3 Boxed convolution of type A**]{}
An operation on formal power series in several noncommuting indeterminates and with complex coefficients is introduced in [@ns96], and called boxed convolution. It is defined as a convolution on the lattices of non-crossing partitions (of type A). We recall here this definition, but for series in only one variable (for simplicity) and with coefficients in any unital complex algebra. This is already the point of view adopted in [@bgn].
\[starbox\] Let $\mathcal{C}$ be a unital commutative algebra over ${\mathbb{C}}$. On $\Theta _{\mathcal{C}}^{(A)}$ we define a binary operation $_{\mathcal{C}}^{(A)}$, as follows. If $$f(z)=\sum_{n=1}^\infty \alpha _nz^n \in \Theta _{\mathcal{C}}^{(A)},$$ and $$g(z)=\sum_{n=1}^\infty \beta _nz^n \in \Theta _{\mathcal{C}}^{(A)},$$ then $f$$_{\mathcal{C}}^{(A)}g$ is the series $\sum_{n=1}^\infty \gamma _nz^n$, where $$\gamma _m=\sum_{\substack{p\in NC^{(A)}(m)\\p:=\{E_1,\ldots ,E_h\}\\\mbox{Kr}(p):=\{F_1,\ldots ,F_l\}}}
(\prod_{i=1}^h\alpha _{\mbox{card}(E_i)}) \cdot (\prod_{j=1}^l\beta _{\mbox{card}(F_j)}).$$
\[multi\] It is obviously possible to define a boxed convolution operation for power series in several noncommuting indeterminates and with coefficients in $\mathcal{C}$. The formulas are the same as in the case of complex coefficients, which first appeared in [@ns96] and can also be found in [@nsbook].
The operation $_{\mathcal{C}}^{(A)}$ is associative, commutative and the series $$\Delta _\mathcal{C}^{(A)}(z):=1_{\mathcal{C}}z$$ is its unit element. There is another important series in $\Theta _{\mathcal{C}}^{(A)}$, namely $$\zeta _\mathcal{C}^{(A)}(z):=\sum_{n=1}^\infty 1_{\mathcal{C}}z^n.$$ Notice that a series $f\in \Theta _{\mathcal{C}}^{(A)}$ is invertible with respect to $_{\mathcal{C}}^{(A)}$ if and only if its coefficient of degree one is itself invertible in the algebra $\mathcal{C}$. In particular, $\zeta _\mathcal{C}^{(A)}$ is invertible with respect to $_{\mathcal{C}}^{(A)}$, and its inverse is called the Möebius series, and denoted by $\mbox{M\"ob} _\mathcal{C}^{(A)}$. The proofs of these claims may be obtained by a straightforward adaptation of the proofs given in [@ns96]. The free moment-cumulant relation of free probability (and its inverse) may be read at the level of power series : more precisely, in a noncommutative probability space $(\mathcal{A} , \varphi )$, the moment series and the R-transform of $a\in \mathcal{A} $ satisfy the following relations: $M_a=R_a$$_{{\mathbb{C}}}^{(A)} \zeta _{{\mathbb{C}}}^{(A)}$, $R_a=M_a$$_{{\mathbb{C}}}^{(A)} \mbox{M\"ob} _{{\mathbb{C}}}^{(A)}$. These formulas have infinitesimal analogues, as stated in the next proposition. It is indeed straightforward to check that, in the particular case of single variables, the formulas and may be read at the level of power series as follows :
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space and consider a $\mathcal{C}_k$-noncommutative random variable $a\in \mathcal{A}$. Then the $\mathcal{C}_k$-valued moment series $\tilde M_a$ and the $\mathcal{C}_k$-valued R-transform $\tilde R_a$ of $a$ are related by the equivalent formulas : $\tilde M_a=\tilde R_a$$_{\mathcal{C}_k}^{(A)} \zeta _{\mathcal{C}_k}^{(A)}$, $\tilde R_a=\tilde M_a$$_{\mathcal{C}_k}^{(A)} \mbox{M\"ob} _{\mathcal{C}_k}^{(A)}.$
The importance of boxed convolution (with complex coefficients) in free probability also comes from the fact, proved in [@ns96], that $_{{\mathbb{C}}}^{(A)}$ provides the combinatorial description for the multiplication of two free noncommutative random variables, in terms of their R-transforms. More precisely, we have, for free $a,b$ in a noncommutative probability space $(\mathcal{A} , \varphi )$ : $R_{a\cdot b}=R_a$$_{{\mathbb{C}}}^{(A)}R_b$.
Interestingly, $_{\mathcal{C}_k}^{(A)}$ also provides the combinatorial description for the multiplication of two infinitesimally free infinitesimal noncommutative random variables, in terms of their R-transforms.
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. Consider $a,b\in \mathcal{A} $ that are infinitesimally free of order $k$, then
In [@ns97], a “Fourier transform” is introduced for multiplicative functions on non-crossing partitions. It is barely a map $\mathcal{F}$ which associates to $f(z)=\sum_{n=1}^\infty \alpha _nz^n\in \Theta _{{\mathbb{C}}}^{(A)}$, with $\alpha _1\not=0$ (to ensure that $f$ is invertible with respect to the composition of formal power series), the series $\mathcal{F}(f)(z):=\frac{1}{z}f^{\langle -1\rangle}(z)$. The map $\mathcal{F}$ has the important property to convert the boxed convolution into the multiplication of formal power series : $\mathcal{F}(f$$_{{\mathbb{C}}}^{(A)}g)=\mathcal{F}(f)\cdot \mathcal{F}(g)$. If $a$ is a noncommutative random variable with non-zero mean and R-transform $R_a$ in a noncommutative probability space, the series $\mathcal{F}(R_a)$ is of great importance : this is a combinatorial approach to Voiculescu’s S-tranform [@voi2]. As noticed in [@pop], the combinatorial proofs remain valid for series with $\mathcal{C}_k$-valued coefficients such that the coefficient of degree one is invertible.
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. The $\mathcal{C}_k$-valued S-transform of an infinitesimal noncommutative random variable $a\in \mathcal{A} $ such that $\tilde \varphi (a)$ is invertible in $\mathcal{C}_k$ is the power series $\tilde S_a\in \Theta ^{(k)}$ defined as follows: $$\tilde S_a(z):=\frac{1}{z}\tilde R_a^{\langle -1\rangle}(z).$$
Let $( \mathcal{A} , \tilde \varphi )$ be a scarce $\mathcal{C}_k$-noncommutative probability space. Consider $a,b\in \mathcal{A} $ that are infinitesimally free of order $k$, and such that $\tilde \varphi (a)$ and $\tilde \varphi (b)$ are invertible in $\mathcal{C}_k$, then the $\mathcal{C}_k$-valued S-transform $\tilde S_{a\cdot b}$ of $a\cdot b$ satisfies: $$\tilde S_{a\cdot b}(z)=\tilde S_a(z) \tilde S_b(z).$$
Practically speaking, the computation of the distribution of the product of two infinitesimally free infinitesimal noncommutative random variables requires a good understanding of the $\mathcal{C}_k$-valued version of the boxed convolution. More precisely, in the notations of Definition \[starbox\], it would be of interest to have a formula for $\gamma _m^{(i)}$ as a function of the $\alpha _n^{(j)}$’s and the $\beta _n^{(j)}$’s.\
As mentioned before, the version of the boxed convolution with scalars in $\mathcal{C}_0={\mathbb{C}}$ is a classical operation in free probability. The version of the boxed convolution with scalars in $\mathcal{C}_1=\mathbb{G}$ has already been considered in [@bgn], where it is shown to coincide with the boxed convolution based on non-crossing partitions of type B, in connection with free probability of type B. This leads to the natural question : does the operation $_{\mathcal{C}_k}^{(A)}$ coincide with a boxed convolution based on a certain set of special non-crossing partitions. In Section 7, we will give a positive answer to this problem, by introducing the non-crossing partitions of type $k$. Before that, we review some background on non-crossing partitions and boxed convolution of type B.
[**5. Non-crossing partitions and boxed convolution of type B**]{}
This section is devoted to some background on non-crossing partitions of type B and on the boxed convolution of type B.
[**5.1 Non-crossing partitions of type B**]{}
As recalled in Section 2, there is a close link between the lattice of non-crossing partitions and the Cayley graph of the symmetric group. Actually, one may interpret the lattices of non-crossing partitions in terms of the root systems of type A, justifying the notation $NC^{(A)}(n)$. This led Reiner to introduce in [@rei] the type B analogue $NC^{(B)}(n)$ of the lattice of non-crossing partitions. To this aim, consider the totally ordered set $$[\pm n]=\{1<2<\ldots <n<-1<-2<\ldots <-n\}.$$ One defines $NC^{(B)}(n)$ to be the subset of $NC^{(A)}([\pm n])$ consisting of non-crossing partitions that are invariant under the inversion map $x\mapsto -x$.\
In such a partition $\pi \in NC^{(B)}(n)$, there is at most one block that is inversion-invariant, called, whenever it exists, the zero-block of $\pi $. The other blocks of $\pi $ come two by two: if $F$ is a block which is not inversion-invariant, then $-F$ is another block (obviously not inversion-invariant).\
It is immediate that $NC^{(B)}(n)$ is a sublattice of $NC^{(A)}([\pm n])$, with the same minimal and maximal elements.\
Moreover, $NC^{(B)}(n)$ is closed under the Kreweras complementation maps $\mbox{Kr}$ and $\mbox{Kr}'$ (considered on $NC^{(A)}([\pm n])$).\
When restricted from $NC^{(A)}([\pm n])$ to $NC^{(B)}(n)$, these maps will then give two anti-isomorphisms of $NC^{(B)}(n)$, inverse to each other, and which will also be called (without ambiguity) Kreweras complementation maps (on $NC^{(B)}(n)$). In this case, the important relation becomes $$| \pi |+| \mbox{Kr}(\pi ) |=2n+1, \forall \pi \in NC^{(B)}(n).$$ As a consequence, for $\pi \in NC^{(B)}(n)$, exactly one of the two partitions $\pi $ and $\mbox{Kr}(\pi )$ has a zero-block. In the description of a non-crossing partition of type B, a role is played by the absolute value map $\mbox{Abs} : [\pm n] \longrightarrow [n]$ sending $\pm i$ to $i$.
Any map $f$ defined from $[m]$ into $[n]$ is naturally extended to a map from $[m]\cup \overline{[m]}$ into $[n]\cup \overline{[n]}$ by simply requiring that $f(\overline{i})=\overline{f(i)}$.\
Moreover, if $Y$ is a subset of $[m]\cup \overline{[m]}$, we will use the notation $f(Y)$ for the set $\{f(y), y\in Y\} \subset [n]\cup \overline{[n]}$.\
Finally, given a collection $\Upsilon $ of subsets of $[m]\cup \overline{[m]}$, we will denote by $f(\Upsilon )$ the collection $\{f(Y), Y\in \Upsilon \}$ of subsets of $[n]\cup \overline{[n]}$.
Let us now state a key result of [@bgn].
$\pi \mapsto \mbox{Abs}(\pi )$ is a $(n+1)$-to-$1$ map from $NC^{(B)}(n)$ onto $NC^{(A)}(n)$.
We refer to the paper [@bgn] for the proof. In the next subsection, we recall the definition of the type B analogue of the boxed convolution operation and give the announced result stating that this operation is a boxed convolution of type A on the algebra $\mathcal{C}_1$.
[**5.2 Boxed convolution of type B**]{}
1. We denote by $\Theta ^{(B)}$ the set of power series of the form $$f(z)=\sum_{n=1}^\infty (\alpha _n',\alpha _n'')z^n,$$ where the $\alpha _n'$’s and $\alpha _n''$’s are complex numbers.
2. Let $f(z):=\sum_{n=1}^\infty (\alpha _n',\alpha _n'')z^n$ and $g(z)=\sum_{n=1}^\infty (\beta _n',\beta _n'')z^n$ be in $\Theta ^{(B)}$. For every $m\geq 1$, consider the numbers $\gamma _m'$ and $\gamma _m''$ defined by $$\gamma _m'=\sum_{\substack{p\in NC^{(A)}(m)\\p:=\{E_1,\ldots ,E_h\}\\\mbox{Kr}(p):=\{F_1,\ldots ,F_l\}}}
(\prod_{i=1}^h\alpha _{\mbox{card}(E_i)}') \cdot (\prod_{j=1}^l\beta _{\mbox{card}(F_j)}'),$$ $$\gamma _m''=\sum_{\substack{p\in NC^{(B)}(m)\,with\,zero-block\\p:=\{Z,X_1,-X_1,\ldots ,X_h,-X_h\}\\
\mbox{Kr}(p):=\{Y_1,-Y_1,\ldots ,Y_l,-Y_l\}}} (\prod_{i=1}^h\alpha _{\mbox{card}(X_i)}') \cdot
\alpha _{\mbox{card}(Z)/2}'' \cdot (\prod_{j=1}^l\beta _{\mbox{card}(Y_j)}')$$ $$+ \sum_{\substack{p\in NC^{(B)}(m)\,without\,zero-block\\p:=\{X_1,-X_1,\ldots ,X_h,-X_h\}\\\mbox{Kr}(p)
:=\{Z,Y_1,-Y_1,\ldots ,Y_l,-Y_l\}}} (\prod_{i=1}^h\alpha _{\mbox{card}(X_i)}') \cdot \beta _{\mbox{card}(Z)/2}''
\cdot (\prod_{j=1}^l\beta _{\mbox{card}(Y_j)}').$$ Then the series $\sum_{n=1}^\infty (\gamma _n',\gamma _n'')z^n$ is called the boxed convolution of type B of $f$ and $g$, and is denoted by $f$ $^{(B)} g$.
[@bgn] Theorem 5.3 $^{(B)}=$$_{\mathcal{C}_1}^{(A)}$
We now introduce the non-crossing partitions of type $k$, generalizing non-crossing partitions of type A and B.
[**6. Non-crossing partitions of type $k$**]{}
This section is devoted to the introduction and study of a set of non-crossing partitions, namely the set of non-crossing partitions of type $k$, which has to be a cover of $NC{(A)}(n)$ related to the version of the boxed convolution with scalars in $\mathcal{C}_k$.
[**6.1 Definition and first properties**]{}
Let $n$ be a positive integer. We call [*reduction mod $n$ map*]{} the map $$\mbox{Red}_n^{(k)} : [(k+1)n] \rightarrow [n]$$ sending each $i\in [(k+1)n]$ to its congruence class mod $n$.
For $k=0$, the map $\mbox{Red}_n^{(0)}$ is simply the identity map on $[n]$.\
For $k=1$, up to identifying $[2n]$ with $[\pm n]$, the map $\mbox{Red}_n^{(1)}$ is identified with $\mbox{Abs}$.
\[redprop\] A non-crossing partition $\pi $ of $[(k+1)n]$ is said to satisfy [*the mod $n$ reduction property*]{} if $\mbox{Red}_n^{(k)}(\pi )$ is a non-crossing partition of $[n]$ and if $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ is a non-crossing partition of $\overline{[n]}$.
[*Non-crossing partitions of type $k$*]{} are the non-crossing partitions of $[(k+1)n]$ satisfying the mod $n$ reduction property.
We write $NC^{(k)}(n)$ for the set of non-crossing partitions of type $k$, that is non-crossing partitions of $[(k+1)n]$ satisfying the mod $n$ reduction property.
All non-crossing partitions of $[n]$ trivially satisfy the mod $n$ reduction property (since $\mbox{Red}_n^{(0)}$ is simply the identity map). Hence $NC^{(0)}(n)=NC^{(A)}(n)$.
The next proposition states that the non-crossing partitions of type $k$ are a generalization of the non-crossing partitions of type B.
If we identify $[\pm n]$ with $[2n]$ and $\mbox{Abs}$ with $\mbox{Red}_n^{(1)}$,\
then $NC^{(B)}(n)=NC^{(1)}(n)$.
That $\pi \in NC^{(B)}(n)$ satisfies the mod $n$ reduction property is a corollary of Proposition 1.3 and Lemma 1 in [@bgn].\
For the converse, let $\pi \in NC^{(1)}(n)$ satisfy the mod $n$ reduction property, and assume that there exist two elements $x,y\in [\pm n]$ such that $x\sim _\pi y$ , $-x\not\sim _\pi -y$. By reduction mod $n$ property, we necessarily have $-y\sim _\pi x\sim _\pi y\sim _\pi -x$, which is a contradiction.
\[B\] The proof above and Lemma 1 in [@bgn] show that, for a non-crossing partition $\pi $ of $[2n]$, the mod $n$ reduction property is equivalent to the only requirement that $\mbox{Red}_n^{(1)}(\pi )$ is a non-crossing partition of $[n]$.
In Definition \[redprop\], the reduction mod $n$ property for a non-crossing partition $\pi $ of $[(k+1)n]$ consists of two requirements : $\mbox{Red}_n^{(k)}(\pi )$ has to be a non-crossing partition of $[n]$ and $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ has to be a non-crossing partition of $\overline{[n]}$. Actually, there is a slightly stronger characterization stated in the next proposition.
\[charac\] A non-crossing partition $\pi $ of $[(k+1)n]$ satisfies the reduction mod $n$ property if and only if $\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))$ is a non-crossing partition of $[n]\cup \overline {[n]}$.
If $\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))$ is a non-crossing partition of $[n]\cup \overline {[n]}$, since $\mbox{Red}_n^{(k)}(\pi )$ is a family of subsets of $[n]$ and $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ is a family of subsets of $\overline{[n]}$, they have to be non-crossing partitions of $[n]$ and $\overline{[n]}$ respectively ; in other words $\pi $ has to satisfy the reduction mod $n$ property.\
We assume now that $\pi $ is a non-crossing partition of $[(k+1)n]$ satisfying the reduction mod $n$ property, and aim at proving that $\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))$ is a non-crossing partition of $[n]\cup \overline {[n]}$.\
By the reduction property, $\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))=
\mbox{Red}_n^{(k)}(\pi )\cup \mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ is the union of a partition of $[n]$ and of a partition of $\overline{[n]}$, and hence a partition of $[n]\cup \overline {[n]}$. To prove that this partition is non-crossing, consider four elements $a<b<c<d$ of $[n]\cup \overline {[n]}$, such that $a\sim _{\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))}c$ and $b\sim _{\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))}d$. We have to show that $a\sim _{\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))}b$.\
Let $1\leq i_0\leq (k+1)n+1$ be minimal with the property that $\mbox{Mix}(\pi ,i_0)$ contains an element $x$ such that $\mbox{Red}_n^{(k)}(x)\in \{a,b,c,d\}$. Choose also the smallest such $x$. We may assume that $\mbox{Red}_n^{(k)}(x)=a$ (the other cases are similar). By assumption, $c\in \mbox{Red}_n^{(k)}(\mbox{Mix}(\pi ,i_0))$ : there is an element $z\in \mbox{Mix}(\pi ,i_0)$ such that $\mbox{Red}_n^{(k)}(z)=c$. Our choice of $x$ ensures that $x<z$ and there is necessarily an element $x<y<z$ such that $\mbox{Red}_n^{(k)}(y)=b$. By minimality of $i_0$, $y\in \mbox{Mix}(\pi ,i_0)$, hence $b\in \mbox{Red}_n^{(k)}(\mbox{Mix}(\pi ,i_0))$ is linked to $a$ in ${\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))}$ and we are done.
When $k=0$, the reduction mod $n$ property is satisfied by any non-crossing partition of $[n]$ and is in particular equivalent to the only empty requirement : $\pi \in NC^{(A)}(n)$ satisfies the reduction mod $n$ property if and only if $\mbox{Red}_n^{(0)}(\pi )$ is a non-crossing partition of $[n]$.\
As explained in Remark \[B\], this is also the case when $k=1$ : $\pi \in NC^{(A)}(2n)$ satisfies the reduction mod $n$ property if and only if $\mbox{Red}_n^{(1)}(\pi )$ is a non-crossing partition of $[n]$.\
Assume now that $k\geq 2$ ; the situation then is different.\
As an example, for $k=2$ and $n=2$, consider the partition $$\pi :=\{\{1,2,3\},\{4,5,6\}\}\in NC^{(A)}(6).$$ It is straightforward to check that $\mbox{Red}_2^{(2)}(\pi )=\{1,2\}$ is a non-crossing partition of $[2]$. However, from the easy computation $\mbox{Kr}(\pi )=\{\{\overline{1}\},\{\overline{2}\},\{\overline{4}\},\{\overline{5}\},\{\overline{3},\overline{6}\}\}$, we deduce that $\mbox{Red}_2^{(2)}(\mbox{Kr}(\pi ))$ is not a partition of $\overline{[6]}$ and consequently that $\pi $ does not satisfy the reduction mod $2$ property.
The following proposition states that the Kreweras complementation maps may be considered as two order-reversing bijections of $NC^{(k)}(n)$.
The restrictions from $NC^{(A)}((k+1)n)$ to $NC^{(k)}(n)$ of $\mbox{Kr}$ and $\mbox{Kr}'$ are order-reversing bijections of $NC^{(k)}(n)$.
The name of Kreweras complementation map and the notations $\mbox{Kr}$, $\mbox{Kr}'$ will be conserved as there is no ambiguity about the meaning of $\mbox{Kr}(\pi )$ or $\mbox{Kr}'(\pi )$ whether $\pi $ is viewed as an element of $NC^{(k)}(n)$ or of $NC^{(A)}((k+1)n)$.
It is clearly sufficient to prove that the non-crossing partition $\mbox{Kr}(\pi )$ of $\overline{[(k+1)n]}$ satisfies the reduction mod $n$ property whenever $\pi $ does. Assume that the non-crossing partition $\pi $ of $[(k+1)n]$ satisfies the reduction mod $n$ property. By assumption, $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ is a non-crossing partition of $\overline{[n]}$. It remains to prove that $\mbox{Red}_n^{(k)}(\mbox{Kr}^2(\pi ))$ is a non-crossing partition of $\overline{\overline{[n]}}$.\
From the geometric description of $\mbox{Kr}^2(\pi )$ given in Section 3, we deduce that $\mbox{Red}_n^{(k)}(\mbox{Kr}^2(\pi ))$ is obtained from $\mbox{Red}_n^{(k)}(\pi )$ by a rotation. By reduction mod $n$ property, $\mbox{Red}_n^{(k)}(\pi )$ is a non-crossing partition of $[n]$, so $\mbox{Red}_n^{(k)}(\mbox{Kr}^2(\pi ))$ is itself a non-crossing partition of $\overline{\overline{[n]}}$. Thus the proof is complete.
Given $\pi \in NC^{(k)}(n)$, $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$ and $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ are thus two non-crossing partitions of $\overline{[n]}$. The following lemma, generalizing Lemma 1 of [@bgn], states that these two partitions coincide.
\[comm\] $\forall \pi \in NC^{(k)}(n), \mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))=\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$.
Let $\pi$ be a non-crossing partition of type $k$. By Proposition \[charac\], $\mbox{Red}_n^{(k)}(\pi )\cup \mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))=
\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))$ is a non-crossing partition of $[n]\cup \overline {[n]}$. Since $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$ is maximal with the property that $\mbox{Red}_n^{(k)}(\pi )\cup \mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$ is non-crossing, it follows that $$\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))\preceq \mbox{Kr}(\mbox{Red}_n^{(k)}(\pi )).$$ There is equality if, for any $\overline x$ having a neighbour $\overline y>\overline x$ in $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$, $\overline y$ is linked to $\overline x$ in $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$. For such elements $\overline x, \overline y\in \overline{[n]}$, we call $V$ the block of $\pi $ containing $x+1$. The reduction property implies that $\mbox{Red}_n^{(k)}(V)$ is a block of the partition $\mbox{Red}_n^{(k)}(\pi )$. By construction of the Kreweras complement, $x+1$ is the smallest element of both $V$ and $\mbox{Red}_n^{(k)}(V)$, and $y$ is the greatest element of $\mbox{Red}_n^{(k)}(V)$. Consider now the greatest element $z$ of $V$. Notice that $x+1\leq \mbox{Red}_n^{(k)}(z)\leq y$. By construction of the Kreweras complement again, $\overline x$ is linked to $\overline z$ in $\mbox{Kr}(\pi )$, then $\overline x$ is linked to $\mbox{Red}_n^{(k)}(\overline z)=\overline {\mbox{Red}_n^{(k)}(z)}$ in $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ and therefore in $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$. This means that, if $\mbox{Red}_n^{(k)}(z)<y$, $\overline y$ would not be the neighbour of $\overline x$ in $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$, which is a contradiction. So $\mbox{Red}_n^{(k)}(z)=y$ or, in other words, $\overline x$ is linked to $\overline y$ in $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$.
A deeper description of non-crossing partitions of type $k$ is given in the next subsection.
[**6.2 Structure of non-crossing partitions of type $k$**]{}
The goal of this subsection is to describe the structure of a non-crossing partition of type $k$. In the next proposition, $t$ denotes the bijection between non-crossing partitions and permutations lying on a geodesic in the Cayley graph of the symmetric group, introduced by Biane in [@bia], and described in Section 3. We warn the reader that we choose to use the same notation $t$ for this bijection, defined either on $NC(n)$ or $NC((k+1)n)$. We hope that this choice, made in the sake of simplicity, will not be a source of confusion in the reader’s mind. The content of this proposition is, roughly speaking, that a type $k$ non-crossing partition $\pi $ is characterized by the two requirements : $\mbox{Red}_n^{(k)}(\pi )$ is a non-crossing partition of $NC^{(A)}(n)$ and the elements of each of the blocks of $\pi $ come in the same order as their congruence classes in its reduction $\mbox{Red}_n^{(k)}(\pi )$.
\[biane\] For $\pi \in NC^{(A)}((k+1)n)$ such that $\mbox{Red}_n^{(k)}(\pi )\in NC^{(A)}(n)$, $\pi \in NC^{(k)}(n)$ if and only if $$\label{shift}
\forall x\in [(k+1)n], \mbox{Red}_n^{(k)}(t(\pi )(x))=t(\mbox{Red}_n^{(k)}(\pi ))(\mbox{Red}_n^{(k)}(x)).$$
Assume first that $\pi \in NC^{(k)}(n)$ and fix $x\in [(k+1)n]$. Set $y:=t(\pi )(x)$.\
By construction of $t$, $y$ is the neighbour of $x$ in $\pi $ and $t(\mbox{Red}_n^{(k)}(\pi ))(\mbox{Red}_n^{(k)}(x))$ is the neighbour of $\mbox{Red}_n^{(k)}(x)$ in $\mbox{Red}_n^{(k)}(\pi )$. By construction of the Kreweras complement, $\overline{x}$ is the neighbour of $\overline{y-1}$ in $\mbox{Kr}(\pi )$, and $\overline{\mbox{Red}_n^{(k)}(x)}$ is the neighbour of $\overline{t(\mbox{Red}_n^{(k)}(\pi ))(\mbox{Red}_n^{(k)}(x))-1}$ in $\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))=\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ (the latter equality holds because of Proposition \[comm\]). By reduction property, $\overline{\mbox{Red}_n^{(k)}(y-1)}$ is linked to $\overline{\mbox{Red}_n^{(k)}(x)}$. It follows that the neighbour of $\mbox{Red}_n^{(k)}(x)$ in $\mbox{Red}_n^{(k)}(\pi )$, $t(\mbox{Red}_n^{(k)}(\pi ))(\mbox{Red}_n^{(k)}(x))$, is the first point coming after $\overline{\mbox{Red}_n^{(k)}(y-1)}$ linked to $\mbox{Red}_n^{(k)}(x)$ : it is $\mbox{Red}_n^{(k)}(y)$ and we are done. For the converse, let $\pi \in NC^{(A)}((k+1)n)$ be such that $\mbox{Red}_n^{(k)}(\pi )\in NC^{(A)}(n)$ and assume that condition ? holds. We have to prove that $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))$ is a non-crossing partition of $\overline{[n]}$. Let $\overline{x}\in \overline{[(k+1)n]}$, its neighbour in $\mbox{Kr}(\pi )$ is $\overline{t(\pi )^{-1}(x+1)}$, by construction of the Kreweras complement. It follows of condition ? that $\mbox{Red}_n^{(k)}(t(\pi )^{-1}(x+1))=t(\mbox{Red}_n^{(k)}(\pi ))^{-1}(\mbox{Red}_n^{(k)}(x+1))$. Hence the congruence class of the neighbour of $\overline{x}$ in $\mbox{Kr}(\pi )$ only depends on the congruence class of $x$, and moreover $\mbox{Red}_n^{(k)}(\mbox{Kr}(\pi ))=\mbox{Kr}(\mbox{Red}_n^{(k)}(\pi ))$ and we are done.
The preceding proposition has some important consequences.
Let $\pi \in NC^{(A)}((k+1)n)$ and $V$ be a block of $\pi\cup \mbox{Kr}(\pi )$. The cardinal of $\mbox{Red}_n^{(k)}(V)$ divides the cardinal of $V$. We call [*multiplicity*]{} of $V$ the quotient $$\mbox{mult}_{\pi\cup \mbox{Kr}(\pi )}(V):=\frac{\mbox{card}(V)}{\mbox{card}(\mbox{Red}_n^{(k)}(V))}.$$ This is a positive integer lower or equal than $k+1$. The blocks of multiplicity $1$ will be called [*simple*]{}.
For $x \in V$, the cardinal of $V$ is the smallest positive $i$ verifying $$(t(\pi ))^i(x)=x.$$ A repeated use of Proposition \[biane\] gives that, for such an $i$, $$\label{cycl}
(t(\mbox{Red}_n^{(k)}(\pi )))^i(\mbox{Red}_n^{(k)}(x))=\mbox{Red}_n^{(k)}(x).$$ Thus $i$ is a multiple of the cardinal of $\mbox{Red}_n^{(k)}(V)$, which is also characterized by the fact that it is the smallest positive $i$ veriying condition .
It is not so difficult to see that, if there is a block of multiplicity $k+1$ in $\pi\cup \mbox{Kr}(\pi )$, for $\pi \in NC^{(A)}((k+1)n)$, the other blocks are necessarily simple, because one cannot link two elements of the same congruence class without crossing the block of multiplicity $k+1$. This is in fact a particular case of the following result :
For $\pi \in NC^{(k)}(n)$, $$\sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi ))} (mult_{\pi \cup \mbox{Kr}(\pi )}(V)-1)=k.$$
This is a simple computation. First notice that $$\begin{aligned}
\label{mult}
\nonumber \sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi ))} (mult_{\pi \cup \mbox{Kr}(\pi )}(V)-1) =
\\\sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi ))} mult_{\pi \cup \mbox{Kr}(\pi )}(V)-|\pi \cup \mbox{Kr}(\pi )|.\end{aligned}$$ The first term in is $$\sum _{W\in \mbox{bl}(\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi )))}
\sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi )): \mbox{Red}_n^{(k)}(V)=W} mult_{\pi \cup \mbox{Kr}(\pi )}(V).$$ But for any block $W$ of $\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))$, one has $$\sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi )): \mbox{Red}_n^{(k)}(V)=W} mult_{\pi \cup \mbox{Kr}(\pi )}(V)=k+1.$$ Applying twice formula , we get $$\begin{aligned}
\nonumber \sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi ))} (mult_{\pi \cup \mbox{Kr}(\pi )}(V)-1)=\ &(k+1)|\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi ))|-|\pi \cup \mbox{Kr}(\pi )|
\\\nonumber =\ &(k+1)(n+1)-((k+1)n+1)
\\\nonumber =\ &k.\end{aligned}$$
For a partition $\pi \in NC^{(k)}(n)$, one may define a vector $\lambda _\pi $ with integer coordinates as follows : $$(\lambda _\pi )_i=\sum _{V\in \mbox{bl}(\pi \cup \mbox{Kr}(\pi )): \mbox{Red}_n^{(k)}(V)
=\mbox{Mix}(\mbox{Red}_n^{(k)}(\pi \cup \mbox{Kr}(\pi )),i)} (mult_{\pi \cup \mbox{Kr}(\pi )}(V)-1).$$ The vector $\lambda _\pi \in \Lambda _{n+1,k}$ is called the [*shape*]{} of $\pi $.
A type B non-crossing partition $\pi $ is determined by its absolute value $p:=\mbox{Abs}(\pi )$ and the choice of the block $Z\in \mbox{bl}(p\cup \mbox{Kr}(p))$, which has to be lifted to the zero-block of $\pi $. This latter choice is encoded in the shape $\lambda _\pi $ of $\pi $. Indeed, type B corresponds to the case $k=1$ of non-crossing partitions of type $k$ and therefore the shape $\lambda _\pi $ belongs to the set $\Lambda _{n+1,1}$ consisting of the $n+1$ vectors $e_i=(\delta _i^j)_{1\leq j\leq n+1} , 1\leq i\leq n+1$. That $\lambda _\pi =e_i$ means exactly that we have to choose the block $Mix(p,i)$ as the absolute value of the zero-block. The conclusion is that a type B non-crossing partition, considered as a non-crossing partition of type $1$, is determined by its reduction (or absolute value in the type B language) and its shape. Unfortunately, this is not the case when $k\geq 2$. It is interesting to ask how to determine a general non-crossing partition of type $k$. This question is investigated in the proof of the next proposition.
\[const\] Let $\lambda \in \Lambda _{n+1,k}$. The number of $\pi \in NC^{(k)}(n)$ having shape $\lambda $ and reduction a fixed non-crossing partition $p\in NC^{(A)}(n)$ is the same for any choice of $p\in NC^{(A)}(n)$. We will denote this quantity by $r(\lambda )$.
As announced, we investigate how to determine a type $k$ non-crossing partition $\pi \in NC^{(k)}(n)$, once its reduction $p\in NC^{(A)}(n)$ and its shape $\lambda \in \Lambda _{n+1,k}$ are given. We know that $\mbox{Mix}(p,1)$ is a singleton of $[n]\cup \overline{[n]}$. For simplicity, we assume that it is a singleton $\{i\}$ of $[n]$. We need to know how to form the blocks of $\pi $ reducing to $\{i\}$. The number of admissible ways to form these blocks depends on the value of $\lambda _1$ but of course not on $p$, because the actual value of $i$ does not come into the game. Assume that these blocks are formed; this gives a decomposition of $[(k+1)n]\setminus \{x\mid \mbox{Red}_n^{(k)}(x)=i\}\cup \overline{[(k+1)n]}$ into $\lambda _1+1$ sets, according to the following process : let us denote by $\{i+l_1n,\ldots ,i+l_mn\}$ the smallest (with respect to $\sqsubset $) of the blocks we have just formed that is not simple (if there is no such block, i.e. when $\lambda _1=0$, our decomposition is trivial) ; each of the $\{\overline{i+l_jn},\ldots ,\overline{i-1+l_{j+1}n}\}$ becomes a set in our decomposition after erasing the $i+ln, l_j\leq l\leq l_{j+1}$, for each $1\leq j\leq m-1$. Then remove all elements $x$ such that $i+l_1n\leq x\leq i+l_mn$ and repeat the process by considering the new smallest block with respect to $\sqsubset $ among the remaining blocks that are not simple. Notice that the sets obtained this way may be identified with sets of the form $[l(n-1)]\cup \overline{[l(n-1)]}$, for some $l\leq k+1$, up to identifying the first and last elements of the sets. This can be done, because these elements are necessarily linked by construction of the Kreweras complement. On each of these sets, $\pi $ induces a non-crossing partition that belongs to $NC^{(l)}(n-1)$. All such induced non-crossing partitions have the same reduction $\tilde p$ obtained by erasing in $p\cup \mbox{Kr}(p)$ the element $i$ and by identifying $\overline{i-1}$ with $\overline{i}$ (which are also necessarily linked in $\mbox{Kr}(p)$). The shapes of the induced partitions sum to the shape $\lambda $ of $\pi $. Hence a non-crossing partition of type $k$ is determined by its reduction $p$, its shape $\lambda $, an admissible way to form the blocks reducing to $\mbox{Mix}(p,1)$, an admissible decomposition of $\lambda $ and the choice of the induced non-crossing partitions in sets $NC^{(l)}(n-1)$, having reduction $\tilde p$ and shape the summands in the decomposition of $\lambda $.\
Our argument goes by induction on $n$. For $n=1$ and any $k$, there is only one possible reduction, because $\# NC^{(A)}(1)=1$ and consequently there is nothing to prove in that case. Assume that, for any $l$, the number of partitions in $NC^{(l)}(n-1)$ with given shape and reduction does not depend on the choice of the reduction. According to our analysis of the first part of the proof, the number of partitions $\pi \in NC^{(k)}(n)$ with given shape $\lambda $ and reduction $p$ does not depend on the choice of the reduction, because we noticed that the number of admissible ways to form the blocks reducing to $\mbox{Mix}(p,1)$ does not depend on $p$, the shape decomposition depend only on $\lambda $ and the way the latter blocks are formed, and by induction, the numbers of choices for the induced partitions only depend on their shapes.
For small values of $k$, one may easily compute the values of $r(\lambda )$ for each $\lambda \in \Lambda _{n+1,k}$.\
In the simplest case $k=0$, $$\Lambda _{n+1,0}=\{(0,\ldots ,0)\},$$ $$r((0,\ldots ,0))=1.$$ For $k=1$, $$\Lambda _{n+1,1}=\{e_i\}_{i=1,\ldots ,n+1},$$ and one has $$\forall 1\leq i\leq n+1, r(e_i)=1.$$ For $k=2$, $$\Lambda _{n+1,2}=\{e_i+e_j\}_{i,j=1,\ldots ,n+1}.$$ The value of $r(e_i+e_j)$ depends on whether $i=j$ or not: $$\forall 1\leq i\leq n+1, r(2e_i)=1.$$ $$\forall 1\leq i<j\leq n+1, r(e_i+e_j)=3.$$
We investigate in the next subsection some properties of the set $NC^{(k)}(n)$.
[**6.3 Study of the poset $NC^{(k)}(n)$**]{}
The set $NC^{(k)}(n)$, being a subset of $(NC^{(A)}((k+1)n),\preceq )$, inherits its partially ordered set (abbreviated poset) structure. Contrary to $NC^{(B)}(n)$, which is a sublattice of $(NC^{(A)}(2n),\preceq )$ (up to the identification $[\pm n]=[2n]$), $(NC^{(k)}(n),\preceq )$ is unfortunately not a sublattice of $(NC^{(A)}((k+1)n),\preceq )$, when $k\geq 2$.
\[cex\] When $k=2$ and $n=2$, consider the partitions $$\pi :=\{\{2,3,4,5\},\{1,6\}\}\in NC^{(2)}(2)$$ and $$\rho :=\{\{1,2\},\{3,4,5,6\}\}\in NC^{(2)}(2).$$ It is an easy exercise to determine the meet of these two partitions in the lattice $(NC^{(A)}(6),\preceq )$: $$\pi \wedge _{NC^{(A)}(6)} \rho =\{\{1\},\{2\},\{3,4,5\},\{6\}\}.$$ It is immediate that $\pi \wedge _{NC^{(A)}(6)} \rho $ is not an element of $NC^{(2)}(2)$ which is consequently not a sublattice of $(NC^{(A)}(6),\preceq )$ ; the same kind of argument would prove that $NC^{(k)}(n)$ is never a sublattice of $(NC^{(A)}((k+1)n),\preceq )$, as soon as $k\geq 2$.
It is natural to ask whether $NC^{(k)}(n)$ is or not a lattice in its own right for the reverse refinement order $\preceq $. We do not know the answer to this question.
We now state and prove the main result of this section.
\[cover\] $\pi \mapsto \mbox{Red}_n^{(k)}(\pi )$ is a $\frac{1}{(k+1)n+1}C_{(n+1)(k+1)}^{k+1}$-to-$1$ map from $NC^{(k)}(n)$ onto $NC^{(A)}(n)$.
We fix $p\in NC^{(A)}(n)$. The shape $\lambda _\pi $ of a $\pi \in NC^{(k)}(n)$ satisfying $\mbox{Red}_n^{(k)}(\pi )=p$ is an element of the set $\Lambda _{n+1,k}$, and for each $\lambda \in \Lambda _{n+1,k}$, there are exactly $r(\lambda )$ non-crossing partitions of type $k$ with reduction $p$ and shape $\lambda $. Hence there are $\sum_{\lambda \in \Lambda _{n+1,k}} r(\lambda )$ non-crossing partitions of type $k$ with reduction $p$, and we know by Proposition \[const\] that this number does not depend on $p$. It remains to prove that $\sum_{\lambda \in \Lambda _{n+1,k}} r(\lambda )=\frac{1}{(k+1)n+1}C_{(n+1)(k+1)}^{k+1}$, by counting the non-crossing partitions of type $k$ with reduction $1_{[n]}$. The set formed by these partitions is precisely the set $NC_n(k)$ of non-crossing partitions of $[(k+1)n]$ having blocks of size divisible by $n$. The latter set appears in [@bbcc], where it is proved that its cardinal is $\frac{1}{(k+1)n+1}C_{(n+1)(k+1)}^{k+1}$.
We end this section by defining a subset of $NC^{(k)}(n)$ that will be used in Section 7.
We write $NC_\ast ^{(k)}(n)$ for the set of non-crossing partitions of type $k$ without non-simple blocks in their Kreweras complement.
In the shape of a non-crossing partition $\pi \in NC_\ast ^{(k)}(n)$, the coordinates corresponding to blocks of $\mbox{Kr}(\pi )$ are zero ; there is therefore a straightforward bijection between the set of shapes of non-crossing partitions $\pi \in NC_\ast ^{(k)}(n)$ satisfying $\mbox{Red}_n^k(\pi )=p$ and the set $\Lambda _{|p|,k}$. Notice also that, given $p\in NC^{(A)}(n)$ and $\lambda \in \Lambda _{|p|,k}$, there are exactly $r(\lambda )$ non-crossing partitions $\pi \in NC_\ast ^{(k)}(n)$ with reduction $p$ and, with a small abuse of language, shape $\lambda $.
Non-crossing partitions of type $k$ give a combinatorial description of the version of the boxed convolution with scalars in $\mathcal{C}_k$, as explained in the next section.
[**7. Boxed convolution of type $k$**]{}
As for type A and B, there is a boxed convolution operation associated to the non-crossing partitions of type $k$. It is defined on formal power series with coefficients in ${\mathbb{C}}^{k+1}$, as follows.
1. We denote by $\Theta ^{(k)}$ the set of power series of the form $$f(z)=\sum_{n=1}^\infty (\alpha _n^{(0)},\ldots ,\alpha _n^{(k)})z^n,$$ where, for each $n\geq 1$ and $0\leq i\leq k$, $\alpha _n^{(i)}$ is a complex number.
2. Let $f(z)=\sum_{n=1}^\infty (\alpha _n^{(0)},\ldots ,\alpha _n^{(k)})z^n$ and $g(z)=\sum_{n=1}^\infty (\beta _n^{(0)},\ldots ,\beta _n^{(k)})z^n$ be in $\Theta ^{(k)}$. For every $m\geq 1$ and every $0\leq i\leq k$, consider the numbers $\gamma _m^{(i)}$ defined by $$\begin{aligned}
\nonumber \gamma _m^{(i)}=\sum_{\pi \in NC^{(i)}(m)} \ &
\frac{C_i^{(\lambda _\pi )_1,\ldots ,(\lambda _\pi )_{m+1}}}{r(\lambda _\pi )}
\prod_{j=1}^{ | \mbox{Red}_m^{(i)}(\pi ) | }
\alpha _{\mbox{card}(\mbox{Sep}(\mbox{Red}_m^{(i)}(\pi ),j))}^{((\lambda _\pi )_j)}\cdot
\\\nonumber \ &\prod_{j= | \mbox{Red}_m^{(i)}(\pi ) | +1}^{m+1}
\beta _{\mbox{card}(\mbox{Sep}(\mbox{Red}_m^{(i)}(\pi )),j)}^{((\lambda _\pi )_j)}.\end{aligned}$$ Then the series $\sum_{n=1}^\infty (\gamma _n^{(0)},\ldots ,\gamma _n^{(k)})z^n$ is called the boxed convolution of type $k$ of $f$ and $g$, and is denoted by $f$ $^{(k)} g$.
It turns out that, up to identifying the two sets $\Theta ^{(k)}$ and $\Theta _{\mathcal{C}_k}^{(A)}$, the two operations $^{(k)}$ and $_{\mathcal{C}_k}^{(A)}$ are actually the same, as stated in the next theorem.
\[starboxthm\] $^{(k)}=$$_{\mathcal{C}_k}^{(A)}$
Let $f(z)=\sum_{n=1}^\infty (\alpha _n^{(0)},\ldots ,\alpha _n^{(k)})z^n$ and $g(z)=\sum_{n=1}^\infty (\beta _n^{(0)},\ldots ,\beta _n^{(k)})z^n$ be in $\Theta ^{(k)}$.\
Write $f$ $^{(k)} g=\sum_{n=1}^\infty (\gamma _n^{(0)},\ldots ,\gamma _n^{(k)})z^n$ and $f$ $_{\mathcal{C}_k}^{(A)} g=\sum_{n=1}^\infty (\delta _n^{(0)},\ldots ,\delta _n^{(k)})z^n$. We fix a positive integer $n$, for which we will show that $$(\gamma _n^{(0)},\ldots ,\gamma _n^{(k)})=(\delta _n^{(0)},\ldots ,\delta _n^{(k)}).$$ Let us look at $\gamma _n^{(i)}$. First, we have $$\begin{aligned}
\nonumber \gamma _n^{(i)}=\sum_{\pi \in NC^{(i)}(n)} \ &
\frac{C_i^{(\lambda _\pi )_1,\ldots ,(\lambda _\pi )_{n+1}}}{r(\lambda _\pi )}
\prod_{j=1}^{ | \mbox{Red}_m^{(i)}(\pi ) | }
\alpha _{\mbox{card}(\mbox{Sep}(\mbox{Red}_m^{(i)}(\pi ),j))}^{((\lambda _\pi )_j)}\cdot
\\\nonumber \ &\prod_{j= | \mbox{Red}_m^{(i)}(\pi ) | +1}^{n+1}
\beta _{\mbox{card}(\mbox{Sep}(\mbox{Red}_m^{(i)}(\pi )),j)}^{((\lambda _\pi )_j)}.\end{aligned}$$ For every $\pi \in NC^{(i)}(n)$, $1\leq j\leq n+1$ and $0\leq \lambda \leq k$, we put $p=\mbox{Red}_n^{(i)}(\pi )$ and $$\theta ^{(\lambda )}(p,j):=\left\{\begin{array}{ccc}
\alpha _{\mbox{card}(\mbox{Sep}(p,j))}^{(\lambda )}&if&j\leq | p | ,
\\\beta _{\mbox{card}(\mbox{Sep}(p,j))}^{(\lambda )}&if&j> | p | ,\\\end{array}\right.$$ The summation over $NC^{(i)}(n)$ can be reduced to one over $NC^{(A)}(n)$, by using the cover $\mbox{Red}_n^{(i)} : NC^{(i)}(n)\rightarrow NC^{(A)}(n)$. When doing so, and taking into account the explicit description of $(\mbox{Red}_n^{(i)})^{-1}(p), p\in NC^{(A)}(n)$ provided by the proof of Theorem \[cover\], one gets $$\gamma _n^{(i)}=\sum_{p\in NC^{(A)}(n)}
\sum_{\lambda \in \Lambda _{n+1,i}} C_i^{\lambda _1,\ldots ,\lambda _{n+1}} \prod_{j=1}^{n+1}\theta ^{(\lambda _j)}(p,j).$$ On the other hand, by recalling the definition of the operation $_{\mathcal{C}_k}^{(A)}$, we see that $\delta _n^{(i)}$ equals $$\sum_{p\in NC^{(A)}(n)} \sum_{\lambda \in \Lambda _{n+1,i}}
C_i^{\lambda _1,\ldots ,\lambda _{n+1}} \prod_{j=1}^{n+1}\theta ^{(\lambda _j)}(p,j).$$ By comparing, we obtain $(\gamma _n^{(0)},\ldots ,\gamma _n^{(k)})=(\delta _n^{(0)},\ldots ,\delta _n^{(k)})$, as desired.
The operation $^{(k)}$ is associative, commutative and the series $\Delta ^{(k)}(z)=\Delta _{\mathcal{C}_k}^{(A)}(z)$ is its unit element. A series $f\in \Theta ^{(k)}$ is invertible with respect to $^{(k)}$ if and only if its coefficient of degree one has a non-zero first component.
Theorem \[starboxthm\] tells us that the operation $^{(k)}$ is a boxed convolution of type A, for which it is noticed in Remark \[multi\] that one may define a generalization to power series in several noncommuting indeterminates. This means that there exists an operation $^{(k)}$ on power series in several noncommuting indeterminates. We do not find interesting to record here the formulas involved in this operation.
Non-crossing partitions of type $k$ are thus the combinatorial objects describing the version of the boxed convolution of type A with scalars in the algebra $\mathcal{C}_k$.
It is now easy to rewrite the main formulas involving infinitesimal non-crossing cumulants with sums indexed by the set of non-crossing partitions of type $k$. This is the content of the next proposition :
Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$. The infinitesimal non-crossing cumulant functionals satisfy, for every $n\geq 1$, every $0\leq i\leq k$ and every $a_1,\ldots ,a_n\in \mathcal{A}$ : $$\varphi ^{(i)}(a_1\cdots a_n) = \sum_{\pi \in NC_\ast ^{(i)}(n)}
\frac{C_i^{(\lambda _\pi )_1,\ldots ,(\lambda _\pi )_{|\pi |}}}{r(\lambda _\pi )}
\kappa _{\mbox{Red}_n^{(i)}(\pi )}^{(\lambda _\pi )}(a_1,\ldots ,a_n).$$
Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$. Consider subsets $\mathcal{M}_1,\mathcal{M}_2$ of $\mathcal{A}$ that are infinitesimally free of order $k$. Then, one has, for each $n\geq 1$, each $n$-tuples $(a_1,\ldots ,a_n)\in \mathcal{M}_1^n,(b_1,\ldots ,b_n)\in \mathcal{M}_2^n$ and each $0\leq i\leq k$ : $$\kappa _n^{(i)}(a_1\cdot b_1,\ldots ,a_n\cdot b_n)=\sum_{\pi \in NC^{(i)}(n)}
\frac{C_i^{(\lambda _\pi )_1,\ldots ,(\lambda _\pi )_{n+1}}}{r(\lambda _\pi )}
\kappa _{\mbox{Red}_n^{(i)}(\pi \cup \mbox{Kr}(\pi ))}^{(\lambda _\pi )}(a_1,b_1,\ldots ,a_n,b_n).$$
We move to the main application of infinitesimal freeness.
[**8. Application to derivatives of the free convolution**]{}
In this final section, we give an application of infinitesimal freeness of order $k$. We consider the situation already examined in [@bs] : let $\{a_u^v(t)\mid 1\leq v\leq m_u\}_{t\in K}$ be $s$ families of noncommutative random variables in a (usual) noncommutative probability space $(\mathcal{A},\varphi )$. These families are indexed by a subset $K$ of ${\mathbb{R}}$ having zero as an accumulation point, and we are interested in the joint distribution $\mu _t$ of $\{a_u^v(t)\mid 1\leq v\leq m_u, 1\leq u\leq s\}$ when $t$ is going to $0$, in other words for infinitesimal values of $t$. Recall that $\mu _t$ is the linear functional on ${\mathbb{C}}\langle X_u^v , 1\leq v\leq m_u , 1\leq u\leq s \rangle$ defined by : $$\mu _t(P((X_u^v)_{1\leq v\leq m_u , 1\leq u\leq s}))=
\varphi (P((a_u^v(t))_{1\leq v\leq m_u , 1\leq u\leq s})).$$ In what follows, we will consider a family $\{\mu _t\}_{t\in K}$ of linear functionals on ${\mathbb{C}}\langle X_u^v , 1\leq v\leq m_u , 1\leq u\leq s \rangle$ without any further reference to the variables $\{a_u^v(t)\mid 1\leq v\leq m_u, 1\leq u\leq s\}_{t\in K}$. For each value of $t\in K$, one may obviously define the non-crossing cumulant functionals $((\kappa _t)_n:
({\mathbb{C}}\langle X_u^v , 1\leq v\leq m_u , 1\leq u\leq s \rangle )^n \rightarrow {\mathbb{C}})_{n=1}^\infty $ associated to the noncommutative probability space $({\mathbb{C}}\langle X_u^v , 1\leq v\leq m_u , 1\leq u\leq s \rangle , \mu _t)$. A way to capture the behavior of $\mu _t$ for infinitesimal values of $t$ is to introduce recursively its derivatives at $0$ by : $$\label{zeroderiv}
\mu ^{(0)}:=\lim_{t\rightarrow 0} \mu _t,$$ $$\label{ithderiv}
\frac{\mu ^{(i)}}{i!}:=\lim_{t\rightarrow 0} \frac{1}{t^i}(\mu _t-\sum_{j=0}^{i-1} \frac{t^j}{j!}\mu ^{(j)}),1\leq i\leq k.$$ We will assume that the limits in formulas and exist and use the notation $\mu ^{(i)}=\frac{d^i}{dt^i}_{|t=0}\mu _t$. Notice that, in [@bs], only $\mu ^{(0)}$ and $\frac{d}{dt}_{|t=0}\mu _t$ were studied. It follows from formulas and that $$\mu _t=\sum_{i=0}^k \frac{\mu ^{(i)}}{i!} t^i+o(t^k).$$ Notice that $(\mu ^{(i)})_{0\leq i\leq k}$ is an infinitesimal law (of order $k$) on $\sum_{u=1}^s m_u$ variables and therefore $({\mathbb{C}}\langle X_u^v , 1\leq v\leq m_u , 1\leq u\leq s \rangle , (\mu ^{(i)})_{0\leq i\leq k})$ is an infinitesimal noncommutative probability space of order $k$. Associated to this infinitesimal noncommutative probability space of order $k$, we have infinitesimal non-crossing cumulant functionals $(\kappa _n^{(i)}: \mathcal{A}^n \rightarrow {\mathbb{C}}, 0\leq i\leq k)_{n=1}^\infty $, as defined by formula . These infinitesimal cumulant functionals are linked to $((\kappa _t)_n)_{n=1}^\infty $ as follows :
\[propcumtime\] For every $n\geq 1$ and $0\leq i\leq k$, $$\kappa _n^{(i)}=\frac{d^i}{dt^i}_{|t=0}(\kappa _t)_n.$$
By the inverse of the free moment-cumulant formula, one has $$\label{timecum}
\forall t\in K, (\kappa _t)_n=\sum_{p\in NC^{(A)}(n)} \mbox{M\"ob}(p,1_n) (\mu _t)_p.$$ By the assumption made above, the right-hand side of formula has $k$ derivatives at $0$, hence $\frac{d^i}{dt^i}_{|t=0}(\kappa _t)_n$ is well-defined and, using linearity of derivation and Leibniz rule, one obtains : $$\frac{d^i}{dt^i}_{|t=0}(\kappa _t)_n=\sum_{\substack{p\in NC^{(A)}(n)\\p:=\{V_1,\ldots ,V_h \} }}
\sum_{\lambda \in \Lambda _{h,i}} \mbox{M\"ob}(p,1_n) C_i^{\lambda _1,\ldots ,\lambda _h} \mu _{p}^{(\lambda )}.$$ One recognizes in the right-hand side above the right-hand side of formula , and we are done.
This proposition will be the main tool to characterize infinitesimal freeness of order $k$ in terms of moments in Theorem \[inffmomentsthm\]. We first give a recipe to deduce the infinitesimal behaviour of the free convolution of two families of distributions from their individual infinitesimal behaviours.
Let $\{\mu _t\}_{t\in K}$ (resp. $\{\nu _t\}_{t\in K}$) be a family of linear functionals on ${\mathbb{C}}\langle X_u , 1\leq u\leq m \rangle$ (resp ${\mathbb{C}}\langle Y_u , 1\leq u\leq m \rangle$) such that $\mu ^{(i)}=\frac{d^i}{dt^i}_{|t=0}\mu _t$ (resp. $\nu ^{(i)}=\frac{d^i}{dt^i}_{|t=0}\nu _t$) exist for $0\leq i\leq k$. Set : $$(\eta ^{(i)})_{0\leq i\leq k}:=(\mu ^{(i)})_{0\leq i\leq k} \boxplus ^{(k)}(\nu ^{(i)})_{0\leq i\leq k},$$ $$(\theta ^{(i)})_{0\leq i\leq k}:=(\mu ^{(i)})_{0\leq i\leq k} \boxtimes ^{(k)}(\nu ^{(i)})_{0\leq i\leq k}.$$ Then $\eta ^{(i)}=\frac {d^i}{dt^i}\mid _{t=0} \mu _t \boxplus \nu _t$ and $\theta ^{(i)}=\frac {d^i}{dt^i}\mid _{t=0} \mu _t \boxtimes \nu _t$.
For each $t\in K$, we consider the free product $$({\mathbb{C}}\langle X_u, Y_u , 1\leq u\leq m \rangle,\mu _t\star \nu _t).$$ Since $\frac{d^i}{dt^i}_{|t=0}\mu _t$ and $\frac{d^i}{dt^i}_{|t=0}\nu _t$ exist by assumption for each $0\leq i\leq k$, we obtain the existence of $\frac{d^i}{dt^i}_{|t=0}(\mu _t\star \nu _t)$ for each $0\leq i\leq k$ and these functionals are completely determined by the $\mu ^{(i)}$’s and the $\nu ^{(i)}$’s. In the infinitesimal noncommutative probability space $({\mathbb{C}}\langle X_u, Y_u , 1\leq u\leq m \rangle ,
(\frac{d^i}{dt^i}_{|t=0}(\mu _t\star \nu _t))_{0\leq i\leq k})$, the unital subalgebras $\mathcal{A}_1={\mathbb{C}}\langle X_u , 1\leq u\leq m \rangle$ and $\mathcal{A}_2={\mathbb{C}}\langle Y_u , 1\leq u\leq m \rangle$ are infinitesimally free of order $k$ : indeed, if $n\geq 1 , 0\leq i\leq k$ and $P_1\in \mathcal{A}_{i_1},\ldots ,P_n\in \mathcal{A}_{i_n}$ are such that $i_1,\ldots ,i_n$ are not all equal, then $$\kappa _n^{(i)}(P_1,\ldots ,P_l)=\frac{d^i}{dt^i}_{|t=0}(\kappa _t)_n(P_1,\ldots ,P_l),$$ where $(\kappa _t)_n$ is the $n$-th non-crossing cumulant functional in the noncommutative probability space $({\mathbb{C}}\langle X_u, Y_u , 1\leq u\leq m \rangle,\mu _t\star \nu _t)$, by Proposition \[propcumtime\]. But it follows from the construction of the free product that $(\kappa _t)_n(P_1,\ldots ,P_l)=0$ for each $t\in K$. In particular $\kappa _n^{(i)}(P_1,\ldots ,P_l)=0$. The infinitesimal distribution of the $m$-tuple $(X_1+Y_1,\ldots ,X_m+Y_m)$ (resp. $(X_1\cdot Y_1,\ldots ,X_m\cdot Y_m)$) is, on the one hand $(\frac{d^i}{dt^i}_{|t=0}(\mu _t\boxplus \nu _t))_{0\leq i\leq k}$ (resp. $(\frac{d^i}{dt^i}_{|t=0}(\mu _t\boxtimes \nu _t))_{0\leq i\leq k}$) by construction of the free product and, on the other hand, $(\eta ^{(i)})_{0\leq i\leq k}$ (resp. $(\theta ^{(i)})_{0\leq i\leq k}$) by the argument above.
We conclude by a characterization of infinitesimal freeness of order $k$ in terms of moments. Its formulation and proof rely on the Proposition \[propcumtime\].
\[inffmomentsthm\] Let $(\mathcal{A}, (\varphi ^{(i)})_{0\leq i\leq k})$ be an infinitesimal noncommutative probability space of order $k$, and $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ be unital subalgebras of $\mathcal{A}$. Then $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ are infinitesimally free of order $k$ if and only if for any positive integer $l\in {\mathbb{N}}^*$, and any $a_1\in \mathcal{A}_{i_1},\ldots ,
a_l\in \mathcal{A}_{i_l}$, one has $$\label{inffmomentstime}
\varphi _t((a_1-\varphi _t(a_1))\cdots (a_l-\varphi _t(a_l)))=o(t^k),$$ whenever $i_1\not=\ldots \not=i_l$, where $\varphi _t:=\sum_{i=0}^k \frac{\varphi ^{(i)}}{i!} t^i$. The condition translates into $k+1$ requirements : $$\label{inffmoments}
\forall i\in \{0,\ldots ,k\}, \sum_{j=0}^i \sum_{\lambda \in \Lambda _{l,i-j}} (-1)^{\#\{m, \lambda _m>0\}}
\mu ^{(j)}(\hat \mu ^{(\lambda _1)}(P_1)\cdots \hat \mu ^{(\lambda _{l})}(P_l))=0,$$ where $\hat \mu ^{(\lambda )}(P):=P-\mu ^{(0)}(P)$ if $\lambda =0$, and $\hat \mu ^{(\lambda )}(P):=\mu ^{(\lambda )}(P)$ else.\
We assume that condition holds and have to prove that $\mathcal{A}_1,\ldots ,\mathcal{A}_n$ satisfy the vanishing of mixed infinitesimal cumulants condition. Using Proposition \[propcumtime\], it is equivalent to prove that for $l\geq 2$, and $a_1\in \mathcal{A}_{i_1},\ldots ,
a_l\in \mathcal{A}_{i_l}$ $$\label{vmcumtime}
(\kappa _t)_l(a_1,\ldots ,a_l)=o(t^k)$$ whenever $\exists r\not=s, i_r\not=i_s$, where $(\kappa _t)_l$ is the $l$-th non-crossing cumulant functional in $(\mathcal{A}, \varphi _t)$.\
We proceed by induction on $l\geq 2$.\
It is easy to see that $$\label{fmc2}
(\kappa _t)_2(a_1,a_2)=\varphi _t((a_1-\varphi _t(a_1))((a_2-\varphi _t(a_2)).$$ If $a_1\in \mathcal{A}_{i_1},
a_2\in \mathcal{A}_{i_2}$ with $i_1\not=i_2$, the right-hand side of is $o(t^k)$ by assumption. We assume then that the vanishing of mixed infinitesimal cumulants is proved for $2,3,\ldots ,l-1$ variables, and consider $(\kappa _t)_l(a_1,\ldots ,a_l)$ with $a_1\in \mathcal{A}_{i_1},\ldots ,
a_l\in \mathcal{A}_{i_l}$ such that $\exists r\not=s, i_r\not=i_s$. By Propositions \[cwsae\], \[cwpae\] and the induction hypothesis, we may assume that $\varphi _t(a_1)=\ldots =\varphi _t(a_l)=0$ and $i_1\not=\ldots \not=i_l$. Write then the free moment-cumulant formula : $$\forall t\in K, (\varphi _t)(a_1\cdots a_l)-\sum_{\substack{p\in NC^{(A)}(l)\\p\not=1_l}}
(\kappa _t)_p(a_1,\ldots ,a_l) = (\kappa _t)_l(a_1,\ldots ,a_l).$$ By assumption, $(\varphi _t)(a_1\cdots a_l)=o(t^k)$. Any non-crossing partition $p\not =1_l$ owns an interval-block $V_0$, as noticed in Section 3. If $V_0$ is a singleton, $$(\kappa _t)_p(a_1,\ldots ,a_l) =
(\varphi _t)_{|V_0|}((a_1,\ldots ,a_l) \mid V_0) \prod_{V\not =V_0} (\kappa _t)_{|V|}((a_1,\ldots ,a_l) \mid V)=0.$$ Otherwise, $V_0$ contains two following, hence distinct, indices, and, by induction hypothesis, $$(\kappa _t)_{|V_0|}((a_1,\ldots ,a_l) \mid V_0)=o(t^k).$$ Since, for each $V\in \mbox{bl}(p)$, $(\kappa _t)_{|V|}((a_1,\ldots ,a_l) \mid V)$ is bounded in a neighborhood of $0$, one may affirm that $$(\kappa _t)_p(a_1,\ldots ,a_l) = o(t^k).$$ We conclude that $$(\kappa _t)_l(a_1,\ldots ,a_l)=o(t^k),$$ as required.\
For the converse, we assume that the vanishing of mixed infinitesimal cumulants is satisfied, or equivalently that equation holds. We write then the free moment-cumulant formula : $$\begin{aligned}
\label{fmcumtimebis}
\forall t\in K, (\varphi _t)(a_1-\varphi _t(a_1)\cdots a_l-\varphi _t(a_l)) =
\\\label{fmcumtimeter}
\sum_{p\in NC^{(A)}(l)} (\kappa _t)_p(a_1-\varphi _t(a_1),\ldots ,a_l-\varphi _t(a_l)).\end{aligned}$$ If $a_1\in \mathcal{A}_{i_1},\ldots ,
a_l\in \mathcal{A}_{i_l}$ with $i_1\not=\ldots \not=i_n$, the same argument as above gives that is $o(t^k)$. This concludes the proof.
[**Acknowledgements.**]{} The author would like to express all his gratitude to Serban Teodor Belinschi for suggesting this problem, and for many useful discussions. This work was initiated during a stay at University of Saskatchewan, that the author thanks for the warm welcome and the excellent work conditions provided. This work was partially supported by the [*A*gence Nationale de la Recherche]{} grant ANR-08-BLAN-0311-03.
[9]{}
T. [<span style="font-variant:small-caps;">Banica</span>]{}, S.T. [<span style="font-variant:small-caps;">Belinschi</span>]{}, M. [<span style="font-variant:small-caps;">Capitaine</span>]{}, B. [<span style="font-variant:small-caps;">Collins</span>]{}. Free bessel laws, preprint 2007. Available online at www.arxiv.org under reference arXiv:0710.5931.
S.T. [<span style="font-variant:small-caps;">Belinschi</span>]{}, D. [<span style="font-variant:small-caps;">Shlyakhtenko</span>]{}. Free probability of type B: analytic aspects and applications, preprint 2009. Available online at www.arxiv.org under reference arXiv:0903.2721.
P. [<span style="font-variant:small-caps;">Biane</span>]{}, F. [<span style="font-variant:small-caps;">Goodman</span>]{}, A. [<span style="font-variant:small-caps;">Nica</span>]{}. Non-crossing cumulants of type B, *Trans. Amer. Math. Soc* **355**, 2263-2303 (2003).
P. [<span style="font-variant:small-caps;">Biane</span>]{}. Some properties of crossings and partitions, *Discret. Math.* **175**, 41-53 (1997).
G. [<span style="font-variant:small-caps;">Birkhoff</span>]{}. *Lattice theory*, AMS (1979).
M. [<span style="font-variant:small-caps;">Février</span>]{}, A. [<span style="font-variant:small-caps;">Nica</span>]{}. Infinitesimal non-crossing cumulants and free probability of type B, *J. Funct. Anal.* **258**, 2983-2023 (2010).
F. [<span style="font-variant:small-caps;">Hiai</span>]{}, D. [<span style="font-variant:small-caps;">Petz</span>]{}. *The semicircle law, free random variables and entropy*, AMS (2000).
G. [<span style="font-variant:small-caps;">Kreweras</span>]{}. Sur les partitions non croises d’un cycle, *Discrete Mathematics* **1**, 333-350 (1972).
A. [<span style="font-variant:small-caps;">Nica</span>]{}, R. [<span style="font-variant:small-caps;">Speicher</span>]{}. On the multiplication of free $n$-tuples of noncommutative random variables, *Amer. J. Math.* **118**, 799-837 (1996).
A. [<span style="font-variant:small-caps;">Nica</span>]{}, R. [<span style="font-variant:small-caps;">Speicher</span>]{}. A “Fourier transform” for multiplicative functions on non-crossing partitions, *J. Alg. Comb.* **6**, 141-160 (1997).
A. [<span style="font-variant:small-caps;">Nica</span>]{}, R. [<span style="font-variant:small-caps;">Speicher</span>]{}. *Lectures on the combinatorics of free probability*, Cambridge University Press (2006).
I. [<span style="font-variant:small-caps;">Oancea</span>]{}. Posets of non-crossing partitions of type B and applications, [*University of Waterloo PhD Thesis*]{}, October 2007, available online at uwspace.uwaterloo.ca/bitstream/10012/3402/1/BDTEZA.pdf.
M. [<span style="font-variant:small-caps;">Popa</span>]{}. Freeness with amalgamation, limit theorems and S-transform in noncommutative probability spaces of type B, preprint, 2007. Available online at www.arxiv.org under reference arXiv:0709.0011.
V. [<span style="font-variant:small-caps;">Reiner</span>]{}. Non-crossing partitions for classical reflection groups, *Discrete Mathematics* **177**, 195-222 (1997).
R. [<span style="font-variant:small-caps;">Speicher</span>]{}. Multiplicative functions on the lattice of non-crossing partitions and free convolution, *Math. Annalen* **298**, 611-628 (1994).
D.V. [<span style="font-variant:small-caps;">Voiculescu</span>]{}, K.J. [<span style="font-variant:small-caps;">Dykema</span>]{}, A. [<span style="font-variant:small-caps;">Nica</span>]{}. *Free random variables*, AMS (1992).
D.V. [<span style="font-variant:small-caps;">Voiculescu</span>]{}. Multiplication of certain non-commuting random variables, *J. Operator Theory* **18**, 223-235 (1987).
D.V. [<span style="font-variant:small-caps;">Voiculescu</span>]{}. Limit laws for random matrices and free products, *Inventiones Math.* **104**, 202-220 (1991).
[^1]: Institut de Mathématiques de Toulouse, Equipe de Statistique et Probabilités, F-31062 Toulouse Cedex 09. E-mail: [email protected]
[^2]: This work was partially supported by the [*A*gence Nationale de la Recherche]{} grant ANR-08-BLAN-0311-03.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Gamow states describe the quasinormal modes of quantum systems. It is shown that the resonance amplitude associated with the Gamow states is given by the complex delta function. It is also shown that under the near-resonance approximation of neglecting the lower bound of the energy, such resonance amplitude becomes the Breit-Wigner amplitude. This result establishes the precise connection between the Gamow states, Nakanishi’s complex delta function and the Breit-Wigner amplitude. In addition, this result provides another theoretical basis for the phenomenological fact that the almost-Lorentzian peaks in cross sections are produced by intermediate, unstable particles.'
author:
- |
Rafael de la Madrid\
\
title: '**The resonance amplitude associated with the Gamow states**'
---
\#1 \#1 \#1\#2
[*Keywords*]{}: Gamow states; resonances; rigged Hilbert space; complex delta function; Breit-Wigner lineshape
PACS: 03.65.-w, 03.65.Bz; 03.65.Ca; 03.65.Db
Introduction
============
\[sec:Intro\]
Resonances appear in all areas of quantum physics, in both the relativistic and non-relativistic regimes. Resonances are intrinsic properties of a quantum system, and they describe the system’s preferred ways of decaying. Experimentally, resonances appear as sharp peaks in the cross section that resemble the Breit-Wigner (Lorentzian) lineshape.
The Gamow states are the natural wave functions of resonances, and they were introduced by Gamow in his paper on $\alpha$-decay of radioactive nuclei [@GAMOW]. Since then, they have been used by a number of authors, see e.g. [@SIEGERT; @PEIERLS; @HUMBLET; @ZELDOVICH; @BERGGREN; @GASTON; @BERGGREN78; @SUDARSHAN; @MONDRAGON83; @CURUTCHET; @BL; @LIND; @BERGGREN96; @BOLLINI; @FERREIRA; @BETAN; @MICHEL1; @AJP02; @KAPUSCIK1; @MONDRAGON03; @MICHEL2; @KAPUSCIK2; @MICHEL3; @MICHEL4; @MICHEL5; @URRIES; @MICHEL6; @TOMIO]. Likewise the bound states, the Gamow states are properties of the Hamiltonian, and they are associated with the natural frequencies of the system. The usefulness of the Gamow states is attested by the remarkable success of the Gamow Shell Model [@MICHEL1; @MICHEL2; @MICHEL3; @MICHEL4; @MICHEL5; @MICHEL6] and similar nuclear-structure formalisms [@FERREIRA; @BETAN].
Because resonances leave a quasi-Lorentzian fingerprint in the cross section, and because the Gamow states are the natural wave functions of resonances, the resonance amplitude associated with the Gamow states must be related to the Breit-Wigner amplitude. The purpose of this paper is to show that the resonance amplitude associated with the Gamow states is proportional to the complex delta function, $\delta (E-z_{\rm R})$, and that such amplitude can be approximated in the near-resonance region by the Breit-Wigner amplitude. More precisely, we will show that the transition amplitude from a resonance state of energy $z_{\rm R}$ to a scattering state of energy $E\geq 0$, ${\cal A}(z_{\rm R} \to E)$, is given by $${\cal A}(z_{\rm R} \to E)= {{\rm i}}\sqrt{2\pi}{\cal N}_{\rm R}
\delta (E-z_{\rm R})
\simeq -\frac{{\cal N}_{\rm R}}{\sqrt{2\pi}} \frac{1}{E-z_{\rm R}} \, ,
\quad E\geq 0 \, ,
\label{aeqconsBW}$$ where $${\cal N}_{\rm R}^2 \equiv {{\rm i}}\, {\rm res}[S(E)]_{E=z_{\rm R}} \equiv
{{\rm i}}\, {\rm r}_{\rm R} \, ,
\label{normafac}$$ $S$ denotes the $S$ matrix, and ${\rm r}_{\rm R}$ denotes the residue of $S$ at $z_{\rm R}$. In addition, we will see that the lower bound of the energy (threshold) is the reason why this amplitude is not exactly but only approximately given by the Breit-Wigner amplitude.
Section \[sec:basics\] provides a quick summary of the most important properties of the Gamow states, along with some basic phenomenological properties of resonances. The proof of (\[aeqconsBW\]) is provided in Sec. \[sec:proof\]. The conclusions are included in Sec. \[sec:conclusions\].
For the sake of clarity, we shall prove Eq. (\[aeqconsBW\]) using the example of the spherical shell potential for zero angular momentum. However, as explained in Appendix \[sec:appendix-gener\], the result is valid for any partial wave and for spherically symmetric potentials that fall off faster than exponentials. Finally, in Appendix \[sec:appendix-cdf\], we provide a thorough characterization of the complex delta function and its associated functional, since they have rarely appeared in the literature.
Basics of resonances and Gamow states
=====================================
\[sec:basics\]
Resonance peaks are characterized by the energy $E_{\rm R}$ at which they occur and by their width $\Gamma _{\rm R}$. The resonance peak is related to a pole of the $S$ matrix at the complex number $z_{\rm R}= E_{\rm R} - {{\rm i}}\, \Gamma _{\rm R}/2$, because the theoretical expression of the cross section in terms of the $S$ matrix fits the experimental cross section in the neighborhood of $E_{\rm R}$, see Eqs. (\[Smex\]) and (\[crossS\]) below.
When the peak is too narrow and its width cannot be measured, one measures the lifetime $\tau _{\rm R}$ of the decaying particle. Decaying systems follow the exponential decay law, except for short- and long-term deviations.
Although a decaying particle has a finite lifetime, it is otherwise assigned all the properties that are attributed to stable particles, like angular momentum, charge, spin and parity. For example, a radioactive nucleus has a finite lifetime, but otherwise it possesses all the properties of stable nuclei; in fact, it is included in the periodic table of the elements along with the stable nuclei. Similarly, most elementary particles are unstable, and they are listed along with the stable ones in the Particle Data Table [@PDT] and attributed values for the mass, spin and width (or lifetime). Thus, stable particles differ from unstable ones by the value of their width, which is zero in the case of stable particles and different from zero in the case of unstable ones. Hence, phenomenologically, unstable particles are not less fundamental than the stable ones.
A priori, resonances and decaying particles are different entities. A resonance refers to the energy distribution of the outgoing particles in a scattering process, and it is characterized by its energy and width. A decaying state is described in a time-dependent setting by its energy and lifetime. Yet the difference is quantitative rather than qualitative, and both concepts are related by $$\Gamma _{\rm R}=\frac{\hbar}{\tau _{\rm R}} \, ,
\label{lifewidre}$$ though in most systems one can measure either $\tau _{\rm R}$ or $\Gamma _{\rm R}$, but not both.
Theoretically, however, the relation (\[lifewidre\]) is usually justified as an approximation, $\tau _{\rm R}\Gamma _{\rm R} \sim \hbar$, as a kind of time-energy uncertainty relation. For a long time, it was not possible to experimentally check whether the relation (\[lifewidre\]) is exact or approximate, since the lifetime and width could not be measured in the same system. This changed with the measurements of the width [@OATES] and lifetime [@VOLZ] of the $3p\ ^2P_{3/2}$ state of Na, which provide a firm experimental basis that Eq. (\[lifewidre\]) holds exactly, not just approximately. Thus, resonances and decaying systems are two sides of the same phenomenon.
Although the resonance peaks in the cross section resemble the Lorentzian, the resonance lineshape does not coincide exactly with the Lorentzian. Two features of the cross section reveal so. First, the maximum of the resonance peak never occurs at $E=E_{\rm R}$, whereas the maximum of the Lorentzian occurs exactly at $E=E_{\rm R}$. And second, the Laurent expansion of the $S$ matrix around the resonance pole, $$S(E)=\frac{{\rm r}_{\rm R}}{E-z_{\rm R}}+B(E) \, ,
\label{Smex}$$ which produces the Lorentzian peak in the cross section [@NOTE1], $$\sigma \sim \frac{1}{(E-E_{\rm R})^2+(\Gamma _{\rm R}/2)^2} \, ,
\label{crossS}$$ is valid only in the vicinity of the resonance pole. Because (\[Smex\]) and (\[crossS\]) are valid only in the vicinity of the resonance energy, the Lorentzian lineshape is just a near-resonance approximation to the exact resonance lineshape.
Because the Lorentzian does not coincide exactly with the resonance lineshape, the Breit-Wigner amplitude cannot coincide exactly with the resonance amplitude. One can reach the same conclusion by using the point of view of decaying states as follows. The Breit-Wigner amplitude yields the exponential decay law only when it is defined over the whole of the energy real line $(-\infty , \infty )$ rather than just over the scattering spectrum (see e.g. [@FONDA]). Because in quantum mechanics the scattering spectrum has a lower bound, the Breit-Wigner amplitude would yield the exponential decay law only if it was defined also at energies that do not belong to the scattering spectrum. Thus, the Breit-Wigner amplitude is incompatible with the exponential decay law, and therefore cannot coincide with the exact resonance/decay amplitude.
Mathematically, the Gamow states are eigenvectors of the Hamiltonian with a complex eigenvalue $z_{\rm R}=E_{\rm R} -{{\rm i}}\, \Gamma _{\rm R}/2$, $$H|z_{\rm R}\rangle = z_{\rm R} |z_{\rm R}\rangle \, ,
\label{tiSeq}$$ and, in the radial position representation, they satisfy a “purely outgoing boundary condition” (POBC) at infinity: $$\langle r|z_{\rm R}\rangle \sim
{{\rm e}}^{{\rm i}\sqrt{(2m/\hbar ^2)z_{\rm R}} \, r}
\, , \quad \mbox{as} \ r \to \infty \, .
\label{POBC}$$ The time-independent Schrödinger equation (\[tiSeq\]) subject to the POBC (\[POBC\]) is equivalent to the following integral equation of the Lippmann-Schwinger type: $$|z_{\rm R}\rangle =\frac{1}{z_{\rm R}-H_0+{{\rm i}}\hskip0.2mm 0}
V |z_{\rm R}\rangle \, ,
\label{inteGam}$$ where $H_0$ is the free Hamiltonian and $V$ is the potential. Since Eq. (\[inteGam\]) also yields the bound states, the Gamow states are a natural generalization to resonances of the wave functions of bound states. The bound and resonance energies obtained by solving (\[inteGam\]) coincide with the poles of the $S$ matrix.
The time evolution of a Gamow state is given by $${{\rm e}}^{-{{\rm i}}Ht/\hbar}|z_{\rm R}\rangle =
{{\rm e}}^{-{{\rm i}}E_{\rm R}t/\hbar} {{\rm e}}^{-\Gamma _{\rm R}t/(2\hbar)}
|z_{\rm R}\rangle \, ,
\label{tdSeq}$$ and therefore the Gamow states abide by the exponential decay law. Because the eigenvalue of Eq. (\[tiSeq\]) is also a pole of the $S$ matrix, Eq. (\[tdSeq\]) implies that Eq. (\[lifewidre\]) holds. In this way, the Gamow states unify the concepts of resonance and decaying particle, and they provide a “particle status” for them.
Furthermore, since one can obtain both the bound and the resonance energies from Eq. (\[inteGam\]), or from the poles of the $S$ matrix, resonances are qualitatively the same as bound states. The only difference is quantitative: The Gamow states have a non-zero width (i.e., finite lifetime), whereas the bound states have a zero width (i.e., infinite lifetime).
An important feature of the Gamow states is that they form a basis that expands any wave packet $\varphi ^+$, see e.g. review [@05CJP]. The basis formed by the Gamow states is not complete though, and one has to add an additional set of kets to complete the basis. In a system with several resonances, we have that $$\varphi ^+(t)=\sum_n {{\rm e}}^{-{{\rm i}}z_nt/\hbar}
|z_n\rangle \langle z_n|\varphi ^+\rangle
+\int_0^{-\infty}{{\rm d}}E \, {{\rm e}}^{-{{\rm i}}Et/\hbar}
|E^+\rangle \langle ^+E|\varphi ^+ \rangle \, ,
\label{resoexpan}$$ where $z_n=E_n-{{\rm i}}\, \Gamma _n/2$ denotes the $n$th resonance energy. In this equation, the sum contains the resonance contribution, whereas the integral contains the background. For simplicity, we have omitted the contribution from the bound states. The main virtue of resonance expansions is to isolate each resonance’s contribution to the wave packet.
Resonance expansions allow us to understand the deviations from exponential decay [@RAIZEN]. In the energy region where one resonance ${\rm R}$ is dominant, the expansion (\[resoexpan\]) can be written as $$\varphi ^+(t)= {{\rm e}}^{-{{\rm i}}z_{\rm R}t/\hbar}
|z_{\rm R}\rangle \langle z_{\rm R}|\varphi ^+\rangle +
{\rm background(R)} \, ,
\label{resoexpanR}$$ where the term “background(R)” contains all contributions not associated with the resonance ${\rm R}$, including those from other resonances. Because “background(R)” will always be nonzero, there will always be deviations from exponential decay. The magnitude of these deviations depends on how well we tune the system around the resonance energy: The better we tune the system around the Gamow state $|z_{\rm R}\rangle$, the smaller “background(R)” will be. Note that “background(R)” is the analog to the background $B(E)$ of the expansion (\[Smex\]).
Proof
=====
\[sec:proof\]
Preliminaries {#sec:prelmi}
-------------
The proof of (\[aeqconsBW\]) presented below is a straightforward application of the theory of distributions. Rather than working in a general setting, we will use the example of the spherical shell potential, $$V({\bf x})= V(r)=\left\{ \begin{array}{ll}
0 &0<r<a \\
V_0 &a<r<b \\
0 &b<r<\infty \, ,
\end{array}
\right.
\label{potential}$$ and restrict ourselves to the s partial wave.
In order to prove (\[aeqconsBW\]), we need to recall that in quantum mechanics, the transition amplitude from one state to another is given by the scalar product of those states: $${\cal A}(z_{\rm R} \to E)= \langle ^-E|z_{\rm R}\rangle \, ,\quad
E\geq 0 \, ,
\label{tramaid}$$ where $\langle ^-E|$ is the “out” bra solution of the Lippmann-Schwinger equation.
For the potential (\[potential\]), the “out” Lippmann-Schwinger eigenfunction reads, in the radial position representation, as $$\langle r|E^{-}\rangle \equiv
\chi ^{-}(r;E)= N(E) \,
\frac{\chi (r;E)}{{\cal J}_{-}(E)}\, , \qquad E\in [0,\infty ) \, ,
\label{pmeigndu}$$ where $N(E)$ is a delta-normalization factor, $$N(E)=\sqrt{\frac{1}{\pi}
\frac{2m/\hbar ^2}{\sqrt{2m/\hbar ^2 \, E \,}\,}\,} \, ,
\label{Nfactor}$$ $\chi (r;E)$ is the regular solution, $$\chi (r;E) = \left\{
\begin{array}{ll}
\sin (kr) & 0<r<a \\
{\cal J}_1(E){{\rm e}}^{{{\rm i}}Qr}
+{\cal J}_2(E){{\rm e}}^{-{{\rm i}}Qr} & a<r<b \\
{\cal J}_3(E) {{\rm e}}^{{{\rm i}}kr}
+{\cal J}_4(E) {{\rm e}}^{-{{\rm i}}kr}
& b<r<\infty \, ,
\end{array}
\right.
\label{LSchi}$$ the wave numbers $k$ and $Q$ are given by $$k=\sqrt{ \frac{2m}{\hbar^2}E\,} \, , \quad \,
Q=\sqrt{\frac{2m}{\hbar ^2}(E-V_0)\,} \, ,$$ and ${\cal J}_{\pm}(E)$ are the Jost functions, $${\cal J}_{+}(E)= -2{{\rm i}}{\cal J}_4(E) \, , \quad \,
{\cal J}_-(E)=2{{\rm i}}{\cal J}_3(E) \, .$$ The resonance energies $z_{\rm R}$ produced by the potential (\[potential\]) coincide with the zeros of ${\cal J}_{+}$. With each resonance energy $z_{\rm R}$, we associate a Gamow eigenfunction $u(r;z_{\rm R})$: $$\langle r|z_{\rm R} \rangle =u(r;z_{\rm R})=
\sqrt{\frac{m}{\hbar ^2 k_{\rm R}}\,} {\cal N}_{\rm R} \times
\left\{ \begin{array}{ll}
\frac{1}{{\mathcal J}_3(k_{\rm R})}\sin(k_{\rm R}r) &0<r<a \\ [1ex]
\frac{{\mathcal J}_1(k_{\rm R})}{{\mathcal J}_3(k_{\rm R})} \,
{{\rm e}}^{{{\rm i}}Q_{\rm R}r}
+\frac{{\mathcal J}_2(k_{\rm R})}{{\mathcal J}_3(k_{\rm R})} \,
{{\rm e}}^{-{{\rm i}}Q_{\rm R}r} &a<r<b
\\ [1ex]
{{\rm e}}^{{{\rm i}}k_{\rm R}r} &b<r<\infty \, ,
\end{array}
\right.
\label{dgv0p}$$ where $$k_{\rm R}= \sqrt{\frac{2m}{\hbar ^2}z_{\rm R}\,} \, , \quad
Q_{\rm R}= \sqrt{\frac{2m}{\hbar ^2}(z_{\rm R}-V_0)\,} \, ,$$ and where ${\cal N}_{\rm R}$ is given by Eq. (\[normafac\]). From Eqs. (\[pmeigndu\]) and (\[dgv0p\]) we obtain $$\chi ^-(r;z_{\rm R}) = \frac{1}{{{\rm i}}\sqrt{2\pi}{\cal N}_{\rm R}}
u(r;z_{\rm R}) \, .
\label{relaniodripp}$$
As shown in [@LS2], the analytic continuations of the Lippmann-Schwinger kets –and therefore also the Gamow kets– are well defined as antilinear functionals over the space of test functions $\psi ^-$ for which the following quantities are finite: $$\| \psi ^{-}\|_{n,n'} := \sqrt{\int_{0}^{\infty}{{\rm d}}r \,
\left| \frac{nr}{1+nr}\, {{\rm e}}^{nr^2/2} (1+H)^{n'}
\psi ^{-}(r) \right|^2 \, }
\, , \quad n,n'=0,1,2, \ldots
\label{normsLS}$$ The action of the Gamow ket $|z_{\rm R}\rangle$ on the test functions $\psi ^-$ is explicitly given by $$\langle \psi ^-|z_{\rm R} \rangle =
\int_0^{\infty}{{\rm d}}r \, \langle \psi ^-|r \rangle
\langle r|z_{\rm R} \rangle =
\int_0^{\infty}{{\rm d}}r \, \psi ^-(r)^* u(r;z_{\rm R}) \, .
\label{wendned}$$
For the sake of clarity, we need to introduce a special notation that will specify when we are working in the energy representation: Whenever we work in such representation, we will add a hat to the corresponding quantity. For example, the energy representation of a wave function $\psi ^-(r)$ will be denoted by $\widehat{\psi}^-(E)$.
By using the operator $U_-$ of [@LS1], one can obtain the energy representation of $\psi ^-(r)$, $$\widehat{\psi}^-(E)=(U_-\psi ^-)(E)=
\int_0^{\infty}{{\rm d}}r \, \psi ^-(r) \chi ^-(r;E)^* \, .
\label{-LSinteexpre}$$ As shown in [@LS2], when $\psi ^-(r)$ satisfies (\[normsLS\]), the analytic continuation of $[\widehat{\psi}^-(E)]^*$ exists. We shall denote such analytic continuation by $\widehat{\psi}^-(z^*)^*$: $$\widehat{\psi}^-(z^*)^*=
\int_0^{\infty}{{\rm d}}r \, \psi ^-(r)^* \chi ^-(r;z) \, .
\label{anaconcoc}$$ Not only $\widehat{\psi}^-(z^*)^*$ exist, all the test functions $[\widehat{\psi}^-(z^*)]^*$ are analytic at the resonance energies.[^1] As explained in Appendix \[sec:appendix-cdf\], this means that the antilinear complex delta functional at the resonance energies $z_{\rm R}$ can be defined as $$\begin{array}{rcl}
\widehat{\delta}_{z_{\rm R}} : \widehat{\Phi}_{-\rm exp} &
\longmapsto &
{\mathbb C} \\
\widehat{\psi}^- & \longmapsto &
\langle \widehat{\psi}^-|\widehat{\delta}_{z_{\rm R}} \rangle \equiv
\widehat{\delta}_{z_{\rm R}}(\widehat{\psi}^-) :=
\widehat{\psi}^- (z_{\rm R}^*)^* \, ,
\end{array}
\label{Cdelta-anti-mt}$$ where $\widehat{\Phi}_{-\rm exp}$ is the space of test functions $\widehat{\psi}^-$. That is, $\widehat{\delta}_{z_{\rm R}}$ associates a test function $\widehat{\psi}^-$ with the value that the analytic continuation of $[\widehat{\psi}^-]^*$ takes at $z_{\rm R}$.
The Gamow state and the complex delta function {#sec:GSandcdf}
----------------------------------------------
In order to obtain the equality of Eq. (\[aeqconsBW\]), we are going first to denote the energy representation of the Gamow ket as $$|\widehat{z}_{\rm R}^{-}\rangle \equiv
U_{-}|z_{\rm R}\rangle \, .$$ Then, it follows that $$\begin{aligned}
\langle \widehat{\psi}^-|\widehat{z}_{\rm R}^-\rangle &=&
\langle \widehat{\psi}^-|U_-|z_{\rm R} \rangle \nonumber \\
&=& \langle U_-^{\dagger}\widehat{\psi}^-|z_{\rm R} \rangle \nonumber \\
&=& \langle \psi ^-|z_{\rm R} \rangle \nonumber \\
&=& \int_0^{\infty}{{\rm d}}r \, [\psi ^-(r)]^* u(r;z_{\rm R})
\hskip2.97cm \mbox{by~(\ref{wendned})} \nonumber \\
&=& {{\rm i}}\sqrt{2\pi \, } \, {\cal N}_{\rm R} \int_0^{\infty}{{\rm d}}r \,
\psi ^-(r)^* \chi ^-(r;z_{\rm R})
\hskip1.20cm \mbox{by~(\ref{relaniodripp})} \nonumber \\
&=& {{\rm i}}\sqrt{2\pi \, } \, {\cal N}_{\rm R} \,
\widehat{\psi}^-(z_{\rm R}^*)^*
\hskip3.94cm \mbox{by~(\ref{anaconcoc})} \nonumber \\
&=& {{\rm i}}\sqrt{2\pi \, } \, {\cal N}_{\rm R}
\langle \widehat{\psi}^-|\widehat{\delta}_{z_{\rm R}} \rangle
\hskip3.96cm \mbox{by~(\ref{Cdelta-anti-mt})} \, .
\label{profosjd}\end{aligned}$$ This equation proves that the energy representation of the Gamow functional, $|\widehat{z}_{\rm R}^- \rangle$, is proportional to the antilinear complex delta functional, $|\widehat{\delta}_{z_{\rm R}} \rangle$.
In the energy representation, the identity $\widehat{\mathbb I}$ can be written as $$\int_0^{\infty}{{\rm d}}E \, |\widehat{E}^-\rangle \langle ^-\widehat{E}| =
\widehat{\mathbb I} \, ,
\label{identity}$$ where $|\widehat{E}^-\rangle$ denotes the energy representation of $|E^-\rangle$. By inserting (\[identity\]) into the first and the last terms of (\[profosjd\]), we obtain $$\int_0^{\infty}{{\rm d}}E \, \langle \widehat{\psi}^-|\widehat{E}^-\rangle
\langle ^-\widehat{E}|\widehat{z}_{\rm R}^- \rangle =
\int_0^{\infty}{{\rm d}}E \, {{\rm i}}\sqrt{2\pi \, } {\cal N}_{\rm R}
\langle \widehat{\psi}^-|\widehat{E}^-\rangle
\langle ^-\widehat{E}|\widehat{\delta}_{z_{\rm R}}\rangle \, .
\label{identity2}$$ Because Eq. (\[identity2\]) holds for any $\widehat{\psi}^-$, it follows that $$\langle ^-\widehat{E}|\widehat{z}_{\rm R}^- \rangle =
{{\rm i}}\sqrt{2\pi \, } {\cal N}_{\rm R}
\langle ^-\widehat{E}|\widehat{\delta}_{z_{\rm R}}\rangle \equiv
{{\rm i}}\sqrt{2\pi \, } {\cal N}_{\rm R} \, \delta(E-z_{\rm R}) \, ,
\label{identity3}$$ which, after dropping the hat notation and using Eq. (\[tramaid\]), becomes the equality in Eq. (\[aeqconsBW\]).
The Gamow state and the Breit-Wigner amplitude {#sec:GSandBW}
----------------------------------------------
In order to obtain the approximation of Eq. (\[aeqconsBW\]), we are going to obtain first the transition amplitude $\widetilde{ {\cal A}}(z_{\rm R} \to E)$ from a resonance state of energy $z_{\rm R}$ to a scattering state of energy $-\infty < E< \infty$, $$\widetilde{ {\cal A}}(z_{\rm R} \to E)=
\langle ^-E|z_{\rm R}\rangle \, , \quad E\in (-\infty , \infty ) \, .
\label{tildeamp}$$ Even though a quantum system can only decay to a scattering state of energy $E\geq 0$, we are going to ask the system to pretend that it can also decay to negative energies. Mathematically, this is equivalent to ask the system to pretend that its scattering spectrum runs from $-\infty$ to $+\infty$. Physically, it is equivalent to ignore the effect of the lower bound of energy $E=0$. Calculating (\[tildeamp\]), i.e., forcing the system to decay also to negative energies, needs a regulator. The regulator we will use is ${{\rm e}}^{-{{\rm i}}\alpha E}$, where $\alpha >0$ and $E$ has zero or negative imaginary part. The reason why we use this regulator is that for complex $z$, the wave functions $\widehat{\psi}^-(z^*)^*$ grow slower than ${{\rm e}}^{|{\rm Im}(\sqrt{2m/\hbar ^2 \, z})|^2}$ in the lower half plane of the second sheet [@LS2]. More precisely, Proposition 3 in [@LS2] shows that for each $n=1,2,\ldots$ and for each $\beta >0$, there is a $C>0$ such that in the lower half plane of the second sheet, $\widehat{\psi}^-(z^*)^*$ is bounded by $$|\widehat{\psi}^-(z^*)^*| \leq C \,
\frac{1} {|z|^{1/4}|1+z|^{n}} \,
{{\rm e}}^{\frac{\, \, |{\rm Im}(\sqrt{2m/\hbar ^2 \, z\,})|^2}{2\beta}} \, .
\label{boundinwhvalp+}$$ This estimate implies that ${{\rm e}}^{-{{\rm i}}\alpha z} \widehat{\psi}^-(z^*)^*$ tends to zero in the infinite arc of the lower half of the second sheet, $$\lim_{z\to \infty} {{\rm e}}^{-{{\rm i}}z\alpha}\widehat{\psi}^-(z^*)^* = 0 \, ,
\quad \alpha >0 \, .
\label{indifli}$$ In its turn, the limit (\[indifli\]) enables us to apply Cauchy’s theorem to obtain $$\widehat{\psi}^-(z_{\rm R}^*)^*=
\lim _{\alpha \to 0}- \frac{1}{2\pi {{\rm i}}} \int_{-\infty}^{\infty}{{\rm d}}E \,
{{\rm e}}^{-{{\rm i}}\alpha E} \widehat{\psi}^-(E)^* \frac{1}{E-z_{\rm R}} \, ,
\label{CAUSK}$$ where the integral is performed infinitesimally below the real axis of the second sheet. By multiplying both sides of (\[CAUSK\]) by ${{\rm i}}\sqrt{2\pi}{\cal N}_{\rm R}$, and by recalling (\[profosjd\]), we obtain $$\langle \widehat{\psi}^-|\widehat{z}_{\rm R}^-\rangle =
\lim _{\alpha \to 0} \int_{-\infty}^{\infty}{{\rm d}}E \,
{{\rm e}}^{-{{\rm i}}\alpha E} \psi ^-(E)^* (-1)
\frac{{\cal N}_{\rm R}}{\sqrt{2\pi}} \frac{1}{E-z_{\rm R}} \, .
\label{itenrmind}$$ In the bra-ket notation, Eq. (\[itenrmind\]) reads as $$\langle \widehat{\psi}^-|\widehat{z}_{\rm R}\rangle =
\lim _{\alpha \to 0} \int_{-\infty}^{\infty}{{\rm d}}E \,
{{\rm e}}^{-{{\rm i}}\alpha E} \langle \widehat{\psi}^-|\widehat{E}^-\rangle
\langle ^-\widehat{E}|\widehat{z}_{\rm R}\rangle \, .
\label{itenrmindb-k}$$ Comparison of (\[itenrmind\]) with (\[itenrmindb-k\]) yields the following expression for the amplitude (\[tildeamp\]): $$\widetilde{ {\cal A}}(z_{\rm R} \to E) =
-\frac{{\cal N}_{\rm R}}{\sqrt{2\pi}} \frac{1}{E-z_{\rm R}} \, ,
\ E\in (-\infty , \infty ) \, .$$ Thus, if the scattering spectrum was the whole real line, the resonance amplitude would be exactly the Breit-Wigner amplitude. However, because the scattering spectrum has a lower bound, the resonance amplitude is not exactly the Breit-Wigner amplitude. Only when we can neglect the effect of the threshold, the resonance amplitude coincides with the Breit-Wigner amplitude: $${\cal A}(z_{\rm R} \to E) \simeq \widetilde{ {\cal A}}(z_{\rm R} \to E)
\, ,$$ which is the approximation on the right-hand side of (\[aeqconsBW\]). In particular, when the threshold can be ignored, the complex delta function becomes for all intends and purposes the Breit-Wigner amplitude.
It should be stressed that the amplitude (\[tildeamp\]) is not physical, because in (\[tildeamp\]) the energy $E$ runs over the whole real line rather than over the scattering spectrum. However, such unphysical amplitude helps us understand what the physical amplitude –the complex delta function– is, by allowing us to see how the resonance would decay if the scattering spectrum was the whole real line.
Further remarks {#sec:fR}
---------------
Aside from phase space factors, cross sections are determined by the transition amplitude from an “in” to an “out” state, ${\cal A}(E_i\to E_f )$. If $|E^{\pm}\rangle$ denote the “in” and “out” solutions of the Lippmann-Schwinger equation, then $${\cal A}(E_i\to E_f ) = \langle ^-E_f|E_i^+\rangle =
S(E_i) \delta (E_f -E_i) \, .$$ If we imagine now that instead of an initial state $|E_i^+\rangle$ we had an unstable particle $|z_{\rm R}\rangle$, the transition (decay) amplitude ${\cal A}(z_{\rm R} \to E_f )$ would be given by (\[aeqconsBW\]). Using the approximate decay amplitude of (\[aeqconsBW\]), one obtains the following approximate decay probability: $$|{\cal A}(z_{\rm R} \to E_f )|^2 \simeq
\frac{|{\cal N}_{\rm R}|^2}{2\pi}
\frac{1}{(E_f-E_{\rm R})^2+ (\Gamma _{\rm R}/2)^2} \, ;
\label{almoslpro}$$ that is, the decay probability of a resonance is given by the Lorentzian when the effect of the threshold can be ignored.[^2] Because the almost-Lorentzian decay probability (\[almoslpro\]) coincides with the almost-Lorentzian peaks in cross sections, resonances can be interpreted as intermediate, unstable particles.
Finally, it is worthwhile to compare the Gamow states with the states introduced by Kapur and Peierls [@KAPUR]. As mentioned above, the Gamow states are eigenfunctions of the Hamiltonian that satisfy the POBC (\[POBC\]) at infinity; the wave numbers involved in the POBC (\[POBC\]) are complex and proportional to the square root of the complex eigenenergies of the Gamow states; such complex eigenenergies are the same as the poles of the $S$ matrix, and they do not depend on any external parameter or energy. By contrast, the Kapur-Peierls states are eigenfunctions of the Hamiltonian that satisfy a POBC at a finite radial distance $r_0$, where $r_0$ is such that the potential vanishes for $r>r_0$; the wave numbers involved in the POBC satisfied by the Kapur-Peierls states are real and proportional to the square root of the real energy of the incoming particle; the POBC satisfied by the Kapur-Peierls states makes them and their associated complex eigenenergies depend on $r_0$ and on the real energy of the incoming particle; also, the complex eigenenergies of the Kapur-Peierls states are not the same as the poles of the $S$ matrix. Thus, the Kapur-Peierls states do not seem to be related to the standard Breit-Wigner amplitude, because such amplitude does not depend on $r_0$ and its complex energy does not depend on the energy of the incoming particle.
Conclusions {#sec:conclusions}
===========
Since resonances leave an almost-Lorentzian fingerprint in the cross section, and since the Gamow states are the wave functions of resonances, the decay amplitude provided by a Gamow state should be linked to the Breit-Wigner amplitude. In this paper, we have found that the precise link is given by Eq. (\[aeqconsBW\]), and we have interpreted this result by saying that the resonance amplitude associated with a Gamow state is exactly given by the complex delta function, and that the Breit-Wigner amplitude is an approximation to such resonance amplitude, which approximation is valid when we can neglect the effect of the threshold. Thus, Eq. (\[aeqconsBW\]) establishes the precise relation between the Gamow state, Nakanishi’s complex delta function and the Breit-Wigner amplitude. In addition, Eq. (\[aeqconsBW\]) affords another theoretical argument in favor of interpreting the almost-Lorentzian peaks in cross sections as intermediate, unstable particles—resonances are real (as opposed to virtual) particles, in accordance with resonance phenomenology.
As is well known, the actual resonance lineshape of cross sections can be very different from a quasi-Lorentzian one, due to the effect of thresholds, other resonances, or extra channels. The usefulness of (\[aeqconsBW\]) does not lie in predicting the exact shape of the cross section, but rather in identifying what contribution to the cross section comes from each pole of the $S$ matrix. In particular, although the equality in Eq. (\[aeqconsBW\]) is always exact, for practical purposes the approximation in Eq. (\[aeqconsBW\]) is useful only for narrow resonances.
When we add Eq. (\[aeqconsBW\]) to the other known properties of the Gamow states, we see that such states have all the necessary properties to describe resonance/unstable particles:
- They are associated with poles of the $S$ matrix.
- They exhibit the correct phenomenological signatures of both resonances (almost-Lorentzian lineshape) and unstable particles (exponential decay), and they provide a firm theoretical basis for (\[lifewidre\]).
- They are basis vectors that isolate each resonance’s contribution to a wave packet.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks Casey Koeninger for encouragement, and Alfonso Mondragón for enlightening criticisms. The author thanks Alfonso Mondragón also for his precise explanation of the relation between the Gamow and the Kapur-Peierls states.
Generalizations
===============
\[sec:appendix-gener\]
Equation (\[aeqconsBW\]) is not valid only for the spherical shell potential (\[potential\]) but actually holds for a quite large class of potentials. The reason can be found in well-known results of scattering theory [@TAYLOR; @NUSSENZVEIG]. As explained in [@TAYLOR], page 191, partial wave analysis is valid whenever the spherically symmetric potential satisfies the following requirements:
1. $V(r)=O(r^{-3-\epsilon})$ as $r\to \infty$.
2. $V(r)=O(r^{-3/2+\epsilon})$ as $r\to 0$.
3. $V(r)$ is continuous for $0<r<\infty$, except perhaps at a finite number of finite discontinuities.
These conditions are, however, not sufficient to guarantee that the $S$ matrix $S(E)$, the Jost functions ${\cal J}_{\pm}(E)$ and the Lippmann-Schwinger eigenfunction $\chi ^{-}(r;E)$ can be analytically continued into the whole complex plane. Such analytic continuation is guaranteed when we replace condition $\widetilde{\rm I}$ by the more stringent
1. $V(r)$ falls off faster than exponentials as $r\to \infty$,
as stated throughout Chapters 11 and 12 of [@TAYLOR], and in Chapter 5 of [@NUSSENZVEIG], especially in Theorem 5.3.2. Thus, when $V(r)$ satisfies I-III, even though we may not know their exact analytic expressions, we know that $S(E)$, ${\cal J}_{\pm}(E)$ and $\chi ^{-}(r;E)$ can be analytically continued into the whole complex plane and that the Gamow eigenfunction $u(r;z_{\rm R})$ is well defined. Moreover, since in the asymptotic region $r\to \infty$ the expressions of $u(r;z_{\rm R})$ and $\chi ^-(r;E)$ for any potential satisfying I-III are the same as the expressions of $u(r;z_{\rm R})$ and $\chi ^-(r;E)$ for the spherical shell potential in the region $r>b$ (with different expressions for the Jost functions), the general proof goes through exactly the same lines as the proof for the spherical shell potential. Finally, the argument extends without difficulty to higher angular momentum.
The complex delta functional
============================
\[sec:appendix-cdf\]
In quantum mechanics, the complex delta function was originally introduced by Nakanishi [@NAKANISHI] to describe resonances in the Lee model [@LEE]. In mathematics, the complex delta function was introduced by Gelfand and Shilov [@GELFAND]. The purpose of this appendix is to introduce the precise mathematical definition of the complex delta function and to show that, when the test functions are analytic, such definition coincides with the one given by Nakanishi.
Three definitions of the (linear) complex delta functional
----------------------------------------------------------
The complex delta functional has different forms depending on the properties of the test functions on which it acts. We shall review the three most important forms, namely when the complex delta functional acts on analytic functions (this form is used in this paper and and was introduced in [@GELFAND]), when it acts on meromorphic functions (this is the form used by Nakanishi [@NAKANISHI]), and when it acts on non-meromorphic functions (this form was introduced in [@GELFAND]). When the space of test functions are analytic, as is our case, these three forms coincide (as they should) and can be written as in Eq. (\[Cdelta-anti-mt\]).
### First definition—the test functions are analytic {#sec:firsdief}
According to page 1 of Volume I of Ref. [@GELFAND], a distribution is a function that associates a complex number with each function belonging to a vector space: $$\begin{array}{rcl}
{\rm distribution} :\{ {\rm Space \ of \ functions}\} & \longmapsto &
{\mathbb C}
\\
{\rm function} & \longmapsto & {\rm complex \ number} \, .
\end{array}
\label{distributions}$$ The functions in the “$\{ {\rm Space\ of\ functions}\}$” are usually called test functions. Because a distribution maps functions into complex numbers, they are usually called functionals. Such functionals can be linear or antilinear.
A more precise definition is the following. If $\Phi$ is a vector space of test functions endowed with a topology, a linear (antilinear) distribution $F$ is a function from $\Phi$ to $\mathbb C$ $$\begin{array}{rcl}
F : \Phi & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & F(\phi)
\end{array}
\label{F}$$ such that
- $F$ is well defined,
- $F$ is linear (antilinear),
- $F$ is continuous.
A very important example of distribution is the (linear) Schwartz delta functional at a real number $E$. Such functional associates with each test function $\phi$ the value that $\phi$ takes at $E$: $$\begin{array}{rcl}
\delta _E : \Phi_{\rm Schw} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \delta _E(\phi)= \phi (E) \, ,
\end{array}
\label{SDdelta}$$ where the test functions of $\Phi _{\rm Schw}$ are infinitely differentiable and of polynomial falloff. It is straightforward to show that definition (\[SDdelta\]) satisfies the above requirements ([*i*]{})-([*iii*]{}).
The (linear) complex delta functional is defined in a completely analogous way. As stated by Gelfand and Shilov [@GELFAND Vol. 2, page 85], the point $E$ in Eq. (\[SDdelta\]) may be complex in the spaces of analytic functions. If $\Phi _{\rm anal}$ denotes a vector space of [*analytic*]{} functions at the complex point $z_0$, then the linear complex delta functional at $z_0$ is defined as a function that associates with each test function $\phi$ the value that the analytic continuation of $\phi$ takes at $z_0$: $$\begin{array}{rcl}
\delta _{z_0} : \Phi _{\rm anal} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \delta _{z_0}(\phi)=
\phi (z_0) \, .
\end{array}
\label{Cdelta}$$ Two important comments are in order here. First, the test functions of $\Phi _{\rm anal}$ must be analytic at $z_0$; that is, $z_0$ is not a singularity (e.g., a pole) of any $\phi$, otherwise definition (\[Cdelta\]) makes no sense. And second, the complex delta functional is completely specified by Eq. (\[Cdelta\]) because the test functions are analytic at $z_0$, and therefore one does not need to introduce any contour in the definition of (\[Cdelta\]), even though one could use such a contour, as in Eq. (\[Cdeltameroph\]) below.
Definition (\[Cdelta\]) actually fulfills the requirements ([*i*]{})-([*iii*]{}).[^3] The only property that is conceptually challenging is ([*i*]{}). Because we are assuming that the test functions are analytic at $z_0$, $\phi (z_0)$ exists and is unique, which grants requirement ([*i*]{}). Thus, definition (\[Cdelta\]) completely, rigorously and unambiguously defines the complex delta functional.
### Second definition—the test functions are meromorphic {#sec:second-linear}
Many functions are not analytic but just meromorphic. That is, when we analytically continue them, they have isolated singularities (“poles”) in the complex plane. At such poles, definition (\[Cdelta\]) makes no sense, and one has to extend it. If $\Phi _{\rm mero}$ is a vector space of meromorphic functions at $z_0$, the (linear) complex delta functional at $z_0$ is defined as $$\begin{array}{rcl}
\delta _{z_0} : \Phi _{\rm mero} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \delta _{z_0}(\phi)=
\frac{1}{2\pi {{\rm i}}} \oint {{\rm d}}z \, \frac{\phi (z)}{z-z_0} \, .
\end{array}
\label{Cdeltameroph}$$ One can again check very easily that definition (\[Cdeltameroph\]) satisfies requirements ([*i*]{})-([*iii*]{}). Note that because in definition (\[Cdeltameroph\]) the test functions are meromorphic, such definition depends on Cauchy’s theorem and on the contour used.[^4]
If we denote by $a_0$ the zeroth term of the Laurent expansion of $\phi (z)$ around $z_0$, then definition (\[Cdeltameroph\]) associates $a_0$ with each test function $\phi$, since $$a_0=\frac{1}{2\pi {{\rm i}}} \oint {{\rm d}}z \, \frac{\phi (z)}{z-z_0} \, .$$ Thus, we may write definition (\[Cdeltameroph\]) as $$\begin{array}{rcl}
\delta _{z_0} : \Phi _{\rm mero} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \delta _{z_0}(\phi)= a_0 \, .
\end{array}
\label{Cdeltameroph-alter}$$ Obviously, both (\[Cdeltameroph\]) and (\[Cdeltameroph-alter\]) define the same functional, because both associate the same complex number with the same function, even though in (\[Cdeltameroph-alter\]) no contour integral has been explicitly used.
Now, when $\phi (z)$ is analytic at $z_0$, $a_0$ is simply $\phi (z_0)$. Thus, when the test functions are not just meromorphic but also [*analytic*]{} at $z_0$, definitions (\[Cdeltameroph\]) and (\[Cdeltameroph-alter\]) become definition (\[Cdelta\]), [because in such case all these definitions associate each function $\phi$ with one and the same complex number $\phi (z_0)$]{}. This is why, when the test functions $\phi$ are all [*analytic*]{} at $z_0$, one can define the complex delta functional by way of Eq. (\[Cdelta\]), as Gelfand and Shilov do in page 85, Vol. II of [@GELFAND].
### Third definition—the test functions are not meromorphic {#sec:third-linear}
When the test functions are not meromorphic, definitions (\[Cdelta\]), (\[Cdeltameroph\]) and (\[Cdeltameroph-alter\]) make no sense. One can still define a complex delta functional at the origin following the prescription of Gelfand and Shilov [@GELFAND Vol. I, Appendix B]. When the functions are meromorphic, such definition of the complex delta functional at the origin becomes (\[Cdeltameroph\]) and (\[Cdeltameroph-alter\]).
However, because in this paper we use test functions that are analytic at the resonance energies, we do not need to use this general definition or definition (\[Cdeltameroph\]), because all these definitions actually become (\[Cdelta\]).
Three definitions of the (antilinear) complex delta functional {#sec:anti}
--------------------------------------------------------------
In this paper, we have used antilinear (rather than linear) functionals. We will therefore briefly explain how one defines such functionals for the cases considered in the previous section.
The (antilinear) Schwartz delta functional at a real number $E$ associates with each test function $\phi$, the complex conjugate of the value that $\phi$ takes at $E$: $$\begin{array}{rcl}
\widehat{\delta}_E : \Phi_{\rm Schw} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \widehat{\delta}_E(\phi)= \phi (E)^* \, .
\end{array}
\label{SDdelta-anti}$$ When we write the action of $\widehat{\delta}_E$ as an integral operator, the kernel of such integral operator is Dirac’s delta function: $$\widehat{\delta}_E(\phi )= \int_0^{\infty}{{\rm d}}E' \,
\delta (E'-E) \phi (E')^*
=\phi (E)^* \, .
\label{SDdeltaIO}$$
If $\Phi _{\rm anal}$ denotes a vector space of test functions $\phi$ such that $\phi ^*$ are all [*analytic*]{} at $z_0$, then the antilinear complex delta functional at $z_0$ is a function that associates with each test function $\phi$, the value that the analytic continuation of $\phi ^*$ takes at $z_0$: $$\begin{array}{rcl}
\widehat{\delta}_{z_0} : \Phi _{\rm anal} & \longmapsto &
{\mathbb C} \\
\phi & \longmapsto & \widehat{\delta}_{z_0}(\phi)=
\phi (z_0^*)^* \, .
\end{array}
\label{Cdelta-anti}$$ When we write the expression for $\widehat{\delta}_{z_0}$ as an integral operator, the kernel of such integral operator is the complex delta function: $$\widehat{\delta}_{z_0}(\phi )= \int_0^{\infty}{{\rm d}}E' \,
\delta (E'-{z_0}) \phi (E')^* =
\phi (z_0^*)^*\, .
\label{decdf}$$
When the test functions are only meromorphic and $z_0$ is one of their poles, definition (\[Cdelta-anti\]) needs to be changed to $$\begin{array}{rcl}
\widehat{\delta}_{z_0} : \Phi _{\rm mero} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \widehat{\delta}_{z_0}(\phi)=
\frac{1}{2\pi {{\rm i}}} \oint {{\rm d}}z \, \frac{\phi (z^*)^*}{z-z_0} \, .
\end{array}
\label{Cdeltameroph-anti}$$ If we denote by $a_0^*$ the zeroth term of the Laurent expansion of $\phi (z^*)^*$ around $z_0$, then definition (\[Cdeltameroph-anti\]) associates $a_0^*$ with each test function $\phi$, and therefore we can write $$\begin{array}{rcl}
\widehat{\delta}_{z_0} : \Phi _{\rm mero} & \longmapsto & {\mathbb C} \\
\phi & \longmapsto & \widehat{\delta}_{z_0}(\phi)= a_0^* \, .
\end{array}
\label{Cdeltameroph-anti-a8}$$
If the functions are not even meromorphic, we need to use the prescription of Gelfand and Shilov [@GELFAND Vol. I, Appendix B].
The same conclusions as in the previous section apply to the antilinear complex delta functional. When $\phi (z^*)^*$ are all [*analytic*]{} at $z_0$, $a_0^*$ is simply $\phi (z_0^*)^*$. Thus, when the test functions are all [*analytic*]{} at $z_0$, definition (\[Cdeltameroph-anti\]) becomes definition (\[Cdelta-anti\]), and we are allowed to use (\[Cdelta-anti\]).
Nakanishi’s definition {#sec:Nakadef}
----------------------
Nakanishi [@NAKANISHI] uses a slightly different version of the complex delta function. When he writes $\delta _{\rm N}(\phi )$ as an integral operator, Nakanishi uses the following expression: $$\delta _{\rm N}(\phi )= \int_{\gamma} {{\rm d}}E \,
\phi (E^*)^* \delta _{\rm N}(E-z_{\rm R}) \, ,
\label{Nakscon1}$$ where $$\delta _{\rm N}(E-z_{\rm R}) = \frac{1}{2\pi {{\rm i}}} \left(
\frac{1}{E^{(-)}-z_{\rm R}} - \frac{1}{E^{(+)}-z_{\rm R}}
\right) \, ,$$ and where the contour $\gamma$ is such that the integral in Eq. (\[Nakscon1\]) decomposes into two terms. The end points of the integration paths are the same for the two terms, namely, $0$ and $+\infty$. The integration path for the first term, $\frac{1}{E^{(-)}-z_{\rm R}}$, passes below $z_{\rm R}$, whereas the integration path for the second term, $\frac{1}{E^{(+)}-z_{\rm R}}$, passes above $z_{\rm R}$. Adding the two terms we obtain $$\int_{\gamma}{{\rm d}}E \, \phi (E^*)^* \delta _{\rm N}(E-z_{\rm R})=
\frac{1}{2\pi {{\rm i}}} \oint {{\rm d}}E \, \frac{\phi (E^*)^*}{E-z_{\rm R}} =
\phi (z_{\rm R}^*)^* \, .
\label{Nakscon}$$ Thus, the distributional definition (\[Cdelta-anti\]) is equivalent to Nakanishi’s definition (\[Nakscon1\])-(\[Nakscon\]), because both approaches associate the same complex number, $\phi (z_{\rm R}^*)^*$, with the same test function, $\phi$.
[99]{}
G. Gamow, Z. Phys. [**51**]{}, 204 (1928).
A.F.J. Siegert, Phys. Rev. [**56**]{}, 750 (1939).
R.E. Peierls, Proc. R. Soc. (London) A [**253**]{}, 16 (1959).
J. Humblet, L. Rosenfeld, Nucl. Phys. [**26**]{}, 529 (1961).
Ya.B. Zeldovich, Sov. Phys. JETP [**12**]{}, 542 (1961).
T. Berggren, Nucl. Phys. A [**109**]{}, 265 (1968).
G. García-Calderón, R. Peierls, Nucl. Phys. A [**265**]{}, 443 (1976).
T. Berggren, Phys. Lett. B [**73**]{}, 389 (1978).
G. Parravicini, V. Gorini, E.C.G. Sudarshan, J. Math. Phys. [**21**]{}, 2208 (1980).
E. Hernández, A. Mondragón, Phys. Rev. C [**29**]{}, 722 (1984).
P. Curutchet, T. Vertse, R.J. Liotta, Phys. Rev. C [**39**]{}, 1020 (1989).
T. Berggren, P. Lind, Phys. Rev. C [**47**]{}, 768 (1993).
P. Lind, Phys. Rev. C [**47**]{}, 1903 (1993).
T. Berggren, Phys. Lett. B [**373**]{}, 1 (1996).
C.G. Bollini, O. Civitarese, A.L. De Paoli, M.C. Rocca, Phys. Lett. B[**382**]{}, 205 (1996).
L.S. Ferreira, E. Maglione, R.J. Liotta, Phys. Rev. Lett. [**78**]{}, 1640 (1997).
R. Id Betan, R.J. Liotta, N. Sandulescu, T. Vertse, Phys. Rev. Lett. [**89**]{}, 042501 (2002).
N. Michel, W. Nazarewicz, M. Ploszajczak, K. Bennaceur, Phys. Rev. Lett. [**89**]{}, 042502 (2002).
R. de la Madrid, M. Gadella, Am. J. Phys. [**70**]{}, 626 (2002).
E. Kapuscik, P. Szczeszek, Czech. J. Phys. [**53**]{}, 1053 (2003).
E. Hernández, A. Jáuregui, A. Mondragón, Phys. Rev. A [**67**]{}, 022721 (2003).
N. Michel, W. Nazarewicz, M. Ploszajczak, J. Okolowicz, Phys. Rev. C [**67**]{}, 054311 (2003).
E. Kapuscik, P. Szczeszek, Found. Phys. Lett. [**18**]{}, 573 (2005).
N. Michel, W. Nazarewicz, M. Ploszajczak, J. Rotureau, Phys. Rev. C [**74**]{}, 054305 (2006); [nucl-th/0609016]{}.
J. Rotureau, N. Michel, W. Nazarewicz, M. Ploszajczak, J. Dukelsky, Phys. Rev. Lett. [**97**]{}, 110603 (2006); [nucl-th/0603021]{}.
N. Michel, W. Nazarewicz, M. Ploszajczak, Phys. Rev. C [**75**]{}, 031301 (2007); [nucl-th/0702021]{}.
J. Julve, F.J. de Urr[í]{}es, [arXiv:quant-ph/0701213]{}.
N. Michel, W. Nazarewicz, M. Ploszajczak, Nuc. Phys. A [**794**]{}, 29 (2007); [arXiv:0707.0767]{}.
N. Hatano, K. Sasada, H. Nakamura, T. Petrosky, Prog. Theo. Phys. [**119**]{}, 187 (2008); [arXiv:0705.1388]{}.
S. Eidelman [*et al.*]{} “Review of Particle Physics,” Phys. Lett. B[**592**]{}, 1 (2004).
C.W. Oates, K.R. Vogel, J.L. Hall, Phys. Rev. Lett. [**76**]{}, 2866 (1996).
U. Volz, M. Majerus, H. Liebel, A. Schmitt, H. Schmoranzer, Phys. Rev. Lett. [**76**]{}, 2862 (1996).
The cross section will actually show a distinctive, almost-Lorentzian peak if, in addition to $z_{\rm R}$ being a pole of the $S$ matrix, the resonance is not too broad ($\Gamma _{\rm R}/E_{\rm R} << 1$), and if other possible resonances are not in the vicinity of $z_{\rm R}$, so their interference with $z_{\rm R}$ can be ignored.
L. Fonda, G.C. Ghirardi, A. Rimini, Rep. Prog. Phys. [**41**]{}, 587 (1978).
R. de la Madrid, G. Garc[í]{}a-Calderón, J.G. Muga, Czech. J. Phys. [**55**]{}, 1141 (2005).
S.R. Wilkinson [*et al.*]{} Nature [**387**]{}, 575 (1997).
R. de la Madrid, J. Phys. A: Math. Gen. [**39**]{}, 3981 (2006); [quant-ph/0603177]{}.
R. de la Madrid, J. Phys. A: Math. Gen. [**39**]{}, 3949 (2006); [quant-ph/0603176]{}.
P.L. Kapur, R. Peierls, Proc. R. Soc. (London) A [**166**]{}, 277 (1938).
H.M. Nussenzveig, [*Causality and Dispersion Relations*]{}, Academic Press, New York and London (1972).
J.R. Taylor, [*Scattering Theory*]{}, John Wiley & Sons, Inc., New York (1972).
N. Nakanishi, Prog. Theo. Phys. [**19**]{}, 607 (1958).
T.D. Lee, Phys. Rev. [**95**]{}, 1329 (1954).
I.M. Gelfand, G.E. Shilov [*Generalized Functions, Vol. 1-3*]{}, Academic Press, New York (1964). I.M. Gelfand, N.Ya. Vilenkin [*Generalized Functions Vol. 4*]{}, Academic Press, New York (1964).
[^1]: The poles of $[\widehat{\psi}^-(z^*)]^*$ are located on the first sheet of the Riemann surface, that is, on the upper half of the wave-number plane.
[^2]: A much more detailed study of the dependence of the cross section (and expectation values of observables) on the Breit-Wigner amplitude can be found in e.g. [@SIEGERT; @BERGGREN78; @BERGGREN96].
[^3]: In this paper, we omit any explicit discussion on the continuity requirement ([*iii*]{}). The reason is that first, the continuity of the complex delta function is guaranteed by the results of [@LS2], and second, continuity is not essential to our main discussion.
[^4]: The contour used in Eq. (\[Cdeltameroph\]) is assumed to be a circle around $z_0$ such that the test function $\phi$ is analytic inside such circle except perhaps at $z_0$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Pascal O. Vontobel$^1$ and Ralf Koetter$^2$\
$^1$ Dept. of EECS, MIT, Cambridge, MA 02139, USA, `[email protected]`.\
$^2$ CSL and Dept. of ECE, University of Illinois, Urbana, IL 61801, USA, `[email protected]`.
title: 'Towards Low-Complexity Linear-Programming Decoding'
---
[^1]
Introduction {#sec:introduction:1}
============
Linear-programming (LP) decoding [@Feldman:03:1; @Feldman:Wainwright:Karger:05:1] has recently emerged as an interesting option for decoding low-density parity-check (LDPC) codes. Indeed, the observations in [@Koetter:Vontobel:03:1; @Vontobel:Koetter:05:1:subm; @Vontobel:Koetter:04:2] suggest that the LP decoding performance is very close to the message-passing iterative (MPI) decoding performance. Of course, one can use any general-purpose LP solver to solve the LP that appears in LP decoding, however in this paper we will argue that one should take advantage of the special structure of the LP at hand in order to formulate efficient algorithms that provably find the optimum of the LP.
Feldman et al. [@Feldman:Karger:Wainwright:02:1] briefly mention the use of sub-gradient methods for solving the LP of an early version of the LP decoder (namely for turbo-like codes). Moreover, Yang et al. [@Yang:Wang:Feldman:05:1] present a variety of interesting approaches to solve the LP where they use some of the special features of the LP at hand. However, we belive that one can take much more advantage of the structure that is present: this paper shows some results in that direction.
So far, MPI decoding has been successfully used in applications where block error rates on the order of $10^{-5}$ are needed because for these block error rates the performance of MPI decoding can be guaranteed by simulation results. However, for applications like magnetic recording, where one desires to have block error rates on the order of $10^{-15}$ and less, it is very difficult to guarantee that MPI decoding achieves such low block error rates for a given signal-to-noise ratio. The problem is that simulations are too time-consuming and that the known analytical results are not strong enough. Our hope and main motivation for the present work is that efficient LP decoders, together with analytical results on LP decoding (see e.g. [@Vontobel:Koetter:04:1; @Chaichanavong:Siegel:05:1; @Vontobel:Smarandache:05:1]), can show that efficient decoders exist for which low block error rates can be guaranteed for a certain signal-to-noise ratio.
This paper is structured as follows. We start off by introducing in Sec. \[sec:primal:linear:program:1\] the primal LP that appears in LP decoding. In Sec. \[sec:dual:linear:program:1\] we formulate the dual LP and in Secs. \[sec:softened:dual:linear:program:1\] and \[sec:primal:softened:dual:linear:program:1\] we consider a “softened” version of this dual LP. Then, in Secs. \[sec:decoding:algorithm:1\] and \[sec:decoding:algorithm:2\] we propose some efficient decoding algorithms and in Sec. \[sec:simulation:results:1\] we show some simulation results. Finally, in Sec. \[sec:conclusions:1\] we offer some conclusions and in the appendix we present the proofs and some additional material.
Before going to the main part of the paper, let us fix some notation. We let ${\mathbb{R}}$, ${\mathbb{R}_{+}}$, and ${\mathbb{R}_{++}}$ be the set of real numbers, the set of non-negative real numbers, and the set of positive real numbers, respectively. Moreover, we will use the canonical embedding of the set ${\mathbb{F}_{2}} = \{ 0, 1 \}$ into ${\mathbb{R}}$. The convex hull of a set ${\mathcal{A}} \subseteq {\mathbb{R}}^n$ is denoted by ${\operatorname{conv}}({\mathcal{A}})$. If ${\mathcal{A}}$ is a subset of ${\mathbb{F}_{2}}^n$ then ${\operatorname{conv}}({\mathcal{A}})$ denotes the convex hull of the set ${\mathcal{A}}$ after ${\mathcal{A}}$ has been canonically embedded in ${\mathbb{R}}^n$. The $i$-th component of a vector ${{\mathbf{x}}}$ will be called $[{{\mathbf{x}}}]_i$ and the element in the $j$-th row and $i$-th column of a matrix ${\mathbf{A}}$ will be called $[{\mathbf{A}}]_{j,i}$.
Moreover, we will use Iverson’s convention, i.e. for a statement $A$ we have $[A] = 1$ if $A$ is true and $[A] = 0$ otherwise. From this we also derive the notation $\big\llbracket A \big\rrbracket {\triangleq}-\log [A]$, i.e. $\big\llbracket A \big\rrbracket = 0$ if $A$ is true and $\big\llbracket
A \big\rrbracket = +\infty$ otherwise. Let ${\mathcal{A}}$ and ${\mathcal{X}}$ be some arbitrary sets fulfilling ${\mathcal{A}} \subseteq {\mathcal{X}}$. A function like ${\mathcal{X}} \to {\mathbb{R}_{+}}: {{\mathbf{x}}}\mapsto [{{\mathbf{x}}}\in {\mathcal{A}}]$ is called an indicator function for the set ${\mathcal{A}}$, whereas a function like ${\mathcal{X}} \to {\mathbb{R}_{+}}: {{\mathbf{x}}}\mapsto \big\llbracket {{\mathbf{x}}}\in {\mathcal{A}} \big\rrbracket$ is called a neglog indicator function for the set ${\mathcal{A}}$. Of course, this second function can also be considered as a cost or penalty function.
Throughout the paper, we will consider a binary linear code ${\mathcal{C}}$ that is defined by a parity-check matrix ${\mathbf{H}}$ of size $m$ by $n$. Based on ${\mathbf{H}}$, we define the sets ${\mathcal{I}} {\triangleq}{\mathcal{I}}({\mathbf{H}}) {\triangleq}\{ 1,
\ldots, n \}$, ${\mathcal{J}} {\triangleq}{\mathcal{J}}({\mathbf{H}}) {\triangleq}\{ 1, \ldots, m \}$, ${\mathcal{I}}_j {\triangleq}{\mathcal{I}}_j({\mathbf{H}}) {\triangleq}\{ i \in {\mathcal{I}} \ | \
[{\mathbf{H}}]_{j,i} = 1 \}$ for each $j \in {\mathcal{J}}$, ${\mathcal{J}}_i {\triangleq}{\mathcal{J}}_i({\mathbf{H}}) {\triangleq}\{ j \in {\mathcal{J}} \ | \ [{\mathbf{H}}]_{j,i} = 1 \}$ for each $i \in {\mathcal{I}}$, and ${\mathcal{E}} {\triangleq}{\mathcal{E}}({\mathbf{H}}) {\triangleq}{\mathcal{\{}}
(i,j) \in {\mathcal{I}} \times {\mathcal{J}} \ | \ i \in {\mathcal{I}}, j \in {\mathcal{J}}_i \} =
{\mathcal{\{}} (i,j) \in {\mathcal{I}} \times {\mathcal{J}} \ | \ j \in {\mathcal{J}}, i \in {\mathcal{I}}_j
\}$. Moreover, for each $j \in {\mathcal{J}}$ we define the codes ${\mathcal{C}}_j
{\triangleq}{\mathcal{C}}_j({\mathbf{H}}) {\triangleq}\{ {{\mathbf{x}}}\in {\mathbb{F}_{2}}^n \ | \ {\mathbf{h}}_j
{{\mathbf{x}}}^{\mathsf{T}}= 0 \text{ (mod $2$)} \}$, where ${\mathbf{h}}_j$ is the $j$-th row of ${\mathbf{H}}$. Note that the code ${\mathcal{C}}_j$ is a code of length $n$ where all positions not in ${\mathcal{I}}_j$ are unconstrained.
We will express the linear programs in this paper in the framework of Forney-style factor graphs (FFG) [@Kschischang:Frey:Loeliger:01; @Forney:01:1; @Loeliger:04:1], sometimes also called normal graphs. For completeness we state their formal definition. An FFG is a graph ${\mathsf{G}}(V,E)$ with vertex set $V$ and edge set $E$. To each edge $e$ in the graph we associate a variable $x_e$ defined over a suitably chosen alphabet ${\mathcal{X}}_e$. Let $v$ be a node in the FFG and let $E_v$ be the set of edges incident to $v$. Any node $v$ in the graph is associated with a function $f_v$ with domain ${\mathcal{X}}_{e_1} \times {\mathcal{X}}_{e_2} \times \cdots \times
{\mathcal{X}}_{e_\ell}$ where $\{e_1,e_2,\ldots,e_\ell\}=E_v$. The co-domain of $f_v$ is typically ${\mathbb{R}}$ or ${\mathbb{R}}_+$.
FFGs typically come in two flavors, either representing the factorization of a function into a product of terms or a decomposition of an additive cost function. In our case we will exclusively deal with the latter case. The global function $g(x_{e_1},x_{e_2},\ldots,x_{e_{|E|}})$ represented by an FFG is then given by the sum $g(x_{e_1},x_{e_2},\ldots,x_{e_{|E|}}) {\triangleq}\sum_{v\in V} f_v$.
The Primal Linear Program {#sec:primal:linear:program:1}
=========================
The code ${\mathcal{C}}$ is used for data transmission over a binary-input memoryless channel with channel law $P_{{\mathbf{Y}} | {\mathbf{X}}}({\mathbf{y}}
| {\mathbf{x}}) = \prod_{i \in {\mathcal{I}}} P_{Y | X}(y_i | x_i)$. Upon observing ${\mathbf{Y}} = {\mathbf{y}}$, the maximum-likelihood decoding (MLD) rule decides for $\hat {{\mathbf{x}}}({{\mathbf{y}}}) = \arg \max_{{{\mathbf{x}}}\in {\mathcal{C}}}
P_{{\mathbf{Y}} | {\mathbf{X}}}({\mathbf{y}} | {\mathbf{x}})$. This can also be written as
[\
\
]{}
It is clear that instead of $P_{{{\mathbf{Y}}}|{{\mathbf{X}}}}({{\mathbf{y}}}| {{\mathbf{x}}})$ we can also maximize $\log P_{{{\mathbf{Y}}}|{{\mathbf{X}}}}({{\mathbf{y}}}| {{\mathbf{x}}}) = \sum_{i \in {\mathcal{I}}} \log P_{Y |
X}(y_i | x_i)$. Introducing $\lambda_i {\triangleq}\lambda_i(y_i) {\triangleq}\log \bigl( \frac{P_{Y | X}(y_i|0)}{P_{Y | X}(y_i | 1)} \bigr)$, $i \in
{\mathcal{I}}$, and noting that $\log P_{Y | X}(y_i | x_i) = - \lambda_i x_i + \log
P_{Y | X}(y_i | 0), \textbf{MLD1}$ can then be rewritten to read
[\
\
]{}
Because the cost function is linear, and a linear function attains its minimum at the extremal points of a convex set, this is essentially equivalent to
[\
\
]{}
Although this is a linear program, it can usually not be solved efficiently because its description complexity is usually exponential in the block length of the code.
However, one might try to solve a relaxation of **MLD3**. Noting that ${\operatorname{conv}}({\mathcal{C}}) \subseteq {\operatorname{conv}}({\mathcal{C}}_1) \cap \cdots \cap
{\operatorname{conv}}({\mathcal{C}}_m)$ (which follows from the fact that ${\mathcal{C}} =
{\mathcal{C}}_1 \cap \cdots \cap {\mathcal{C}}_m$), Feldman, Wainwright, and Karger [@Feldman:03:1; @Feldman:Wainwright:Karger:05:1] defined the (primal) linear programming decoder (PLPD) to be given by the solution of the linear program
[\
\
]{}
The inequalities that are implied by the expression ${{\mathbf{x}}}\in
{\operatorname{conv}}({\mathcal{C}}_j)$ can be found in [@Feldman:03:1; @Feldman:Wainwright:Karger:05:1; @Koetter:Vontobel:03:1; @Vontobel:Koetter:05:1:subm]. Although **PLPD1** is usually suboptimal compared to MLD, it is especially attractive for LDPC codes for two reasons: firstly, for these codes the description complexities of ${\operatorname{conv}}({\mathcal{C}}_j)$, $j \in {\mathcal{J}}$, turn out to be low [@Feldman:Wainwright:Karger:05:1; @Vontobel:Koetter:05:1:subm] and, secondly, the relaxation is relatively benign only if the weight of the parity checks is low. There are many ways of reformulating this **PLPD1** rule by introducing auxiliary variables: one way that we found particularly useful is shown as **PLPD2** below. The reason for its usefulness is that there is a one-to-one correspondence between parts of the program and the FFG shown in Fig. \[fig:ffg:ldpc:code:binary:1:1\], as we will discuss later on. Indeed, while the notation may seem heavy at first glance, it precisely reflects the structure of the constraints that are summarily folded into the seemingly simpler constraint ${{\mathbf{x}}}\in {\operatorname{conv}}({\mathcal{C}}_j) \ (j \in {\mathcal{J}})$ of **PLPD1**.
[\
\
]{}
Here we used the following codes, variables and vectors. The code ${\mathcal{A}}_i \subseteq \{ 0, 1 \}^{|\{ 0 \} \cup {\mathcal{J}}_i|}$, $i \in {\mathcal{I}}$, is the set containing the all-zeros vector and the all-ones vector of length $|{\mathcal{J}}_i| + 1$, and ${\mathcal{B}}_j \subseteq \{ 0, 1 \}^{|{\mathcal{I}}_j|}$, $j \in
{\mathcal{J}}$, is the code ${\mathcal{C}}_j$ shortened at the positions ${\mathcal{I}}
\setminus {\mathcal{I}}_j$.[^2] For $i \in {\mathcal{I}}$ we will also use the vectors ${{\mathbf{u}}}_i$ where the entries are indexed by $\{ 0 \} \cup {\mathcal{J}}_i$ and denoted by $u_{i,j}
{\triangleq}[{{\mathbf{u}}}_i]_j$, and for $j \in {\mathcal{J}}$ we will use the vectors ${{\mathbf{v}}}_j$ where the entries are indexed by ${\mathcal{I}}_j$ and denoted by $v_{j,i} {\triangleq}[{{\mathbf{v}}}_j]_i$. Later on, we will use a similar notation for the entries of ${{\mathbf{a}}}_i$ and ${{\mathbf{b}}}_j$, i.e. we will use $a_{i,j} {\triangleq}[{{\mathbf{a}}}_i]_j$ and $b_{j,i}
{\triangleq}[{{\mathbf{b}}}_j]_i$, respectively.
The above optimization problem is elegantly represented by the FFG shown in Fig. \[fig:ffg:ldpc:code:binary:1:1\]. In order to express the LP itself in an FFG we have to express the constraints as additive cost terms. This is easily accomplished by assigning the cost $+\infty$ to any configuration of variables that does not satisfy the LP constraints. The above minimization problem is then equivalent to the (unconstrained) minimization of the augmented cost function $$\begin{aligned}
\sum_{i \in {\mathcal{I}}}
\lambda_i x_i
&
+
\sum_{i \in {\mathcal{I}}}
\big\llbracket
x_i = u_{i,0}
\big\rrbracket
+
\sum_{(i,j) \in {\mathcal{E}}}
\big\llbracket
u_{i,j} = v_{j,i}
\big\rrbracket
\nonumber \\
&
+
\sum_{i \in {\mathcal{I}}}
A_i({{\mathbf{u}}}_i)
+
\sum_{j \in {\mathcal{J}}}
B_j({{\mathbf{v}}}_j),
\label{eq:primal:lp:augmented:cost:function:1}\end{aligned}$$ where for all $i \in {\mathcal{I}}$ and all $j \in {\mathcal{J}}$, respectively, we introduced $$\begin{aligned}
A_i({{\mathbf{u}}}_i)
&{\triangleq}\left\llbracket
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\alpha_{i, {{\mathbf{a}}}_i} {{\mathbf{a}}}_i
= {{\mathbf{u}}}_i
\right\rrbracket \\
&\quad\quad
+
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\big\llbracket
\alpha_{i, {{\mathbf{a}}}_i}
\geq 0
\big\rrbracket
+
\left\llbracket
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\alpha_{i, {{\mathbf{a}}}_i}
= 1
\right\rrbracket, \\
B_j({{\mathbf{v}}}_j)
&{\triangleq}\left\llbracket
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\beta_{j, {{\mathbf{b}}}_j} {{\mathbf{b}}}_j
= {{\mathbf{v}}}_j
\right\rrbracket \\
&\quad\quad
+
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\big\llbracket
\beta_{j, {{\mathbf{b}}}_j}
\geq 0
\big\rrbracket
+
\left\llbracket
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\beta_{j, {{\mathbf{b}}}_j}
= 1
\right\rrbracket.\end{aligned}$$ With this, the global function of the FFG in Fig. \[fig:ffg:ldpc:code:binary:1:1\] equals the augmented cost function in and we have represented the LP in terms of an FFG.[^3]
Of course, any reader who is familiar with LDPC codes will have no problem to make a connection between the FFG of Fig. 1 and the standard representation as a Tanner graph. Indeed, a node $A_i$ corresponds to a variable node in a Tanner graph and a node $B_j$ takes over the role of a parity check node. However, instead of simply assigning a variable to node $A_i$ we assign a local set of constraints corresponding to the convex hull of a repetition code. These are the equations $\sum_{{{\mathbf{a}}}_i\in {\cal{A}}_i} \alpha_{i,{{\mathbf{a}}}_i}
{{\mathbf{a}}}_i ={{\mathbf{u}}}_i$, $\alpha_{i,{{\mathbf{a}}}_i}\geq 0$, $\sum_{{{\mathbf{a}}}_i\in {\cal{A}}_i}
\alpha_{i,{{\mathbf{a}}}_i} =1$. Similarly, the equations for the convex hull of a simple parity-check code can be identified for nodes $B_j$.
The Dual Linear Program {#sec:dual:linear:program:1}
=======================
The dual linear program [@Bertsimas:Tsitsiklis:97:1] of **PLPD2** is
[\
\
]{}
Expressing the constraints as additive cost terms, the above maximization problem is equivalent to the (unconstrained) maximization of the augmented cost function $$\begin{aligned}
&
\sum_{i \in {\mathcal{I}}}
A'_i({{\mathbf{u}}}'_i)
+
\sum_{j \in {\mathcal{J}}}
B'_j({{\mathbf{v}}}'_j)
-
\sum_{(i,j) \in {\mathcal{E}}}
\llbracket
u'_{i,j}
= - v'_{j,i}
\rrbracket
\nonumber \\
&
-
\sum_{i \in {\mathcal{I}}}
\llbracket
u'_{i,0}
= - x'_i
\rrbracket
-
\sum_{i \in {\mathcal{I}}}
\llbracket
x'_i
= \lambda_i
\rrbracket,
\label{eq:dual:lp:augmented:cost:function:1}\end{aligned}$$ with $$\begin{aligned}
A'_i({{\mathbf{u}}}'_i)
&= \phi'_i
-
\left\llbracket
\phi'_i
\leq
\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
\right\rrbracket, \\
B'_j({{\mathbf{v}}}'_j)
&= \theta'_j
-
\left\llbracket
\theta'_j
\leq
\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle
\right\rrbracket.\end{aligned}$$ The augmented cost function in is represented by the FFG in Fig. \[fig:dual:ffg:ldpc:code:binary:1:1\].[^4] (For deriving **DLPD2** we used the techniques introduced in [@Vontobel:02:2; @Vontobel:Loeliger:02:2]; note that the techniques presented there can also be used to systematically derive the dual function of much more complicated functions that are sums of convex functions. Alternatively, one might also use results from monotropic programming, cf. e.g. [@Bertsekas:99:1].)
Because for each $i \in {\mathcal{I}}$ the variable $\phi'_i$ is involved in only one inequality, the optimal solution does not change if we replace the corresponding inequality signs by equality signs in **DLPD2**. The same comment holds for all $\theta'_j$, $j \in {\mathcal{J}}$.
Let ${\mathcal{A}} {\triangleq}{\mathcal{A}}_1 \times \cdots \times {\mathcal{A}}_n$ and let ${\mathcal{B}} {\triangleq}{\mathcal{B}}_1 \times \cdots \times {\mathcal{B}}_m$. For ${{\mathbf{a}}}{\triangleq}({{\mathbf{a}}}_1, \ldots, {{\mathbf{a}}}_n) \in {\mathcal{A}}$ and ${{\mathbf{b}}}{\triangleq}({{\mathbf{b}}}_1,
\ldots, {{\mathbf{b}}}_m) \in {\mathcal{B}}$ define $$\begin{aligned}
g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')
&{\triangleq}\sum_{i \in {\mathcal{I}}}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
+
\sum_{j \in {\mathcal{J}}}
\langle
+{{\mathbf{u}}}'_j, {{\mathbf{b}}}_j
\rangle,
\end{aligned}$$ where, with a slight abuse of notation, ${{\mathbf{u}}}'_j$ is such that $[{{\mathbf{u}}}'_j]_i =
u'_{i,j}$ for all $(i,j) \in {\mathcal{E}}$. Moreover, we call $({{\mathbf{a}}},{{\mathbf{b}}}) \in
{\mathcal{A}} \times {\mathcal{B}}$ consistent if $a_{i,j} = b_{j,i}$ for all $(i,j) \in
{\mathcal{E}}$. [$\square$]{}
Obviously, $g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')$ is a linear function in ${{\mathbf{u}}}'$. With the above definition, **DLPD2** can be rewritten to read
[\
\
]{}
\[lemma:consistent:assignment:1\]
Let ${{\mathbf{u}}}'$ be such that $u'_{i,0} = -\lambda_i$, $i \in {\mathcal{I}}$. If $({{\mathbf{a}}},{{\mathbf{b}}}) \in {\mathcal{A}} \times {\mathcal{B}}$ is consistent then $g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')$ is constant in ${{\mathbf{u}}}'$. Moreover, $g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}') =
\langle {\boldsymbol{\lambda}}, {{\mathbf{x}}}\rangle$, where ${{\mathbf{x}}}$ is such that $x_i = a_{i,0}$, $i
\in {\mathcal{I}}$. If $({{\mathbf{a}}},{{\mathbf{b}}}) \in {\mathcal{A}} \times {\mathcal{B}}$ is not consistent then $g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')$ is not a constant function for at least one $u_{i,j}$, $(i,j) \in {\mathcal{E}}$.
See Sec. \[sec:proof:lemma:consistent:assignment:1\].
A Softened Dual Linear Program {#sec:softened:dual:linear:program:1}
==============================
For any $\kappa \in {\mathbb{R}_{++}}$, we define the soft-minimum operator to be $$\begin{aligned}
{\min_{\ell}}^{(\kappa)}
z_{\ell}
&{\triangleq}-
\frac{1}{\kappa}
\log
\left(
\sum_{\ell}
{\mathrm{e}}^{-\kappa z_{\ell}}
\right).\end{aligned}$$ (Note that $\kappa$ can be given the interpretation of an inverse temperature.) One can easily check that ${\min_{\ell}}^{(\kappa)} z_{\ell}
\leq \min_{\ell} \{ z_{\ell} \}$ with equality in the limit $\kappa \to
+\infty$. Replacing the minimum operators in **DLPD2** by soft-minimum operators, we obtain the modified optimization problem
[\
\
]{}
In the following, unless noted otherwise, we will set $\kappa_i
{\triangleq}\kappa$, $i \in {\mathcal{I}}$, and $\kappa_j {\triangleq}\kappa$, $j \in
{\mathcal{J}}$, for some $\kappa \in {\mathbb{R}_{++}}$. It is clear that in the limit $\kappa
\to +\infty$ we recover **DLPD2**.
A Comment on the Dual of the Softened Dual Linear Program {#sec:primal:softened:dual:linear:program:1}
=========================================================
Let $$\begin{aligned}
H(\alpha_i)
&{\triangleq}-
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\alpha_{i,{{\mathbf{a}}}_i}
\log(\alpha_{i,{{\mathbf{a}}}_i})\end{aligned}$$ be the entropy of of a random variable whose pmf takes on the values $\{
\alpha_{i,{{\mathbf{a}}}_i} \}_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}$. Similarly, let $$\begin{aligned}
H(\beta_j)
&{\triangleq}-
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\beta_{j,{{\mathbf{b}}}_j}
\log(\beta_{j,{{\mathbf{b}}}_j}).\end{aligned}$$ The dual of **SDLDP2** can then be written as
[\
\
]{}
We note that this is very close to the following Bethe free energy optimization problem, cf. e.g. [@Yedidia:Freeman:Weiss:05:1]
[\
\
]{}
which, in turn, can also be written as
[\
\
]{}
Without going into the details we note that the term $+\frac{1}{\kappa} \sum_{i \in {\mathcal{I}}} (|{\mathcal{J}}_i|-1) H(\alpha_i)$ is responsible for the fact that the cost function in **BFE2** is usually non-convex for FFGs with cycles.
Decoding Algorithm 1 {#sec:decoding:algorithm:1}
====================
In the following, we assume that $u'_{i,j}$ and $v'_{j,i}$ are “coupled”, i.e. we always have $u'_{i,j} = - v'_{j,i}$ for all $(i,j) \in {\mathcal{E}}$.
The first algorithm that we propose is a coordinate-ascent-type algorithm for solving **SDLPD2**. The main idea is to select edges $(i,j) \in {\mathcal{E}}$ according to some update schedule: for each selected edge $(i,j) \in {\mathcal{E}}$ we then replace the old values of $u'_{i,j}$, $\phi'_i$, and $\theta'_j$ by new values such that the dual cost function is increased (or at least not decreased). Practically, this means that we have to find an $\overline{u}'_{i,j}$ such that $h'(\overline{u}'_{i,j}) \geq h'(u'_{i,j})$, where $$\begin{aligned}
h'(u'_{i,j})
&{\triangleq}{\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}}^{(\kappa)}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
+
{\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}}^{(\kappa)}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle.\end{aligned}$$ A simple way to achieve this is by setting $$\begin{aligned}
\overline{u}'_{i,j}
&{\triangleq}\arg \max_{u'_{i,j}}
h'(u'_{i,j}).
\label{eq:coordinate:ascent:edge:update:1:1}\end{aligned}$$ The variables $\phi'_i$ and $\theta'_j$ are then updated accordingly so that we obtain a new (dual) feasible point.
\[lemma:coordinate:ascent:update:rule:1\]
The value of $\overline{u}'_{i,j}$ in is given by $$\begin{aligned}
\overline{u}'_{i,j}
&= \frac{1}{2}
\bigg(
+
\big(
S'_{i,0}
-
S'_{i,1}
\big)
-
\big(
T'_{j,0}
-
T'_{j,1}
\big)
\bigg),
\end{aligned}$$ where $$\begin{aligned}
S'_{i,0}
&{\triangleq}-
{\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i \atop a_{i,j} = 0}}^{(\kappa)}
\langle
-{\tilde{{\mathbf{u}}}}_i, {\tilde{{\mathbf{a}}}}_i
\rangle, \\
S'_{i,1}
&{\triangleq}-
{\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i \atop a_{i,j} = 1}}^{(\kappa)}
\langle
-{\tilde{{\mathbf{u}}}}_i, {\tilde{{\mathbf{a}}}}_i
\rangle, \\
T'_{j,0}
&{\triangleq}-
{\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j \atop b_{j,i} = 0}}^{(\kappa)}
\langle
-
{\tilde{{\mathbf{v}}}}_j, {\tilde{{\mathbf{b}}}}_j
\rangle, \\
T'_{j,1}
&{\triangleq}-
{\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j \atop b_{j,i} = 1}}^{(\kappa)}
\langle
-{\tilde{{\mathbf{v}}}}_j, {\tilde{{\mathbf{b}}}}_j
\rangle.
\end{aligned}$$ Here the vectors ${\tilde{{\mathbf{u}}}}$ and ${\tilde{{\mathbf{a}}}}$ are the vectors ${{\mathbf{u}}}$ and ${{\mathbf{a}}}$, respectively, where the $j$-th position has been omitted. Similarly, the vectors ${\tilde{{\mathbf{v}}}}$ and ${\tilde{{\mathbf{b}}}}$ are the vectors ${{\mathbf{v}}}$ and ${{\mathbf{b}}}$, respectively, where the $i$-th position has been omitted. Note that the differences $S'_{i,0} - S'_{i,1}$ and $T'_{i,0} - T'_{i,1}$, which are required for computing $\overline{u}'_{i,j}$, can be obtained very efficiently by using the sum-product algorithm [@Kschischang:Frey:Loeliger:01].
See Sec. \[sec:proof:lemma:coordinate:ascent:update:rule:1\].
In the introduction we wrote that we would like to use the special structure of the primal/dual LP at hand; Lemma \[lemma:coordinate:ascent:update:rule:1\] is a first example how this can be done. Please note that when computing the necessary quantities (for the case $\kappa = 1$) one has do computations that are (up to some flipped signs) equivalent to computations that are done during message updates while performing sum-product algorithm decoding of the LDPC code at hand.
\[lemma:coodinate:ascent:convergence:1\]
Assume that all the rows of the parity-check matrix ${\mathbf{H}}$ of the code ${\mathcal{C}}$ have Hamming weight at least $3$.[^5] Then, updating cyclically all edges $(i,j) \in {\mathcal{E}}$, the above coordinate-ascent algorithm converges to the maximum of **SDLPD2**.
See Sec. \[sec:proof:lemma:coodinate:ascent:convergence:1\]
As we mentioned in the proof of Lemma \[lemma:coodinate:ascent:convergence:1\], the above algorithm can be seen as a Gauss-Seidel-type algorithm. Let us remark that there are ways to see sum-product algorithm decoding as applying a Gauss-Seidel-type algorithm to the dual of the Bethe free energy, see e.g. [@Tan:Rasmussen:05:1; @Walsh:Regalia:Johnson:05:1]; in light of the observations in Sec. \[sec:primal:softened:dual:linear:program:1\] it is not surprising that there is a tight relationship between our algorithms and the above-mentioned algorithms.
\[lemma:coordinate:ascent:update:rule:2\]
For $\kappa \to \infty$, the function $h'(u'_{i,j})$ is maximized by any value $u'_{i,j}$ that lies in the closed interval between $$\begin{aligned}
\big(
S'_{i,0}
-
S'_{i,1}
\big)
\quad \text{ and } \quad
-
\big(
T'_{j,0}
-
T'_{j,1}
\big),
\end{aligned}$$ where $$\begin{aligned}
S'_{i,0}
&{\triangleq}-
\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i \atop a_{i,j} = 0}
\langle
-{\tilde{{\mathbf{u}}}}_i, {\tilde{{\mathbf{a}}}}_i
\rangle, \\
S'_{i,1}
&{\triangleq}-
\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i \atop a_{i,j} = 1}
\langle
-{\tilde{{\mathbf{u}}}}_i, {\tilde{{\mathbf{a}}}}_i
\rangle, \\
T'_{j,0}
&{\triangleq}-
\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j \atop b_{j,i} = 0}
\langle
-
{\tilde{{\mathbf{v}}}}_j, {\tilde{{\mathbf{b}}}}_j
\rangle, \\
T'_{j,1}
&{\triangleq}-
\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j \atop b_{j,i} = 1}
\langle
-{\tilde{{\mathbf{v}}}}_j, {\tilde{{\mathbf{b}}}}_j
\rangle.
\end{aligned}$$
See Sec. \[sec:proof:lemma:coordinate:ascent:update:rule:2\].
\[conj:coodinate:ascent:convergence:2\]
Again, we can cyclically update the edges $(i,j) \in {\mathcal{E}}$ whereby the new $\overline{u}'_{i,j}$ is chosen randomly in the above interval. Although the objective function for $\kappa \to +\infty$ is concave, it is not everywhere differentiable. This makes a convergence proof in the style of Lemma \[lemma:coodinate:ascent:convergence:1\] difficult. We think that we can again use the special structure of the LP at hand to show that the algorithm cannot get stuck at a suboptimal point. However, so far we do not have a proof of this fact. Sec. \[sec:considerations:conj:coodinate:ascent:convergence:2\] discusses briefly why a convergence proof is not a trivial extension of Lemma \[lemma:coodinate:ascent:convergence:1\]. [$\square$]{}
Before ending this section, let us briefly remark how a codeword decision is obtained from a solution of **DLPD2**. Assume that $\hat {{\mathbf{x}}}$ is the pseudo-codeword that is the solution to **PLPD1** or to **PLPD2**.[^6] Knowing the solution of **DLPD2** we cannot directly find $\hat {{\mathbf{x}}}$, however, we can find out at what positions $\hat {{\mathbf{x}}}$ is $0$ and at what positions $\hat {{\mathbf{x}}}$ is $1$. Namely, letting $\check {{\mathbf{x}}}\in \{ 0, ?,
1 \}^n$ have the components $$\begin{aligned}
\check x_i
{\triangleq}\begin{cases}
0 & \text{ if }
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{0}}}
\ < \
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{1}}} \\
? & \text{ if }
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{0}}}
\ = \
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{1}}} \\
1 & \text{ if }
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{0}}}
\ > \
\langle
- {{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
|_{{{\mathbf{a}}}_i = {\mathbf{1}}}
\end{cases},\end{aligned}$$ we have $\check x_i = \hat x_i$ when $\hat x_i$ equals $0$ or $1$ and $\check
x_i = \ ?$ when $\hat x_i \in (0,1)$. In other words, with the solution to **DLPD2** we do not get the exact $\hat {{\mathbf{x}}}$ in case $\hat {{\mathbf{x}}}$ is not a codeword. However, as a side remark, because ${\operatorname{supp}}(\hat {{\mathbf{x}}}) = {\operatorname{supp}}(\check
{{\mathbf{x}}})$ (where ${\operatorname{supp}}$ is the set of all non-zero positions) we can use $\check
{{\mathbf{x}}}$ to find the stopping set [@Di:Proietti:Telatar:Richardson:Urbanke:02:1] associated to $\hat
{{\mathbf{x}}}$.
Decoding Algorithm 2 {#sec:decoding:algorithm:2}
====================
Again, we assume that $u'_{i,j}$ and $v'_{j,i}$ are “coupled”, i.e. we always have $u'_{j,i} = - v'_{i,j}$ for all $(i,j) \in {\mathcal{E}}$.
While the iterative solutions of the coordinate-ascent methods that we presented in the previous section resemble the traditional min-sum algorithm decoding rules (and sum-product algorithm decoding rules) relatively closely, other methods for solving the linear program also offer attractive complexity/performance trade-offs. We would like to point out one such algorithm which is well suited for the linear programming problem arising from the decoding setup. Indeed, observing the formulation of the dual linear program **DLPD2**, sub-gradient methods[^7] are readily available to perform the required maximization. However, in order to exploit the structure of the problem we focus our attention to incremental sub-gradient methods [@Nedic:02:1]. Algorithms belonging to this family of optimization procedures allow us to exploit the fact that the objective function is a sum of a number of terms and we can operate on each term, i.e. each constituent code in the FFG, individually. In order to derive a concise formulation of the procedure we start by considering a check node $j \in {\mathcal{J}}$. For a particular choice of dual variables ${{\mathbf{v}}}'_j$ the contribution of node $j$ to the overall objective function is $$\begin{aligned}
f'_j({{\mathbf{v}}}'_j)
&= \min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle.\end{aligned}$$ Let a function ${{\mathbf{g}}}'({{\mathbf{v}}}'_j)$ be defined as ${{\mathbf{g}}}'_j({{\mathbf{v}}}'_j) {\triangleq}- \arg
\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j} \langle -{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j \rangle$ where, if ambiguities exist, ${{\mathbf{g}}}'_j({{\mathbf{v}}}'_j)$ is the negative of an arbitrary combination of the set of ambiguous vectors ${{\mathbf{b}}}'_j$. Note that for obtaining ${{\mathbf{g}}}'_j({{\mathbf{v}}}'_j)$ we can again take advantage of the special structure of the LP at hand.
Using the defining property of sub-gradient ${{\mathbf{d}}}'_j$ at ${{\mathbf{v}}}'_j$, namely, $$\begin{aligned}
f'({{\mathbf{\breve{v}}}}'_j)
&\leq f({{\mathbf{v}}}'_j)
+
\langle
{{\mathbf{d}}}'_j,
{{\mathbf{\breve{v}}}}'_j - {{\mathbf{v}}}'_j)\end{aligned}$$ it can be seen that ${{\mathbf{g}}}'_j({{\mathbf{v}}}'_j)$ is a sub-gradient. We can then update ${{\mathbf{v}}}'_j$ as follows: $$\begin{aligned}
{{\mathbf{v}}}'_j
&\leftarrow
{{\mathbf{v}}}'_j
+
\mu_{\ell} {{\mathbf{g}}}'_j({{\mathbf{v}}}'_j),\end{aligned}$$ where $\mu_{\ell} \in {\mathbb{R}_{++}}$. Given this, one can formulate the following algorithm: at iteration $\ell$ update consecutively all check nodes $j \in {\mathcal{J}}$ and then, in an analogous manner, update all variable nodes $i
\in {\mathcal{I}}$.
For this algorithm we cannot guarantee that the value of the objective function increases for each iteration (not even for small $\mu_{\ell}$). Nevertheless, its convergence to the maximum can be guaranteed for a suitably chosen sequence $\{ \alpha_\ell \}_{\ell \geq
1}$ [@Nedic:02:1].
Let us point out that gradient-type methods have also been used to decode codes in different contexts, see e.g. the work by Lucas et al. [@Lucas:Bossert:Breitbach:98:1]. However, the setup in [@Lucas:Bossert:Breitbach:98:1] has some significant differences to the setup here: firstly, the objective function of the optimization problem in [@Lucas:Bossert:Breitbach:98:1] does not depend on the observed log-liklihood ratio vector ${\boldsymbol{\lambda}}$, secondly, the starting point in [@Lucas:Bossert:Breitbach:98:1] is chosen as a function of ${\boldsymbol{\lambda}}$.
Simulation Results {#sec:simulation:results:1}
==================
As a proof of concept we show some simulation results for a randomly generated $(3,6)$-regular $[1000,500]$ LDPC code where four-cycles in the Tanner graph have been eliminated. Fig. \[fig:simulation:results:1:1\] shows the decoding results based on Decoding Algorithm $1$ with update rule Lemma \[lemma:coordinate:ascent:update:rule:2\] compared with standard min-sum algorithm decoding [@Kschischang:Frey:Loeliger:01].
Conclusions {#sec:conclusions:1}
===========
We have discussed some initial steps towards algorithms that are specially targeted for efficiently solving the LP that appears in LP decoding. It has been shown that algorithms with memory and time complexity similar to min-sum algorithm decoding can be achieved. There are many avenues to pursue this topic further, e.g. by improving the update schedule, by studying how to design codes that allow efficient hardware implementation of the proposed algorithms, or by investigating other algorithms that use the structure of the LP that appears in LP decoding. We hope that this paper raises the interest in exploring these research directions.
Finally, without going into the details, let us remark that the algorithms here can also be used to solve certain linear programs whose value can be used to obtain lower bounds on the minimal AWGNC pseudo-weight of parity-check matrices, cf. [@Vontobel:Koetter:04:1 Claim 3]. (Actually, one does not really need to solve the linear program in [@Vontobel:Koetter:04:1 Claim 3] in order to obtain a lower bound on the minimum AWGNC pseudo-weight, any dual feasible point is good enough for that purpose.)
Proof of Lemma \[lemma:consistent:assignment:1\] {#sec:proof:lemma:consistent:assignment:1}
------------------------------------------------
If $({{\mathbf{a}}},{{\mathbf{b}}})$ is consistent then $$\begin{aligned}
g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')
&= \sum_{i \in {\mathcal{I}}}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
+
\sum_{j \in {\mathcal{J}}}
\langle
+{{\mathbf{u}}}'_j, {{\mathbf{b}}}_j
\rangle \\
&= -
\sum_{i \in {\mathcal{I}}}
u'_{i,0} a_{i,0}
-
\sum_{(i,j) \in {\mathcal{E}}}
u'_{i,j} a_{i,j} \\
&\quad\,
+
\sum_{(i,j) \in {\mathcal{E}}}
u'_{i,j} b_{j,i} \\
&= -
\sum_{i \in {\mathcal{I}}}
u'_{i,0} a_{i,0} \\
&= \sum_{i \in {\mathcal{I}}}
\lambda_i x_i.\end{aligned}$$ On the other hand, if $({{\mathbf{a}}},{{\mathbf{b}}})$ is not consistent and $(i,j) \in {\mathcal{E}}$ is such that $a_{i,j} \neq b_{j,i}$ then $g'_{{{\mathbf{a}}},{{\mathbf{b}}}}({{\mathbf{u}}}')$ is non-constant in $u'_{i,j}$.
Proof of Lemma \[lemma:coordinate:ascent:update:rule:1\] {#sec:proof:lemma:coordinate:ascent:update:rule:1}
--------------------------------------------------------
This result is obtained by taking the derivative of $h'(u'_{i,j})$, setting it equal to zero, and solving for $u'_{i,j}$. Let us go through this procedure step by step. Using the fact that $u'_{i,j} = -v'_{j,i}$, the function $h'(u'_{i,j})$ can be written as $$\begin{aligned}
h'(u'_{i,j})
&{\triangleq}{\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}}^{(\kappa)}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
+
{\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}}^{(\kappa)}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle \\
&= -
\frac{1}{\kappa}
\log
\left(
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
{\mathrm{e}}^
{
+
\kappa
\langle
{{\mathbf{u}}}_i, {{\mathbf{a}}}_i
\rangle
}
\right) \\
&\quad\,
-
\frac{1}{\kappa}
\log
\left(
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
{\mathrm{e}}^
{
+
\kappa
\langle
{{\mathbf{v}}}_j, {{\mathbf{b}}}_j
\rangle
}
\right) \\
&= -
\frac{1}{\kappa}
\log
\left(
\sum_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
{\mathrm{e}}^
{
+
\kappa u_{i,j} a_{i,j}
+
\kappa
\langle
\tilde {{\mathbf{u}}}_i, {\tilde{{\mathbf{a}}}}_i
\rangle
}
\right) \\
&\quad\,
-
\frac{1}{\kappa}
\log
\left(
\sum_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
{\mathrm{e}}^
{
-
\kappa u_{i,j} b_{j,i}
+
\kappa
\langle
\tilde {{\mathbf{v}}}_j, {\tilde{{\mathbf{b}}}}_j
\rangle
}
\right) \\
&= -
\frac{1}{\kappa}
\log
\left(
{\mathrm{e}}^{\kappa S'_{i,0}}
+
{\mathrm{e}}^{+\kappa u'_{i,j}}
{\mathrm{e}}^{\kappa S'_{i,1}}
\right) \\
&\quad\,
-
\frac{1}{\kappa}
\log
\left(
{\mathrm{e}}^{\kappa T'_{j,0}}
+
{\mathrm{e}}^{-\kappa u'_{i,j}}
{\mathrm{e}}^{\kappa T'_{j,1}}
\right)\end{aligned}$$ Setting the derivative of $h'(u'_{i,j})$ with respect to $u'_{i,j}$ equal to zero we obtain $$\begin{aligned}
0
&\overset{!}{=}
\frac{\partial h'(u'_{i,j})}
{\partial u'_{i,j}} \\
&= -
\frac{1}{\kappa}
\cdot
\frac{+\kappa {\mathrm{e}}^{+\kappa u'_{i,j}} {\mathrm{e}}^{\kappa S'_{i,1}}}
{{\mathrm{e}}^{\kappa S'_{i,0}}
+
{\mathrm{e}}^{+\kappa u'_{i,j}}
{\mathrm{e}}^{\kappa S'_{i,1}}
} \\
&\quad\,
-
\frac{1}{\kappa}
\cdot
\frac{-\kappa {\mathrm{e}}^{-\kappa u'_{i,j}} {\mathrm{e}}^{\kappa T'_{j,1}}}
{
{\mathrm{e}}^{\kappa T'_{j,0}}
+
{\mathrm{e}}^{-\kappa u'_{i,j}}
{\mathrm{e}}^{\kappa T'_{j,1}}
}.\end{aligned}$$ Multiplying out we get $$\begin{aligned}
&
{\mathrm{e}}^{+\kappa u'_{i,j}} {\mathrm{e}}^{\kappa (S'_{i,1} + T'_{j,0})}
+
{\mathrm{e}}^{\kappa (S'_{i,1} + T'_{j,1})} \\
&\quad
= {\mathrm{e}}^{-\kappa u'_{i,j}} {\mathrm{e}}^{\kappa (S'_{i,0} + T'_{j,1})}
+
{\mathrm{e}}^{\kappa (S'_{i,1} + T'_{j,1})}.\end{aligned}$$ This yields $$\begin{aligned}
\overline{u}'_{i,j}
&= \frac{1}{2}
\bigg(
+
\big(
S'_{i,0}
-
S'_{i,1}
\big)
-
\big(
T'_{j,0}
-
T'_{j,1}
\big)
\bigg),\end{aligned}$$ which is the promised result.
Proof of Lemma \[lemma:coodinate:ascent:convergence:1\] {#sec:proof:lemma:coodinate:ascent:convergence:1}
-------------------------------------------------------
We can use results from [@Bertsekas:99:1 Sec. 2.7], where the following setup is considered.[^8] Consider the optimization problem
[\
\
]{}
where ${\mathcal{X}} {\triangleq}{\mathcal{X}}_1 \times \cdots \times {\mathcal{X}}_m$. The set ${\mathcal{X}}_i$ is assumed to be a closed convex subset of ${\mathbb{R}}^{n_i}$ and $n =
n_1 + \cdots + n_m$. The vector ${{\mathbf{x}}}$ is partitioned as ${{\mathbf{x}}}= ({{\mathbf{x}}}_1, \ldots,
{{\mathbf{x}}}_m)$ where each ${{\mathbf{x}}}_i \in {\mathbb{R}}^{n_i}$. So the constraint ${{\mathbf{x}}}\in {\mathcal{X}}$ is equivalent to ${{\mathbf{x}}}_i \in {\mathcal{X}}_i$, $i \in \{ 1, \ldots, m \}$.
The following algorithm, known as block coordinate-ascent or non-linear Gauss-Seidel method, generates the next iterate ${{\mathbf{x}}}^{k+1} {\triangleq}({{\mathbf{x}}}_1^{k+1}, \ldots, {{\mathbf{x}}}_m^{k+1})$, given the current iterate ${{\mathbf{x}}}^k {\triangleq}({{\mathbf{x}}}_1^k, \ldots, {{\mathbf{x}}}_m^k)$ according to the iteration $$\begin{aligned}
{{\mathbf{x}}}_i^{k+1}
&{\triangleq}\arg \max_{{\boldsymbol{\xi}}_i \in {\mathcal{X}}_i}
\nonumber \\
&\quad\quad
f({{\mathbf{x}}}_1^{k+1}, \ldots, {{\mathbf{x}}}_{i-1}^{k+1},
{\boldsymbol{\xi}}_i,
{{\mathbf{x}}}_{i+1}^k, \ldots, {{\mathbf{x}}}_m^k).
\label{eq:bertsekas:2}\end{aligned}$$
Suppose that $f$ is continuously differentiable over the set ${\mathcal{X}}$. Furthermore, suppose that for each $i$ and ${{\mathbf{x}}}\in {\mathcal{X}}$, the maximum below $$\begin{aligned}
\max_{{\boldsymbol{\xi}}_i \in {\mathcal{X}}_i}
f({{\mathbf{x}}}_1, \ldots, {{\mathbf{x}}}_{i-1}, {\boldsymbol{\xi}}_i, {{\mathbf{x}}}_{i+1}, \ldots, {{\mathbf{x}}}_m)
\end{aligned}$$ is uniquely attained. Let $\{ {{\mathbf{x}}}^k \}$ be the sequence generated by the block coordinate-ascent method . Then every limit point of $\{ {{\mathbf{x}}}^k \}$ is a stationary point. [$\square$]{}
We turn our attention now to our optimization problem. The fundamental polytope (which is the set $\bigcap_{j \in {\mathcal{J}}} {\operatorname{conv}}({\mathcal{C}}_j)$), has dimension $n$ if and only if the parity-check matrix has no rows of Hamming weight $1$ and $2$. This type of non-degeneracy of **PLPD2** implies the strict concavity of the function that we try to optimize in **SDLPD2**. Based on ${\boldsymbol{\lambda}}$ one can then without loss of generality define suitable closed intervals for each variable so that one can apply the above proposition to our algorithm.
Proof of Lemma \[lemma:coordinate:ascent:update:rule:2\] {#sec:proof:lemma:coordinate:ascent:update:rule:2}
--------------------------------------------------------
Define the functions $$\begin{aligned}
s'(u'_{i,j})
&{\triangleq}\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle
\quad \text{ and } \\
t'(u'_{i,j})
&{\triangleq}\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle\end{aligned}$$ such that $h'(u'_{i,j}) = s'(u'_{i,j}) + t'(u'_{i,j})$. Then $$\begin{aligned}
s'(u'_{i,j})
&{\triangleq}\min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
\langle
-{{\mathbf{u}}}'_i, {{\mathbf{a}}}_i
\rangle \\
&= \min_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}
-
u'_{i,j} a_{i,j}
-
\langle
\tilde {{\mathbf{u}}}'_i, {\tilde{{\mathbf{a}}}}_i
\rangle \\
&= \min
\big(
-S'_{i,0}, \ - u'_{i,j} {-} S'_{i,1}
\big), \\
t'(u'_{i,j})
&{\triangleq}\min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
\langle
-{{\mathbf{v}}}'_j, {{\mathbf{b}}}_j
\rangle \\
&= \min_{{{\mathbf{b}}}_j \in {\mathcal{B}}_j}
+
u'_{i,j} b_{j,i}
-
\langle
\tilde {{\mathbf{v}}}'_j, {\tilde{{\mathbf{b}}}}_j
\rangle \\
&= \min
\big(
-T'_{i,0}, \ + u'_{i,j} {-} T'_{i,1}
\big).\end{aligned}$$ As can be seen from Fig. \[fig:critical:points:1:1\], the functions $s'(u'_{i,j})$ and $t'(u'_{i,j})$ are both piece-wise linear functions. Whereas the function $s'(u'_{i,j})$ is flat up to $u'_{i,j} =
S'_{i,0} - S'_{i,1}$ and then has slope $-1$, the function $t'(u'_{i,j})$ increases with slope $+1$ up to $u'_{i,j} = T'_{i,1} - T'_{i,0}$ and is then flat. From Fig. \[fig:critical:points:1:1\] is can also be seen that, independently if $S'_{i,0} - S'_{i,1}$ is larger or smaller than $T'_{i,1} -
T'_{i,0}$, the function $h'(u'_{i,j})$ always consists of three parts: first it increases with slope $+1$, then it is flat, and finally it decreases with slope $-1$. From this observations, the lemma statement follows.
Comment to Conjecture \[conj:coodinate:ascent:convergence:2\] {#sec:considerations:conj:coodinate:ascent:convergence:2}
-------------------------------------------------------------
This section briefly discusses a concave function where a coordinate-ascent approach does not find the global maximum. Let $0 < a < 1$ and let $$\begin{aligned}
f(x_1, x_2)
{\triangleq}\min \big(\!
& - x_1 + x_2 + a (x_1 + x_2), \\
& + x_1 - x_2 + a (x_1 + x_2)
\big).\end{aligned}$$ The level curves of $f(x_1, x_2)$ are shown in Fig. \[fig:problematic:contour:1:1\]. By choosing $(x_1, x_2) {\triangleq}(\alpha, \alpha)$ and letting $\alpha$ go to $\infty$ we see that this function is unbounded.
Consider now the optimization problem
[\
\
]{}
where ${\mathcal{X}}$ is some suitably chosen closed convex subset of ${\mathbb{R}}^2$. Assume that a coordinate-ascent-type method has e.g. found the point $(x_1, x_2) = (0,0)$ with $f(0,0) = 0$. (Of course, we assume that $(0,0) \in
{\mathcal{X}}$.) Unfortunately, at this point the coordinate-ascent-type method cannot make any progress because $f(x_1,0) = \min \big(-(1-a)x_1, (1+a) x_1
\big) < 0$ for all $x_1 \neq 0$ and $f(0,x_2) = \min \big((1+a)x_2, -(1-a)x_2
\big) < 0$ for all $x_2 \neq 0$.
However, defining $$\begin{aligned}
f^{(\kappa)}(x_1, x_2)
{\triangleq}{\min}^{(\kappa)}
\big(\!
& - x_1 + x_2 + a (x_1 + x_2), \\
& + x_1 - x_2 + a (x_1 + x_2)
\big),\end{aligned}$$ where $\kappa \in {\mathbb{R}_{++}}$ is arbitrary, a coordinate-ascent method can successfully be used for the “softened” optimization problem
[\
\
]{}
[10]{}
J. Feldman, [*Decoding Error-Correcting Codes via Linear Programming*]{}. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 2003. Available online under `http://www.columbia.edu/~jf2189/pubs.html`.
J. Feldman, M. J. Wainwright, and D. R. Karger, “Using linear programming to decode binary linear codes,” [*IEEE Trans. on Inform. Theory*]{}, vol. IT–51, no. 3, pp. 954–972, 2005.
R. Koetter and P. O. Vontobel, “Graph covers and iterative decoding of finite-length codes,” in [*Proc. 3rd Intern. Symp. on Turbo Codes and Related Topics*]{}, (Brest, France), pp. 75–82, Sept. 1–5 2003.
P. O. Vontobel and R. Koetter, “Graph-cover decoding and finite-length analysis of message-passing iterative decoding of [LDPC]{} codes,” [ *submitted to IEEE Trans. Inform. Theory, available online under*]{} `http://www.arxiv.org/abs/` `cs.IT/0512078`, Dec. 2005.
P. O. Vontobel and R. Koetter, “On the relationship between linear programming decoding and min-sum algorithm decoding,” in [*Proc. Intern. Symp. on Inform. Theory and its Applications (ISITA)*]{}, (Parma, Italy), pp. 991–996, Oct. 10–13 2004.
J. Feldman, D. R. Karger, and M. J. Wainwright, “Linear programming-based decoding of turbo-like codes and its relation to iterative approaches,” in [*Proc. 40th Allerton Conf. on Communications, Control, and Computing*]{}, (Allerton House, Monticello, Illinois, USA), October 2–4 2002. Available online under `http://www.columbia.edu/~jf2189/` `pubs.html`.
K. Yang, X. Wang, and J. Feldman, “Non-linear programming approaches to decoding low-density parity-check codes,” in [*Proc. 43rd Allerton Conf. on Communications, Control, and Computing*]{}, (Allerton House, Monticello, Illinois, USA), Sep. 28–30 2005.
P. O. Vontobel and R. Koetter, “Lower bounds on the minimum pseudo-weight of linear codes,” in [*Proc. IEEE Intern. Symp. on Inform. Theory*]{}, (Chicago, IL, USA), p. 70, June 27–July 2 2004.
P. Chaichanavong and P. H. Siegel, “Relaxation bounds on the minimum pseudo-weight of linear block codes,” in [*Proc. IEEE Intern. Symp. on Inform. Theory*]{}, (Adelaide, Australia), pp. 805–809, Sep. 4–9 2005. Available online under `http://www.arxiv.org/abs/cs.IT/0508046`.
P. O. Vontobel and R. Smarandache, “On minimal pseudo-codewords of [T]{}anner graphs from projective planes,” in [*Proc. 43rd Allerton Conf. on Communications, Control, and Computing*]{}, (Allerton House, Monticello, Illinois, USA), Sep. 28–30 2005. Available online under `http://www.arxiv.org/` `abs/cs.IT/0510043`.
F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and the sum-product algorithm,” [*IEEE Trans. on Inform. Theory*]{}, vol. IT–47, no. 2, pp. 498–519, 2001.
G. D. [Forney, Jr.]{}, “Codes on graphs: normal realizations,” [*IEEE Trans. on Inform. Theory*]{}, vol. IT–47, no. 2, pp. 520–548, 2001.
H.-A. Loeliger, “An introduction to factor graphs,” [*IEEE Sig. Proc. Mag.*]{}, vol. 21, no. 1, pp. 28–41, 2004.
D. Bertsimas and J. N. Tsitsiklis, [*Linear Optimization*]{}. Belmont, MA: Athena Scientific, 1997.
P. O. Vontobel, [*Kalman Filters, Factor Graphs, and Electrical Networks*]{}. Post-Diploma Project, ETH Zurich, 2002. Available online under `http://``www.isi.ee.ethz.ch/` `publications`.
P. O. Vontobel and H.-A. Loeliger, “On factor graphs and electrical networks,” in [*Mathematical Systems Theory in Biology, Communication, Computation, and Finance, IMA Volumes in Math. & Appl.*]{} (D. Gilliam and J. Rosenthal, eds.), Springer Verlag, 2003.
D. Bertsekas, [*Nonlinear Programming*]{}. Belmont, MA: Athena Scientific, second ed., 1999.
J. S. Yedidia, W. T. Freeman, and Y. Weiss, “Constructing free-energy approximations and generalized belief propagation algorithms,” [*IEEE Trans. on Inform. Theory*]{}, vol. IT–51, no. 7, pp. 2282–2312, 2005.
P. H. Tan and L. K. Rasmussen, “The serial and parallel belief propagation algorithms,” in [*Proc. IEEE Intern. Symp. on Inform. Theory*]{}, (Adelaide, Australia), pp. 729 – 733, Sep. 4–9 2005.
J. M. Walsh, P. Regalia, and C. R. [Johnson Jr.]{}, “A convergence proof for the turbo decoder as an instance of the [G]{}auss-[S]{}eidel iteration,” in [ *Proc. IEEE Intern. Symp. on Inform. Theory*]{}, (Adelaide, Australia), pp. 734–738, Sep. 4–9 2005.
C. Di, D. Proietti, [I]{}. E. Telatar, T. J. Richardson, and R. L. Urbanke, “Finite-length analysis of low-density parity-check codes on the binary erasure channel,” [*IEEE Trans. on Inform. Theory*]{}, vol. IT–48, no. 6, pp. 1570–1579, 2002.
A. Nedi[ć]{}, [*Subgradient Methods for Convex Minimization*]{}. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 2002.
R. Lucas, M. Bossert, and M. Breitbach, “On iterative soft-decision decoding of linear binary block codes and product codes,” [*IEEE J. Sel. Areas Comm.*]{}, vol. JSAC–16, no. 2, pp. 276–296, 1998.
[^1]: P.O.V.’s research was supported by NSF Grants TF 05-14801, ATM-0296033, DOE SciDAC, and ONR Grant N00014-00-1-0966. The research for this paper was partly done while being at the Dept. of ECE, University of Wisconsin-Madison, Madison, WI 53706, USA. R.K.’s research was supported by NSF Grants CCR 99-84515, CCR 01-05719, and TF 05-14869.
[^2]: For the codes ${\mathcal{C}}$ under consideration this means that ${\mathcal{B}}_j$ contains all vectors of length $|{\mathcal{I}}_j|$ of even parity.
[^3]: Note that instead of drawing function nodes for the terms that appear in the definition of $A_i({{\mathbf{u}}}_i)$ and an edge for the variables $\{ \alpha_{i,{{\mathbf{a}}}_i} \}_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}$, we preferred to simply draw a box for $A_i$, $i \in {\mathcal{I}}$. A similar comment applies to $B_j$, $j \in {\mathcal{J}}$. An alternative approach would have been to apply the concept of “closing the box” by Loeliger, cf. e.g. [@Loeliger:04:1], where $A_i({{\mathbf{u}}}_i)$ would be defined as the minimum over $\{ \alpha_{i,{{\mathbf{a}}}_i}
\}_{{{\mathbf{a}}}_i \in {\mathcal{A}}_i}$ of the above $A_i({{\mathbf{u}}}_i)$ function. Here we preferred the first approach because we wanted to keep variables like ${{\mathbf{u}}}_i$ and $\alpha_{i,{{\mathbf{a}}}_i}$ at the “same level”. \[footnote:closing:the:box:comment:1\]
[^4]: A similar comment applies here as in Footnote \[footnote:closing:the:box:comment:1\]. Here, the $\phi'_i$ and $\theta'_j$ have to be seen as dual variables that would appear as edges in a more detailed drawing of the boxes $A'_i({{\mathbf{u}}}'_i)$ and $B'_j({{\mathbf{v}}}'_j)$, respectively.
[^5]: Note that any interesting code has a parity-check matrix whose rows have Hamming weight at least $3$.
[^6]: We assume here that there is a unique optimal solution $\hat {{\mathbf{x}}}$ to **PLPD1** or to **PLPD2**; more general statements can be made for the case when there is not a unique optimal solution.
[^7]: The use of sub-gradients is necessary since the objective function is concave but not everywhere differentiable, cf. e.g. [@Bertsekas:99:1].
[^8]: We have adapted the text for maximizations instead of minimizations.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The paper develops the dynamics and thermodynamics of Stirling engines that run with temperature differences below $100~^0$C. The working gas pressure is analytically expressed using an alternative thermodynamic cycle. The shaft dynamics is studied using its rotational equation of motion. It is found that the initial volumes of the cold and hot working gas play a non-negligible role in the functioning of the engine.'
author:
- Alejandro Romanelli
title: Stirling engine operating at low temperature difference
---
Introduction
============
In the field of energy efficiency, the use of waste energy is one of the keys to improve the performance of facilities, whether industrial or domestic. In general the waste energy arises as heat, from some thermal process, that it is necessary to remove. Therefore the use of the waste energy is usually conditioned by the difficulty of converting heat into other forms of energy.[@tesis4; @tesis3] The Stirling engines, being external combustion machines, have the potential to take advantage of any source of thermal energy to convert it into mechanical energy. This makes them candidates to be used in heat recovery systems.
The Stirling engine is essentially a two-part hot-air engine which operates in a closed regenerative thermodynamic cycle, with cyclic compressions and expansions of the working fluid at different temperature levels.[[@Reid; @Sier]]{} The flow of the working fluid is controlled only by the internal volume changes; there are no valves and there is a net conversion of heat into work or vice-versa.
A Stirling-cycle machine can be constructed in a variety of different configurations.[@Walker; @Darlington] In this work we focus on the study of low temperature difference (LTD) Stirling engines, that is, operating on a temperature difference below $100~^0$C and, in general, in a Gamma configuration. These engines use a heat source that excludes direct combustion, which occurs at temperatures of several hundred degrees. The temperature range where the Stirling engine works determines its geometry and proportions.[@Senft3] Stirling engines, with a high temperature difference need a relatively long distance between the hot and cold chambers to avoid an excessive heat loss between the chambers, while the size of the heating and cooling surface area is less critical. On the other hand, LTD Stirling engines require gas chambers with a large surface area to facilitate the heat transfer with the environment, but there is also less heat conduction from the hot to the cold chambers so the distance between them can be smaller.
The first reference to LTD Stirling engines is related to the work developed by Ivo Kolin of the University of Zagreb during the $1970$s and $1980$s.[@Kolin1; @Kolin2] Subsequently James R. Senft of the University of Wisconsin also developed engines capable of operating with a LTD.[@Senft3; @Senft1; @Senft2] These two pioneers show what a sustained work of research and development can do with a concept: their engines that initially operated at temperature difference of $44~^0$C evolved to run at only $0.5~^0$C.
At the present time research on Stirling engines is one of the lines that contribute both to the rational use of energy and to sustainable development.[@Kong; @Barreto; @Jokar] In particular the solar thermal conversion systems based on these engines are amongst the most interesting and promising research lines. [@tesis4; @tesis3; @tesis1; @tesis2; @tesis5; @tesis6]
To achieve an adequate theoretical description of the Stirling thermodynamic cycle it is necessary to adopt certain simplifications. Usually this cycle is modeled by alternating two isothermal and two isometric processes.[@Zeman] In previous papers we developed an alternative approach to the Stirling cycle that provides analytical expressions for the pressure and temperatures of the working gas, and the work and heat exchanged with the reservoirs. The theoretical pressure-volume diagram achieved a closer agreement with the experimental one than the standard analysis.[@Romanelli1; @Romanelli2] Due to the generality of the analytical expressions obtained they can be adapted to any type of Stirling engine. In the present paper we use the mentioned model to study the dynamics and thermodynamics of a particular Stirling engine that runs at a low temperature difference.
The paper is organized as follows. In the next section a simple model of LTD Stirling engine is presented. In the third section we introduce some aspects of our alternative thermodynamics for the Stirling engine. In section four we introduce the dynamics of the engine in order to complete the model and make some numerical calculation. In the last section we present the main conclusions.
![Schematic picture of LTD Stirling engine from two perpendicular points of view. $V_1$ and $V_2$ are the cold and the hot volumes respectively, $x_1+x_2=2r$. Open arrows indicate the movements of the gas. The flywheel moves counterclockwise and (in the instant photo) both power piston and displacer are descending. Several experimental models, whether homemade or commercial, found in the Internet fit in this scheme.[]{data-label="f1"}](Romanellifig01.eps)
Simple model of a Stirling engine
=================================
In this section we present a simple model of LTD Stirling engine that corresponds to a Gamma configuration. Figure \[f1\] provides a useful insight from head-on and side of the engine running with the power piston modifying the gas volume and the displacer, continually sweeping the gas up and down. This configuration has two cylinders joined to form a single connected space with the same pressure. The cylinder with the larger cross-section contains the displacer and the other the power piston. The piston and the displacer are joined to a shaft through rods which move in parallel planes with a relative phase of $\phi\equiv\pi/2$ between the cranks. The shaft has also a flywheel.
The engine is on the top of a cup of boiling water and the remainder of the engine is in contact with the environment at room temperature. Note that the engine also would operate on the top of a cup of cold water, but in this case the flywheel would rotate in the opposite direction. The gas pressure remains uniform during the entire operation but it changes due to the motion of the piston and the action of the heat reservoirs, represented by the cup of boiling water and the room atmosphere.
In the total gas volume two zones, separated by the displacer, can be distinguished by their temperatures. Our mathematical model assumes uniform temperatures for these zones. This assumption becomes reasonable when the characteristic times associated with the movement of the pistons are much larger than those associated with the mean free path of the gas molecules. The hot zone is below the displacer where the gas is in contact with the hot reservoir at the external temperature $T_h$, and the cold zone is above the displacer where the gas is in contact with the cold reservoir at the external temperature $T_c$. However both zones have the same pressure $P$ because they are connected through the loose fit of the displacer with its cylinder. When the displacer is in its lowest position all the gas is in the cold zone, but in the rest of the cycle, the gas is never completely in either the hot or cold zone of the engine. The piston is also subjected to the atmospheric pressure $P_0$ from the outside.
In order to start the engine we must set in motion the flywheel externally and immediately let it move by the action of the engine. Then the mechanism of energy transfer between the hot and cold zones begins to operate and it can be qualitatively understood as follows. When the gas is swept by the displacer to the hot zone it expands, the gas pressure increases and the piston pushes up. When the gas is swept by the displacer to the cold zone it contracts, and the momentum of the machine, usually enhanced by the flywheel, pushes the piston down to compress the gas. The phase difference between displacer and piston is a crucial parameter for the efficiency and power delivered by the engine. The optimum phase is around $\phi=\pi/2$, as it has been shown.[@Senft3; @Romanelli1]
The piston and the displacer are connected to the shaft by rods of length $R$ as shown in Fig. \[f1\]. The piston and displacer cylinders have cross-section areas $a$ and $A$ respectively, and the crank has a rotating radius $r$. The relation between the cold and hot gas volumes, $V_1$ and $V_2$, and the flywheel angle $\theta$ can be obtained from the geometric analysis of Fig. \[f1\], *i.e.* $${V_2}=\left\{1+z\left[1-\cos(\theta+\phi)\right]-\sqrt{1-z^{2}\sin^{2}(\theta+\phi)}\right\}RA,
\label{v2g}$$ $${V_1}=2rA-{V_2}+{V_p}, \label{v1g}$$ where $z={r}/{R}$ and $V_p$ is the volume swept by the piston run: $${V_p}=\left\{1+z\left(1-\cos\theta\right)-\sqrt{1-z^{2}\sin^{2}\theta}\right\}Ra. \label{v0g}$$ The total volume of gas $V$ is given by $$V={V_1}+{V_2}=2rA+{V_p}. \label{v}$$
Thermodynamics of an LTD Stirling engine
========================================
![Dimensionless pressure-volume diagram for the initial conditions $\theta_0=1.46~\pi$ (red thick curve). The thin black curve, corresponding to the initial condition $\theta_0=0.46~\pi$, has been moved in the diagram to the current position for the sake of comparison. The arrows indicate the evolution with growing $\theta$, and the parameters are: $\phi=0.5~\pi$, $\alpha=1.25$, $\beta=1.23$, $a/A=0.016$ and $z=0.067$.[]{data-label="f2"}](Romanellifig02.eps)
In previous work we have obtained an analytical generic solution for the pressure and temperatures inside of any Stirling engine.[@Romanelli1] That follows from the application of the first law of thermodynamics to an ideal gas subjected to polytropic processes. We have used the mentioned solution to study a particular type of Stirling engine known as Fluidyne.[@Romanelli2] Here we use the expression of the gas pressure as a function of the volumes (Eq. (20) of Ref. ), *i.e.* $$P=P_{0}\left(\frac{\alpha\,V_{10}+V_{20}}{\alpha\,V_{1}+V_{2}}\right)\left(\frac{V_{0}}{V}\right)^{\beta-1},
\label{pe}$$ where $\alpha$ is the ratio of reservoir temperatures: $$\alpha\equiv{T_{h}}/{T_{c}}, \label{alfa1}$$ $V_{10}$, $V_{20}$ and $V_0$ are the initial values of $V_{1}$, $V_{2}$ and $V$ respectively (Eqs. (\[v2g\]), (\[v1g\]) and (\[v\])), and $\beta$ is a parameter called polytropic index.[@romanelli] The typical values of $\beta$ are such that $1\leq\beta\leq \gamma$ with $\gamma=c_p/c_v$ the quotient of the specific heats of the gas at constant pressure $c_p$ and constant volume $c_{v}$. For the isothermal and adiabatic cases, $\beta=1$ and $\beta=\gamma$ respectively, Eq. (\[pe\]) reduces to the well known solutions found in the specialized literature of Stirling engines.[@Schmidt; @Bercho; @Formosa]
Equation (\[pe\]) depends on $\alpha$ and it has a finite asymptotic limit when $\alpha\rightarrow\infty$. This means that after a certain value of $\alpha$, no matter how much we increase the temperature ratio the pressure and the absorbed heat are bounded, and this in turn explains why the useful work and the efficiency in any Stirling engine are asymptotically bounded[@Romanelli1].
The pressure-volume diagram of the LTD Stirling engine is obtained starting from Eqs. (\[v0g\]), (\[v\]) and (\[pe\]) and implementing numerical calculation. Such diagrams are shown in Fig. \[f2\] for two characteristic initial values of $\theta$ which, as we shall see, correspond to the maximum and minimum engine work. The area enclosed by the curves represents the total work $W$ of the cycle, whose expression is $$W=\oint P\,dV=\int_{\theta_0}^{\theta_0+2\pi}P{\frac{dV}{d\theta}}\,d\theta
=\int_{0}^{2\pi}P{\frac{dV}{d\theta}}\,d\theta, \label{work}$$ where $\theta_0$ is the initial angular position that determines the initial conditions for the cold and hot gas volumes. The last equality in Eq. (\[work\]) is obtained with a change of variable that contemplates the fact that the integrand is a periodic function of $\theta$. $W$ is the work available for overcoming mechanical friction losses and for providing useful power to the engine shaft. $W$ depends on $\theta_0$ because $P$ depends on $\theta_0$ through $V_{10}$, $V_{20}$ and $V_{0}$ (see Eq. (\[pe\])). Here we point out that this dependence, as far as we know, has not been explored experimentally and it would be interesting to do so.
![Relative work per cycle as a function of the initial angular condition $\theta_0$, in black thick line, $\Delta W=W-W_i$ where $W_i$ is the work associated with $\theta_0=0$. The dimensionless initial gas volume of the piston cylinder is shown with the blue thin line. The dimensionless initial gas volume of the hot zone, is shown with the red dashed line. The values of the parameters are the same as in Fig. \[f2\].[]{data-label="f3"}](Romanellifig03.eps)
We have integrated numerically Eq. (\[work\]) with Eqs. (\[pe\]), (\[v\]) and (\[v0g\]) as functions of $\theta_0$, using the standard Simpson’s rule. Figure \[f3\] shows that the initial volumes of cold and hot gas play a non-negligible role in the functioning of the engine. Observe that there is a difference up to $20\%$ between the maximum and minimum relative work for $\theta_0=1.46~\pi$ and $\theta_0=0.46~\pi$. It is clear that the maximum work is obtained when initially all the gas is in the cold zone, and the minimum work is obtained when initially almost all the gas is in the hot zone.
We have also integrated numerically the heat equation (Eq. (22) of Ref. ) as in previous works to obtain the absorbed heat $Q_{in}$ and rejected heat $Q_{out}$, in a cycle.[@Romanelli1; @Romanelli2]
![Dimensionless angular velocity as a function of time. Initial conditions are $\omega_0=10~
\omega_{th}$ ($\omega_{th}=3\pi{b}/{I}$) for the black curve and $\omega_0=\omega_{th}$ for the red curve, and in both cases $\theta_0=1.46~\pi$. The inset shows a magnification of both curves when the steady states are already reached. It is seen that $\langle\omega\rangle=5.24~\omega_{th}$, (dashed line), and that both curves have the same asymptotic period, $\mathcal{T'}\simeq~0.25$ seconds. The parameters are: $\phi=0.5~\pi$, $\alpha=1.25$, $\beta=1.23$, $a/A=0.016$, $z=0.067$, $b/I=0.5~s^{-1}$ and $I=1.2~{10}^{-4}kg~m^2$.[]{data-label="f4"}](Romanellifig04.eps)
Dynamics of LTD Stirling engine
===============================
In order to simplify the mechanical model we neglect the masses of piston, displacer and rods. The conservation of energy imposes that all the power generated by the gas must be transferred to the the shaft, this means that $$(P-P_0)\frac{dV}{dt}=\tau \frac{d\theta}{dt}, \label{pote}$$ where $\tau$ is the external torque over the shaft and, ${d\theta}/{dt}$ is the angular velocity of the flywheel. Then $\tau$ is given by $$\tau=(P-P_0)\frac{dV}{d\theta}. \label{torque}$$ The inevitable mechanical friction of the system, as well as a possible additional load incorporated to extract energy from it, generate an additional friction torque on the shaft, modeled as $-b~{d\theta}/{dt}$, where the sign indicates loss of power and $b$ is the damping coefficient.
![Dimensionless angular velocity as a function of time. Initial conditions are $\theta_0=0.46~\pi$ for the black curve, $\theta_0=1.46~\pi$ for the red curve and for both curves, $\omega_0=\omega_{th}$. The inset shows a magnification of both curves when the steady states are already reached. It is seen that for the black curve $\langle\omega\rangle=4.20~\omega_{th}$ (dashed line) and the asymptotic period is $\mathcal{T'}\simeq~0.33$ s. Similarly for the red curve $\langle\omega\rangle=5.24~\omega_{th}$ and $\mathcal{T'}\simeq~0.25$ s. The values of the parameters are the same as in Fig. \[f4\].[]{data-label="f5"}](Romanellifig05.eps)
![Dimensionless gas power (in a logarithmic scale) as a function of a generalized friction coefficient. For the upper red curve $\theta_0=1.46~\pi$, and for the lower black curve $\theta_0=0.46~\pi$, and for both curves $\omega_0=40~\omega_{th}$.[]{data-label="f6"}](Romanellifig06.eps)
Therefore the rotational equation of motion indicates that the net external torque on the shaft determines the rate of change of its angular momentum, *i.e.* $$I\frac{d^{2}\theta}{dt^{2}}=-b\frac{d\theta}{dt}+(P-P_0)\frac{dV}{d\theta}, \label{newton}$$ where $I$ is the flywheel moment of inertia. In order to solve Eq. (\[newton\]) we need $P$ and $V$ as functions of $\theta$ given by Eqs. (\[v\]) and (\[pe\]); as a result we get a nonlinear equation of $\theta$.
Multiplying Eq. (\[newton\]) by the angular velocity $\omega\equiv{d\theta}/{dt}$ and integrating the resulting equation from $t=0$ to the arbitrary time $t$ we obtain the energy equation: $$\frac{1}{2}I\omega^2-\frac{1}{2}I{\omega_0}^2=-b\int_{0}^{t}{\omega}^2\,dt^{'}+
\int_{\theta_0}^{\theta}(P-P_0){\frac{dV}{d\theta^{'}}}\,d\theta^{'}, \label{energy}$$ where $\omega_0$ is the initial angular velocity. The first term on the right-hand side is the dissipative work and the second is associated with the engine power. From Eq. (\[pe\]) together with Eqs. (\[v2g\]) and (\[v1g\]) it is seen that $P$ is a periodic function of $\theta$, then the second term of the right-hand side satisfies $$\int_{\theta_0}^{\theta}(P-P_0){\frac{dV}{d\theta}}\,d\theta=
n\int_{0}^{2\pi}P{\frac{dV}{d\theta}}\,d\theta+
\int_{\theta_0}^{\theta_r}(P-P_0){\frac{dV}{d\theta}}\,d\theta, \label{ww}$$ where $n=(\theta-\theta_0)[mod~2\pi]$ is the number of complete cycles and $\theta_r=(\theta-\theta_0)-2\pi n$.
Equations (\[energy\]) and (\[ww\]) predict that if $b=0$ the kinetic energy grows with the number of turns $n$, however in real situations we always have $b\neq0$. As it will be shown later our model predicts that $\omega$ has a finite asymptotic average value, with a little oscillation $\Delta\omega$ around it.
Dividing both sides of Eq. (\[energy\]) by $t$ and taking the limit for $t\rightarrow\infty$ we obtain the relation between the dissipated power by the friction and power delivered by the gas $$b\langle\omega^2\rangle=\frac{1}{\mathcal{T}}\int_{0}^{2\pi}P{\frac{dV}{d\theta}}\,d\theta,
\label{wm}$$ where the flywheel period $\mathcal{T}$ is defined as $$\begin{aligned}
\mathcal{T}\equiv &\lim{}\frac{t}{n} \cr
&t\rightarrow \infty \, \label{asym0}\end{aligned}$$ and the mean quadratic angular velocity is defined as $$\begin{aligned}
\langle\omega^2\rangle\equiv&\lim{}{\frac{1}{t}}{\int_{0}^{t}{\omega}^2\,dt} .\cr &t\rightarrow
\infty \, \label{asym}\end{aligned}$$ Finally, using Eq. (\[wm\]) we can approximate the asymptotic average value of $\omega$, $\langle\omega\rangle\equiv{2\pi}/{\mathcal{T}}$, as $$\langle\omega\rangle\simeq\frac{1}{2\pi b}\int_{0}^{2\pi}P{\frac{dV}{d\theta}}\,d\theta,
\label{wm2}$$ if $|\Delta\omega/\langle\omega\rangle|\ll1$.
As mentioned, it is necessary to start the engine with an initial angular velocity $\omega_0$ to overcome the initial static friction and the dynamic friction that follows it. This velocity must exceed a certain threshold value $\omega_{th}$ that depends on the nonlinear dynamics of Eq. (\[newton\]) which in turn depends on the parameters $I$, $b$, $\alpha$, $\beta$, $z$, $a/A$ and $\theta_0$. Considering the order of magnitude of the parameters used in this paper, we estimate a practical value of $\omega_{th}$ solving Eq. (\[newton\]) with $P-P_0=0$. In this case, the solution of Eq. (\[newton\]) is $$\theta=\omega_0\frac{I}{b}(1-e^{-\frac{b}{I}t}). \label{theta}$$ We choose arbitrarily $\theta=3\pi$ in order that the flywheel can asymptotically complete at least one and half cycles. This leads to $$\omega_{th}\sim\omega_{0}=3\pi\frac{b}{I}\ll\langle\omega\rangle. \label{threshold}$$ Now we study numerically the dynamics of the shaft, and in particular its dependence on both the engine parameters and the initial conditions. Using a standard Runge-Kutta method for the approximate solutions of ordinary differential equations, we solved Eq. (\[newton\]) with the constraints Eqs. (\[v\]) and (\[pe\]).
![Dimensionless gas power as a function of the temperature ratio of heat reservoirs $\alpha$. For $\alpha=T_h/T_c\sim1.34$ the thermal difference $T_h-T_c\sim100~^0C$. For the upper red curve $\theta_0=1.46~\pi$, and for the lower black curve $\theta_0=0.46~\pi$, and for both curves $\omega_0=40~\omega_{th}$.[]{data-label="f7"}](Romanellifig07.eps)
Figures \[f4\] and \[f5\] show the flywheel angular velocity as a function of time for two different initial conditions. Both figures show that the steady state is reached for a sufficiently large time. These asymptotic states are characterized by a constant average value with a fast small oscillation around it. Therefore our toy model predicts an asymptotic time dependent angular velocity with a well defined period $\mathcal{T'}$, which, in principle, could be different from $\mathcal{T}$, the flywheel period.
In the case of Fig. \[f4\] we use the same value of $\theta_0$ but with two different initial angular velocities $\omega_0$. This figure shows that in the steady state the angular velocity does not keep any correlation with its initial value; such a behavior is characteristic of dissipative systems. In the inset we see that the period $\mathcal{T'}$ is the same for both curves, that is independent of $\omega_0$.
In Fig. \[f5\] we keep the same value for $\omega_0$ but with two different values for $\theta_0$. Now the asymptotic average angular velocities have different values, this means that they depend on $\theta_0$, and the same is true for the periods $\mathcal{T'}$.
At this point we emphasize that the numerical calculation shows that the flywheel period $\mathcal{T}={2\pi}/{\langle\omega\rangle}$ and the asymptotic period $\mathcal{T'}$ of $\omega(t)$ coincide numerically, that is $$\mathcal{T}=\mathcal{T'}. \label{times}$$ This coincidence has no obvious theoretical explanation, all the more so since the dynamical differential equation is nonlinear. In order to show that the relation given by Eq. (\[times\]) could be different, we reason as follows: When the system reaches the steady state, Figs. \[f4\] and \[f5\] suggest that $\omega(t)$ can be approximated by $$\omega(t)=\langle\omega\rangle+\zeta\cos{\frac{2\pi}{\mathcal{T'}}t}, \label{wmdt}$$ where $\zeta$ is a small amplitude. However, since the flywheel angular position $\theta(t)$ is characterized by the period $\mathcal{T}$, the angular velocity (that satisfies $\omega(t)\equiv
d\theta/dt$) must have also the same period. Therefore, $\mathcal{T}$ and $\mathcal{T'}$ must satisfy the relation $$\mathcal{T}= m \mathcal{T'}, \label{tytp}$$ where $m$ is an integer. Our numerical calculation indicates that $m=1$ but we have no additional theoretical argument for this value.
Returning to our central development, let us define the power of the engine $\Pi$ by the right-hand side of Eq. (\[wm\]), *i.e.* $$\Pi\equiv\frac{1}{\mathcal{T}}\int_{0}^{2\pi}P{\frac{dV}{d\theta}}\,d\theta, \label{power}$$ of course it is also true that $\Pi=b\langle\omega^2\rangle$.
In Fig. \[f6\] we show $\Pi$ as a function of $b/I$ for the initial angular conditions associated with the maximum and minimum work. The upper limit of friction (*i.e.* the maximum value of $b/I$) is determined by the vanishing of the angular velocity, hence the vanishing of $\Pi$. The minimum value of $b/I$ has been chosen to ensure an adequate numerical convergence because asymptotically $\omega\rightarrow\infty$, the period $\mathcal{T}\rightarrow0$ and $\Pi\rightarrow\infty$. The figure also shows that the power for $\theta_0=1.46~\pi$ is larger than for $\theta_0=0.46~\pi$, independently of $b/I$, as it is the case for the work.
Figure \[f7\] shows the growth of $\Pi$ with the the temperature ratio $\alpha$ of the heat reservoirs. It also shows the dependence of $\Pi$ on the initial condition $\theta_0$.
Conclusions
===========
This paper developed the dynamics and thermodynamics for a Stirling engine that operates at low temperature difference. The working gas pressure is expressed analytically using an alternative thermodynamic model developed in a previous paper. This pressure depends on the hot and cold gas volumes as well as their initial conditions and the temperature ratio of the heat reservoirs.
The rotational dynamical equation for the shaft includes terms related to the energy delivered by the engine and to the energy dissipated by friction. This nonlinear equation, which we solved numerically, has an explicit dependence on the initial gas volumes through the initial angular position $\theta_0$.
We showed numerically that both the maximum shaft work and power are obtained when at the start all the gas is in the cold zone. Conversely the minimum shaft work and power are obtained when initially almost all the gas is in the hot zone. These results are independent of: (a) the shaft initial angular velocity, (b) the shaft friction intensity, (c) the temperature ratio of the heat reservoirs.
In order to choose initial conditions we estimated a minimal initial flywheel speed to start the engine. Our model predicts that the flywheel velocity achieves a finite asymptotic average value with a small periodic oscillation around it. Our system has a dissipative dynamics, however its steady state showed an unexpected dependence on the initial angular position. This means that the angular velocity evolves to an attractor independently of its initial value but dependent on the initial angular position. This asymmetric behavior is due the fact that the model incorporated the initial angular position in the dynamical equation but not the initial angular velocity.
Summarizing, this paper describes the operation of the LTD Stirling engine in a simple and understandable way, highlighting some unknown aspects of its behavior.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
I acknowledge the stimulating discussions with Víctor Micenmacher, and the support from ANII and PEDECIBA (Uruguay).
[10]{}
D. García Menéndez, “Desarrollo de motores de Stirling para aplicaciones solares,” PhD thesis, (Universidad de Oviedo, 2013).
Caleb C. Lloyd, “A low temperature differential Stirling engine for power generation,” Master of Engineering thesis, (University of Canterbury, 2009).
J.S. Reid, “Stirling Stuff”(2016), https://arxiv.org/abs/1604.02362.
R. Sier, *Hot Air Caloric and Stirling Engines: A History*, (L. A. Mair, Chelmsford, U.K., 1999), Vol. 1.
G. Walker, J.R. Senft, *Free Piston Stirling Engines*, (Springer-Verlag, Berlin, 1980).
R. Darlington, K. Strong, *Stirling and Hot Air Engines*, (The Crowood Press, Marlborough, 2005).
J.R. Senft, “Optimum Stirling engines geometry,” Int. J. Energy Res. **26**, 1087-1101 (2002).
Kolin Ivo, “Stirling Motor - History, Theory, Practice,” (Zagreb University Publications, Dubrovnik, 1991).
Kolin Ivo, Koscak-Kolin, Sonja and Golub, Miroslav, “Geothermal Electricity Production by Means of the Low Temperature Difference Stirling Engine,” (Kyushu: Faculty of Mining, Geology and Petroleum Engineering, World Geothermal Congress, 2000)
J.R. Senft, *An Introduction to Low Temperature Differential Stirling Engines*, (Moriya Press., 1996).
J.R. Senft, “Theoretical limits on the performance of Stirling engines,” Int. J. Energy Res. **22**, 991-1000 (1998).
B. Kongtragool, S. Wongwises, “Thermodynamic analysis of a Stirling engine including dead volumes of hot space, cold space and regenerator,” Renewable Energy, **31**, 345-359 (2006).
G. Barreto, P. Canhoto, “Modelling of a Stirling engine with parabolic dish for thermal to electric conversion of solar energy,” Energy Convers. Manage. **132**, 119-135 (2017).
H. Jokar, A. R. Tavakolpour-Salej, “A novel solar-powered active low temperature differential Stirling pump,” Renewable Energy, **81**, 319-337 (2015).
L.G. Thieme, S. Qiu, M.A. White, “Technology development for a Stirling radioisotope power system,” NASA TM-2000-209791 (2000).
R.G. Lange, W. P. Carroll, “Review of recent advances of radioisotope power systems,” Energy Convers. Manage., **49**, 393-401 (2008).
I. M. Santos Ráez, “Estudio de un motor Stirling con absorbedor interno alimentado con energía solar,” PhD thesis, (Universidad de Málaga, 2015).
G. Barreto, P. Canhoto, “Modelling of a Stirling engine with parabolic dish for thermal to electric conversion of solar energy,” Energy Convers. Manage., **132**, 119-135 (2017).
M.W. Zemansky, R.H. Dittman, *Heat and Thermodynamics*, (McGraw-Hill, London, 1997).
A. Romanelli, “Alternative thermodynamic cycle for the Stirling machine,” Am. J. Phys. **85**, 926-931 (2017).
A. Romanelli, “The Fluidyne engine,” Am. J. Phys. **87**, 33-37 (2019).
A. Romanelli, I.Bove, F.González, “Air expansion in a water rocket,” Am. J. Phys. **81**, 762-766 (2013).
G. Schmidt, “Theorie der Lehmann’schen kalorischen Maschine,” Z. Ver. Deut. Ing. **15**, 1 (1871).
I.Urieli, D.Berchowitz, *Stirling Cycle Engine Analysis*, (Adam Hilger, Bristol, 1983).
F.Formosa, G.Despesse, “Analytical model for Stirling cycle machine design,” Energy Convers. Manage., **51**, 1855-1863 (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Realizing topological superconductivity and Majorana zero modes in the laboratory is one of the major goals in condensed matter physics. We review the current status of this rapidly-developing field, focusing on semiconductor-superconductor proposals for topological superconductivity. Material science progress and robust signatures of Majorana zero modes in recent experiments are discussed. After a brief introduction to the subject, we outline several next-generation experiments probing exotic properties of Majorana zero modes, including fusion rules and non-Abelian exchange statistics. Finally, we discuss prospects for implementing Majorana-based topological quantum computation in these systems.'
author:
- 'R. M. Lutchyn'
- 'E. P. A. M. Bakkers'
- 'L. P. Kouwenhoven'
- 'P. Krogstrup'
- 'C. M. Marcus'
- 'Y. Oreg'
title: 'Realizing Majorana zero modes in superconductor-semiconductor heterostructures'
---
Introduction
============
The search for topological phases of matter has generated significant interest in physics, chemistry, and material science [@Wilczek'09; @SternPerspective; @Brouwer_Science; @Lee14]. The discovery of topological superconductors supporting Majorana zero-energy modes is of fundamental scientific importance and has profound technological applications for quantum information processing. Majorana zero modes correspond to exotic [*neutral*]{} excitations of a superconductor comprising of an equal superposition of an electron and a hole. In this sense, a solid-state Majorana quasiparticle which is its own antiparticle is closely related to a Majorana fermion that was first introduced by Ettore Majorana in the context of elementary particle physics [@Majorana2008]. A [*neutral*]{} excitation in a superconductor has a very special property due to the inherent particle-hole symmetry - it is bound to zero energy so that there is no cost to occupy such a state. Therefore, the presence of a large number of localized Majorana zero-energy modes leads to a massive ground-state degeneracy which can be exploited for topological quantum computing purposes [@kitaev2003fault; @Nayak08]. Theory [@Moore1991; @ReadGreen] predicts that Majorana zero-energy modes (MZMs) localized at defects obey an exotic exchange statistics, similar to that of non-Abelian anyons. This is by far the most interesting property of MZMs. Indeed, exchanging the position of MZMs corresponds to a nontrivial transformation within the degenerate ground-state manifold, and represents a non-commutative operation which does not depend on the way and the details of its execution. Therefore, such an operation is topologically protected (i.e. only depends on exchange statistics of the quasiparticles) and may be used to implement quantum gates. These ideas are at the heart of the topological approach to quantum computing [@kitaev2003fault; @Nayak08]. Indeed, by exploiting topological materials, which by their nature minimize errors [@kitaev2003fault], one may overcome the largest barrier to building scalable quantum computing - decoherence.
A number of platforms for realizing MZMs in the laboratory were recently put forward [@Qi'08; @Beenakker13a; @Alicea12a; @Leijnse12; @Stanescu13b; @Franz'15; @DasSarma15; @MasatoshiSato2016; @Aguado2017]. Most of them propose to engineer an appropriate model Hamiltonian at the interface of a conventional superconductor and some other materials such as topological insulators , semiconductors with strong spin-orbit coupling [@Sau2010; @Alicea10; @Lutchyn10; @Oreg10; @SukBum'11; @Duckheim'11; @Potter'12] and magnetic atom chains [@Choy'11; @nadj2013; @klinovaja2013; @braunecker2013; @vazifeh2013; @pientka2013; @Nakosai2013; @Kim14; @brydon2014; @Li14; @Kotetes2014; @Ojanen2015; @nadjperge2014; @Ruby15; @Pawlak'16; @J.Zhang'16]. Among the most promising ones are the proximitized nanowire proposals [@Lutchyn10; @Oreg10], discussed in details in this review. There is mounting experimental evidence for the presence of MZMs in proximitized nanowires [@Mourik12; @Rokhinson12; @Deng12; @Churchill13; @Das12; @Finck12; @Albrecht16; @Zhang16; @Chen2016; @Deng2016; @Suominen2017; @Nichele2017; @Zhang2017; @Zhang2017a; @Sestoft2017; @Deng2017; @Geresdi2017].
A paradigmatic model for a one-dimensional (1D) topological superconductor involves spinless fermions that hop along the chain and experience proximity-induced p-wave pairing [@kitaev01]. Such a system supports an odd number of localized MZMs at the opposite ends of the chain. As explained above, these modes are neutral and, thus, a single MZM cannot accommodate a conventional (Dirac) fermion. At least two MZMs are needed to form a fermionic state with a well-defined occupation number. As a result, a 1D topological superconductor supports a non-local fermionic mode comprising of two MZMs localized at the opposite ends and separated by a distance which can be much larger than the superconducting coherence length. This non-local entanglement in a gapped system is the hallmark of topological superconductors and represents an instance of electron fractionalization [@kitaev01]. Note that the above argument relies on having an odd number of MZMs per wire end since an even number of Majorana modes can pair up and form a conventional Andreev state locally. In the latter case, the fermionic mode generically resides at a finite energy and, as a result, there is no non-local entanglement in the ground-state. Thus, there is a profound difference between conventional gapped superconductors, which have a unique ground state with even fermion parity, and topological superconductors, which instead have a highly degenerate ground state due to the presence of many MZMs [@ReadGreen; @kitaev01].
![a) Energy spectrum as a function of the momentum along the nanowire $k$. Spin-orbit coupling shifts the parabolas describing electron spectrum sideways by $k_{so}$ and introduces new energy scale $E_{so}$. Zeeman coupling $V_Z$ due to applied external magnetic field perpendicular to the direction of the spin-orbit coupling opens a gap at $k\!=\!0$ [@Lutchyn10; @Oreg10]. Arrows indicate approximate spin orientation for different momenta. b) Topological quantum phase diagram. By changing chemical potential $\mu$ in the nanowire or Zeeman splitting $V_Z$ one can drive the system into a topological phase. Here $\Delta_0$ is the induced pairing potential in the wire.[]{data-label="fig:fig1a"}](Fig1a){width="0.9\columnwidth"}
Spinless p-wave superconductivity is a key element for realizing separated MZMs. However, electrons in conventional materials have spin $\frac{1}{2}$, and, thus, the notion of a spinless superconductor does not seem immediately relevant to real physical systems. An elegant way to overcome this difficulty is to use spin-orbit materials, where spin and orbital degrees of freedom are correlated. Indeed, spinless superconductivity can effectively emerge in a semiconductor nanowire with strong spin-orbit interaction proximity-coupled to a conventional (s-wave) superconductor [@Lutchyn10; @Oreg10]. The corresponding energy spectrum for the nanowire with spin-orbit coupling is shown in Fig. \[fig:fig1a\]a. The spin-orbit interaction shifts the momentum parabolas sideways, the Zeeman term, which does not commute with the spin-orbit term, opens a gap in the spectrum at zero momentum, see Box \[box\] for details. When the chemical potential $\mu$ is tuned to be in the gap (see Fig. \[fig:fig1a\]a), the spectrum has a single band crossing Fermi points and, thus, the system in this regime is effectively described by the “spinless” degrees of freedom. However, due to the spin-orbit-induced rotation of the spins at the opposite Fermi points, see Fig.\[fig:fig1a\]a, proximity-induced s-wave interaction opens a pairing gap in the spectrum. The resulting state is closely related to a spinless p-wave superconductor as explained in the Box \[box\].
One of the virtues of the above model is that the proximitized nanowire can be driven into a topological phase by tuning magnetic field or chemical potential, see Fig.\[fig:fig1a\]b. The emergence of MZMs at a certain critical value of a control parameter is necessarily accompanied by the closing of the bulk gap [@ReadGreen; @kitaev01]. This phenomenon corresponds to a topological quantum phase transition (i.e., a quantum phase transition between topologically trivial and non-trivial states, see Fig.\[fig:fig1a\]b). The topological phase is stable with respect to small perturbations (e.g., disorder) as long as these perturbations do not collapse the bulk gap in the spectrum. The latter depends on the effective spin-orbit energy $E_{so}$, the proximity-induced gap $\Delta_0$, and the effective Zeeman energy $V_Z$ in the heterostructure. Therefore, when engineering such materials, an obvious first choice is to pick material components that individually have required properties. So far, the heavy-element semiconductors InAs and InSb have received much attention due to the strong spin-orbit coupling as well as large Lande $g$-factor, whereas Al and NbTiN have been primarily used for the superconducting components. Relevant bulk properties of these materials are listed in Table \[Table1\]. The large $g$-factor in the semiconductor allows one to achieve the desired effect, opening a large effective Zeeman gap in the hybrid nanowire using an in-plane magnetic field at fields below the critical field of the s-wave superconductor. High-quality interfaces between the two components lead to a significant hybridization of the semiconducting and metallic states resulting in a large induced pairing potential. The interface should be smooth since disorder scattering tends to suppress p-wave superconductivity and eventually leads to the collapse of the topological phase [@Motrunich01; @Brouwer11; @Stanescu11; @Akhmerov2011; @Potter12; @Lobos12; @Potter_Lee2011; @lutchyn_momentum; @DeGottardi2013; @Takei13; @Adagideli14; @Hui15; @Cole16; @Hegde2016; @DongLiu2017]. All these requirements are important for the observation of MZMs and pose a challenging material science task.
Semiconductors InAs InSb
--------------------------------------------------------------------------------------- --------------------------- -------------------------
$g$-factor 8-15 40-50
effective mass $m^*$ 0.023 $m_e$ 0.014 $m_e$
spin-orbit energy $E_{so}=\frac{m^* \alpha^2}{2\hbar^2}$ 0.05 - 1 meV 0.05 - 1 meV
spin-orbit coupling $\alpha$ 0.2 - 0.8 eV$\cdot$ [Å]{} 0.2 - 1 eV$\cdot$ [Å]{}
spin-orbit length $\lambda_{so}\!\equiv\! k^{-1}_{so}\!=\!\frac{\hbar^2}{\alpha m^*}$ 180 - 40 nm 230 - 50 nm
Superconductors Al NbTiN
superconducting gap $\Delta$ 0.2 meV 3 meV
critical field $B_c$ 10 mT 10 T
critical temperature $T_c$ 1.2 K 15 K
: Bulk properties of typically used semiconductors and superconductors [@levinshteinhandbook; @Winkler2003; @Cochran1958; @Mourik12]. Rashba spin-orbit coupling strength was measured in nanoscale structures in Refs. [@Shabani16; @Weperen2015].[]{data-label="Table1"}
In this article, we provide a perspective on the current state of the realization of MZMs in superconductor (SC)-semiconductor (SM) heterostructures. In the next section, we will provide an overview of the recent material science developments in growing high-mobility semiconductors and preparing high-quality interfaces between superconductors and semiconductors. We then summarize robust signatures of MZMs and discuss the best experimental evidence for the presence of MZMs in proximitized nanowires. Finally, we conclude with an outlook and discuss the perspective for topological quantum computation with MZMs. Due to space constraints, we will not cover topological insulator and magnetic-atom-chain Majorana proposals. There are several excellent reviews on this subject [@Qi'08; @Hasan'10; @Murakami'11; @Ando'13; @bernevig2013topological; @Bansil'16] to which we refer an interested reader.
Material science aspects
========================
Semiconductor growth
--------------------
![Growth of bottom-up nanowires. a) Vapor-Liquid-Solid (VLS) nanowire growth mechanism, b) SEM image of an InP-InSb nanowire grown along the (111)B crystal direction, c) TEM image showing the defect-free nature of the InSb wire, d) HRTEM image demonstrating the (111) growth direction, e) generic method to fabricate nanowire network with a pre-defined number of wires. Trenches are etched from a (100)-oriented InP substrate such that (111)B facets are exposed. E-beam lithography is used to define catalyst on the (111)B facets with a controlled position. SEM images of f) InP nanowire grown from the catalysts on the sloped facets, g) InP-InSb nanowires forming a cross, h) InSb nanowire hashtags. Scale bars are $1 \mu m$. Panels e)-h) reproduced with permission from Ref. [@Gazibegovic2017].[]{data-label="fig:fig2"}](Fig2new){width="\columnwidth"}
The fabrication of one-dimensional confinement in the InAs and InSb semiconducting materials has been obtained by two different methods. One is top-down (lithography-based) definition of 1D structures in two-dimensional (2D) quantum wells grown by Molecular Beam Epitaxy (MBE). Here the quantum wells are grown close to the surface allowing for the deposition or epitaxial growth of a superconducting material in close proximity to the well (see section \[episuper\]). The growth of 2D semiconductor materials has been studied for many years and is now well understood for many different material combinations. However, for SM-SC heterostructures new challenges emerge. To increase the mobility in conventional 2D materials, the first thing would be to grow a thick insulating, and, ideally, lattice-matched layer on top of the 2DEG to protect it from the surface disorder. However, this will also suppress induced superconductivity which relies on electron exchange between SM and SC. Thus, a thinner and low barrier material is grown as an intermediate layer, and the task is to find the barrier height and thickness that give the best combination of mobility and induced superconductivity [@Shabani16]. Moreover, as the 1D confinement will be defined by top-down methods using either mesa etch or selective etch of the superconductor, there will be additional material/etching requirements for allowing a controlled definition of the nanowires. This is likely to be an ongoing research topic for the coming years.
The other way to produce nanowires is to use the bottom-up Vapor-Liquid-Solid (VLS) growth mechanism. This mechanism has been developed over the last 50 years and can be used to fabricate nanowires of a variety of semiconductor materials [@Wagner64; @Caroff09; @Lugani10; @Nilsson10; @Vogel11; @Plissard12]. InAs and InSb materials have also shown promising behavior in this fabrication scheme. Typically, a catalyst particle is used to collect precursor material and to establish a local supersaturation, as shown in Fig. \[fig:fig2\]a. The catalyst becomes liquid, and when a critical supersaturation level is reached, the material is precipitated at the liquid/substrate interface by a layer-by-layer growth mechanism. This mechanism results in a crystalline nanowire with the radius determined by the catalyst dimensions. Precursor material can directly impinge from the gas phase onto the catalyst surface, or it can first land on the substrate surface and subsequently diffuse towards the catalyst. While diffusing, these adatoms can be trapped, and when this happens on the side wall of the nanowire it will lead to radial growth. Since we want to fabricate 1D nanowires, it is important to suppress this effect. In the case of InSb nanowire growth, a problem is the low surface energy of Sb, which acts as a surfactant and, therefore, makes it difficult to initiate nanowire growth because the wetting layer tends to form a 2D layer. One way to overcome these problems is by growing a heterostructure. For example, as shown in Fig. \[fig:fig2\]b, an InP/InAs stem wire can be used to grow an InSb wire [@Caroff09; @Lugani10; @Nilsson10; @Vogel11; @Plissard12]. These wires can be a few microns long and have a diameter of 80-100 nm.
The crystal structures of VLS grown nanowires are relatively easily investigated by transmission electron microscopy (TEM). Fig. \[fig:fig2\]d shows a defect-free cubic (zincblende) structure of a InSb nanowire grown in the \[111\]B direction. On the other hand, defect-free InAs nanowires are typically grown in a hexagonal (wurtzite) crystal phase with a \[001\]B axis. It is important to mention that nanowires have a strong tendency to grow in \[111\]B/\[001\]B direction, which can be used to direct the orientation of the wires as discussed below. These uniform wires have been used for the first-generation Majorana experiments [@Mourik12], and their electronic properties are discussed in Fig. \[fig:fig4\]. For next-generation Majorana experiments more complex nanowire structures are required, for instance, T-junctions as explained in Sec. \[sec:perspective\]. An important requirement is that the junction between the branches has small reflection amplitude for an incoming electron. While such structures can be processed ex-situ using top-down lithography in 2D materials, VLS grown structures can be fabricated in-situ by using the crystallography of the substrate [@plissard13; @Kang13; @Car14; @Dalacu13; @Heedt16; @Rieger16; @Krizek17]. For example, instead of a (111) substrate we can use a (001) substrate where the growth direction can be turned into a non-vertical direction. Instead of a (111) substrate one can, for example, use a (001) substrate, in which trenches are etched exposing the (111)B facets. By using e-beam lithography the catalysts are defined on the sloped facets as depicted in Fig. \[fig:fig2\]e. Nanowires will grow perpendicular from these facets into two equivalent $< 111 >$B directions as shown in Fig. \[fig:fig2\]f. After growth of the InP stem, the InSb segment is grown. Since all wires are epitaxially connected to the substrate, they will, by definition, form a single crystal when they merge. For the merging process, there are two important aspects to be considered. First, there should be a junction offset in catalyst position as indicated in Fig. \[fig:fig2\]e. If the offset is too small, the catalyst particles will merge together, and a nanowire bridge is formed; the large Au particle crawls on a nanowire side facet. If the offset is too large the wires will miss each other. An optimum offset equals the wire radius. Second, the merging process itself is governed by radial growth, which, as mentioned above, always takes place as a competitive reaction besides axial growth [@Car14]. This radial growth results in an epitaxial shell and is responsible for the merging of two nearby wires. In the SEM images in Fig. \[fig:fig2\]g and h it is clear how wires can merge to form crosses or hashtags. With this method, a predefined number of wires can be connected into a network, and the dimensions and symmetry of the resulting structures can be tuned by the spacings b and c as indicated in Fig. \[fig:fig2\]e. TEM studies reveal the single crystalline nature of the crossed wires.
Superconducting interfaces {#episuper}
--------------------------
![High symmetry epitaxial Al/InAs interfaces. a) For planar InAs-based materials, the dominating Al fcc crystal orientation at low temperatures is \[111\] out of plane. Unlike the nanowire crystals, the planar epitaxial heterostructure shows better performance with an InGaAs barrier layer between the Al and InAs. b) A typical example of a layer stack grown on a lattice mismatched substrate like InP is shown. Panels a)-b) adapted with permission from Ref. [@Shabani16]. Here the superlattice traps the impurities and the graded buffer makes the transition to the lattice constant of the active region c) Thick Al layers on the nanowire facets typically lead to a gradual transition to a \[11-2\] out-of-plane crystal reconstruction. The image shows dark and white contrasts which denote two different degenerate twin orientations. The resulting Al can form a complete coherent InAs/Al bicrystal phase with a near lattice matched 3:2 ratio as shown in d). e) SEM micrograph of an array of InAs/Al nanowires where the Al sits on two facets. The nanowires bend towards the Al covered side. f) A thin uniform Al phase is obtained with \[111\] Al crystal orientation out-of-plane, as seen in the high-resolution zoom on the interface in g). Panels c)-g) adapted with permission from Ref. [@Krogstrup15]. []{data-label="fig:fig3"}](Fig3){width="\columnwidth"}
First-generation Majorana experiments [@Mourik12; @Deng12; @Churchill13; @Das12; @Finck12] showed a large subgap conductance indicating unintended subgap states, which are detrimental to topological quantum computation as they degrade topological protection. A likely source of the so-called soft gap, or continuum of subgap states, is disorder at the semiconductor/superconductor interface [@Takei13]. For these devices, the native oxide forming when the wires are exposed to ambient air has to be removed prior to the deposition of the superconductor material. There has been significant progress in improving these interfaces by gently etching the oxide and optimizing the superconducting deposition [@Krogstrup15; @Zhang16; @Zhang2017a]. An important step forward was taken with epitaxial growth of thin aluminum films on pristine facets of the nanowires, without breaking vacuum [@Krogstrup15]. As shown in Fig. 4c, this method of in-situ superconductor growth yields further significant suppression of the subgap density of states [@Chang14]. Subsequently, high-quality epitaxial superconductor-semiconductor interfaces have been realized in other heterostructures [@Shabani16; @Gazibegovic2017; @Beidenkopf2017; @Gusken2017; @Sestoft2017]. Typical values of the proximity-induced superconducting gap in different semiconductor-superconductor heterostructures are listed in Table \[Table2\].
The basic requirements for the Al film, besides the ability to induce superconductivity, are, first, it should stay thin enough to withstand high parallel magnetic fields without being driven normal, and second, it should have a uniform morphology along the nanowire, ensuring translational invariance in order to avoid disorder-induced subgap states. In-situ shadow deposition can be used to define junctions in the superconductor without additional lithography and etching steps, avoiding detrimental disorder from top-down processing [@Gazibegovic2017; @Krizek17].
---------------------------------------------- ------------ --------- ---------
semiconductor-superconductor heterostructure InSb-NbTiN InSb-Al InAs-Al
$\Delta_0$, meV 3 0.2-0.3 0.2-0.3
max\[$\Delta_{\rm ind}$\], meV 1 0.2 0.2
---------------------------------------------- ------------ --------- ---------
: Typical values for the bulk $\Delta_0$ and maximum values for the proximity-induced $\Delta_{\rm ind}$ superconducting gaps in different semiconductor-superconductor heterostructures [@Zhang16; @Deng2016; @Zhang2017]. SC gap of Al depends on the film thickness [@Court2008]. Note that $\Delta_{\rm ind}$ in aluminum heterostructures is close to the bulk superconducting gap indicating on strong tunneling regime between semiconductor and superconductor.[]{data-label="Table2"}
The equilibrium shape of Al on a large rigid InAs surface is not a continuous film, but a dewetted droplet with a contact angle of about 60-80 degrees. However, if the characteristic migration length of the Al adatoms is much shorter than the thickness of the nominally deposited Al, it can be kinetically locked to a continuous film. The key to obtaining a short adatom migration length is lowering the temperature sufficiently below room temperature. Moreover, because the native oxide stabilizes the film morphology, it is important to take the sample out of vacuum before it reaches dewetting temperatures (the specific transition temperature depends on the type of interface), which for Al on InAs is typically below room temperature. Unlike the InAs growth, where the crystal structure and morphology are locked once it is formed, the Al phase can easily recrystallize through grain growth transitions, which, at the right temperatures, moves the grain boundaries while keeping film continuous.
The grain growth process is driven exponentially by thermodynamic excesses such as interfacial and strain-free energies. The measure of the thermodynamic driving force is the excess chemical potential, which for a given grain M in the thin film limit can be divided into four terms corresponding to the surface, SC-SM interface, grain boundary (GB) and strain contributions [@Krogstrup15; @Shabani16]: $$\label{eq:1}
\mu_M\propto \frac{\gamma_{\rm surface}}{h}+\frac{\gamma_{\rm SC-SM}}{h}+\frac{\gamma_{\rm GB}}{R}+\frac{S \varepsilon^2 }{1-\nu}.$$ Here $\gamma_i$ is the energy density of the $i$’th interface, $h$ is the film thickness, $S$, $\varepsilon$ and $\nu$ are the shear modulus, strain and Poisson ratio of the thin film material, respectively; $R$ is the mean in-plane radius of curvature of $M$.
The epitaxial growth of Al on the facets of free-standing InAs nanowires is a more complex and dynamical process than on a planar InAs substrate, as was described subsequently for a two-dimensional platform in Ref. [@Shabani16]. In the single interface case, the dominating term for thin uniform layers (when $h \ll R$ ) is the surface term, because $\gamma_{\rm surface} > \gamma_{\rm SC-SM}$, which fixes the out-of-plane orientation to the lowest energy facet, which is the (111) facet, see Fig. \[fig:fig3\]a. The in-plane orientation appears to be driven primarily by the SC-SM interfacial bi-crystal symmetries which are governed by the second term. As shown in Ref. [@Shabani16], only two degenerate in-plane grain orientations of Al on planar InAs (100) were observed, indicating that these orientations are local minima of the interface energy with the (111) out-of-plane constraint. The size of the grains, however, is driven by minimizing the grain boundary energies, governed by the third term in Eq. . The last term in Eq. could play an important role if a particular bicrystal lattice match at the SC-SM interface has a substantial residual lattice mismatch but a significant minimum in the respective interfacial energy for the particular match. However, for a critical value of the film thickness $h \approx 5$nm Al phase on a planar InAs substrate was measured to be fully relaxed [@Shabani16], which means that the critical thickness for relaxation is presumably much less, and therefore does not play an important role in the growth kinetics for this particular material combination. For free-standing nanowires with Al growing on multiple facets, the grain boundary energies play a dominant role in the overall orientation, especially at thicker phases. As the film gets thicker the Al recrystallizes from a (111) out-of-plane orientation, similar as in the case of a single plane, to an orientation that lowers the total amount of grain boundary energy. In other words, the Al orientation changes so that it matches the symmetry of the nanowires and minimizes the incoherent grain boundary excesses across the adjacent facets, see Fig. \[fig:fig3\]c and d. The surface of the Al gets more uneven because it still has a tendency to form non-planar (111) facets. However, thin Al, as needed to withstand high magnetic fields, can stay extremely flat and uniform, see, for example, Fig. \[fig:fig3\]e-g.
Characterization of superconductor-semiconductor heterostructures
-----------------------------------------------------------------
![Characterization of different materials. a) Quantized conductance through an InSb wire at zero magnetic field. Inset: SEM image of an InSb nanowire with ohmic contacts. Panel a) adapted with permission from Ref. [@Kammhuber16]. b) Aharonov-Bohm oscillations in a nanowire hashtag demonstrating phase coherent transport. Inset: SEM image of an InSb nanowire hashtag. Panel b) adapted with permission from Ref. [@Gazibegovic2017]. c) Differential conductance as a function of the source-drain voltage of an epitaxial full-shell InAs/Al nanowire (blue) compared with an ex-situ evaporated Al shell (red) at $B = 0$ (solid line) and above the critical field $B > B_c$ (dashed line). Panel c) adapted with permission from Ref. [@Chang14]. d) Tunneling spectra show hard induced gap beyond 1.5 T in an InAs nanowire with a two-facet epitaxial Al film. Panel d) adapted with permission from Ref. [@Deng2016]. The wire is depleted by a gate and shows no subgap states in this regime. e) Weak antilocalization measurements on a hall bar with a 5nm InGaAs barrier and Al selectively etched away. The data allows one to extract the magnitude of the SOC in the semiconductor. Panel e) adapted with permission from Ref. [@Shabani16]. f) Tunneling conductance in the high density (black line) and low density (red line) regimes. The latter provides an estimate for the induced superconducting gap. Panel f) adapted with permission from Ref. [@Kjaergaard2016]. []{data-label="fig:fig4"}](Fig4){width="\columnwidth"}
Together with strong spin-orbit coupling and induced superconductivity, a key requirement for MZMs is quasi-ballistic electron transport with a controlled odd number of occupied modes. In the ballistic regime, the motion of 1D confined electrons is restricted to discrete energy bands resulting in quantized conductance plateaus. The observation of quantized conductance in the nanowires, therefore, provides direct evidence for the quasi-ballistic transport in these nanowires and allows one to control the mode occupation. Quantized conductance is difficult to achieve in nanowires because impurities, as well as scattering at the nanowire/contact interface, result in an increased probability of reflection in quasi-1D geometry, obscuring the observation of quantized conductance. In order to have highly transparent contacts, InSb wires have been exposed to a sulfur etch to remove the oxide layer before Cr/Au contact deposition. The resulting nanowire device, shown in the inset of Fig. \[fig:fig4\]a, consists of bottom gates on which an h-BN flake is deposited as a high-quality dielectric [@Kammhuber16]. The samples are mounted in a dilution refrigerator with a base temperature of 15 mK. The differential conductance $G$ is measured as a function of gate voltage and bias voltage, see Fig. \[fig:fig4\]a. At zero bias voltage, an extended plateau is visible at $ G_0=2 e^2/h$, demonstrating ballistic transport through the lowest nanowire subband. From these measurements, one can extract a mean-free-path $l \approx 300$nm. An additional small plateau at $2G_0$ is observed, indicating that the device has a small, but finite energy splitting between the second and third sub-bands. At finite bias voltage, the two-terminal conductance will only be quantized in integer values of $G_0$ if both $\mu_{\rm source}$ and $\mu_{\rm drain}$ occupy the same subband. This creates diamond shaped regions of constant conductance indicated by black dotted lines in Fig. \[fig:fig4\]a. At the tip of the diamond, the two dotted lines cross when $V_{\rm bias}$ is equal to the subband energy spacing $\Delta E_{sub}$ which can be estimated to be $\Delta E_{sub}\approx 16$ meV.
Phase-coherent transport is a basic requirement for the interferometric read-out schemes of the topological data, see Sec. \[sec:perspective\]. In order to study the transport properties of nanowire networks, the structures are transferred from the growth chip via a nanomanipulator in a scanning electron microscope (SEM) to the desired chip for device fabrication [^1]. An SEM image of the contacted nanowire hashtag structure is shown in the inset of Fig. \[fig:fig4\]b. The magnetoconductance of the device is shown in Fig. \[fig:fig4\]b where Aharonov-Bohm oscillations can be clearly seen. As expected, the period of the oscillations scales with the area of the hashtag devices demonstrating the phase coherent transport through the device. From temperature-dependent measurements, a phase coherence length of a few micrometers has been estimated.
As mentioned in the introduction, it is important to eliminate unintended subgap states in order to achieve a hard superconducting gap, which protects against thermal quasiparticle excitations. The density of states in the proximitized nanowire can be probed with tunneling spectroscopy by measuring the differential conductance as a function of source-drain voltage. A comparison between conventionally evaporated Al and epitaxially grown Al on InAs nanowires is shown in Fig. \[fig:fig4\]c. With epitaxially grown Al, the ratio of above-gap to subgap conductance is about 1/100. Interestingly, for thin Al on two facets, the superconducting gap remains hard, as shown in Fig. \[fig:fig4\]d. Furthermore, the size of the gap increases with decreasing Al thickness up to several nanometers [@Court2008]. The devices with thin Al layer also withstand high magnetic fields (up to $\sim$ 3 T) as shown in Fig. \[fig:fig4\]d satisfying an important requirement for reaching the topological phase transition [@Lutchyn10; @Oreg10].
Another important material requirement is a strong spin-orbit coupling (SOC) in the semiconductor. SOC is difficult to measure, especially in ballistic quasi-one-dimensional nanowires. A standard approach is via the quenching of weak antilocalization (WAL) by a small magnetic field. The field scale where the conductance peak at zero magnetic field associated with quantum-suppressed backscattering is quenched yields an estimate for the strength of SOC, within a rather complicated theoretical model [@Iordanskii1994] that includes different scattering processes. Fig. \[fig:fig4\]e shows a fit to a WAL peak from a 2DEG, see Ref. [@Shabani16] for details, from which the phase coherence length and Rashba SOC $\alpha$ can be extracted. The values measured in this InAs 2DEG are comparable to that measured in InAs and InSb nanowires [@Shabani16; @Weperen2015]. Similarly to the situation in nanowires, the density of states in 2DEGs can be measured via a tunnel probe, defined in the present case with a quantum point contact formed by top gates, see inset in Fig. \[fig:fig4\]f. The gate-controlled constriction can be tuned from high transparency regime (i.e., a single conducting channel showing $4e^2/h$ two-terminal conductance) to the weak tunneling regime where the conductance measurement is proportional to the local density of states.
Experimental signatures of Majoranas
====================================
Even when all material requirements are satisfied, there remains a question of how MZMs can be identified experimentally, and how unambiguous are various signatures. One should appreciate that even when the system is in a trivial phase, without well-separated MZMs, fine-tuning of some system parameters can lead to experimental observations with signatures that resemble those expected for MZMs. A system in the topological phase, however, shows these signatures in a robust way, i.e. they should persist under a moderate evolution of system parameters. It is this robustness that is the most characteristic trait of topological phases, and ultimately what makes these states useful for quantum information processing. In the following, we will discuss a [*set*]{} of experiments showing Majorana signatures, which increases our confidence that proximitized nanowires can be driven into a topological superconducting phase.
The simplest experiment involves tunneling conductance measurement of the local density of states at the nanowire ends [@ZeroBiasAnomaly0; @ZeroBiasAnomaly2; @ZeroBiasAnomaly3; @ZeroBiasAnomaly31; @ZeroBiasAnomaly4; @ZeroBiasAnomaly6; @ZeroBiasAnomaly61; @PradaPRB'12; @RainisPRB'13], see Fig. \[fig:fig1\]a. The dependence of the differential tunneling conductance on magnetic field is shown in Fig. \[fig:fig1\]b. As one can see, the characteristic zero-bias peak in tunneling conductance appears in a finite field (i.e. topological phase). The height of the zero-bias peak is predicted to be $2e^2/h$ at zero temperature [@ZeroBiasAnomaly3; @Fidkowski12; @lutchyn_andreev13]. In addition to the zero-bias tunneling spectroscopy, a number of current correlations measurements probing the presence of MZMs has been proposed [@Bolech2007; @ZeroBiasAnomaly2; @ZeroBiasAnomaly3; @Golub2011; @Haim2014; @Haim2015; @Liu2015a; @Liu2015b].
Fractional Josephson effect is another peculiar feature of MZMs. This effect corresponds to coherent tunneling of single electrons between two topological superconductors forming a Josephson junction. As a function of magnetic field, the spectrum of Andreev states should change from being $2\pi$-periodic in units of SC flux quantum to $4\pi$-periodic, see Fig. \[fig:fig1\]c. The former corresponds to tunneling of Cooper pairs whereas the latter appears due to the coherent single-electron tunneling processes. The characteristic crossing of two Andreev states at zero energy is protected by fermion parity in the junction [@kitaev01]. When parity is conserved, there is a true level crossing in the energy spectrum leading to a $4\pi$ periodicity of the supercurrent. In practice, finite-size effects lead to a small avoided level crossing, dynamical change of fermion parity may also occur due to unpaired electrons in the system. Therefore, supercurrent measurements should be performed at the sufficiently high frequency that these effects are avoided [@kitaev01; @Lutchyn10; @Jiang'11; @Pikulin2012; @Houzet2013; @PabloSanJose2013; @Setiawan2017].
Another signature of MZMs appears when semiconductor nanowire is proximitized by a floating superconductor (rather than grounded superconductor discussed so far) so that the overall system has a finite charging energy [@Fu10; @Zazunov2011; @Hutzen2012; @Hassler2015; @Heck16; @Lutchyn_charge'16; @LutchynGlazman2017]. The corresponding Hamiltonian reads $$\begin{aligned}
H_C=E_C (\hat {N}-{\cal N}_g)^2,\end{aligned}$$ where $\hat {N}$ is the total number of electrons in the mesoscopic system (island), $E_C=e^2/2C_{\Sigma}$ with $e$ and $C_{\Sigma}$ being the electron charge and the total capacitance of the nanowire and its superconducting shell, ${\cal N}_g=C_g V_g/e$ is the dimensionless gate voltage with $C_g$ and $V_g$ being the gate capacitance and gate voltage, respectively.
![ a) Simulated local density of states $\nu$ as a function of energy and coordinate along the nanowire for the proximitized nanowire in a topological phase with localized Majorana zero energy states at its ends. b) Differential tunneling conductance as a function of an in-plane magnetic field and applied voltage bias. Each line cut corresponds to a different magnetic field $B$. At the critical value of $B$-field, Majorana-induced zero-bias peak emerges. This feature is stuck to zero bias for a significant range of a magnetic field. Panels a) and b) adapted with permission from Ref. [@Stanescu11]. c) The energy spectrum of Andreev bound states in SC-NW-SC junction as a function of SC phase difference $\Phi$ across the junction. Red and blue lines correspond to the energy spectrum for trivial ($B=0$) and topological ($B>B_c$) superconductors. In the trivial phase, there is an avoided level crossing at $\Phi=\pi, 3\pi$ due to doubly degenerate spectrum enforced by the time-reversal symmetry, whereas in the topological state this crossing is protected by particle-hole symmetry [@kitaev01]. Dashed and solid blue energy states have different fermion parity. d) Charging energy of a proximitized nanowire as a function of the dimensionless gate voltage ${\cal N}_g$. Panel d) adapted with permission from Ref. [@Albrecht16]. Blue and green lines correspond to the odd-charge states of a conventional (trivial) superconductor with $\Delta > E_C$ and $\Delta < E_C$, respectively. The red line represents the ground-state energy of an odd-charge state in a topological superconductor, in which case even- and odd-charge states are (almost) degenerate due to the presence of Majorana zero-energy modes. []{data-label="fig:fig1"}](Fig4a){width="\columnwidth"}
In a conventional (s-wave) superconductor all electrons are paired in the ground state when the number of particles in a finite-size superconductor is even. Adding one more electron costs electrostatic energy as well as pairing energy $\Delta(B)$. Therefore, if $\Delta (B) > E_C$ the ground-state energy of the proximitized nanowire corresponds to an even-charge sector with $2e$-periodic dependence on ${\cal N}_g$, see Fig. \[fig:fig1\]d. The presence of a magnetic field leads to the suppression of the quasiparticle gap $\Delta(B)$, and the parity of the ground-state may have even- and odd-charge sectors, but the ground-state remains $2e$-periodic as long as $B < B_c$. In the topological phase ($B>B_c$), the periodicity of the ground-state energy changes due to the presence of MZMs and becomes $1e$-periodic [@Fu10] as shown in Fig. \[fig:fig1\]d. The combination of this fact and elastic transport through a Coulomb-blockaded island (Majorana island) provides strong evidence for MZMs in the proximitized nanowires [@Heck16]. Note that the corresponding transport through a metallic island (i.e. no MZMs) is dominated by inelastic processes [@aleiner2002].
An interesting scenario appears when Coulomb-blockaded island contains several pairs of MZMs. When such an island is coupled to several normal leads, Majorana-induced ground-state degeneracy leads to an exotic multi-channel Kondo effect [@Beri2012; @Altland2013; @Michaeli2016].
We emphasize again that all these properties characterize a topological phase and should persist in a finite parameter range. It may be possible to explain each of these physical observations separately within certain non-MZM scenarios by fine-tuning the system parameters. However, as discussed below, a large body of experimental data accumulated in different physical systems allows one to rule out many of the false-positive interpretations in favor of a MZM scenario.
![Tunneling conductance in InSb/Al hybrid structures. a) The layout of the device: upper panel - false-color electron micrograph, lower panel - device schematics. Here the side gates and the contacts are Cr/Au (10 nm/100 nm) and the Al shell thickness is 10 nm. The substrate is p-doped Si which acts as a global back-gate. The barrier between the normal lead and the proximitized nanowire can be controlled by the two tunnel gate voltages. The scale bar is 500 nm. b) Schematic plot of Andreev reflection processes at a normal metal/superconductor interface. c) Tunneling conductance as a function of an in-plane magnetic field (up to 1 Tesla). The induced gap in this device is $\Delta_{\rm ind}= 0.2$meV; the zero-bias peak is visible between 0.7 and 1Tesla. Lower panel: zero-bias conductance line-cut as a function of magnetic field. Here the temperature is $T=20mK$. d) Dependence of the tunneling conductance on the tunnel-gate voltage at a fixed $B = 0.8$T. Lower panels show horizontal line-cuts at zero and finite bias. The zero-bias conductance is robust in large parameter regime and remains close to the quantized value of $2e^2/h$. Panels a), c) and d) adapted with permission from Ref. [@Zhang2017]. []{data-label="fig:fig5"}](Fig5){width="\columnwidth"}
Zero-bias tunneling conductance measurements
--------------------------------------------
Tunneling spectroscopy is a conceptually simple method to measure local density of states and is well-suited to detect reconstruction of the energy spectrum across a topological phase transition. Given that MZMs are localized at the ends of the nanowire, the differential conductance measurement should exhibit robust zero-bias tunneling conductance peak in the topological phase. Typical device layout for tunneling spectroscopy measurement is shown in Fig. \[fig:fig5\]a. Low-bias transport in the normal-metal/superconductor junction is predominantly determined by the Andreev processes, during which an incident electron is reflected as a hole and a Cooper pair is added to the superconducting condensate, see Fig. \[fig:fig5\]b. In the non-topological phase ($B<B_c$) such processes are suppressed in a weakly-transparent junction. In contrast, the transmission probability becomes close to one in the topological phase [@ZeroBiasAnomaly3; @ZeroBiasAnomaly4; @ZeroBiasAnomaly6; @ZeroBiasAnomaly31; @ZeroBiasAnomaly61; @ZeroBiasAnomaly7; @PradaPRB'12; @RainisPRB'13; @Liu2017]. The presence of MZMs leads to a [*resonant*]{} Andreev reflection, and, as a result, differential tunneling conductance at low temperatures (compared to tunnel broadening) acquires a universal value of $2e^2/h$ [@Law09; @Fidkowski12; @lutchyn_andreev13].
Shortly after the theoretical proposals for proximitized nanowires [@Lutchyn10; @Oreg10], the emergence of the zero-bias peak in a finite magnetic field was observed in Ref. [@Mourik12]. Similar results were published in several subsequent studies [@Deng12; @Das12; @Finck12; @Churchill13; @Chen2016] in InSb and InAs nanowires. First-generation experiments [@Mourik12; @Deng12; @Das12; @Finck12; @Churchill13] reported the appearance of the zero-bias peak together with the closing of the bulk gap. The zero-bias feature persisted over a substantial range of magnetic fields and gate voltages and disappeared with the suppression of the superconductivity. Furthermore, the zero-bias peak disappeared when the in-plane magnetic field was aligned transverse to the wire axis, indicating that the effect depends on the interplay of Zeeman and spin-orbit couplings. All these observations are consistent with the Majorana scenario. At the same time, first-generation experiments revealed a large subgap density of states due to the interface disorder which co-existed with the zero-bias feature. It is challenging to separate the two contributions and various alternative scenarios have been proposed [@Law2012; @Bagrets2012; @Pikulin2012; @Kondo_Aguado; @Lee_arxiv2013].
Recent experiments with improved InSb/NbTiN [@Zhang16; @Chen2016] deposited interfaces, epitaxial InAs/Al [@Deng2016] nanowires and 2D InA/Al heterostructures [@Suominen2017; @Nichele2017], and epitaxial InAsSb/Al [@Sestoft2017] and InSb/Al [@Zhang2017] interfaces have a much lower subgap density of states, see Fig. \[fig:fig5\]c, and exhibit ballistic transport properties allowing one to rule out the scenario of the disorder-induced zero-bias conductance peak. It is the material science progress reviewed in this work that enabled significant improvement in electric transport and allowed one to investigate subtle transport properties such as quantization of the zero-bias tunneling conductance. Indeed, previous experiments [@Mourik12; @Deng12; @Das12; @Finck12; @Churchill13] reported zero-bias peak much smaller than the predicted value $2e^2/h$. Recent improvements of the samples quality led to the increase of the zero-bias conductance peak, and $G$ of order $e^2/h$ can be observed in InAs/Al [@Vaitiekenas2017] and InSb/Al [@Zhang2017] devices. Using improved InSb/NbTiN samples Chen [*et al.*]{} [@Chen2016] mapped out the topological phase diagram through the dependence of zero-bias peak on the chemical potential and magnetic field. These measurements allow one to rule out many false-positive scenarios. However, it was pointed out in Ref. [@Liu2017] that Andreev bound states may exhibit similar signatures. In order to distinguish between local Andreev bound states and delocalized MZMs, one has to carry out a detailed analysis of how the observed signal depends on temperature, tunnel coupling [@Nichele2017; @Setiawan2017], which was reported experimentally recently in 1D and 2D heterostructure devices [@Nichele2017; @Zhang2017], or consider more complex geometries, for instance with a quantum dot coupled to the end of a nanowire [@Leijnse2011; @Liu2011; @Lee2013; @Kondo_Aguado; @Lee_arxiv2013; @Cheng2014a; @Clarke2017; @Prada2017], which allows a measurement of the non-locality of MZMs [@Deng2017]. The authors of Ref. [@Zhang2017] showed that tunneling conductance due to local Andreev states depends on the tunnel barrier height, see Fig. \[fig:fig5\]a, whereas Majorana-induced zero-bias peak is robust with respect to the tunnel coupling and, in fact, exhibits close to $2e^2/h$ conductance plateau [@Zhang2017].
Taken altogether, recent tunneling spectroscopy measurements provide a strong evidence for MZMs in semiconductor-superconductor heterostructures. The high quality of materials, as well as improved control in devices, open a possibility to study more exotic properties of MZMs.
Coulomb blockade experiments with proximitized nanowires
--------------------------------------------------------
The significant improvement of the interface quality in epitaxially-grown proximitized nanowires opens a possibility to study an interplay of topological superconductivity and Coulomb blockade phenomena [@Higginbotham15; @Albrecht16]. The device consists of a mesoscopic proximitized nanowire with a finite charging energy $E_C$ and is coupled to two normal leads, see Fig. \[fig:fig6\]a. In the Coulomb blockade regime, the charge on the island can be controlled by changing the gate voltage $V_g$, as well as electrostatically tuning single-channel barriers. Both these knobs enable one to investigate the crossover from strong to weak Coulomb blockade regimes and study quantum charge fluctuations in the island. Quantum charge fluctuations allow one to identify the topological phase since they have an imprint of the physical state of the system [@Lutchyn_charge'16].
![Coulomb blockade experiment with proximitized nanowires. a) Electron micrograph (false color) of a device that is lithographically similar to the measured devices. Gold contacts (yellow), InAs nanowire (green), and two-facet Al shell of length $L$ (light blue). Applied voltage bias, $V_{SD}$, source-drain current, $I$, and gate voltage, $V_G$, are indicated. Inset: Cross section of a hexagonal InAs nanowire, showing orientation of Al shell and field directions $B_{||}$ and $B_{\perp}$. b) Two-terminal conductance, $g$, as a function of an external magnetic field $B_{||}$ and gate voltage for $L=1.5\mu$m. Coulomb peaks become dim at field $B^*$ and brighten at field $B_c$. The field $B^*$ is the characteristic scale for the crossover from coherent Cooper-pair to incoherent single-electron tunneling regime whereas the field $B_c$ corresponds to the topological phase transition. At $B>B_c$ transport is dominated by the coherent single-electron tunneling processes due to MZMs [@Fu10; @Heck16]. c) Zero-bias conductance, $g$, as a function of gate voltage, $V_G$, and parallel magnetic field, $B_{||}$, for $L=0.9\mu$m device, showing a series of $2e$-periodic Coulomb peaks below $100$mT and $1e$ nearly-periodic peaks above 100 mT. d) and e) Average peak spacing for even and odd Coulomb valleys, $S_{e/o}$, as a function of magnetic field for L = 0.9 $\mu$m and L = 1.5 $ \mu$m, respectively. Inset: Even and odd peak spacings, $S_{e/o}$, are indicated by arrows. f) Oscillatory amplitude, $A$, plotted against the proximitized nanowire length, $L$, for 5 devices from 330 nm to 1.5 $\mu$m (black dots) and exponential fit to $A = A_0 \exp(-L/\xi)$ with $A_0$= 300 $\mu$eV and $\xi = 260$ nm. Error bars indicate uncertainties propagated from lever arm measurements and fits to peak maxima. Panels a) - f) adapted with permission from Ref. [@Albrecht16].[]{data-label="fig:fig6"}](Fig6){width="\columnwidth"}
As explained above, physical observables change the dependence on the dimensionless gate voltage ${\cal N}_g$ across the topological phase transition [@Fu10] - charge periodicity changes from $2e$ to $1e$. Therefore, two-terminal conductance $G({\cal N}_g)$ through the island at zero-bias should exhibit the crossover between these two limits as a function of magnetic field. Theory [@Heck16] predicts that the conductance is determined by different processes as magnetic field is increased from zero: (I) Cooper-pair dominated regime for $B < B^*$; (II) sequential tunneling regime ($B^* < B < B_c$); (III) Majorana-dominated regime for $B>B_c$. Here $B^*$ is the field at which ground-state can have different parity, i.e. $\Delta(B^*)=E_C$. In the first limit, the ground state of the system involves only even-parity sectors. Therefore, the zero-bias conductance peaks appear at the degeneracy points between nearest even-charge states, see Fig. \[fig:fig6\]b. In the case (II), ground state as a function of ${\cal N}_g$ changes its parity between even- and odd-charge sectors. Zero-bias conductance peaks now correspond to the change in ground-state parity. As a result, each conductance peak splits into two at $B=B^*$, see Fig. \[fig:fig6\]b, and this splitting is increasing with magnetic field until it reaches $B=B_c$. Note that at these degeneracy points the dominant conduction mechanism is the resonant transfer of single electrons rather than Cooper pairs. In this intermediate regime, the peak conductance at low temperatures becomes suppressed, see Fig. \[fig:fig6\]b, and, in fact, should vanish in the thermodynamic limit $L \rightarrow \infty$ [@Averin'93; @Heck16]. Finally, in the topological regime (III) peak positions occur at half-integer values of ${\cal N}_g$, independent of the magnetic field (up to finite-size corrections). In this limit, the dominant contribution to the peak conductance originates from the resonant tunneling via MZMs. Due to the non-local nature of the fermionic mode shared between two MZMs, the height of conductance peak is much larger than in the case (II) and is independent of $L$.
In order to investigate the non-local nature of the topological ground-state degeneracy, one may study the splitting energy originating from an exponentially-small Majorana wavefunction overlap: $\delta E \propto \exp(-L/\xi)$. This difference for even- and odd-parity states results in a small shift of the charge degeneracy points. Thus, the peak spacing within the Majorana scenario in the even- ($S_e$) and odd-charge ($S_o$) sectors should be different, whereas in the normal-island case $A=S_e-S_o$ is negligibly small. As shown in Figs. \[fig:fig6\]c-f, odd-even spacing $A$ does depend on the nanowire length and, thus, the experimental results [@Albrecht16] are consistent with the MZM scenario. The splitting measurement provides important information about the physical system (e.g., superconducting coherence length), and it would be quite interesting to analyze this issue more systematically in the future and to compare experimental results with a microscopic simulation of the proximitized nanowires [@Chiu2017].\
Perspective for topological quantum computation {#sec:perspective}
===============================================
During the last five years we witnessed the birth of a new field of research - mesoscopic topological superconductivity [@Fu10; @Alicea11; @Zazunov2011; @Hutzen2012; @Hyart13; @Beri2012; @Altland2013; @Altaland2014; @Cheng2014; @Dong2015; @Albrecht16; @Landau16; @Hoffman2016; @Michaeli2016; @Heck16; @Aasen16; @Plugge16a; @Vijay2016; @Plugge16b; @Karzig16; @Lutchyn_charge'16; @LutchynGlazman2017], and saw an impressive growth of experimental explorations of various superconducting heterostructures [@Mourik12; @Deng12; @Das12; @Finck12; @Churchill13; @Chang14; @Krogstrup15; @Higginbotham15; @Kammhuber16; @Shabani16; @Zhang16; @Deng2016; @Albrecht2017; @Gazibegovic2017]. The interplay of mesoscopic and topological physics enables one to manipulate quantum information stored in topological (non-local) degrees of freedom in a way that minimizes decoherence effects. Remarkable experimental progress, rapid improvement in the materials quality as well as our enhanced theoretical understanding of the microscopic details are encouraging developments for the prospect of building topological qubits in superconductor-semiconductor heterostructures.
Quantum gates within the topological quantum computation approach rely on non-local transformations within the degenerate ground-state manifold, which can be performed either by adiabatically changing physical parameters of the system [@Alicea11; @Hyart13; @Bonderson2013; @Barkeshli2015; @Aasen16], or using the projective measurements [@Bonderson08b; @Bonderson2009; @Plugge16b; @Karzig16]. We now outline several immediate directions, which should be explored both experimentally and theoretically in order to test the necessary ingredients for a successful operation of a futuristic quantum computer, see Fig. \[fig:fig7\].
![Proposed future experiments probing key properties of topological superconductors for quantum computing applications. a) Majorana fusion experiment. The device, consisting of two mesoscopic Majorana islands (orange) with gate-tunable tunnel barriers (valves). Charging energy of the island can be quenched by lowering the barrier to a bulk superconductor (dark blue). The middle barrier connects the islands. When the middle valve is completely open, one effectively achieves single-island configuration. The specific protocol of opening (closing) the valves allows one to investigate Majorana fusion rules. b) Topological T-junction for braiding experiment. Panels a) - b) reproduced with permission from Ref. [@Aasen16]. Anyonic state teleportation, equivalent to physically braiding Majoranas, can be achieved by closing (opening) the valves as explained in Ref. [@Alicea11]. c) Majorana interferometer - a basic element of the Majorana surface code [@Landau16; @Vijay2016; @Plugge16a]. Coherent transport through the interferometer probes non-local nature of Majorana zero modes. The position of MZMs $\gamma_i$ is represented as red dots. d) Minimal setup for a quasiparticle-poisoning-protected topological qubit [@Plugge16b; @Karzig16]. Mesoscopic superconducting island hosting four MZMs (red dots) is in a Coulomb blockade regime with fixed total charge. The island has two-fold ground-state degeneracy and represents the simplest topological qubit. Semiconductor quantum dot, coupled to the qubit via gate-tunable barriers (i.e. the barriers can be controlled with gate voltages $V_{Gi}$ ), is used to perform projective measurements. Parity-dependent charge fluctuations between SM dot and SC island shift energy level in QD, allowing one to distinguish between different degenerate states of the qubit. The energy difference between even- and odd-charge states can be tuned with $V_{QD}$. e) Scalable architecture for Majorana-based quantum computing. Topological qubits are coupled together via semiconducting quantum dots. Quantum computation is performed using measurement-only protocol, which is facilitated by tunable couplings between Majorana zero modes and nearby semiconductor quantum dots [@Bonderson08b]. Panel e) reproduced with permission from Ref. [@Karzig16]. []{data-label="fig:fig7"}](Fig7){width="\columnwidth"}
As a first step, we need to establish that the exotic non-local properties of MZMs are indeed realized in the experimental system. A hallmark of topological states of matter is ground-state degeneracy which so far has not been tested experimentally. Fusion and braiding rules associated with zero modes define a topological phase and are of fundamental importance for topological quantum computing. The fusion and braiding experiments can be performed in a small network of nanowires as shown in Fig. \[fig:fig7\]a and b. The fusion experiment is simpler than braiding as it does not require an exchange of MZM position but nevertheless reveals topological properties of the system.
In order to test fusion rules, we follow the ideas outlined in Refs. [@Alicea11; @Aasen16] and consider a device consisting of two Majorana islands connected together and to the superconducting leads by gate-tunable tunnel barriers (valves), see Fig. \[fig:fig7\]a. The effective charging energy of the left (right) island can be reduced by lowering the left (right) barrier to a superconducting lead. Let us denote open (low barrier) and closed (high barrier) valve as “1” and “0”, respectively. The string ‘010’, for example, represents the configuration where the left and right valves are closed whereas the middle valve is open. This configuration of valves effectively merges the left and right islands to a single island disconnected from the superconducting reservoirs. As shown in Ref. [@Aasen16], two distinctive sequences of open and closed valves configurations, differing only by one intermediate configuration, should eventually lead to different final charge configurations on the islands. The latter reflects the fusion rules of non-Abelian MZMs. While the sequence $a_1$: $010 \rightarrow 000 \rightarrow 101 \rightarrow 111$ leads to a definite deterministic charge configuration on the left and right islands; the sequence $a_2$: $010\rightarrow 111 \rightarrow 101 \rightarrow 111$ results in an equal superposition of two distinct charge distributions. A measurement of the charge on the island following an execution of sequence $a_1$ should be deterministic. In contrast, after the execution of sequence $a_2$ one would expect to find two charge configurations with equal probability.
In order to perform braiding operations three Majorana wires forming a T-junction, see Fig. \[fig:fig7\]b, are necessary [@Alicea11; @Aasen16]. The elementary sequence requires controlling five valves, i.e. keeping the leftmost and the rightmost valves open and tuning the remaining three valves [@Aasen16]. Following the aforementioned notations, the sequence $b_1$ - $00000 \rightarrow 01100 \rightarrow 00110 \rightarrow 00000$ - results in braiding of the two MZMs next to the operative left and right valves. Here the numbering of valves goes from left to right with the middle number representing the coupling of the middle island to the lead. Note that double braiding operation should always lead to a flip of the parity of the qubit which is another non-trivial test.
At the heart of measurement-only topological quantum computation is a projective measurement of the joint fermion parity encoded in a pair of MZMs. In this scheme, an effective braiding between two non-Abelian anyons is achieved by measuring the topological charge of anyonic pairs (i.e. fermion parity) rather than physically exchanging their position [@Bonderson08b; @Bonderson2009]. Interferometric measurement is a promising way to distinguish between even and odd fermion parities, see, for example, the device shown in Fig. \[fig:fig7\]c. As explained above, transport through a proximitized nanowire in a Coulomb blockade regime changes its periodicity with the gate charge from $2e$ and $1e$ as a function of applied magnetic field. In the topological regime (large $B$-field), charge transport is dominated by the coherent single-electron transmission through MZMs [@Fu10; @Heck16; @Lutchyn'17] which is very different from the inelastic transport in a normal-metal island although both systems exhibit $1e$ periodic signal. Thus, the interferometric measurement should be able to distinguish between different tunneling processes: transmission via MZMs or tunneling through some other delocalized subgap states. The former preserves electron coherence and, therefore, should lead to the dependence on the enclosed magnetic flux $\Phi$ whereas the latter does not. Although this setup cannot access properties of topological qubits (since the MZM parities are fixed by the charging energy), such an experiment is a crucial test of the concept of measuring MZM states using coherent links and represents a basic building block of Majorana-based surface [@Landau16; @Vijay2016; @Plugge16a] and color [@Litinski2017] codes.
In order to perform quantum information processing the mesoscopic superconducting island should support at least four MZMs maintaining ground-state degeneracy with a fixed number of electrons due to a large charging energy. The simplest topological qubit is depicted in Fig. \[fig:fig7\]d. The projective measurement can be performed by coupling MZMs to a central semiconductor quantum dot, see Fig. \[fig:fig7\]d, and measuring, for example, the quantum capacitance of the system. Indeed, as shown in Refs. [@Flensberg2011a; @Plugge16b; @Karzig16], virtual exchange of electrons between the semiconductor quantum dot and Majorana island results in a parity-dependent contribution to the ground-state energy which ultimately leads to a change of QD charge distribution. Majorana islands supporting six MZMs open the possibility to perform a measurement-only braiding operation by the sequential measurement of the fermion parity of different pairs of MZMs [@Karzig16].
Finally, an array of topological qubits coupled together via semiconducting quantum dots represents a scalable architecture for Majorana-based quantum computing [@Karzig16]. An example of such architecture is shown in Fig. \[fig:fig7\]e. Quantum computation is performed using the measurement-only protocol which is facilitated by tunable couplings between MZMs and nearby semiconductor quantum dots. The detailed protocol for implementation of Clifford gates using a limited set of projective measurements in this architecture is discussed in Ref. [@Karzig16]. Combining the topologically-protected Clifford gates with the ability to produce and distill magic states [@Bravyi'05; @Bravyi'06; @SauWireNetwork; @JiangKanePreskill; @TopologicalQuantumBus; @Karzig'15; @Clarke2016] should ultimately lead to universal quantum computation with MZMs.
Summary
=======
The search for Majorana zero modes in superconductor-semiconductor heterostructures is evolving at a rapid pace and now encompasses a broad range of disciplines: from material science to condensed matter physics and quantum information science. In less than a decade since the theoretical proposals [@Lutchyn10; @Oreg10], we have accumulated a compelling evidence for MZMs in proximitized nanowires. Most of the experiments searching for Majoranas focused on tunneling spectroscopy. The recent work on InSb/Al proximitized nanowires [@Zhang2017] reported robust $2e^2/h$ conductance consistent with the Majorana scenario. Chen et al. [@Chen2016] mapped out the phase diagram through the dependence of zero-bias peak on chemical potential and magnetic field. The experiments with floating proximitized nanowires [@Albrecht16] set the stage for the exploration of Coulomb blockade effect in mesoscopic topological superconductors. A measurement of $4\pi$ fractional Josephson effect was recently reported in Ref. [@Geresdi2017]. While these experiments strongly support the existence of the emergent MZMs in proximitized nanowires, a direct observation of the non-Abelian properties associated with Majoranas is still lacking. In Sec. \[sec:perspective\], we outlined future experiments for the detection of more exotic properties of MZMs such as braiding and fusion rules. Observing two MZMs at the ends of a single wire simultaneously is a necessary condition for that.
Experiments with proximitized nanowires do not only contribute to our understanding of topological phases of matter but also have significant theoretical and practical ramifications that span beyond condensed matter physics. The next-generation experiments with Majorana nanowires will set the stage for the manipulation of quantum information in topological systems. Scalable designs for topological quantum computation based on superconducting islands hosting multiple pairs of MZMs have been recently put forward [@Plugge16b; @Karzig16].
Experimental demonstration of the exotic physics associated with MZMs, such as non-Abelian statistics, would constitute a breakthrough for all of fundamental physics, as well as pave the way for the next important milestone - a validation of a topological qubit operating in the exponentially-protected regime and exhibiting long coherence times. These accomplishments would lead to the exploration of different ideas for topological quantum memory and fault-tolerant universal quantum computation.
Box 1: A minimal model for 1D topological superconductor {#box}
========================================================
The minimal model for 1D topological superconductor involves semiconductor nanowire with strong spin-orbit coupling proximity-coupled to a conventional (s-wave) superconductor [@Lutchyn10; @Oreg10]. The effective low-energy Hamiltonian for such a system is given by $$\begin{aligned}
H&=H_{\rm SM}+H_{\rm P},\label{eq:H0}\\
\!\!H_{\rm SM}\!&=\sum_{\sigma, \sigma'}\!\int_{0}^{L} \!\!\!d x \psi_{\sigma}^\dag(x)\!\left(\!-\!\frac{\hbar^2\partial_x^2}{2m^*}\!-\!\mu\!+\!i \hbar \alpha \hat{\sigma}_y \partial_x\!+\!V_Z \hat{\sigma}_x\!\right)_{\sigma\sigma'}\!\!\!\!\!\psi_{\sigma'}(x),\nonumber\\
H_{\rm P}&= \int_{0}^{L} \!d x \left(\Delta_{\rm ind} \psi_{\uparrow}^\dag(x) \psi^\dag_{\downarrow}(x)+h.c. \right).\nonumber\end{aligned}$$ Here $\hbar$ is the Planck constant; $m^*$, $\mu$ and $\alpha$ are the effective mass, chemical potential, and Rashba spin-orbit coupling in the semiconductor nanowire of length $L$, respectively. $V_Z$ is the Zeeman splitting due to the applied magnetic field $B_x$: $V_Z\!=\!g \mu_B B_x$ with $g$ and $\mu_B$ being the Lande g-factor and Bohr magneton. $B_x$ is an external magnetic field applied along the nanowire, and $\hat{\sigma}_i$ are Pauli matrices. The proximity to the s-wave superconductor is effectively described by the Hamiltonian $H_P$ with $\Delta_{\rm ind}$ being the induced pairing gap. Typical material parameters for InAs and InSb semiconductors are provided in Table \[Table1\] whereas values for the induced gap are given in Table \[Table2\].
Although semiconductor nanowire is coupled to an s-wave superconductor, the presence of the spin-orbit coupling leads to spin-momentum locking and results in a mixed Cooper-pair wave function consisting of the singlet and triplet pairings [@GorkovRashba2001]. When Zeeman term opens a large gap in the spectrum (i.e. $V_Z > \sqrt{\mu^2+\Delta_0^2}$), the singlet pairing component is suppressed, and the superconducting state forming at the interface has p-wave symmetry of the order parameter. This state is adiabatically connected to Kitaev’s model [@kitaev01] and supports MZMs at the opposite ends of the wire.
[189]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1126/science.1223302) [****, ()](\doibase 10.1126/science.1260282) [****, ()](\doibase 10.1007/BF02961314) @noop [****, ()]{} [****, ()](\doibase 10.1103/RevModPhys.80.1083), [****, ()](\doibase 10.1016/0550-3213(91)90407-O) [****, ()](\doibase 10.1103/PhysRevB.61.10267) [****, ()](\doibase 10.1103/RevModPhys.83.1057) [****, ()](\doibase
10.1146/annurev-conmatphys-030212-184337), [****, ()](\doibase 10.1088/0034-4885/75/7/076501), [****, ()](\doibase 10.1088/0268-1242/27/12/124003), [****, ()](\doibase 10.1088/0953-8984/25/23/233201), [****, ()](\doibase 10.1103/RevModPhys.87.137) [**** ()](http://10.1038/npjqi.2015.1), [****, ()](\doibase 10.7566/JPSJ.85.072001), [****, ()](\doibase 10.1393/ncr/i2017-10141-9) [****, ()](\doibase 10.1103/PhysRevLett.100.096407) [****, ()](\doibase 10.1103/PhysRevB.79.161408) [****, ()](\doibase 10.1103/PhysRevB.84.201105) [****, ()](\doibase
10.1103/PhysRevLett.116.257003) [****, ()](\doibase
10.1103/PhysRevLett.104.040502) [****, ()](\doibase 10.1103/PhysRevB.81.125318) [****, ()](\doibase 10.1103/PhysRevLett.105.077001), [****, ()](\doibase 10.1103/PhysRevLett.105.177002), [****, ()](\doibase
10.1103/PhysRevB.84.060510) [****, ()](\doibase 10.1103/PhysRevB.83.054513) [****, ()](\doibase 10.1103/PhysRevB.85.094516) [****, ()](\doibase 10.1103/PhysRevB.84.195442) @noop [****, ()]{} @noop [****, ()]{}, @noop [****, ()]{}, @noop [****, ()]{}, [****, ()](\doibase 10.1103/PhysRevB.88.155420) [****, ()](\doibase 10.1103/PhysRevB.88.180503) [****, ()](\doibase
10.1103/PhysRevB.90.060401) [****, ()](\doibase 10.1103/PhysRevB.91.064505), [****, ()](\doibase
10.1103/PhysRevB.90.235433) [****, ()](\doibase 10.1103/PhysRevB.90.060507) [****, ()](\doibase 10.1103/PhysRevLett.114.236803) [ (), 10.1126/science.1259327](\doibase 10.1126/science.1259327) [****, ()](\doibase
10.1103/PhysRevLett.115.197204) [****, ()](\doibase 10.1038/npjqi.2016.35), [****, ()](\doibase
10.1103/PhysRevB.93.024507) [****, ()](\doibase 10.1126/science.1222360), [****, ()](\doibase 10.1038/nphys2429), [****, ()](\doibase
10.1021/nl303758w), [****, ()](\doibase 10.1103/PhysRevB.87.241401), [****, ()](\doibase 10.1038/nphys2479), [****, ()](\doibase 10.1103/PhysRevLett.110.126406) [****, ()](\doibase 10.1038/nature17162), @noop [ ()]{}, @noop [ ()]{}, @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.119.176805), [****, ()](\doibase
10.1103/PhysRevLett.119.136803), @noop [ ()]{}, @noop [ ****, ()]{} @noop [ ()]{}, @noop [ ()]{}, @noop [ ()]{}, [****, ()](\doibase 10.1070/1063-7869/44/10S/S29), [****, ()](\doibase 10.1103/PhysRevB.63.224204) [****, ()](\doibase
10.1103/PhysRevB.84.144526) [****, ()](\doibase 10.1103/PhysRevB.84.144522), [****, ()](\doibase 10.1103/PhysRevLett.106.057001) [****, ()](\doibase 10.1103/PhysRevLett.109.267002) [****, ()](\doibase 10.1103/PhysRevLett.109.146403) [****, ()](\doibase 10.1103/PhysRevB.83.184520) [****, ()](\doibase 10.1103/PhysRevB.85.140513) [****, ()](\doibase 10.1103/PhysRevLett.110.146404) [****, ()](\doibase
10.1103/PhysRevLett.110.186803) [****, ()](\doibase 10.1103/PhysRevB.89.144506) [****, ()](\doibase 10.1103/PhysRevB.92.174512) [****, ()](\doibase 10.1103/PhysRevB.94.140505) [****, ()](\doibase 10.1103/PhysRevB.94.115166) @noop [ ()]{}, @noop [**]{} (, ) [**](https://cds.cern.ch/record/684956), Springer tracts in modern physics (, , ) [****, ()](\doibase 10.1103/PhysRev.111.132) [****, ()](\doibase 10.1103/PhysRevB.93.155402), [****, ()](\doibase
10.1103/PhysRevB.91.201413) [****, ()](\doibase 10.1103/RevModPhys.82.3045) [****, ()](http://stacks.iop.org/1367-2630/13/i=10/a=105007) [****, ()](\doibase 10.7566/JPSJ.82.102001), @noop [**]{} (, ) [****, ()](\doibase
10.1103/RevModPhys.88.021004) [****, ()](\doibase 10.1038/nature23468), @noop [**** ()]{} [****, ()](http://stacks.iop.org/0957-4484/20/i=49/a=495606) [****, ()](\doibase
10.1021/cg1006814) [****, (), http://dx.doi.org/10.1063/1.3402760](\doibase
http://dx.doi.org/10.1063/1.3402760) [****, ()](\doibase
10.1021/cg200066q) [****, ()](\doibase
10.1021/nl203846g) @noop [****, ()]{} [****, ()](\doibase
10.1021/nl402571s) [****, ()](\doibase
10.1002/adma.201400924) [****, ()](\doibase 10.1021/nl400820w) [****, ()](\doibase
10.1002/aelm.201500460), [****, ()](\doibase
10.1021/acs.nanolett.5b05157) [****, ()](\doibase
10.1021/acs.nanolett.7b02604) [****, ()](\doibase 10.1038/nmat4176), [****, ()](\doibase 10.1038/nnano.2014.306), [****, ()](\doibase
10.1021/acs.nanolett.7b03444), [****, ()](\doibase 10.1039/C7NR03982D) [****, ()](\doibase 10.1088/0953-2048/21/01/015013), [****, ()](\doibase
10.1021/acs.nanolett.6b00051), [****, ()](\doibase
10.1038/ncomms12841), @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevB.63.144531) [****, ()](\doibase 10.1103/PhysRevLett.101.120403) [****, ()](\doibase 10.1103/PhysRevLett.103.237001) [****, ()](\doibase
10.1103/PhysRevB.82.214509) [****, ()](\doibase 10.1103/PhysRevB.82.180516) [****, ()](\doibase 10.1088/1367-2630/13/5/053016) [****, ()](\doibase 10.1103/PhysRevB.84.144522) [****, ()](\doibase 10.1103/PhysRevB.86.180503), [****, ()](\doibase
10.1103/PhysRevB.87.024515) [****, ()](\doibase 10.1103/PhysRevB.85.245121), [****, ()](\doibase 10.1103/PhysRevB.88.024511) [****, ()](\doibase 10.1103/PhysRevLett.98.237002) [****, ()](\doibase 10.1103/PhysRevB.83.153415), [****, ()](\doibase 10.1103/PhysRevLett.114.166406), [****, ()](\doibase 10.1103/PhysRevB.92.245112), [****, ()](\doibase 10.1103/PhysRevB.91.081405), [****, ()](\doibase 10.1103/PhysRevB.92.205422), [****, ()](\doibase
10.1103/PhysRevLett.107.236401) [****, ()](http://stacks.iop.org/1367-2630/14/i=12/a=125011) [****, ()](\doibase 10.1103/PhysRevLett.111.046401) [****, ()](http://stacks.iop.org/1367-2630/15/i=7/a=075019) [****, ()](\doibase 10.1103/PhysRevB.95.060501) [****, ()](\doibase 10.1103/PhysRevLett.104.056402), [****, ()](\doibase 10.1103/PhysRevB.84.165440) [****, ()](\doibase
10.1103/PhysRevLett.109.166403) [****, ()](\doibase 10.1103/PhysRevB.92.075443) [****, ()](\doibase 10.1103/PhysRevB.93.235431), [****, ()](\doibase 10.1103/PhysRevB.94.125407) [****, ()](\doibase 10.1103/PhysRevLett.119.057002), @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.109.156803) [****, ()](\doibase 10.1103/PhysRevLett.110.196401), [****, ()](\doibase
10.1103/PhysRevB.96.205403) @noop [ ()]{} [****, ()](\doibase 10.1103/PhysRevB.96.075161), [****, ()](\doibase 10.1103/PhysRevLett.103.237001), [****, ()](\doibase 10.1103/PhysRevLett.109.267002) [****, ()](\doibase 10.1103/PhysRevLett.109.227005) [****, ()](\doibase
10.1103/PhysRevLett.109.186802) @noop [****, ()]{} @noop [ ()]{}, [****, ()](\doibase 10.1103/PhysRevB.84.140501), [****, ()](\doibase 10.1103/PhysRevB.84.201308), [****, ()](\doibase 10.1103/PhysRevB.87.241402) [****, ()](\doibase 10.1103/PhysRevX.4.031051), [****, ()](\doibase 10.1103/PhysRevB.96.201109) [****, ()](\doibase 10.1103/PhysRevB.96.085418) [****, ()](\doibase
10.1038/nphys3461), [****, ()](\doibase 10.1103/PhysRevLett.69.1993) [****, ()](\doibase 10.1103/PhysRevB.96.054504), [****, ()](\doibase 10.1038/nphys1915), [****, ()](\doibase
10.1103/PhysRevB.88.035121), [****, ()](\doibase 10.1103/PhysRevLett.113.076401), [****, ()](\doibase
10.1103/PhysRevX.4.031051) [****, ()](\doibase 10.1103/PhysRevB.92.205422) [****, ()](\doibase
10.1103/PhysRevLett.116.050501), [****, ()](\doibase
10.1103/PhysRevB.94.045316) [****, ()](\doibase 10.1103/PhysRevX.6.031016), [****, ()](\doibase 10.1103/PhysRevB.94.174514) [****, ()](\doibase 10.1088/0031-8949/T168/1/014002), [****, ()](\doibase 10.1088/1367-2630/aa54e1), [****, ()](\doibase 10.1103/PhysRevB.95.235305), [****, ()](\doibase 10.1103/PhysRevLett.118.137701) [****, ()](\doibase 10.1103/PhysRevB.87.035113) @noop [ ()]{}, [****, ()](\doibase 10.1103/PhysRevLett.101.010501), [****, ()](\doibase 10.1016/j.aop.2008.09.009), @noop [ ()]{}, [****, ()](\doibase 10.1103/PhysRevX.7.031048), [****, ()](\doibase 10.1103/PhysRevLett.106.090503) [****, ()](\doibase 10.1103/PhysRevA.71.022316) [****, ()](\doibase 10.1103/PhysRevA.73.042313) [****, ()](\doibase 10.1103/PhysRevA.82.052322) [****, ()](\doibase 10.1103/PhysRevLett.106.130504) [****, ()](\doibase 10.1103/PhysRevLett.106.130505) [****, ()](\doibase
10.1103/PhysRevX.6.031019) [****, ()](\doibase 10.1103/PhysRevX.6.021005), [****, ()](\doibase 10.1103/PhysRevLett.87.037004)
[^1]: see movie at http://pubs.acs.org/doi/suppl/10.1021/acs.nanolett.7b00797
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Miaojing Shi$^1$, Zhaohui Yang$^2$, Chao Xu$^2$, Qijun Chen$^3$\
$^1$Univ Rennes, Inria, CNRS, IRISA\
$^2$Key Laboratory of Machine Perception, Cooperative Medianet Innovation Center, Peking University\
$^3$Department of Control Science and Engineering, Tongji University\
bibliography:
- 'bibtex/eccv2018.bib'
title: Revisiting Perspective Information for Efficient Crowd Counting
---
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
A. B. Yanovski$^1$ and T. I. Valchev$^2$\
$^1$ Department of Mathematics & Applied Mathematics,\
University of Cape Town, Rondebosch 7700, Cape Town, South Africa\
$^2$ Institute of Mathematics and Informatics,\
Bulgarian Academy of Sciences, Acad. G. Bonchev Str., 1113 Sofia, Bulgaria\
E-mails: [email protected],[email protected]
title: |
Hermitian and Pseudo-Hermitian Reduction\
of the GMV Auxiliary System. Spectral Properties\
of the Recursion Operators
---
Introduction. The GMV System {#sec:1}
============================
We are going to study the auxiliary linear problem $$\begin{aligned}
\label{eq:gsys}
&&\tilde{L}^0\psi=(\ri\partial_x - \lambda S)\psi=0,\qquad \lambda\in\mathbb{C}, \qquad S = \left(\begin{array}{ccc} 0&u&v\\ \epsilon u^*&0&0\\v^*&0&0
\end{array}\right), \quad \epsilon=\pm 1\end{aligned}$$ and the theory of expansions over its adjoint solutions. In the above, the potential functions $(u,v)$ are smooth complex valued functions on $x\in\mathbb{R}$ and $^*$ stands for the complex conjugation. In addition, $u$ and $v$ satisfy the relations: $$\label{eq:uvcon}
\epsilon |u|^2+|v|^2=1, \qquad \lim_{x\to\pm\infty} u(x)=u_{\pm}, \qquad \lim_{x\to\pm\infty} v(x)=v_{\pm}\, .$$ We shall call $\rm{GMV}_{\epsilon}$ system or $\rm{GMV}_{\pm}$ system[^1]. Thus, $\rm{GMV}_{+}$ is the original Gerdjikov-Mikhailov-Valchev system [@GMV1] obtained after putting $\epsilon=+1$ in .
As demonstrated in [@GMV1; @GMVSIGMA; @YanVal2017], the $\rm GMV_{\pm}$ system arises naturally when one looks for integrable systems whose Lax operators are subject to Mikhailov-type reductions. Indeed, the Mikhailov reduction group $G_0$ [@Mikh1; @Mikh2] acting on the fundamental solutions of is generated by $g_1$ and $g_2$ defined in the following way: $$\label{eq:MG}
\begin{array}{l}
g_1(\psi)(x,\lambda)=\left[Q_{\epsilon}\psi(x,\lambda^{*})^{\dag}Q_{\epsilon}\right]^{-1},\qquad Q_{\epsilon}=\diag(1,\epsilon,1)\; ,\\[4pt]
g_2(\psi)(x,\lambda)=H\psi(x,-\lambda)H \; , \qquad H=\diag(-1,1,1)
\end{array}$$ where $\psi$ is any fundamental solution to and $\dag$ denotes Hermitian conjugation. Since $g_1g_2=g_2g_1$ and $g_1^2=g_2^2=\id$, we have that $G_0=\mathbb{Z}_2\times \mathbb{Z}_2$. Reduction conditions will be called Hermitian when $\epsilon =1$ and pseudo-Hermitian when $\epsilon =-1$. The requirement that $G_0$ is a reduction group for the soliton equations related to implies that the coefficients of $\tilde{L}^0$ and the coefficients of the $A$-operators $$\label{eq:LApairs}
\tilde{A}=\ri\partial_t + \sum_{k=0}^n\lambda^k \tilde{A}_k \; ,\qquad \tilde{A}_k\in \asl(3,\mathbb{C})\; ,$$ forming $L-A$ pairs for these soliton equations, must satisfy: $$\label{eq:Arestr}
\begin{array}{l}
HSH = -S,\qquad H\tilde{A}_kH = (-1)^{k}\tilde{A}_k\, ,\\[4pt]
Q_{\epsilon}S^{\dag}Q_{\epsilon} = S,\qquad Q_{\epsilon}\tilde{A}_{k}^{\dag}Q_{\epsilon} = \tilde{A}_{k}\, .
\end{array}$$
It can be checked that $S$ is diagonalizable, indeed, one has $$g^{-1}Sg=J_0
\label{orbit}$$ where $$\begin{aligned}
\label{eq:g-tr}
g=\frac{1}{\sqrt{2}}\left(\begin{array}{ccc} 1&0&-1\\ \epsilon u^*&\sqrt{2}v& \epsilon u^*\\v^*&-\sqrt{2}u&v^*\end{array}\right), \qquad J_0=\diag(1,0-1)\, .\end{aligned}$$ Following [@YanVal2017], we shall write $SU(\epsilon)$ referring to $SU(3)$ when $\epsilon = 1$ and $SU(2,1)$ when $\epsilon = -1$. Similar convention will apply to the corresponding Lie algebras.
Since $g(x)\in SU(\epsilon)$ (\[orbit\]) means that $S(x)$ belongs to the adjoint representation orbit $$\mathcal{O}_{J_0}(\SU(\epsilon)) := \{ \tilde{X}\in\ri\su(\epsilon) : \tilde{X} = gJ_0g^{-1},\quad g\in\SU(\epsilon)\}$$ of $\SU(\epsilon)$ passing through $J_0$.
Our approach to the $\rm GMV_{\pm}$ system will be based on its gauge equivalence to a generalized Zakharov-Shabat auxiliary system (GZS system) on the algebra $\asl(3,\mathbb{C})$. The auxiliary system $$\label{eq:CBC}
L\psi=\left( \ri\partial_x + q(x) -\lambda J \right) \psi = 0, \qquad \lambda\in\mathbb{C}$$ where $q(x)$ and $J$ belong to some irreducible representation of a simple Lie algebra $\g$ is called generalized Zakharov-Shabat system (for that representation of $\g$) in canonical gauge. The element $J$ must be such that the kernel of $\ad_J(.):= [J, .]$ is a Cartan subalgebra $\mathfrak{h}_J\subset\g$ while $q(x)$ belongs to the orthogonal complement $\mathfrak{h}_J^{\perp}$ of $\mathfrak{h}_J$ with respect to the Killing form: $$\label{eq:Killing}
\langle X, Y \rangle = \tr(\ad_X \ad_Y),\qquad X, Y \in \g\, .$$ It is also assumed that the smooth function $q(x)$ vanishes sufficiently fast as $x\to\pm\infty$. System is gauge equivalent to the system $$\label{eq:BCGgauge}
\tilde{L}\tilde{\Psi} = \left(\ri\partial_x - \lambda
S(x) \right) \tilde{\Psi} = 0, \qquad S(x) \in {\cal O}_J(G)$$ where $G$ is the Lie group corresponding to $\g$. Usually it is also required that $$\lim_{x\to \pm\infty} S(x) = J$$ where the convergence is sufficiently fast but as we shall see in our case it will be different. The concept of gauge transformation and gauge equivalent auxiliary problems was applied for the first time in the case of the Heisenberg ferromagnet equation [@ZaTakh79] and its gauge equivalent — the nonlinear Schrödinger equation. Later, the integrable hierarchies, the conservation laws and Hamiltonian structures associated with and have extensively been studied by using the so-called gauge-covariant theory of the recursion operators related to the GZS systems in canonical and pole gauge [@GerYan85; @G86; @Yan87]. That approach provides a generalization of classical AKNS approach [@AKNS]. We recommend the monograph book [@GerViYa2008] for further reading.
So for GZS system in pole gauge most of the essential issues could be reformulated from the canonical gauge. The main difficulty is technical — to explicitly express all the quantities depending on $q$ and its derivatives through $S$ and its derivatives. A clear procedure of how to do that is described in [@Yan87]. The reader who is interested in that subject can find more details in [@Yan93] regarding GZS related to $\asl(3,\mathbb{C})$ with no reductions imposed and in [@YanVi2012SIGMA] regarding the geometry of the recursion operators for $\asl(3,\mathbb{C})$ in general position. We also refer to [@YanVi2012JNMP] for the case of GMV system.
In the present report we intend to construct expansions over the eigenfunctions of the recursion operators related to with arbitrary constant asymptotic values of the potential functions $(u,v)$. In doing this, we are taking into account the whole spectrum of the scattering operator $\tilde{L}^0$, thus extending in a natural way some of the results published in [@YanVal2017]. Next, we are showing how these expansions modify depending on the symmetries of the functions we expand. We would like to stress on the following:
- We shall be dealing with both $\rm GMV_{\pm}$ systems simultaneously;
- Our approach will be based on the gauge equivalence we mentioned in the above. Consequently, we shall be able to consider general asymptotic conditions — constant limits $\lim_{x\to \pm\infty}u$ and $\lim_{x\to \pm\infty}v$ ;
- Our point of view on the recursion operators when reductions are present is somewhat different from that adopted in [@GMVSIGMA];
- We show some new algebraic features in the spectral theory.
Gauge-Equivalent Systems {#sec:2}
========================
As mentioned, our approach to the $\rm GMV_{\pm}$ system will be based on the fact that it is gauge equivalent to a GZS system on $\asl(3,\mathbb{C}$) and will follow some of the ideas of [@GYaSAM2015; @Yan2015SPT]. We presented our results for the case of the continuous spectrum in [@YanVal2017]. Here, we shall include also the discrete spectrum. In fact, we have the following basic result:
\[th:GEq\] The $\rm GMV_{\pm}$ system is gauge equivalent to a canonical [GZS]{} linear problem on $\asl(3,\mathbb{C})$ $$\label{eq:GZSours}
L^0\psi=(\ri\partial_x +q-\lambda J_0)\psi = 0$$ subject to a Mikhailov reduction group generated by the two elements $h_1$ and $h_2$. For a fundamental solution $\psi$ of system we have: $$\nonumber\label{eq:MGnew}
\begin{array}{l}
{h_1(\psi)(x,\lambda)=\left[Q_{\epsilon}\psi(x,\lambda^{*})^{\dag}Q_{\epsilon}\right]^{-1},
\qquad Q_{\epsilon}=\diag(1,\epsilon,1),\quad \epsilon=\pm 1},\\[4pt]
{h_2(\psi)(x,\lambda)=K\psi(x,-\lambda)K}\, .
\end{array}$$ Since $h_1^2=h_2^2=\id$ and $h_1h_2=h_2h_1$ we have again a $\mathbb{Z}_2\times \mathbb{Z}_2$ reduction. In the above $$\begin{aligned}
\nonumber\label{eq:K}
K=\left(\begin{array}{ccc} 0&0&1\\ 0&1& 0\\1&0&0\end{array}\right)\, .\end{aligned}$$
[**Proof:**]{} Indeed, it is enough to put $q=\ri\psi_0^{-1}(\psi_0)_x$ where $$\nonumber\label{eq:psi0}
\psi_0 = \exp\left[- \ri J'\int_{-\infty}^xb(y)\rd y\right] g^{-1}\, .$$ In the above expression $J'=\diag (1,-2,1)$, $b(x)= \ri(\epsilon uu_x^*+vv_x^*)/2$ (note that this expression is real) and $g$, $J_0$ are the same as in . Then $\psi_0$ is a solution to for $\lambda=0$ and $\rm GMV_{\epsilon}$ is gauge-equivalent to GZS.$\quad\blacksquare$
Also, one gets the following important formulas: $$\label{eq:sympsio}
K\psi_0=\psi_0H, \qquad \psi_0^{-1}K=H\psi_0^{-1}\, .$$ In order to continue we shall need some simple algebraic facts.
Algebraic Preliminaries {#sec:3}
=======================
The reductions we introduced have clear algebraic meaning: $h(X)=HXH$, $k(X)=KXK$ are obviously involutive automorphisms of the algebra $\asl(3,\mathbb{C})$ and $\sigma_{\epsilon}X=-Q_{\epsilon}X^{\dag}Q_{\epsilon}$ defines a complex conjugation of the same algebra. As it is known $\asl(3,\mathbb{C})$ is a simple Lie algebra, the canonical choice for its Cartan subalgebra $\mathfrak{h}$ is the subalgebra of the diagonal matrices. It is also equal to $\mathfrak{h}_J=\ker\ad_J$ where $J$ is any diagonal matrix $\diag(\lambda_1, \lambda_2, \lambda_3)$ with distinct $\lambda_i$’s. In that case we shall call $\mathfrak{h}_J$ *the Cartan subalgebra* and denote it simply by $\mathfrak{h}$ (in particular, we have $\mathfrak{h}_{J_0}=\mathfrak{h}$). More generally, if $S$ is diagonalizable with distinct eigenvalues then $\mathfrak{h}_S=\ker\ad_S$ is also a Cartan subalgebra. We shall denote the projection onto the orthogonal complement $\mathfrak{h}^{\perp}=\mathfrak{h}_J^{\perp}$ of $\mathfrak{h}=\mathfrak{h}_J$ (with respect to the Killing form) by $\pi_0=\pi_J$ when $J$ is diagonal and the projection onto the orthogonal complement $\mathfrak{h}_S^{\perp}$ of $\mathfrak{h}_S$ by $\pi_S$ when $S$ is diagonalizable. One can introduce the system of roots $\Delta$, the systems of positive and negative roots $\Delta_{\pm}$ in a canonical way. The set $\Delta_{+}$ contains $\alpha_1,\alpha_2$ and $\alpha_3=\alpha_1+\alpha_2$, the Cartan-Weil basis shall be denoted by $E_{\pm \alpha_i}, H_1,H_2$ etc. We use the notation and normalizations used in the well known monograph on semisimple Lie algebras [@GoGr]. The matrices $H_{1}, H_{2}$ span $\mathfrak{h}$ and the matrices $E_{\alpha}$, $\alpha\in\Delta$ span $\mathfrak{h}^{\perp}$. The complex conjugation $\sigma_{\epsilon}$ defines the real form $\su(3)$ ($\epsilon=+1$) or the real form $\su(2,1)$ ($\epsilon=-1$) of $\asl(3,\mathbb{C})$. If we introduce the spaces $$\label{eq:INsp}
{\tilde{\g}^{[j]}=\{X: h(X)=(-1)^{j}X\}, \qquad j=0;1~\mod(2)}$$ then we shall have the orthogonal splittings $$\label{eq:spltH}
\begin{array}{l}
{\asl(3,\mathbb{C}) = \tilde{\g}^{[0]}\oplus \tilde{\g}^{[1]}}\, ,\\[4pt]
{\su(\epsilon) = (\tilde{\g}^{[0]}\cap\su(\epsilon))\oplus ( \tilde{\g}^{[1]}\cap\su(\epsilon))} \; .
\end{array}$$ In order to explain our results we shall also need the action ${\cal K}$ of $k:k(X)=KXK$ on the roots: $$\begin{aligned}
\nonumber
&&{\cal K}(\pm \alpha_{1})={\mp\alpha_2} \; , \qquad {\cal K}(\pm \alpha_{3})=\mp\alpha_3 \; , \qquad {\cal K}(\pm \alpha_{2})=\mp \alpha_1\end{aligned}$$ so we have [$k(E_{\alpha})=E_{{\cal K}\alpha}$]{}. We also note that we have the following relations which are used in all calculations: $$\label{eq:f-spaces}
h\circ\ad_S=-\ad_S\circ h \; ,\qquad \sigma_{\epsilon}\circ\ad_S=-\ad_S\circ \sigma_{\epsilon}\, .$$ Another issue we must discuss is the relation between $h$ from one side and $\ad_S^{-1}$ and $\pi_S$ from the other. Here $\ad_S^{-1}$ is defined only on the space $\mathfrak{h}_S^{\perp}$ but one could extend it as zero on $\mathfrak{h}_S$ which we shall always assume. One obtains that $$\begin{aligned}
\label{eq:symSh-s}
&&\ad_S^{-1}\circ h=-h\circ \ad_S^{-1},\qquad \pi_S\circ h=h\circ \pi_S\, ,\\
&&\ad_S^{-1}\circ \sigma_{\epsilon} =- \sigma_{\epsilon}\circ \ad_S^{-1},\qquad \pi_S\circ \sigma_{\epsilon}=\sigma_{\epsilon}\circ \pi_S\, .\end{aligned}$$
Recursion Operators of $\rm GMV_{\epsilon}$ {#sec:4}
===========================================
Recursion operators (also called generating operators or $\Lambda$-operators) are theoretical tools that permit:
- To describe the hierarchies of the nonlinear evolution equations (NLEEs) related to the auxiliary linear problems of GZS type (the AKNS approach [@AKNS]);
- To describe the hierarchies of conservation laws for these NLEEs;
- To describe the hierarchies of compatible Hamiltonian structures of these NLEEs;
- The expansions over their eigenfunctions permit to interpret the inverse scattering problems for GZS systems as generalized Fourier transforms, [**see**]{} [@G86; @GYa94; @IKhKi94];
- Recursion operators have important geometric interpretation — the NLEEs could be viewed as fundamental fields of a Poisson-Nijenhuis structure on the infinite dimensional manifold of “potentials”, a concept introduced in [@Mag78].
For all these aspects of recursion operators see also the monograph book [@GerViYa2008] which contains an extensive bibliography for publications prior to 2008.
The recursion operators $\tilde{\Lambda}_{\pm}$ arise naturally when one tries to find the hierarchy of Lax pairs related to a particular auxiliary GZS linear problem. Assume this problem has the form $\tilde{L}=\ri\partial_x-\lambda S$ where $S$ is in the orbit of the element $J_0$ with no additional assumptions on $S$ and we have that $\tilde L=\psi_0^{-1}L^0\psi_0$ where $L^0$ is a GZS system with $J=J_0$ and $\psi_0$ is a solution to for $\lambda=0$. We have $$\nonumber
{\tilde{\Lambda}_{\pm}=\Ad(\psi^{-1}_0)\circ \Lambda_{\pm}\circ \Ad(\psi_0)}$$ where $\Lambda_{\pm}$ are the recursion operators for $L^0$, see [@Yan93]. The explicit form of $\tilde{\Lambda}_\pm$ is $$\begin{aligned}
\nonumber\label{eq:ROGMV}
&&{\tilde{\Lambda}_\pm({Z})=} \ri {\ad}_S^{-1} {\pi}_S
\left\{ \partial_x{Z} + \frac{S_x}{12}\int\limits_{\pm\infty}^x \langle {Z}, S_y \rangle {\rd} y + \frac{S_{1x}}{4}\int\limits_{\pm\infty}^x
\langle{Z}, S_{1y} \rangle {\rd} y \right\}\end{aligned}$$ where $S_1=S^2- 2/3$ and $S_{1x}=(S_1)_x$. $\rm GMV_{\epsilon}$ is a particular case of a $\asl(3)$ problem so the operators $\tilde{\Lambda}_\pm$ are the recursion operators for $\rm GMV_{\pm}$ system and give the corresponding NLEEs. However, one must be a little more cautious here if one wants to obtain those NLEEs that are compatible with the reduction group. Indeed, the Lax pairs that obey the reductions give hierarchies of equations that have the form: $$\begin{aligned}
\label{eq:NLEEs1}
&&\ad_{S}^{-1}\partial_t S = \sum\limits_{k=0}^r a_{2k}(\tilde{\Lambda}_{\pm})^{2k} \ad_S^{-1}(S_x)+
\sum\limits_{k=1}^m a_{2k-1}(\tilde{\Lambda}_{\pm})^{2k-1} \ad_S^{-1}(S_{1x})\end{aligned}$$ where $a_i$ are some real constants. We shall not enter in more details here, see [@YanVal2017] for this, but one can see that when one considers the hierarchies of the NLEEs the next equation in the hierarchy is obtained not using $\tilde{\Lambda}_\pm$ but $\tilde{\Lambda}^2_\pm$. So one needs to understand what happens with the expansions that play role of generalized Fourier transform.
Spectral Theory of the Recursion Operators {#sec:5}
==========================================
The properties of the fundamental analytic solutions (FAS) of the GZS systems play a paramount role in the spectral theory of such systems. In fact, from the canonical FAS (denoted by $\chi^{\pm}$) in canonical gauge one immediately obtains FAS in the pole gauge $\tilde{\chi}^{\pm}$ (with the same analytic properties) [@GerViYa2008]. In our case we have $\tilde{\chi}^{\pm}(x,\lambda)=\psi_0^{-1}\chi^{\pm}(x,\lambda)$. The superscripts $\pm$ mean that the corresponding solution is analytic in $\mathbb{C}_{\pm}$ (upper and lower half-plane). For these solutions one has
\[fas\_sym\] The FAS $\tilde{\chi}^{\pm}(x,\lambda)$ corresponding to the $GMV_{\epsilon}$ system satisfy: $$Q_{\epsilon}(\tilde{\chi}^{\pm}(x,\lambda^*))^{\dag}Q_{\epsilon}=(\tilde{\chi}^{\mp}(x,\lambda))^{-1},\qquad H\tilde{\chi}^{\pm}(x,\lambda)H = \tilde{\chi}^{\mp}(x,-\lambda)KH\, .$$
Further, one builds the so-called adjoint solutions (or generalized exponents) for the GZS systems:
- GZS system in canonical gauge: ${\bf e}^{\pm}_{\alpha}=\pi_0\chi^{\pm}E_{\alpha}(\chi^{\pm})^{-1}$;
- GZS system in pole gauge: $\tilde{\bf e}^{\pm}_{\alpha}=\pi_S\tilde{\chi}^{\pm}E_{\alpha}(\tilde{\chi}^{\pm})^{-1}$ .
One sees that $\tilde{\bf e}^{\pm}_{\alpha}=\Ad(\psi^{-1}_0){\bf e}^{\pm}_{\alpha}$ and then the fact that they are eigenfunctions of $\tilde{\Lambda}_{\pm}$ and the completeness relations for them become immediate from the classical results for the recursion operators in canonical gauge. Indeed, first $$\nonumber\label{eq:eigentilde}
\begin{array}{c}
{\tilde{\Lambda}_{-}(\tilde{\bf e}^{+}_{\alpha}(x,\lambda))=\lambda \tilde{\bf e}^{+}_{\alpha}(x,\lambda)\, , \qquad \tilde{\Lambda}_{-}(\tilde{\bf e}^{-}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf e}^{-}_{-\alpha}(x,\lambda)}\, ,\\[4pt]\nonumber
{\tilde{\Lambda}_{+}(\tilde{\bf e}^{+}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf e}^{+}_{-\alpha}(x,\lambda)\, , \qquad \tilde{\Lambda}_{+}(\tilde{\bf e}^{-}_{\alpha}(x,\lambda))=\lambda \tilde{\bf e}^{-}_{\alpha}(x,\lambda)}
\end{array}$$ and the completeness relations could be written into the following useful form [@GYa94]: $$\begin{aligned}
\nonumber\label{eq:expanpoleds}
&&{\delta(x-y)\tilde{P}_0 ={\rm DSC_p}+}\\ \nonumber
&&{\frac{1}{2\pi} \int\limits_{-\infty}^\infty \left[
\sum_{\alpha
\in \Delta_{+}} \tilde{\bf e}^+_\alpha(x,\lambda)\otimes \tilde{\bf e}^+_{-\alpha}(y,\lambda) -
\tilde{\bf e}^{-}_{-\alpha}(x,\lambda)\otimes\tilde{\bf e}^{-}_{\alpha}(y,\lambda)\right] \rd\lambda}
\end{aligned}$$ where $\rm DSC_p$ is the discrete spectrum contribution. The second term is the continuous spectrum contribution which we denote by $\rm CSC_p$. Also, in the above $$\begin{aligned}
\nonumber
&&\tilde{P}_0=\sum_{\alpha\in \Delta}\frac{1}{ \alpha(J_0)}(\tilde{E}_{\alpha}\otimes \tilde{E}_{-\alpha})\, , \qquad \tilde{E}_{\alpha}=\Ad(\psi^{-1}_0)E_{\alpha} = \psi^{-1}_0 E_{\alpha}\psi_0\, ,\\ \nonumber
&& \tilde{\bf e}^{\pm}_{\alpha}=\Ad(\psi^{-1}_0){\bf e}^{\pm}_{\alpha}\,.\end{aligned}$$ For $\rm DSC_p$, assuming that one has $N^+$ poles $\lambda_i^+$, $1\leq i\leq N^+$ in the upper half-plane $\mathbb{C}_+$ and $N^-$ poles $\lambda_i^-$, $1\leq i\leq N^-$ in the lower half-plane $\mathbb{C}_-$, we get $$\begin{aligned}
\label{eq:DSC}
&&{\rm DSC}_p= \\ \nonumber
&& -\ri \sum\limits_{\alpha\in \Delta_+}\sum\limits_{k=1}^{N^+}\Res(\tilde{Q}^+_{\alpha}(x,y,\lambda) ;\lambda^+_{k})-\ri \sum\limits_{\alpha\in \Delta_+}\sum\limits_{k=1}^{N^-}\Res(\tilde{Q}^-_{-\alpha}(x,y,\lambda) ;\lambda^-_{k})\, ,\\
\nonumber
&&\tilde{Q}^{+}_{\alpha}(x,y,\lambda)=\tilde{\bf e}^{+}_{\alpha}(x,\lambda)\otimes \tilde{\bf e}^{+}_{-\alpha}(y,\lambda),\qquad {\mathrm{Im}}(\lambda)>0\, ,\\ \nonumber
&&\tilde{Q}^{-}_{-\alpha}(x,y,\lambda)=\tilde{\bf e}^{-}_{-\alpha}(x,\lambda)\otimes \tilde{\bf e}^{-}_{\alpha}(y,\lambda),\qquad {\mathrm{Im}}(\lambda)<0\, .\end{aligned}$$
$\Lambda$-Operators and Reductions {#sec:6}
==================================
Let us see now the implications of the reductions on the expansions over adjoint solutions. We start with the reduction defined by $h$. Since for the FAS we have the properties stated in Theorem \[fas\_sym\], for $\beta\in \Delta$ we obtain $$\nonumber\label{eq:symofe-p}
{h (\tilde{\bf e}_{\beta}^{\pm}(x,\lambda))=\tilde{\bf e}^{\mp}_{\mathcal{K}\beta}(x,-\lambda)} \; .$$ Changing the variable $\lambda$ to $-\lambda$, taking into account that $\mathcal K$ maps the positive roots into the negative ones and vice versa, we obtain after some algebraic transformations $$\begin{aligned}
\nonumber \label{eq:CSCredtenf1}
&&{{\rm CSC_p}=} {\displaystyle\frac{A_h}{2\pi}\int\limits_{-\infty}^\infty\left[
\sum\limits_{\alpha
\in \Delta_{+}} \tilde{\bf e}^+_\alpha(x,\lambda)\otimes\tilde{\bf e}^+_{-\alpha}(y,\lambda) -
\tilde{\bf e}^{-}_{-\alpha}(x,\lambda)\otimes\tilde{\bf e}^{-}_{\alpha}(y,\lambda)\right] \rd\lambda}
\end{aligned}$$ where $$\label{eq:ah}
{A_h=\frac{1}{2}(\id -h\otimes h)}\, .$$ Let us explain what the presence of the “multiplier” $A_h$ means. For simplicity let us first assume we have only continuous spectrum.
Assume $\tilde{Z}(x)$ is such that $h(\tilde{Z})=\tilde{Z}$ and let us make a contraction first to the right followed by integration over $y$ and to the the left followed by integration over $x$. Then taking into account that $h$ is automorphism and the Killing form is invariant under automorphisms we get $$\begin{aligned}
\nonumber\label{eq:newexpviaa}
&&{\tilde{Z}(x) = \displaystyle\frac{1}{2\pi} \displaystyle \int\limits_{-\infty}^\infty \left[
\sum\limits_{\alpha
\in \Delta_{+}}\tilde{\bf s}^{\eta}_\alpha(x,\lambda)\mu^{\eta}_{\alpha} -
\tilde{\bf s}^{-\eta}_{-\alpha}(x,\lambda))\mu^{-\eta}_{\alpha}\right] \rd\lambda}\end{aligned}$$ where $\eta=+$ ($\eta=-$) depending whether we contract to the left or to the right and $$\begin{aligned}
\nonumber
&&{\mu^{\eta}_{\alpha}=\langle\langle \tilde{\bf a}^\eta_{-\alpha},[S,\tilde{Z}]\rangle\rangle},\quad {\mu^{-\eta}_{\alpha}=\langle\langle \tilde{\bf a}^\eta_{\alpha},[S,\tilde{Z}]\rangle\rangle}\, ,\\ \nonumber \label{eq:sfun}
&&{\tilde{\bf s}^{\eta}_{\pm \alpha}(x,\lambda)=\displaystyle \frac{1}{2}\left(\tilde{\bf e}^{\eta}_{\pm\alpha}(x,\lambda)+h(\tilde{\bf e}^{\eta}_{\pm\alpha}(x,\lambda))\right)}\, ,\\ \nonumber\label{eq:afun}
&&{\tilde{\bf a}^{\eta}_{\pm \alpha}(x,\lambda)=\displaystyle \frac{1}{2}\left(\tilde{\bf e}^{\eta}_{\pm\alpha}(x,\lambda)-h(\tilde{\bf e}^{\eta}_{\pm\alpha}(x,\lambda))\right)}\, ,\end{aligned}$$ and for two functions $\tilde{Z}_1(x),\tilde{Z}_2(x)$ with values in $\asl(3)$ we used the notation $$\langle\langle\tilde{Z}_1,\tilde{Z}_2\rangle\rangle=\int\limits_{-\infty}^{+\infty}\langle\tilde{Z}_1(x),\tilde{Z}_2(x)\rangle{\rm d} x \; .$$ If instead of $h(\tilde{Z})=\tilde{Z}$ we assume that $h(\tilde{Z})=-\tilde{Z}$ then in the same manner we shall obtain expansions over the functions $\tilde{\bf a}^{\eta}_\alpha$ and the coefficients are calculated via the functions $\tilde{\bf s}^{\eta}_\alpha$. Since $(\id \pm h)/2$ are in fact projectors onto the $\pm 1$ eigenspaces of $h$ $$\nonumber
{h(\tilde{\bf s}^{\eta}_{\pm \alpha}(x,\lambda))=\tilde{\bf s}^{\eta}_{\pm \alpha}(x,\lambda)\; , \qquad h(\tilde{\bf a}^{\eta}_{\pm \alpha}(x,\lambda))=-\tilde{\bf a}^{\eta}_{\pm \alpha}(x,\lambda)}\; .$$ Thus, in case $h(\tilde{Z})=\tilde{Z}$ or $h(\tilde{Z})=-\tilde{Z}$ the expansions could be written in terms of new sets of adjoint solutions, $\tilde{\bf s}^{\eta}_{\pm \alpha}$ or $\tilde{\bf a}^{\eta}_{\pm \alpha}$ that reflect the symmetry of $\tilde{Z}$. For $\alpha\in \Delta_{+}$ one obtains that $$\nonumber\label{eq:asminus}
\begin{array}{c}
\tilde{\Lambda}_{-}(\tilde{\bf s}^{+}_{\alpha}(x,\lambda))=\lambda \tilde{\bf a}^{+}_{\alpha}(x,\lambda)\; , \qquad \tilde{\Lambda}_{-}(\tilde{\bf s}^{-}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf a}^{-}_{-\alpha}(x,\lambda)\; ,\\[4pt]
\tilde{\Lambda}_{-}(\tilde{\bf a}^{+}_{\alpha}(x,\lambda))=\lambda \tilde{\bf s}^{+}_{\alpha}(x,\lambda)\; , \qquad \tilde{\Lambda}_{-}(\tilde{\bf a}^{-}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf s}^{-}_{- \alpha}(x,\lambda)
\end{array}$$ and also $$\nonumber\label{eq:asplus}
\begin{array}{c}
\tilde{\Lambda}_{+}(\tilde{\bf s}^{+}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf a}^{+}_{-\alpha}(x,\lambda) \; , \qquad \tilde{\Lambda}_{+}(\tilde{\bf s}^{-}_{\alpha}(x,\lambda))=\lambda \tilde{\bf a}^{-}_{\alpha}(x,\lambda)\; , \\[4pt]
\tilde{\Lambda}_{+}(\tilde{\bf a}^{+}_{-\alpha}(x,\lambda))=\lambda \tilde{\bf s}^{+}_{-\alpha}(x,\lambda) \; , \qquad \tilde{\Lambda}_{+}(\tilde{\bf a}^{-}_{\alpha}(x,\lambda))=\lambda \tilde{\bf s}^{-}_{\alpha}(x,\lambda) \; .
\end{array}$$ One sees that the functions in the expansions when we have some symmetry with respect to $h$ are eigenfunctions for $\tilde{\Lambda}^2_{-}$ ($\tilde{\Lambda}^2_{+}$) with eigenvalue $\lambda^2$.This together with the fact that when recursively finding the coefficients for the Lax pairs one effectively uses $\tilde{\Lambda}^2_{+}$ leads to the interpretation that in case we have $\mathbb{Z}_2$ reduction defined by $h$ the role of the generating operator is played by $\tilde{\Lambda}_{\pm}^2$.
All this happens because of the new form of the expansions, involving the “multiplier” $A_h= (\id +h\otimes h)/2$. The point is that the “multiplier” $A_h$ has simple algebraic meaning:
The operator $A_h= (\id +h\otimes h)/2$ (acting on $\g\otimes \g$ where $\g=\asl(3,\mathbb{C})$) is a projector onto the space $$\nonumber
V=\left(\tilde{\g}^{[0]}\otimes \tilde{\g}^{[1]}\right)\oplus \left(\tilde{\g}^{[1]}\otimes \tilde{\g}^{[0]}\right)\, .$$
Consequently, when for $B\in V$ one makes a contraction (from the right or from the left) with $[S,X]$ where $X$ is in $\tilde{\g}^{[s]}$, then $[S,X]\in \tilde{\g}^{[s+1]}$ and $B\,[S,X]\in \tilde{\g}^{[s]}$.
Let us consider now the discrete spectrum term more closely. For a GZS system in pole gauge in general position one has $N^+$ poles $\lambda_i^+$, $1\leq i\leq N^+$ in the upper half-plane $\mathbb{C}_+$ and $N^-$ poles $\lambda_i^-$, $1\leq i\leq N^-$ in the lower half-plane $\mathbb{C}_-$. If we have reduction defined by $h$ we see that we must have $N^+=N^-$ since if $\tilde{\bf e}_{\alpha}^{+}(x,\lambda)$ has a pole of some order at $\lambda=\lambda_s^+$ in $\mathbb{C}_+$ then $\tilde{\bf e}^{-}_{\mathcal{K}\alpha}(x,\lambda) $ will have the same type of singularity at $-\lambda_s^+$ in $\mathbb{C}_-$. In order to simplify the notation we shall put $\lambda_s^+=\lambda_s$, $\lambda_s^-=-\lambda_s$ and $N^+=N^-=N$. Of course, in order to make concrete calculations one needs some assumption on the discrete spectrum. Assume that all the singularities are simple poles, let us consider the contribution from $\tilde{Q}_{\beta}(\lambda)$ for a fixed $\beta$ and two poles: one pole $\lambda=\lambda_0$ located in $\mathbb{C}_+$ and one pole $\lambda=-\lambda_0$ located in $\mathbb{C}_-$. Then for $\beta\in \Delta$ in some discs around $\lambda_0$ and $-\lambda_0$ we have the Laurent expansions that hold uniformly on $x$: $$\begin{aligned}
&&\tilde{\bf e}_{\beta}^{+}(x,\lambda)=\frac{\tilde{A}^{+}_{\beta}(x)}{\lambda-\lambda_0}+\tilde{B}^{+}_{\beta}(x)+\tilde{C}^{+}_{\beta}(x)(\lambda-\lambda_0)+\ldots\, ,\\
&&\tilde{\bf e}_{\mathcal{K}\beta}^{-}(x,\lambda)=
\frac{\tilde{A}^{-}_{\mathcal{K}\beta}(x)}{\lambda+\lambda_0}+\tilde{B}^{-}_{\mathcal{K}\beta}(x)+\tilde{C}^{-}_{\mathcal{K}\beta}(x)(\lambda+\lambda_0)+\ldots\end{aligned}$$ From the properties of the FAS we see that for $\beta\in \Delta$, $h (\tilde{\bf e}_{\beta}^{\pm}(x,\lambda))=\tilde{\bf e}^{\mp}_{\mathcal{K}\beta}(x,-\lambda)$ and consequently $$\begin{aligned}
\label{eq:disspr1-p}
&&h \tilde{A}^+_{\beta}(x)=-\tilde{A}^-_{\mathcal{K}\beta}(x),\quad h \tilde{B}^+_{\beta}(x)=~~\tilde{B}^-_{\mathcal{K}\beta}(x),\quad h \tilde{C}^+_{\beta}(x)=-\tilde{C}^-_{\mathcal{K}\beta}(x)\, .\end{aligned}$$ For $\alpha\in \Delta_+$ the calculation gives $$\Res (\tilde{Q}^+_{\beta}(x,y,\lambda);\,\lambda_0)=\tilde{A}^+_{\beta}(x)\otimes \tilde{B}^+_{-\beta}(y)+
\tilde{B}^+_{\beta}(x)\otimes \tilde{A}^+_{-\beta}(y)$$ but since the singularities occur in pairs we can combine the contributions from $\lambda_0$ and $-\lambda_0$. After performing it, we put the formula for $\rm DSC_p$ into a form in which the poles in the upper and lower half-plane play equal role introducing the notation $$\lambda_i=\lambda_{i}^+,\qquad \lambda_{i+N}=\lambda_{i}^-=-\lambda_{i}^+,\qquad 1\leq i\leq N\, .$$ Then one could write $$\label{eq:DSKsym-p}
{\rm DSC_p}=-A_h\sum\limits_{\alpha\in \Delta_+}\sum\limits_{s=1}^{2N}\ri \Res(\tilde{Q}_{\alpha}(x,y,\lambda)\, ;\lambda_{s})\, .$$ Now, making contractions to the right (left) by $[S,\tilde{Z}]$ and integrating one gets the discrete spectrum contribution to the expansion of a given function $\tilde{Z}$: $$\begin{aligned}
\label{eq:dsred}
&&{\rm DSC_p}([S,\tilde{Z}])=\\ \nonumber
&&-2\ri \sum\limits_{\alpha\in \Delta_{+}} \sum\limits_{k=1}^{N} \tilde{A}_{\epsilon\alpha,k}^{[+;s]}(x)\langle \langle \tilde{B}_{-\epsilon\alpha,k}^{[+;s+1]},[S,\tilde{Z}]\rangle\rangle+\tilde{B}_{\epsilon\alpha,k}^{[+;s]}(x)\langle \langle \tilde{A}_{-\epsilon\alpha,k}^{[+;s+1]},[S,\tilde{Z}]\rangle\rangle\end{aligned}$$ for $\epsilon=\pm 1$ (depending on what side we contracted). In the above $$\tilde{A}_{\beta}^{[+;s]}(y)= \frac{1}{2}(\id+ (-1)^s h) \tilde{A}^+_{\beta}(y)\; , \qquad \tilde{B}_{
\beta}^{[+;s]}(y)= \frac{1}{2}(\id+ (-1)^s h)\tilde{B}^+_{\beta}(y)$$ and $s$ is understood modulo $2$. The action of the recursion operators on the discrete spectrum is not hard to find. For for $\beta\in \Delta$ and the coefficients of the expansion about $\lambda=\lambda_0$ of $\tilde{\bf e}_{\beta}^{+}(x,\lambda)$ we get: $$\begin{aligned}
&& \tilde{\Lambda}_{\pm}\tilde{A}^+_{\beta}(x)=\lambda_0\tilde{A}^+_{\beta}(x) \; , \qquad \tilde{\Lambda}_{\pm}\tilde{B}^+_{\beta}(x)=\lambda_0\tilde{B}^+_{\beta}(x) +\tilde{A}^+_{\beta}(x)\; .\end{aligned}$$ Consider now the space $\tilde{V}^+_{\beta}$ spanned by the vectors $\tilde{A}^+_{\beta}, \tilde{B}^+_{\beta}$. Of course, we must have $\tilde{A}^+_{\beta}\neq 0$, otherwise there is no singularity. One sees that also $\tilde{B}^+_{\beta}\neq 0$. Next, one checks immediately that the above relations could be true only if $\tilde{A}^+_{\beta}$, $\tilde{B}^+_{\beta}$ are linearly independent, so $\tilde{V}^+_{\beta}$ has dimension $2$ and the matrix of $\tilde{\Lambda}_{\pm}$ in the basis $\tilde{A}^+_{\beta}$, $\tilde{B}^+_{\beta}$ consists of $2\times 2$ Jordan block having $\lambda_0$ on the diagonal.
The situation with the spaces $V^{[+;s]}_{\nu, \beta}$ spanned by the vectors $A^{[+;s]}_{\beta}, B^{[+;s]}_{\beta}\neq 0$ is very similar but slightly more complicated. Indeed, since $\tilde{\Lambda}_{\pm}\circ h=-h\circ \tilde{\Lambda}_{\pm}$ we obtain that $$\begin{aligned}
&& \tilde{\Lambda}_{\pm}\tilde{A}^{[+;s]}_{\beta}(x)=\lambda_0\tilde{A}_{\beta}^{[+;s+1]}(x)\, ,\\
&& \tilde{\Lambda}_{\pm}\tilde{B}_{\beta}^{[+;s]}(x)=\lambda_0\tilde{B}_{\beta}^{[+;s+1]}(x) +\tilde{A}_{\beta}^{[+;s+1]}(x)\end{aligned}$$ and therefore $$\begin{aligned}
&& \tilde{\Lambda}_{\pm}^2\tilde{A}^{[+;s]}_{\beta}(x)=\lambda^2_0\tilde{A}_{\beta}^{[+;s]}(x)\, ,\\
&& \tilde{\Lambda}_{\pm}^2\tilde{B}_{\beta}^{[+;s]}(x)=\lambda^2_0\tilde{B}_{\beta}^{[+;s]}(x) +2\lambda_0\tilde{A}_{\beta}^{[+;s]}(x)\, .\end{aligned}$$ We have the following options:
- $\tilde{A}^{[+;s]}_{\beta}\neq 0$. Then one sees that $\tilde{B}^{[+;s]}_{\beta}\neq 0$ and $\tilde{A}^{[+;s]}_{\beta}$, $\tilde{B}^{[+;s]}_{\beta}$ must be linearly independent, $\tilde{V}^{[+;s]}_{\beta}$ has dimension $2$ and the matrix of $\tilde{\Lambda}_{\pm}^2$ in the basis $2\lambda_0\tilde{A}^{[+;s]}_{\beta}$, $\tilde{B}^{[+;s]}_{\beta}$ consists of $2\times 2$ Jordan block having $\lambda_0^2$ on the diagonal;
- $\tilde{A}^{[+;s]}_{\beta}=0$. Then if $\tilde{B}^{[+;s]}_{\beta}\neq 0$ the space $\tilde{V}^{[+;s]}_{\beta}$ is one dimensional and it is an eigenspace with eigenvalue $\lambda_0^2$$\,$ ;
- If $\tilde{A}^{[+;s]}_{\beta}=\tilde{B}^{[+;s]}_{\beta}=0$ then $\tilde{V}^{[+;s]}_{\beta}=0$.
In all the cases we see that for the reduction defined by $h$ the spaces $\tilde{V}^{[+;s]}_{\beta}$ are not invariant under the action of $\tilde{\Lambda}_{\pm}$ but are invariant under the action of $\tilde{\Lambda}^2_{\pm}$. This will happen, of course, when we consider the contribution from all the poles given by the expression .
We have the same effect from the reduction defined by the complex conjugation $\sigma_{\epsilon}$. Both for the continuous and for the discrete spectrum we obtain $$\begin{aligned}
\nonumber
&&{{\rm CSC_p}=A_{\sigma_{\epsilon}}{\rm CSC_p}} \; ,\qquad {{\rm DSC_p}=A_{\sigma_{\epsilon}}{\rm DSC_p}} \end{aligned}$$ where $$\label{eq:asigma}
A_{\sigma_{\epsilon}}=\frac{1}{2}(\id-\sigma_{\epsilon}\otimes \sigma_{\epsilon})\, .$$ Finally, if the reductions defined by $h$ and $\sigma_{\epsilon}$ act simultaneously then $$\begin{aligned}
&&{{\rm CSC_p}=A_h A_{\sigma_{\epsilon}}{\rm CSC_p}} \; , \qquad {{\rm DSC_p}=A_h A_{\sigma_{\epsilon}}{\rm DSC_p}}\end{aligned}$$ where $A_h$ and $A_{\sigma_{\epsilon}}$ are as in and . Note that $A_h$ and $A_{\sigma_{\epsilon}}$ commute. Of course, in the case of the complex conjugation $\sigma_{\epsilon}$ the role of “symmetric” with respect to the action of $h$ is taken by “real” functions with respect to $\sigma_{\epsilon}$ and the role of “anti-symmetric” with respect to the action of $h$ is taken by “imaginary” functions with respect to $\sigma_{\epsilon}$ and so on. However, both in the case of one $\mathbb{Z}_2$ reduction ($h$), and in the case of $\mathbb{Z}_2\times \mathbb{Z}_2$ reduction ($h$ and $\sigma_{\epsilon}$) the role played previously by the operators $\tilde{\Lambda}_{\pm}$ is played by $\tilde{\Lambda}^2_{\pm}$ – the square of these operators.
Conclusion {#sec:7}
==========
We have already discussed that effectively the operators “shifting” the equations along the hierarchies of NLEEs are $\tilde{\Lambda}_{\pm}^{2}$. We have showed that when one uses expansions over adjoint solutions to investigate these evolution equations, then according to the symmetry of the right hand side with respect to $h$ and $\sigma_{\epsilon}$ the expansions modify depending on the symmetries of the functions we expand. So these expansions are over the eigenfunctions of $\tilde{\Lambda}_{\pm}^{2}$ and in the generalized Fourier expansions the role previously played by $\tilde{\Lambda}_{\pm}$ is played now by their squares $\tilde{\Lambda}_{\pm}^{2}$.
Acknowledgement {#acknowledgement .unnumbered}
===============
The work has been supported by the NRF incentive grant of South Africa and grant DN 02–5 of Bulgarian Fund “Scientific Research”.
[99]{} Ablowitz M.J., Kaup D.J., Newell A.C., Segur H., (1974) The Inverse Scattering Problem - Fourier Analysis for Nonlinear Problems, [*Stud. Appl. Math.*]{} **53**, 249–315.
Gerdjikov V.S., (1986) Generalized Fourier Transforms for the Soliton Equations. Gauge-covariant Formulation, [*Inv. Problems*]{} [**2**]{}, 51–74.
Gerdjikov V.S., Grahovski G.G., Mikhailov A.V., Valchev T.I., (2011) Polynomial Bundles and Generalized Fourier Transforms for Integrable Equations on A.III-type Symmetric Spaces, [*SIGMA*]{} [**7**]{}, 096.
Gerdjikov V.S., Mikhailov A.V., Valchev T.I., (2010) Reductions of Integrable Equations on A.III-Symmetric Spaces, [*Jour. Phys. A: Math. Theor.*]{} [**43**]{}, 434015.
Gerdjikov V.S., Vilasi G., Yanovski A.B., [*Integrable Hamiltonian Hierarchies — Spectral and Geometric Methods*]{}, Springer, Heidelberg, 2008.
Gerdjikov V.S., Yanovski A.B., (1986) Gauge-covariant Theory of the Generating Operator. I., [*Commun. Math. Phys.*]{} [**103**]{}, 549–68.
Gerdjikov V.S., Yanovski A.B., (1994) Completeness of the Eigenfunctions for the Caudrey-Beals-Coifman System, [*Jour. Math. Phys.*]{} [**35**]{}, 3687–721.
Gerdjikov V.S., Yanovski A.B., (2014) CBC Systems with Mikhailov Reductions by Coxeter Automorphism: I. Spectral Theory of the Recursion Operators, [*Stud. Appl. Maths.*]{} [**134**]{} Issue 2, 145–180.
Golubchik I.Z., Sokolov V.V., (2000) Multicomponent Generalization of the Hierarchy of the Landau-Lifshitz Equation, [*Theor. Math. Phys.*]{} [**124**]{} n. 1, 909–917.
Goto M., Grosshans F., [*Semisimple Lie Algebras*]{}, Lecture Notes in Pure and Applied Mathematics [**38**]{}, M. Dekker Inc., New-York & Basel, 1978.
Iliev I.D., Khristov E.Kh., Kirchev K.P., [*Spectral Methods in Soliton Equations*]{}, Pitman Monographs and Surveys in Pure and Applied Mathematics **73**, John Wiley & Sons Inc., New-York, 1994.
Magri F., (1978) A Simple Model of the Integrable Hamiltonian Equations, [*Jour. Math. Phys.*]{} [**19**]{}, 1156–1162.
Mikhailov A.V., (1979) Reduction in the Integrable Systems. Reduction Groups, [*Lett. JETF (Letts. Sov. J. Exper. Theor. Phys.)*]{} [**32**]{}, 187–92.
Mikhailov A.V., (1981) The Reduction Problem and Inverse Scattering Method, [*Physica D*]{} [**3**]{}, 73–117.
Yanovski A.B., [*Gauge-covariant Approach to the Theory of the Generating Operators for Soliton Equations*]{}, PhD thesis, Joint Institute for Nuclear Research (JINR) 5–87–222, 1987.
Yanovski A.B., Generating Operators for the Generalized Zakharov-Shabat System and its Gauge Equivalent System in $\asl(3,\mathbb{C})$ Case, Preprint: Universität Leipzig, Naturwissenchaftlich Theoretisches Zentrum Report N20, 1993, http://cdsweb.cern.ch/record/256804/files/P00019754.pdf .
Yanovski A.B., Gauge-covariant Theory of the Generating Operators Associated with Linear Problems of Caudrey-Beals-Coifman Type in Canonical and in Pole Gauge with and without Reductions, In: Slavova, A. (Ed.), Proc. BGSIAM’14, pp. 2–43, Sofia, 2015.
Yanovski A.B., Valchev T.I., (2018) Pseudo-Hermitian Reduction of a Generalized Heisenberg Ferromagnet Equation. I. Auxiliary System and Fundamental Properties, [*Jour. Nonl. Math. Phys.*]{} [**25**]{} (02), 324–350, arXiv:1709.09266v1\[nlin.SI\].
Yanovski A.B., Vilasi G., (2012) Geometry of the Recursion Operators for the GMV System, [*Jour. Nonl. Math. Phys.*]{} [**19**]{}, 1250023-1/18.
Yanovski A.B., Vilasi G., (2012) Geometric Theory of the Recursion Operators for the Generalized Zakharov-Shabat System in Pole Gauge on the Algebra $\asl(n; \mathbb{C})$ with and without Reductions, [*SIGMA*]{}, 087.
Zakharov V.E., Takhtadjan L.A., (1979) Equivalence between Nonlinear Schrödinger Equation and Heisenberg Ferromagnet Equation, [*Theor. Math. Phys.*]{} [**38**]{}, 26–35.
[^1]: A more general system was derived independently by Golubchik and Sokolov [@golsok].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Mitigating the substantial undesirable impact of transportation systems on the environment is paramount. Thus, predicting Greenhouse Gas (GHG) emissions is one of the profound topics, especially with the emergence of intelligent transportation systems (ITS). We develop a deep learning framework to predict link-level GHG emission rate (ER) (in $\text{CO}_{2\text{eq}}$ gram/second) based on the most representative predictors, such as speed, density, and the GHG ER of previous time steps. In particular, various specifications of the long-short term memory (LSTM) networks with exogenous variables are examined and compared with clustering and the autoregressive integrated moving average (ARIMA) model with exogenous variables. The downtown Toronto road network is used as the case study and highly detailed data are synthesized using a calibrated traffic microsimulation and MOVES. It is found that LSTM specification with speed, density, GHG ER, and in-links speed from three previous minutes performs the best while adopting 2 hidden layers and when the hyper-parameters are systematically tuned. Adopting a 30 second updating interval improves slightly the correlation between true and predicted GHG ERs, but contributes negatively to the prediction accuracy as reflected on the increased root mean square error (RMSE) value. Efficiently predicting GHG emissions at a higher frequency with lower data requirements will pave the way to non-myopic eco-routing on large-scale road networks [to alleviate the adverse impact on the global warming]{}.'
author:
- |
Lama Alfaseeh\
Laboratory of Innovations in Transportation (LiTrans)\
Ryerson University\
Toronto, Canada\
`[email protected]` Ran Tu\
University of Toronto\
Toronto, Canada\
`[email protected]` Bilal Farooq\
Laboratory of Innovations in Transportation (LiTrans)\
Ryerson University\
Toronto, Canada\
`[email protected]` Marianne Hatzopoulou\
University of Toronto\
Toronto, Canada\
`[email protected]`\
\
bibliography:
- 'references.bib'
title: Greenhouse Gas Emission Prediction on Road Network using Deep Sequence Learning
---
Introduction {#Introduction}
============
Transportation systems have been consistently ranked as the first largest source of GHG emissions [(mainly carbon dioxide)]{} in the U.S. producing 29% of the the total emissions [@epa2017sources]. [GHG emission is among the main contributors to the global warming and climate change [@liu2020predictions].]{} Hence, topics related to GHG reduction have grabbed the attention of transportation researchers in the past few decades. On the solution side, the employment of the intelligent transportation systems (ITS) has been considered as a most favourable approach [@zegeye2009model] to alleviate the undesirable impact of transportation systems on the environment. The ability of the ITS to capture high resolution information about traffic conditions contributes to a more efficient traffic management process. Furthermore, the availability of real-time data allows the shift to non-myopic routing for [more sustainable transportation systems]{}.
Although prediction has been considered as a crucial topic, most of the studies tackled the matter from an aggregated level. In other words, GHG emissions are modelled or predicted at the national level. Examples include [@pao2011modeling], [@lin2011grey], [@antanasijevic2014forecasting], and [@radojevic2013forecasting]. The main justification is the scarcity of microscopic data and complexity associated with the models. With regards to the models used, Grey Models (GMs) [@dengiz2018grey], the Auto Regressive Integrated Moving Average (ARIMA) [@rahman2017modeling], and the Artificial Neural Networks (ANNs) [@abdullah2015methods] were widely used to forecast GHG emissions. GMs were commonly used due to the fact that they require comparatively small number of data points and can manage the case of limited or missing data [@dengiz2018grey], which can affect adversely the accuracy. ARIMA models are based on historical values of the predicted variable, such as CO$_2$ [@rahman2017modeling]. It is worth mentioning that ARIMA models are associated with a major limitation related to their assumption of linear relationship between variables [@zhang2003time], which restricts ARIMA application in the case of complex non-linear time series. Thus, ANNs have been introduced to incorporate the non-linear relationship between predictors and responses. It has been shown that NNs outperform other models when they are adopted to predict the environmental pollutants [@singh2012linear]. Recurrent neural networks (RNNs), which are a type of the deep neural networks (DNNs), were introduced and were associated with multiple layers between the input and output layers, unlike the artificial neural networks (ANNs). The RNNs define the correct mathematical manipulation to give output from an input, whether it is a linear relationship or a non-linear relationship [@aggarwal2018neural]. The Long-Short Term Memory (LSTM) network is a category of the RNNs, that overcomes the drawback of regular RNNs, the vanishing gradient problem [@amarpuri2019prediction]. The LSTM network has been considered as one of the most powerful RNN architectures where sequential data is involved [@lipton2015critical]. The LSTM deep network consists of three main components: 1) input gate, which controls the information fed to the network, 2) forget gate, which controls whether to keep or forget the information of the previous time step, and 3) output gate, which controls what information to give out of the network [@amarpuri2019prediction]. With reference to the LSTM application, authors mainly utilized it to predict at an aggregated level, spatially and temporally, as in [@ameyaw2019investigating], [@ameyaw2018analyzing], and [@huang2019grey].
While the existing literature forecast GHG emissions at an aggregated level, utilizing generally fuel and economical factors, this study aims to predict GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) at a link level using a deep learning framework, based on LSTM with exogenous variables, while using microscopic data points. An agent-based traffic simulation is utilized to obtain traffic and environmental high resolution information. Furthermore, a comparison is conducted with two other models: ARIMA with exogenous variables, and clustering. The impact of utilizing a deeper LSTM model while systematically tuned is examined. To evaluate the performance, we use the correlation coefficient between observed and predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec), the fit to the ideal straight curve reflecting on the precision, R$^2$ statistics, and the root mean square error (RMSE) reflecting on the accuracy. Finally, a crucial aspect is investigated related to defining the impact of the updating interval of the predictors when LSTM is adopted. Predictors at every 30 seconds and 1 minute are used for the comparison and trade offs are illustrated.
The main contributions of this work are as follows:
1. Development of a deep learning framework based on LSTM to predict GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) at link level in a highly congested urban network, while utilizing microscopic data.
2. Systematic analysis of the importance of various predictors contributing to the models.
3. Comparison of the results with the two commonly used emission prediction models, ARIMA and clustering, and demonstration of the strengths and limitations associated with each.
4. Illustration of the impact of utilizing different lengths of time interval for predictors and demonstrating the effect of the systematic tuning while applying LSTM.
This work is organized as follows: Section \[Literature Review\] presents a brief literature review of studies that predicted GHG. The methodology related to the predictive models utilized is in Section \[Methodology\]. The description of the case study is illustrated in section \[Case study\]. The traffic simulation and the emission model for collecting data are presented in section \[Data collection\] including the details of the scenarios simulated. Discussion and results are in Section \[Discussion and results\]. Finally, concluding remarks and future outlook are in section \[Conclusion\].
Literature Review {#Literature Review}
=================
A large number of studies predicted GHG or CO$_2$ at national level using yearly data points of predictors while adopting different models. [@pao2011modeling] predicted CO$_2$ emissions based on income and energy consumption by using grey prediction model (GM) in Brazil. The authors compared between GM and ARIMA model and found that results are comparable in terms of the forecasting performance. Another study by [@lin2011grey] used GM to predict CO$_2$ emissions in Taiwan. ARIMA model was used by [@rahman2017modeling] and [@tudor2016predicting] to predict CO$_2$ emissions in Bangladesh and Bahrain, respectively. [@antanasijevic2014forecasting] predicted GHG emissions at national level, of European countries, employing a new approach based on ANNs. The predictors considered were agriculture, transportation, energy supply and use, and waste. Another study by [@radojevic2013forecasting] utilized ANN to predict CO$_2$ emissions in Serbia based on the share of renewable sources of energy, the gross domestic product, the gross energy consumption, and energy intensity were selected as the input parameters [@radojevic2013forecasting]. [@sun2017factor] predicted CO$_2$ emissions by utilizing extreme learning machine (ELM), which is a type of ANNs, based on particle swarm (PSO) optimization. The authors used the variables impacting the produced CO$_2$ and classified them into one group. In addition, they found that their proposed approach outperformed the ELM and the back propagation neural network in terms of the RMSE and MAPE. [@grote2018practical] developed Practical Emissions Model for Local Authorities emission (PEMLA) to be used by local government authorities. PEMLA estimates CO$_2$ at a network level based on data collected from inductive loop detectors, which are installed as part of urban traffic control systems. Five traffic variables were taken into consideration: 1) traffic average speed (km/h); 2) traffic density (vehicles/km); 3) traffic average delay rate (seconds/vehicle·km); 4) access density (intersections/km); and 5) the square of traffic average speed (km/h)$^2$. The authors employed Multiple Linear Regression (MLR) and the ANN to define the relationship between variables and predict CO$_2$, respectively. The case study was Southampton, a city on the South coast of the UK with a population of approximately 255,000 and 5 min traffic variables were obtained from detectors to estimate network level CO$_2$ emissions [@grote2018practical].
The Long-Short Term Memory (LSTM) network has been considered as one of the best RNN architectures for the case of sequential data [@lipton2015critical]. With regards to the LSTM application, [@ameyaw2019investigating] utilized it to predict country level CO$_2$. The authors considered several predictors based on their correlation with CO$_2$. Gross fixed capital formation which was measured as a percentage of gross domestic product, total labor force, and trade were also adopted as predictors in addition to gross domestic product per capita [@ameyaw2019investigating]. Similarly, [@ameyaw2018analyzing] predicted CO$_2$ in five West African countries, but considering only one predictor which is the gross domestic product. [@huang2019grey] predicted CO$_2$ in China based on four predictors using LSTM. Clustering is another technique used for prediction in transportation systems, examples include [@poucin2018activity; @Ran2018TRB] and [@gmira2017travel]. In some cases, clustering is used in combination with other models, such as MapReduce Framework as in [@zhao2011mapreduce].
Unlike the aforementioned studies above, which can be classified as top-down approaches, [@dong2019carbon] developed a predictive model for diesel trucks and gasoline passenger cars based on volume to capacity (V/C) ratio as an explanatory variable. Due to time and cost restrictions, V/C range was taken only between 0.15 and 1.1. The authors applied the predictive model on an actual expressway in China [@dong2019carbon]. The authors defined a regression curve that illustrated the relationship between emissions and V/C [@dong2019carbon]. The main shortcoming of their approach is that the density and speed of congested traffic conditions are not captured efficiently when V/C is the only predictive variable and when values out of the defined range are not considered. Hence, their approach did not represent the real situation of traffic and considered a small case study. [@wang2015fine] developed a model to predict 5-min series of carbon monoxide (CO) and fine particulate matter (PM2.5) of an intersection in Shanghai. The proposed model is a hybrid model combining wavelet neural network and genetic algorithm (GA-WNN). In terms of the traffic predictors, delay and queue length of different directions were taken into consideration in addition to other predictors [@wang2015fine]. Another study by [@zhao2011mapreduce] suggested a MapReduce framework for on-road mobile fossil fuel combustion CO$_2$ emissions estimation. Data was obtained from ITS and processed to estimate GHG emissions. Their model had three layers; infrastructure, distributed computing platform, and application layer. In the application layer, data was clustered, outliers were removed, and GHG emissions were estimated based on ERs obtained from speed on links. It is worth mentioning that this approach estimated emission based on aggregated data using regression models of speed [@zhao2011mapreduce]. One more study by [@coelho2009numerical] developed a traffic and emission decision support (TEDS) tool to provide an overall pollution estimate for a traffic interruption. The authors defined the relationship between several pollutants, carbon monoxide (CO), carbon dioxide (CO2), nitric oxide (NO), and hydrocarbons (HC), and queue length and stops. It was assumed that average speed and flow were constant of zones studied, which is a limitation as realistic conditions were not captured. Their case study was two highways to the city of Lisbon (Portugal), namely Freeway A5 and Highway N6 [@coelho2009numerical].
To sum it up, in the existing literature, GHG emissions are mostly predicted at an aggregated level, spatially and temporally, and based mainly on yearly data points of fuel consumption, gross domestic product, or other economical factors. Although transportation systems contribute substantially to the GHG produced [@luo2016real], there is a lack of studies tackling the GHG prediction matter at a microscopic level that will be used in several controlling and management processes, such as routing for more sustainable transport systems. To the best of our knowledge, this study is the first of its kind that employs a deep learning approach i.e. LSTM, to predict GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) at link level for every minute or less, using the most representative predictors for a congested urban network.
Methodology {#Methodology}
===========
Before discussing the LSTM methodology in detail, we describe the two dominantly used models, ARIMA and clustering, for comparison purposes. ARIMA is a statistical model, while clustering is a machine learning algorithm.
[.45]{} ![Methodology for a) ARIMA with exogenous variables, b) clustering, and c) LSTM with exogenous variables[]{data-label="methods"}](ARIMAmethodology.pdf "fig:"){width="3in"} \[ARIMA\_methodology\]
[.45]{} ![Methodology for a) ARIMA with exogenous variables, b) clustering, and c) LSTM with exogenous variables[]{data-label="methods"}](Clusteringmethodology.pdf "fig:"){width="3in"} \[Clustering\_methodology\]
\
\[LSTM\_final\_2\]
ARIMA with exogenous variables {#ARIMAX_methodology_text}
------------------------------
ARIMA (p,d,q) has traditionally been used for time series data that can be made to be “stationary” by differencing of “d” order. In other words there should be no trends in the data in order to predict using the ARIMA model efficiently. Where: “p” and “q” are small integers reflecting on the “autoregressive” (AR) and “moving-average” (MA), respectively. To include $r$ exogenous variables while using ARIMA, Equation \[equation\_ARIMAX\] is applied. $$\hat{y_t}= \mu + \sum_{i=1}^{p}\phi_i y_{t-i} + \sum_{k=1}^{r}\beta_k x_{tk} + e_t+ \sum_{j=1}^{q}\theta_j e_{t-j}
\label{equation_ARIMAX}$$ Where $\hat{y_t}$ is the response at time $t$, $\mu$ is the constant, $\phi_i$ is the AR coefficient at lag $i$, $y_{t-i}$ is the value of the variable in concern for prediction at time $t-i$, $\beta_k$ is a coefficient of exogenous variable $k$, $x_{tk}$ is exogenous variable $k$ at time $t$, $\theta_j$ is the MA coefficient at lag $j$, and $e_{t-j}=y_{t-j}-\hat{y}_{t-j}$ is the forecast error that was made at period $t-j$ [@box2015time].
Figure \[methods\] illustrates the steps followed to develop the ARIMA model and predict the GHG ERs. A ratio of 70% to 30%, training to testing, is considered. To define the optimal parameters (p,d,q), an iterative process took place while considering the auto-correlation, partial auto-correlation plots of the differenced series, and root unit which is a measure that signalizes when the time series is under or overdifferenced. One of the major shortcomings of ARIMA models in our context is that a separate model needs to be estimated for each link, based on the link data, to assure that the time series is stationary. In other words, ARIMA lacks the spatial dimension.
Clustering {#Clustering_methdology_text}
----------
Clustering is an important tool in data mining applications. It mainly groups objects so that in one group the objects are more related to each other than those in other groups [@mann2013review]. To classify the traffic conditions at link level using the most important variables, K-mean clustering has been utilized. It aims to classify data points into K clusters. Every data point belongs to a cluster with the minimum distance to the centroids of that cluster. In K-mean clustering, the optimal cluster is defined when the total intra-cluster variance, or, the squared error is minimized [@poucin2018activity] following Equation \[equation\_clustering\]. $$M= \sum_{m=1}^{k} \sum_{i=1}^{n} || x_i^{(m)}-c_m ||^2
\label{equation_clustering}$$ Where $M$ is the objective function, $k$ is the number of clusters, $n$ is the number of data points (observations), $x_i^{(m)}$ is the observation $i$ being tested for cluster $m$, $c_m$ is the centroid of cluster $m$. The number of clusters should be defined based on the data and a statistical analysis. To define the optimal number of clusters, the elbow method/sum of squared error [@poucin2018activity], which measures the sum of squared distances between the points within a cluster is the guide. [Figure \[methods\]]{} presents the steps followed. For clustering, dataset has been divided into 70%, 10%, and 20% for training, validating, and testing, respectively [@tu2018development]. The optimal GHG ER (g/sec) of each cluster is defined based on the minimum sum of absolute distances between GHG ERs and the centroids of the examined cluster.
LSTM with exogenous variables {#LSTM_methodology_text}
-----------------------------
The main focus of this work is to develop an LSTM based learning framework and compare it to already established models including ARIMA and clustering. LSTM is associated with three gates [@hochreiter1997long], input, forget, and output following the below Equations \[equation\_LSTM1\], \[equation\_LSTM2\], and \[equation\_LSTM3\], respectively. $$i_t=\sigma (w_i[h_{t-1},x_t] + b_i)
\label{equation_LSTM1}$$
$$f_t=\sigma (w_f[h_{t-1},x_t] + b_f)
\label{equation_LSTM2}$$
$$o_t=\sigma (w_o[h_{t-1},x_t] + b_o)
\label{equation_LSTM3}$$
Where: $i_t$ represents the input gate, $f_t$ represents the forget gate, $o_t$ represents the output gate, $\sigma$ represents the sigmoid function, $w_x$ represents the weight for gate $x$ neurons, $h_{t-1}$ represents the output of the previous LSTM block at time $t-1$, $x_t$ represents the input at current time step $t$, and $b_x$ represents biases for respective gates ($x$).
Defining the best network is an iterative process and is dependent on the selection of the predictors, number of sequences, and set of hyper-parameters [@reimers2017optimal; @hutter2015beyond]. It is worth mentioning that increasing the depth of NNs [@hermans2013training; @pascanu2013construct] and effectively tuning the network [@snoek2012practical] may contribute positively to the prediction performance. Thus, one and two hidden layers are investigated in this work and Bayesian optimization [@wu2019hyperparameter] is utilized for the systematic tuning. With reference to hyper-parameters tuning procedure, two stages have taken place. The first is the manual, while, the second is the systematic. [Figure \[methods\]]{} presents the methodology followed in LSTM application. A ratio of 80%-20% has been considered for training to testing, respectively. To compare the performance of LSTM to ARIMA and clustering, four indicators are employed, the correlation coefficient between observed and predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec), the fit to the ideal straight curve, the R$^2$, and the RMSE.
Case study {#Case study}
==========
The road network of downtown Toronto is used as the case study. It was selected due to its high level of recurrent congestion during the morning peak period. The network consists of 223 links and 76 nodes. Links have different characteristics in terms of the number of lanes, free flow speed, and number of directions to assure heterogeneity. [Figure \[Case\_study\]]{} illustrates the area, including the major roads.
{width="\textwidth"}
[.45]{} [![Histogram of a) speed, b) density per lane, c) flow, and d) GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) []{data-label="histograms"}](histogram_speed "fig:"){width="2.5in"} \[histogram speed\]]{}
[.45]{} [![Histogram of a) speed, b) density per lane, c) flow, and d) GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) []{data-label="histograms"}](histogram_density_per_lane "fig:"){width="2.5in"} \[histogram density per lane\]]{}
[.45]{} [![Histogram of a) speed, b) density per lane, c) flow, and d) GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) []{data-label="histograms"}](histogram_total_flow "fig:"){width="2.5in"} \[histogram total flow\]]{}
[.45]{} [![Histogram of a) speed, b) density per lane, c) flow, and d) GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) []{data-label="histograms"}](histogram_CO2_EF "fig:"){width="2.5in"} \[histogram CO2\]]{}
Data collection {#Data collection}
===============
Tackling the GHG prediction problem from a disaggregated perspective, requires high resolution data points. To the best of our knowledge, currently there are no datasets available that report GHG at link level for a large and congested road network. An agent-based traffic and emission simulation has been used in this study to synthesize GHG and traffic characteristics for links at a very high temporal resolution [@Djavadian2018Distributed]. Furthermore, we are of the view that in the near future, with the adoption of smart cities technologies, high resolution datasets from real road network would become available. In that case, the proposed LSTM framework can be retrained on such datasets. The agent based simulator used in this study implemented a calibrated Intelligent Driver Model (IDM) [@Treiber2000] for vehicular movement. Vehicles are dynamically routed on the network based on the real-time traffic information [@Djavadian2018Distributed]. Simulation ends once all of the vehicles reach their destinations. The link level space mean speed, density, and flow information is recorded every minute, which was found to be the optimal updating interval [@alfaseeh2018impact]. [Figure \[histograms\]]{} demonstrates the statistical analysis of the variables reflecting on the traffic conditions and GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) on links. It can be noticed from [Figure \[histograms\]]{} that the mode for speed is 40 km/h. It is crucial to note that the speed limit of 132 and 18 links in our case study is 60 and 80 km/h, respectively. Speed average in the network based on second-by-second data point is 56.16km/h. Nevertheless, speed varies from 0 to 80 km/h. Similarly, density (veh/km.lane) and flow (veh/h) are associated with a wide range reflecting on the different traffic conditions. Finally, GHG ERs (in $\text{CO}_{2\text{eq}}$) start from less than 1 (g/sec) to more than 5 (g/sec).
With regards to the emission modelling, Motor Vehicle Emission Simulator (MOVES), which is developed by the USEPA, is adopted to generate GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) [@MOVES]. MOVES estimates emissions by defining the vehicle operating mode, which is based on the vehicle specific power (VSP) as illustrated in [Equation \[equ:1\]]{}. $$P_{V,t}=\frac{Av_t + Bv_t^2 + Cv_t^3+mv_ta_t}{m}
\label{equ:1}$$
Where:
- $P_{V,t}$ is the vehicle specific power (VSP) at time $t$
- $v_t$ is the speed of vehicle at time $t$ ($m/sec$)
- $a_t$ is the acceleration of vehicle at time $t$ ($m/sec^2$)
- $m$ is the mass of vehicle, usually referred as “weight" (Mg).
- A, B and C are track-road coefficients, representing rolling resistance, rotational resistance and aerodynamic drag, in unit kW-sec/m, kW-sec$^2$/$m^2$ and kW-sec$^3$/m$^3$.
The second-by-second CO$_2$ emissions of every vehicle on every link are then used to estimate the space mean GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) of each link based on a defined updating interval.
In terms of the data points, a variety of scenarios have been simulated to trigger different traffic conditions on links as shown in Table **\[scenarios investigated\]**.
Demand factor No. of Vehicles Departure time distribution
--------------- ----------------- ----------------------------------
0.7 2,437 Exponential, uniform, and normal
1 3,477 Exponential, uniform, and normal
1.3 4,520 Exponential, uniform, and normal
1.5 5,259 Exponential
2 6,988 Exponential
: Scenarios considered for data generation[]{data-label="table1"}
\[scenarios investigated\]
Travel demand is obtained from the Transportation Tomorrow Survey (TTS). The time dependent exogenous demand Origin-Destination (OD) matrices are based on 5 minute intervals from TTS. With regards to the demand, it ranges from 2,437 to 6,988 reflecting on demand factors from 0.7 to 2. To assure more heterogeneity in the traffic conditions generated at link level, different distributions are considered for the departure time, normal, uniform, and exponential. Two datasets are extracted corresponding to two time intervals i.e. 30 seconds and 1 minute, to examine the impact of the two levels of resolution on the prediction performance. The data has been pre-processed to suit each of the models discussed in Section \[Methodology\] and to assure a fair comparison between the models. When 1 minute is the updating interval of predictors for clustering and LSTM, 48,652 and 12,159 data points for training and testing are employed, respectively. For the 30 seconds analysis the number is double. The data have been divided into 80% training and 20% testing for LSTM. With regards to clustering, 70% of the data is randomly selected as the training set, 10% of the total data is taken for validation, and 20% of the data is used for testing, as demonstrated in [Figure \[methods\]]{}. Finally, for the ARIMA model, 4 representative links are considered, which are associated with different characteristics to give an indication of the prediction performance at the network level.
Results and discussion {#Discussion and results}
======================
This section presents the major findings, starting with the detailed correlation analysis outcome. Then we present the results of the ARIMA, clustering, and LSTM models as well as a comparison between them.
Correlation analysis {#Correlation_analysis_discussion}
--------------------
The correlation analysis is the judging factor to define not only the most important predictors for the three models, but also the optimal number of sequences/minutes for the predictors while applying LSTM. Traffic and environmental information of studied links are captured. In order to expand the spatial dimension, the characteristics of in-links (upstream links) are not neglected. Traffic conditions (e.g. speed flow, density, etc.) at time $t$ on the upstream links will give a strong indication of the traffic condition on the studied link downstream at time $t+1$. The variables considered for this analysis are speed, density, flow, delay (difference between free flow travel time and actual travel time), in-links speed, in-links density, in-links flow, and GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec). Five sequences/minutes have been investigated, which are sufficient to capture the changes in traffic conditions on link level in a congested urban network. The main outcome of this analysis is the correlation between the aforementioned indicators and the GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) at the sixth sequence/minute. Linear correlation has been employed for this part of the analysis. With reference to the results, the absolute value of the correlation coefficient between any variable and the GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) at the 6th min, except for delay, increases from minute 1 to minute 5 (top to down) as shown in [Figure \[Correlation\_1\_to\_5\_with\_6thminGHG\_including\_Speed\]]{}.
{width="5in"}
In terms of the importance order, speed, GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec), density, and in-links speed are the top four highly correlated variables with the GHG ER at the 6th minute. Speed is the variable of the highest correlation coefficient with the GHG ER. This is due to the explicit dependency of GHG estimation on speed [@MOVES]. Based on the traffic fundamental relationships, the relationship between speed and density is observed to be monotonically decreasing, which justifies the high correlation between density and the GHG ER [@papacostas1993transportation]. Among the in-links characteristics, in-links speed is the most correlated variable with the GHG ER, which is due to the fact that speed is the variable GHG depends on for estimation. This analysis triggers the choice of not only the predictors for the three models, but also the number of sequences for LSTM.
ARIMA with exogenous variables {#ARIMAX_discussion}
------------------------------
The ARIMA model with exogenous variables does not consider the non-linearity between variables [@zhang2003time], which is a serious drawback when the data and relationships are complex. Furthermore, ARIMA model is not scaleable in which every link requires a model based on its data, which is tedious while dealing with a large number of links. Nevertheless, to compare between the models, a sample of 4 links, which are associated with different characteristics (number of lanes and free flow speed) and conditions (congested and uncongested), are considered. As shown in [Figure \[ARIMA\_link\_20\_24\_71\_185\_DF2\_1min\]]{}, the correlation coefficient and R$^2$ of the ARIMA model of the four links are 0.72 and 0.62, respectively. With regards to the fit to the straight curve, there is a slight overestimation and the RMSE is 0.3998 (g/sec).
{width="2.8in"}
Clustering {#Clustering_discussion}
----------
To define the optimal number of clusters, the sum of squared error (elbow method) is adopted. From [Figure \[Sum\_squared\_error\_final\_15clusters\_30sec\_3pred\]]{}, the optimal number of clusters is 5. However, 10 and 15 clusters are examined for further enhancements. From Figure \[Clustering\], it is noticed that predicted values of GHG ERs are considered as discrete variables triggering dramatic over and underestimation. Values of GHG ERs ranging from 1 to 4 are predicted to be either 1.5, 1.8, 1.9, or 2.3 (g/sec). The correlation coefficient, R$^2$, and RMSE are 0.71, 0.66, and 0.4 (g/sec), respectively. While comparing true versus predicted GHG ERs of 5 clusters to 10 and to 15 clusters, it is illustrated that even when the number of clusters increases, the performance indicators have not improved substantially and the GHG ERs between 2.5 and 4 g/sec are either underestimated to be 2.5 or overestimated to be 4 g/sec.
{width="3in"}
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of clustering where: a) is for 5 clusters, b) is for 10 clusters, and c) is for 15 clusters[]{data-label="Clustering"}](Clustering_5_11sc_normalized "fig:"){width="3in"} \[5\_clusters\]]{}
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of clustering where: a) is for 5 clusters, b) is for 10 clusters, and c) is for 15 clusters[]{data-label="Clustering"}](Clustering_10_11sc_normalized "fig:"){width="3in"} \[10\_clusters\]]{}
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of clustering where: a) is for 5 clusters, b) is for 10 clusters, and c) is for 15 clusters[]{data-label="Clustering"}](Clustering_15_11sc_normalized "fig:"){width="3in"} \[15\_clusters\]]{}
With reference to the fit to the straight curve, it is noticed in Figure \[Clustering\] that GHG ERs between 2.7 to around 4 (g/sec) are either underestimated taking one value of around 2.5(g/sec) or over estimated. It can be concluded that since clustering considers the predicted GHG ER as a discrete variable, the dynamic nature of traffic conditions is not captured efficiently, compared to ARIMA, triggering the aforementioned outcomes. Not to forget that there is a proportional relationship between the number of clusters and the amount of data required for training. Thus, even when the number of clusters is increased the data requirement can be hard to fulfill.
LSTM with exogenous variables results {#Long-Short Term Memory_discussion}
-------------------------------------
[Table \[LSTM scenarios investigated\]]{} shows the specifications of the trained networks including the predictors, number of sequences (previous minutes), number of LSTM layers, updating interval of data points, and the hyper-parameter tuning approach.
------- ------------------------------------------------------------- ------------- -------- ---------- -------------- --
Model Predictors No. of No. of Sequence Hyper-param.
ID sequences/ LSTM interval tuning
minutes layers approach
LSTM1 Speed, density, 3 1 1 min Bayesian
and GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec)
LSTM2 Speed, density, GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec), 3 1 1 min Bayesian
and in-links speed
LSTM3 Speed, density, GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec), 3 2 1 min Bayesian
and in-links speed
------- ------------------------------------------------------------- ------------- -------- ---------- -------------- --
: LSTM specifications considered for application[]{data-label="LSTM scenarios investigated"}
Before demonstrating the results, it is important to note that the networks presented in this section are systematically tuned adopting the Bayesian optimization [@krizhevsky2009learning]. The search range for each of the hyper-parameters is around the optimal values of the manual tuning for each of the networks. For the manual tuning, a comprehensive list of tuning sets is examined to narrow the search range for a more efficient systematic tuning. Two solvers are considered, the stochastic gradient descent with momentum (sgdm) [@robert2014machine] and adaptive moment estimation (Adam) methods [@kingma2014adam]. The sgdm adds a momentum term to the parameter update to reduce oscillation associated with the standard stochastic gradient descent solver along the path of steepest descent towards the optimal solution [@robert2014machine]. Adam is derived from adaptive moment estimation. It adopts different learning rates for the parameters unlike sgdm, but uses a momentum term like the sgdm to further improve network training [@kingma2014adam].
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of a) LSTM1, b) LSTM2, and c) LSTM3[]{data-label="3LSTM_networks_final"}](Iter10_7_makeObjFcn_5_validation_included "fig:"){width="3.2in"} \[Iter10\_7\]]{}
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of a) LSTM1, b) LSTM2, and c) LSTM3[]{data-label="3LSTM_networks_final"}](Iter11_7_makeObjFcn_5_validation_included "fig:"){width="3.2in"} \[Iter11\_7\]]{}
[.45]{} [![True vs. predicted GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) of a) LSTM1, b) LSTM2, and c) LSTM3[]{data-label="3LSTM_networks_final"}](Iter11_10_3_2_makeObjFcn_3_validation_included "fig:"){width="3.2in"} \[Iter11\_10\_3\_2\]]{}
The considered hyper-parameters include: the initial learning rate, momentum, max epochs, learning rate drop factor, learning rate drop period, number of hidden units of the first LSTM (hidden) layer, and the number of hidden units of the second (LSTM) layer when used. Not to neglect that different number of sequences (previous minutes) have been assessed for the best application. Three previous minute sequences have contributed to the best performance as the period is sufficient to capture the changes in traffic conditions and according to the linear correlation analysis as in Figure \[Correlation\_1\_to\_5\_with\_6thminGHG\_including\_Speed\]. [The performance indicators of the best LSTM networks are presented in Table \[Performance indicators of LSTM networks\].]{} Figure \[3LSTM\_networks\_final\] demonstrates the performance indicators of LSTM1, which employs the link’s three highly correlated predictors of the last three minutes. The LSTM1 is associated with a correlation coefficient, R$^2$, and RMSE of 0.75, 0.69, and 0.37 (g/sec), respectively, while underestimating the GHG ERs by 4%. Comparing LSTM2 to LSTM1, a negligible enhancement in the correlation coefficient and R$^2$ is achieved while the RMSE is similar. The LSTM2 underestimates by 5%. This slight improvement stems from the fact that LSTM2 does not only employ the three mostly correlated predictors with the response (GHG ER), but also the in-links highly correlated characteristic (in-links speed) as illustrated in Figure \[Correlation\_1\_to\_5\_with\_6thminGHG\_including\_Speed\]. When speed of in-links (upstream) at time $t$ is very low due to congestion contributing to more GHG emissions, this means that at time $t+1$ these vehicles driving at low speed will be on the downstream studied link. In other words, speed on in-links at time $t$ provides an indication of what the speed will be on the downstream link at time $t+1$. The results of LSTM2 demonstrates the potential of including the highly correlated in-links characteristics to better reflect on the spatial dimension while predicting GHG ERs. LSTM3, which adopts 2 hidden layers to account for the higher order of correlation between the variables, gives the best results in terms of the four performance indicators. The LSTM3 network underestimates by only 3% and is associated with a correlation coefficient, R$^2$, and RMSE of 0.78, 0.7, and 0.36 (g/sec), respectively.
------- ------------- ------- ---------- -------
Model Correlation R$^2$ Linear RMSE
ID coefficient fit
LSTM1 0.749 0.691 Y=0.96.x 0.374
LSTM2 0.750 0.692 Y=0.95.x 0.374
LSTM3 0.767 0.704 Y=0.97.x 0.362
------- ------------- ------- ---------- -------
: Performance indicators of the trained LSTM networks[]{data-label="Performance indicators of LSTM networks"}
A mutual feature in LSTMs is that the models are unable to predict GHG ERs that are higher than 4 g/sec. This is due to the fact that data points reflecting on traffic conditions associated are not sufficient as illustrated in Figure \[histograms\]. That is, more data points representing the traffic conditions associated with GHG ERs higher than 4 (g/sec) may contribute positively.
Investigating the impact of a shorter updating interval is of an added value. Hence, three LSTM models of the same specifications of LSTM1, LSTM2, and LSTM3 have been developed for 30 second intervals. The best model resulted in RSME of 0.48 g/sec, R$^2$ of 0.707, and correlation of 0.78. It is found that increasing the level of resolution has an adverse impact on the RMSE, while improving the correlation coefficient slightly for the three networks. More specifically, for LSTM3 when the updating interval is 30 second both the correlation coefficient and the RMSE increased, by 1.34% and 35%, respectively. The main justification is that the shorter updating interval triggers more noise and local fluctuations that are not well captured by a general LSTM that is not trained for a specific link. Therefore, 1 minute updating interval is sufficient at this point to account for the high level of heterogeneity in link characteristics.
Comparing the LSTM model outcome to the other models, clustering and ARIMA, shows that the three aforementioned LSTM networks outperform the best case of clustering as in Figure \[Clustering\] and ARIMA as in Figure \[ARIMA\_link\_20\_24\_71\_185\_DF2\_1min\] in terms of the four performance indicators. More specifically, the RMSE of LSTM3 is less by 6% and 9% compared to the best case of clustering and ARIMA, respectively. Unlike clustering, LSTM considers GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) as a continuous variable. Compared to ARIMA with exogenous variables, LSTM is scaleable and is applicable to any link in the studied network. It is important to note that the links in downtown Toronto are associated with a high level of heterogeneity in terms of the speed limit and the number of lanes. Thus, the developed LSTM models are network level regardless of the link characteristics, which might contribute to the deterioration in prediction performance when GHG ERs are higher than 4 g/sec.
Conclusion and potential directions {#Conclusion}
===================================
It has been shown that the efficiency of routing is a key tool to reduce the undesirable impact of transportation systems on the environment [@tu2019quantifying]. Thus, reliably predicting GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) is crucial. It can pave the way to develop non-myopic eco-routing [resulting in more sustainable transportation systems]{} and can also be adopted in other applications that can efficiently mitigate the negative effects of GHG emissions. Previous studies are predominantly based on macroscopic data points. Even when microscopic data are used, the case studies involve a single or few intersections/links [@alfaseeh2020multifactor]. Therefore, there is a need for studies tackling the issues related to GHG emissions at a high resolution in terms of both time and space. To predict GHG ERs at link level, a deep learning approach, i,e, LSTM with exogenous variables has been employed in this study. LSTM has been chosen because it has overcome various limitations of other NNs in the context of time-series data, such as the vanishing gradient problem [@amarpuri2019prediction]. To assure satisfying prediction outcome, a sufficient amount of representative data points is generated from MOVE and traffic microsimulation for downtown Toronto. The optimal set of predictors and number of time sequences used are identified based on a comprehensive correlation analysis. A comparison with the commonly used approaches for times series prediction, i.e. clustering and ARIMA, is also conducted.
In this study, the LSTM3 network of two hidden layers, which is systematically tuned using Bayesian optimization, outperforms the best case of clustering and ARIMA models in terms of the four performance indicators. More specifically, the LSTM3 network reduces the RMSE by 6% and 9% compared to the best case of clustering and ARIMA, respectively. While the LSTM can scale up to a network level and considers the GHG ER as a continuous variable, ARIMA requires a model for every link and clustering considers GHG ER as a discrete variable. ARIMA’s major drawbacks are related to scaling and non-linearity consideration between variables. In clustering, the main limitation is that predicted GHG ER (in $\text{CO}_{2\text{eq}}$ g/sec) is considered as a discrete variable, which has affected the fit quality to the straight curve. In other words, GHG ERs (in $\text{CO}_{2\text{eq}}$ g/sec) between 2.5 and 4 g/sec are either underestimated to be 2.5 or overestimated to be 4, even when the number of clusters is 15. This is due to the fact that clustering lacks the temporal dimension compared to the LSTM model. Utilizing a smaller updating interval of 30 seconds in the case of LSTM modelling improves the correlation coefficient slightly compared to the case of 1 minute updating interval, but has a negative impact on the RMSE of the three LSTM models. The shorter updating interval triggers more local fluctuations that are not well captured by the networks level model.
For future work, various applications of the model will be evaluated. One such application is in the non-myopic eco-routing, where future time-step GHG predictions on downstream links are needed in order to decide the dynamic route for a vehicle. Figure \[histograms\] illustrates that the frequency of speed higher than 60 (km/h) or lower than 30 (km/h) is very small. Hence, to enhance the current work, generating more data points reflecting on the aforementioned condition may contribute positively to the current outcomes. With reference to the hyper-parameters tuning, dedicating more time for the Bayeisan optimization may result in further improvement of the model’s prediction capacity. Developing models for categorized links based on their speed limit, and number of lanes may introduce further enhancements. In addition, utilizing physical constraints, which consider suitable traffic characteristics (flow over capacity, density over jam density, etc.), may contribute to higher prediction accuracy. Finally, since the traffic and environmental information are obtained from simulation, employing sensors to collect real data will further enhance the model’s prediction capability.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A systematic algorithm for building integrating factors of the form $\mu(x,y)$, $\mu(x,\y1)$ or $\mu(y,\y1)$ for second order ODEs is presented. The algorithm can determine the existence and explicit form of the integrating factors themselves without solving any differential equations, except for a linear ODE in one subcase of the $\mu(x,y)$ problem. Examples of ODEs not having point symmetries are shown to be solvable using this algorithm. The scheme was implemented in Maple, in the framework of the [*ODEtools*]{} package and its ODE-solver. A comparison between this implementation and other computer algebra ODE-solvers in tackling non-linear examples from Kamke’s book is shown.'
author:
- 'E.S. Cheb-Terrab$^{a,b}$ and A.D. Roche$^a$'
title: Integrating factors for second order ODEs
---
LaTeX[L-.25em.425ex-.075emTeX]{} \#1[Eq.(\[\#1\])]{} ł \#1[sec. \[\#1\]]{}
1 [H]{} \#1[[*abaco$_#1$*]{}]{} ([([height1.29em width0em depth1.29em]{}]{} )[[height1.29em width0em depth1.29em]{})]{} ([([height1em width0em depth1em]{}]{} )[[height1em width0em depth1em]{})]{}
*$^a$Symbolic Computation Group*
*Computer Science Department, Faculty of Mathematics*
*University of Waterloo, Ontario, Canada.*
*$^b$Department of Theoretical Physics*
*State University of Rio de Janeiro, Brazil.*
Introduction
============
Although in principle it is always possible to determine whether a given ODE is exact (a total derivative), there is no known method which is always successful in making arbitrary ODEs exact. For order ODEs - as in the case of symmetries - integrating factors ($\mu$) are determined as solutions of an order linear PDE in $n$+1 variables, and to solve this determining PDE is a major problem in itself.
Despite the fact that the determining PDE for $\mu$ naturally splits into a PDE system, the problem is - as a whole - too general, and to solve it a restriction of the problem in the form of a more concrete ansatz for $\mu$ is required. For example, in a recent work by [@bluman1] the authors explore possible ansatzes depending on the given ODE, which are useful when this ODE has known symmetries of certain type. In another work, [@wolf] explores the use of computer algebra to try various ansatzes for $\mu$, no matter the ODE input, but successively increasing the order of the derivatives (up to the $-1$ order) on which $\mu$ depends; the idea is to try to maximize the splitting so as to increase the chances of solving the resulting PDE system by first simplifying it using differential Groebner basis techniques.
Bearing this in mind, this paper presents a method, for second order explicit ODEs[^1], which systematically determines the existence and the explicit form of integrating factors when they depend on only two variables, that is: when they are of the form $\mu(x,y)$, $\mu(x,\y1)$ or $\mu(y,\y1)$. The approach works without solving any auxiliary differential equations - except for a linear ODE in one subcase of the $\mu(x,y)$ problem - and is based on the use of the forms of the ODE families admitting such integrating factors. It turns out that with this restriction - $\mu$ depends on only two variables - the use of differential Groebner basis techniques is not necessary; these integrating factors, when they exist, can be given directly by identifying the input ODE as a member of one of various related ODE families.
The exposition is organized as follows. In , the standard formulation of the determination of integrating factors is briefly reviewed and the method we used for obtaining the aforementioned integrating factors $\mu(x,y)$, $\mu(x,\y1)$ or $\mu(y,\y1)$ is presented. In , some aspects of the integrating factor and symmetry approaches are discussed, and their complementariness is illustrated with two ODE families not having point symmetries. Sec. \[tests\] contains some statistics concerning the new solving method and the second order non-linear ODEs found in Kamke’s book, as well as a comparison of performances of some popular computer algebra packages in solving a related subset of these ODEs. Finally, the conclusions contain some general remarks about the work.
Aside from this, in the Appendix, a table containing extra information concerning integrating factors for some of Kamke’s ODEs is presented.
Integrating Factors and ODE patterns {#intfactor}
====================================
In this paper we use the term “integrating factor" in connection with the explicit form of an order ODE
$$% {\frac {d^ny}{dx^n}}=\Phi(x,y,\y1,...,\ym)
\yn-\Phi(x,y,\y1,...,\ym) = 0
\label{oh}$$
so that $\mu(x,y,\y1,...,\ym)$ is an integrating factor if
$$% \mu\l({\frac {d^ny}{dx^n}}-\Phi\r) = \frac{d}{dx}R(x,y,\y1,...,\ym)
\mu\l(\yn-\Phi\r) = \frac{d}{dx}R(x,y,\y1,...,\ym)
\label{exact_oh}$$
for some function $R$. The knowledge of $\mu$ is - in principle - enough to determine $R$ by using standard formulas (see for instance Murphy’s book). To determine $\mu$, one can try to solve for it in the exactness condition, obtained applying Euler’s operator to the total derivative $H \equiv \mu\l(\yn-\Phi\r)$:
$${\frac {\pa H}{\pa y}}
-{\frac {d}{dx}}
\l(\frac{\pa H}{\pa \y1}\r)
+{\frac {d^{2}}{d{x}^{2}}}
\l( \frac{\pa H}{\pa \yt} \r)
+ ...
+ (-1)^n
{\frac {d^{n}}{d{x}^{n}}} \l( \frac{\pa H}{\pa y^{(n)}} \r) =0
\label{EC}$$
is of the form
$$A(x,y,\y1,...,y^{(2n-3)}) + y^{(2n-2)} B(x,y,\y1,...,\ym) = 0
\label{PDE_ABn}$$
where $A$ is of degree $n-1$ in $\yn$ and linear in $\yk$ for $n < k
\leq (2n-3)$, so that can be split into a PDE system for $\mu$. In the case of second order ODEs - the subject of this work - is of the form
$$A(x,y,\y1)+ \yt\ B(x,y,\y1)=0
\label{PDE_AB}$$
and the PDE system is obtained by taking $A$ and $B$ equal to zero[^2]:
$$\begin{aligned}
A & \equiv &
% \left ({\frac {\pa ^{2}\mu}{\pa \y1\pa x}}
% + \y1{\frac {\pa^{2}\mu}{\pa \y1\pa y}}
% -{\frac {\pa \mu}{\pa y}}\right)\Phi
% +\left ({\frac { \pa^{2}\Phi}{\pa \y1\pa x}}
% -{\frac {\pa \Phi}{\pa y}}
% +\y1{\frac { \pa ^{2}\Phi}{\pa \y1\pa y}}\right )\mu
% +{\y1}^{2}{\frac {\pa ^{2}\mu}{ \pa {y}^{2}}}
%
\left (y'\mu_{{y'y}}-\mu_{{y}}+\mu_{{y'x}}\right )\Phi+\left (
\Phi_{{y'x}}+y'\Phi_{{y'y}}-\Phi_{{y}}\right )\mu+{y'}^{2}\mu_{{
yy}}
\nonumber
\\*[.1in]
& &
+\left (\mu_{{y}}\Phi_{{y'}}+\mu_{{y'}}\Phi_{{y}}+2\,\mu_{{xy}
}\right )y'+\mu_{{y'}}\Phi_{{x}}+\mu_{{x}}\Phi_{{y'}}+\mu_{{xx}}
= 0
\label{PDE_A}
\\*[.2in]
B
& \equiv &
% 2\,{\frac {\pa \mu}{\pa y}}
% +\Phi\;{\frac { \pa ^{2}\mu}{\pa {\y1}^{2}}}
% +2 \,{\frac {\pa \mu}{\pa \y1}}\;{\frac {\pa \Phi}{\pa \y1}}
% +\mu{\frac {\pa ^{ 2}\Phi}{\pa {\y1}^{2}}}
% +{\frac {\pa ^{2}\mu}{\pa \y1\pa x}}
% +\y1{\frac {\pa^{2}\mu}{\pa \y1 \pa y}}
y'\mu_{{y'y}}+\Phi\,\mu_{{y'y'}}+\mu\,\Phi_{{y'y'}}+2\,\mu_{{y}}
+2\,\mu_{{y'}}\Phi_{{y'}}+\mu_{{y'x}}
= 0
\label{PDE_B}\end{aligned}$$
Regarding the solvability of these equations, unless a more concrete ansatz for $\mu(x,y,\y1)$ is given, the problem is in principle as difficult as solving the original ODE. We then studied the solution for $\mu$ of Eqs.(\[PDE\_A\]) and (\[PDE\_B\]) when $\mu$ depends only on two variables, that is: for $\mu(x,y)$, $\mu(x,\y1)$ and $\mu(y,\y1)$. Concretely, we searched for the existence conditions for such integrating factors, expressed as a set of equations in $\Phi$, plus an algebraic expression for $\mu$ as a function of $\Phi$, valid when the existence conditions hold. Formulating the problem in that manner and taking into account the integrability conditions of the system, Eqs.(\[PDE\_A\]) and (\[PDE\_B\]) turned out to be solvable for $\mu(x,y)$, but appeared to us untractable when $\mu$ depends on two variables one of which is $\y1$.
We then considered a different approach, taking into account from the beginning the form of the ODE family admitting a given integrating factor. As shown in the following sections, it turns out that, using that piece of information ( below), when $\mu$ depends only on two variables the existence conditions and the integrating factors themselves can be systematically determined; and in the cases $\mu(x,\y1)$ and $\mu(y,\y1)$, this can be done without solving any differential equations.
Concerning the ODE families admitting given integrating factors, we note that, from
$$\mu(x,y,\y1,...,\ym)=\fr{\pa\,R\mbox{\hspace{2mm}}}{\pa \ym}$$
and hence the first integral $R$ is of the form
$$R=G(x,y,...,\ymm)+\int \mu\ d\ym
\label{R}$$
for some function $G$. In turn, since $R$ is a first integral, it satisfies
$$% \frac{\pa R}{\pa x}
% + \y1 \frac{\pa R }{\pa y}
% + \Phi \frac{\pa R}{\pa \y1} = 0
%
R_x + \y1 R_y + ... + \Phi R_{y^{(n-1)}} = 0
\label{A}$$
Inserting into the above and solving for $\yn$ leads to the general form of an ODE admitting a given integrating factor:
$$\label{ODE_mu}
\yn = \frac {-1}{\mu}
\l[{
\frac {\pa }{\pa x}\l(\int \mu \, {d{\it \ym}} + G\r)
+ ...
+ \ym \frac {\pa }{\pa \ymm}\l(\int \mu \, d\ym + G\r)
}
\r]$$
Second order ODEs and integrating factors of the form $\mu(x,y)$ {#mu_xy}
----------------------------------------------------------------
We consider first the determination of integrating factors of the form $\mu(x,y)$, which turns out to be straightforward[^3]. The determining equations (\[PDE\_A\]) and (\[PDE\_B\]) for this case are given by:
$$\begin{array}{c}
{y' }^{2}\mu_{{yy}}+2\,\mu_{{xy}}y' +\mu\,\Phi_{{y' x}}+\mu\,\Phi_{
{y' y}}y' -\mu\,\Phi_{{y}}-\mu_{{y}}\Phi+\mu_{{y}}y' \,\Phi_{{y' }}+
\mu_{{xx}}+\mu_{{x}}\Phi_{{y' }} = 0
\\*[.1in]
\mu\,\Phi_{{y' y' }}+2\,\mu_{{y}}=0
\end{array}
\label{sys_mu_xy}$$
Although the use of integrability conditions is enough to tackle this problem, the solving of Eqs.(\[sys\_mu\_xy\]) can be directly simplified if we take into account the ODE family admitting an integrating factor $\mu(x,y)$. From , that ODE family takes the form
$$\label{mu_xy_ode}
\yt = a(x,y)\, \y1 ^2 + b(x,y)\, \y1 + c(x,y),$$
where
$$\label{mu_xy_ode_cond}
a(x,y) = - \frac{\mu_y}{\mu},\ \ \
b(x,y) = - \frac{G_y + \mu_x}{\mu},\ \ \
c(x,y) = - \frac{G_x}{\mu}$$
and $G(x,y)$ is an arbitrary function of its arguments. Hence, as a shortcut to solving Eqs.(\[sys\_mu\_xy\]), one can take as an existence condition - $\Phi$ must be a polynomial of degree two in 1 - and directly solve Eqs.(\[mu\_xy\_ode\_cond\]) for $\mu$. The calculations are straightforward; there are two different cases.
: $\ 2a_x-b_y \neq 0$
Defining the two auxiliary quantities
$$\varphi \equiv c_y - a\,c - b_x,\ \ \ \ \
\Upsilon \equiv a_{xx} + a_x\,b + \varphi_y
\label{auxiliary}$$
an integrating factor of the form $\mu(x,y)$ exists only when
$$\Upsilon_y - a_x = 0,\ \ \ \ \ \
\Upsilon_x + \varphi + b\,\Upsilon - \Upsilon^2 = 0$$
and is then given in solved form, in terms of $a$, $b$ and $c$ by
$$\mu(x,y)=
\exp\l(
\int \l(-\Upsilon+ \fr{\pa}{\pa x} \int a\, dy \r) dx
- \int a\, dy
\r)$$
So, in this case, when an integrating factor of this type exists there is only one[^4] and it can be determined without solving any differential equations.
: $\ 2a_x-b_y = 0$
Redefining $\varphi \equiv c_y - a\,c$, an integrating factor of the form $\mu(x,y)$ exists only when
$$a_{xx} - a_x\,b - \varphi_y = 0,$$
and then $\mu(x,y)$ is given by
$$\mu(x,y)=\nu(x)\ {\rm e}^{{^{-\displaystyle \int a\, dy}}}
\label{mu_xy_2}$$
where $\nu(x)$ is either one of the independent solutions of the second order linear ODE[^5]
$$\nu'' = A(x)\, \nu' + B(x)\, \nu,
\label{lode}$$
and $$A(x) \equiv 2\, {\cal I} - b,\ \ \ \ \ \
B(x) \equiv
\varphi
+ \l(
{\cal I} - \fr{\pa}{\pa x}
\r)
\l(
b- {\cal I}
\r),\ \ \ \ \ \
{\cal I} \equiv \fr{\pa}{\pa x} \int a\, dy$$
So in this case, to transform into an explicit expression for $\mu$ we first need to solve a second order linear ODE. When the attempt to solve is successful, using each of its two independent solutions as integrating factors leads to the general solution of , instead of just a reduction of order.
Second order ODEs and integrating factors of the form $\mu(x,\y1)$ {#theorem}
------------------------------------------------------------------
When the integrating factor is of the form $\mu(x,\y1)$, the determining equations (\[PDE\_A\]) and (\[PDE\_B\]) become
$$\begin{array}{c}
\left (\Phi_{{y}}y' +\Phi_{{x}}\right )\mu_{{y' }}+
\left (-\Phi_{{y}}+\Phi_{{y' x}}+\Phi_{{y' y}}y' \right )\mu+\mu_{{x
x}}+\mu_{{x}}\Phi_{{y' }}+\mu_{{y' x}}\Phi=0
\\*[.1in]
\Phi\,\mu_{{y' y' }}+\mu\,\Phi_{{y' y' }}+2\,\mu_{{y' }}\Phi_{{y' }
}+\mu_{{y' x}}=0
\end{array}
\label{sys_mu_x_y1}$$
As in the case $\mu(x,y)$, the solution we are interested in is an expression for $\mu(x,\y1)$ in terms of $\Phi$, as well as existence conditions for such an integrating factor expressed as equations in $\Phi$. However, differently than the case $\mu(x,y)$, we didn’t find a way to solve the $\mu(x,\y1)$ problem just using integrability conditions, neither working by hand nor using the specialized computer algebra packages [*diffalg*]{} [@boulier] and [*standard\_form*]{} [@reid]. We then considered approaching the problem as explained in the previous subsection, departing from the form of for $\mu =
\mu(x,y')$:
$$\label{reducible1}
\yt = \Phi(x,y,\y1) \equiv
- \frac{F_x + G_x + G_y\, \y1}{F_{y'}}$$
where $G(x,y)$ and $F(x,\y1)$ are arbitrary functions of their arguments and
$$\label{mu_F}
{\mu}(x,\y1) = F_{y'}$$
Now, is not polynomial in either $x$, $y$ or 1, and hence its use to simplify and solve the problem is less straightforward than in the case $\mu(x,y)$. However, in , all the dependence on $y$ comes from $G(x,y)$ in the numerator, and as it is shown below, this fact is a key to solving the problem. Considering ODEs for which $\Phi_y \neq 0$[^6], the approach we used can be summarized in the following three lemmas whose proofs are developed separately for convenience.
.
### Proof of [**Lemma 1**]{}
For to be linear and not missing $y$, either $G_x$ or $G_y$ must be linear in $y$. Both $G_x$ and $G_y$ cannot simultaneously be linear in $y$ since, in such a case, $G_x/F_{y'}$ or $y'G_y/F_{y'}$ would be non-linear in $\{y,y'\}$[^7]; therefore, either $G_{yy}=0$ or $G_{xy}=0$.
: $G_y$ is linear in $y$ and $G_{xy} = 0$
Hence, $G$ is given by
$$G = C_2 y^2 + C_1 y + g(x)$$
where $g(x)$ is arbitrary. From , in order to have $y'G_y/F_{y'}$ linear in $\{y,y'\}$, $F_{y'}$ must of the form $\nu(x) y'$ for some function $\nu(x)$. Also, $F_x/F_{y'}$ can have a term linear in $y'$, and a term proportional to $1/\y1$ to cancel with the one coming from $G_x/F_{y'}=g'/F_{y'}$, so that
$$F_{y'} = \nu \y1,\ \ \ \ \ F_{x} = \frac{\nu' \y1^2}{2} - g'$$
where the coefficient $\nu'/2$ in the second equation above arises from the integrability conditions between both equations. is then of the form
$$\yt=
-{\frac {{\nu}'}{2\;\nu}}\;\y1
-{\frac {2\;C_2\,}{\nu}}\;y
-{\frac {C_1}{\nu}}
\label{lode_mu_x_y1}$$
and hence, a linear ODE $\yt = a(x) \y1 + b(x) y$ has an integrating factor $\mu(x,\y1) = \y1/b$ when $b'/b - 2a = 0$. $\triangle$
: $G_x$ is linear in $y$ and $G_{yy} = 0$
In this case, in order to have linear, $F_{y'}$ cannot depend on $y'$, so that the integrating factor is of the form $\mu(x)$ and hence the case is of no interest: we end up with the standard search for $\mu(x)$ as the solution to the adjoint of the original linear ODE. $\triangle$
### Proof of [**Lemma 2**]{}
It follows from Eqs.(\[reducible1\]) and (\[mu\_F\]) that, given $\F$ satisfying (\[cal\_F\]),
$$\frac {\pa}{\pa y}
\medl(\Phi(x,y,\y1)\ \F(x,\y1)\medr) =
- \frac {G_{y\,x}(x,y) + G_{y\,y}(x,y)\, \y1}{\mut(x)}
\label{G_eq}$$
Hence, by taking coefficients of 1 in the above,
$$\begin{aligned}
\varphi_1 & \equiv &
\Phi_y(x,y,\y1)\, \F(x,\y1) -\y1\, \frac{\pa}{\pa \y1}
\medl( \Phi_y(x,y,\y1)\, \F(x,\y1) \medr)
= - \frac {G_{y\,x}(x,y)}{\mut(x)}
\nonumber\\*[.07 in]
\varphi_2 & \equiv &
\frac{\pa}{\pa \y1}
\medl( \Phi_y(x,y,\y1)\, \F(x,\y1) \medr)
= - \frac {G_{y\,y}(x,y)}{\mut(x)}
\label{varphi_1_2}\end{aligned}$$
where the left-hand-sides can be calculated explicitly since they depend only on $\Phi$ and the given $\F$. Similarly,
$$\begin{aligned}
\varphi_3 & \equiv &
-\frac {\pa}{\pa \y1} \medl( \Phi(x,y,\y1)\ \F(x,\y1) \medr)
=
\frac {F_{y'x}(x,\y1) + G_{y}(x,y)}{\mut(x)}
\nonumber\\*[.07 in]
\varphi_4 & \equiv & \frac {\pa}{\pa \y1} \F(x,\y1)
= \frac {F_{y'y'}(x,\y1)}{\mut(x)}
\label{varphi_3_4}\end{aligned}$$
Now, since in this case the ODE family is nonlinear by hypothesis, either $\varphi_2$ or $\varphi_4$ is different from zero, so that at least one of the pairs $\{\varphi_1,\ \varphi_2\}$ or $\{\varphi_3,\ \varphi_4\}$ can be used to determine $\mut(x)$ as the solution of a first order linear ODE. For example, if $\varphi_2 \neq 0$,
$$\frac {\pa}{\pa y} \l(\varphi_1(x,y)\ {\mut(x)} \r)
= \frac {\pa}{\pa x} \l(\varphi_2(x,y)\ {\mut(x)} \r)$$
from where
$$\mut(x) =
{e^{^{\ \displaystyle
{\int \!
{\frac {1}{\varphi_2}}
\l(
{{\frac {\pa \varphi_1}{\pa y}}}
-{{\frac {\pa \varphi_2}{\pa x}}}
\r)
{dx}}
}}}
\label{mut_2}$$
If $\varphi_2 = 0$ then $\varphi_4 \neq 0$ and we obtain
$$\mut(x) =
{e^{^{\ \displaystyle
{\int \!
{\frac {1}{\varphi_4}}
\l(
{{\frac {\pa \varphi_3}{\pa \y1}}}
-{{\frac {\pa \varphi_4}{\pa x}}}
\r)
{dx}}
}}}
\label{mut_3}$$
When combined with , Eqs.(\[mut\_2\]) and (\[mut\_3\]) alternatively give both an explicit solution to the problem and an existence condition, since a solution $\mut(x)$ - and hence an integrating factor of the form $\mu(x,\y1)$ - exists if the integrand in or only depends on $x$. $\triangle$
### Proof of [**Lemma 3**]{} {#lemma}
We start from by considering the expression
$$\label{Upsilon}
\Upsilon \equiv
\Phi_y =
- \frac {G_{x\,y}(x,y) + G_{y\,y}(x,y)\ \y1}{F_{y'}(x,\y1)}$$
and develop the proof below splitting the problem into different cases. For each case we show how to find $\F(x,\y1)$ satisfying . $\F$ will then lead to the required integrating factor when, in addition to the conditions explained below, the existence conditions for $\mut(x)$ mentioned in the previous subsection are satisfied.
: $G_{x\,y} / G_{y\,y}$ depends on $y$
To determine whether this is the case, we cannot just analyze the ratio $G_{x\,y}/G_{y\,y}$ itself since it is unknown. However, from , in this case the factors of $\Upsilon$ depending on $y$ will also depend on 1, and this condition can be formulated as
$$% \Upsilon_y \neq 0 \ \ \mbox{and}\ \ \
\frac{\partial }{\partial \y1} \l(\frac{\Upsilon_y}{\Upsilon}\r) \neq 0
\label{conditions_A}$$
When this inequation holds, we determine $F_{y'}(x,\y1)$ up to a factor depending on $x$, that is, the required $\F(x,\y1)$, as the reciprocal of the factors of $\Upsilon$ which depend on 1 but not $y$. $\triangle$
Kamke’s ODE 226
This ODE is presented in Kamke’s book already in exact form, so we start by rewriting it in explicit form as
$$\yt={\frac {x^2 y\y1+x y^2}{\y1}} % this was ode 226
% y\yt-{\y1}^{2}+\left (x+x{y}^{2}\right )\y1-y\left (1-{y}^{2}\right ) = 0
% above is ode[123]
\label{k226}$$
We determine $\Upsilon$ () as
$$\Upsilon = \frac {x (x \y1+2 y)}{\y1}$$
The only factor of $\Upsilon$ containing $y$ is:
$$x \y1+2 y$$
and since this also depends on 1, $\F(x,\y1)$ is given by
$$\F(x,\y1) = \y1$$
: either $G_{xy}=0$ or $G_{yy}=0$
When the expression formed by all the factors of $\Upsilon$ containing $y$ does not contain 1, in we will have $\frac{\partial}{\partial y'}(\frac{\Upsilon_y}{\Upsilon}) = 0$, and it is impossible to determine [*a priori*]{} whether one of the functions $\{G_{x\,y},\ G_{y\,y}\}$ is zero, or alternatively their ratio does not depend on $y$. We then proceed by assuming the former, build an expression for $\F(x,\y1)$ as in Case A, and check for the existence of $\mut(x)$ as explained in the previous subsection. If $\mut(x)$ exists, the problem is solved; otherwise we proceed as follows.
: $G_{x\,y} / G_{y\,y}\neq 0 $ and does not depend on $y$
In this case, neither $G_{x\,y}$ nor $G_{y\,y}$ is zero and their ratio is a function of just $x$, so that
$$\begin{aligned}
G_{x\,y} & = &v_1(x)\ w(x,y)
\nonumber\\*[.07 in]
G_{y\,y} & = &v_2(x)\ w(x,y)
\label{rho_v}\end{aligned}$$
for some unknown functions $v_1(x)$ and $v_2(x)$. is then given by
$$\Upsilon = w(x,y)\ \frac{\l(v_1(x) + v_2(x)\ \y1\r)}{F_{y'}(x,\y1)}
\label{case_C}$$
for some function $w(x,y)$, which is made up of the factors of $\Upsilon$ depending on $y$ and not on 1. To determine $F_{y'}(x,\y1)$ up to a factor depending on $x$, we need to determine the ratio $v_1(x)/v_2(x)$. For this purpose, from we build a PDE for $G_y(x,y)$,
$$\label{case_C_pde}
G_{x\,y} = \frac {v_1(x)}{v_2(x)}\, G_{y,y}$$
The general solution of is
$$\label{calG}
G_y = \G\l(y + p(x)\r)$$
where $\G$ is an arbitrary function of its argument and for convenience we introduced
$$p'(x)\equiv v_1(x)/v_2(x)$$
We now determine $p'(x)$ as follows. Taking into account ,
$$\label{case_1_G_2}
v_2(x)\, w(x,y) = \G'(y+p(x))$$
By taking the ratio between this expression and its derivative w.r.t $y$ we obtain
$$\H(y+p(x))\equiv
\frac {\pa \ln(w)}{\pa y}
= \frac {\G''(y + p(x))}{\G'(y + p(x))}
\label{cal_H}$$
that is, a function of $y+p(x)$ only, which we can determine since we know $w(x,y)$. If $\H' \neq 0$, $p'(x)$ is given by
$$\label{p_prime}
p'(x) = \frac{\H_x}{\H_y}
=
\frac {w_{xy} w - w_x w_y } {w_{yy} w - {w_y}^2 }$$
In summary, the conditions for this case are
$$\Upsilon_y \neq 0,\ \ \
\frac{\partial }{\partial \y1} \l(\frac{\Upsilon_y}{\Upsilon}\r)= 0,\ \ \
\frac {\pa^2 }{\pa y \pa x} \ln(w) \neq 0,\ \ \
\frac {\pa^2 }{\pa y \pa y} \ln(w) \neq 0
\label{conditions_C}$$
and then, from , $\F(x,\y1)$ is given by
$$\F(x,\y1)= \frac{(p'+\y1)\ w}{\Upsilon}
\label{F_p_w_Upsilon}$$
where at this point $\Upsilon$, $w(x,y)$ and $p'(x)$ are all known. $\triangle$
Kamke’s ODE 136.
We begin by writing the ODE in explicit form as
$$\yt = \frac {h(\y1)}{x-y}
\label{k136}$$
This example is interesting since the standard search for point symmetries is made difficult by the presence of an arbitrary function of $\y1$. $\Upsilon$ () is determined as
$$\Upsilon = -\frac {h({\it \y1})}{(x-y)^2}$$
and $w(x,y)$ as
$$w(x,y) = \frac {1}{(x-y)^2}$$
Then $\H(y + p(x))$ () becomes
$$\H = \frac {2}{x-y}$$
and hence, from , $p'(x)$ is
$$p'(x) = -1$$
so from :
$$\F(x,\y1)= \frac{1-\y1} {h({\it \y1})}$$
: $\H=0$
We now discuss how to obtain $p'(x)$ when $\H'(y+p(x))=0$. We consider first the case in which $\H=0$. Then, $\G''=0$ and the condition for this case is
$$\Upsilon_{y}=0$$
Recalling , $G$ is given by
$$G(x,y) = C_1\ (y + p(x))^2 + C_2\ (y + p(x)) + g(x)$$
for some function $g(x)$ and some constants $C_1$, $C_2$. From , $\Phi(x,y,\y1)$ takes the form
$$\label{case_1_subcase}
\Phi(x,y,\y1) = - \frac {F_x(x,\y1) + g'(x) + (2 C_1\ (y + p(x)) + C_2)
(\y1+p'(x))}{F_{y'}(x,\y1)}$$
We now determine $p'(x)$ as follows. First, from the knowledge of $\Upsilon$ and $\Phi$ we build the two explicit expressions:
$$\label{Lambda1}
\Lambda \equiv \frac {1}{\Upsilon} = - \frac {F_{y'}} {2 C_1\ (\y1+p'(x))}$$
and
$$\label{Psi1}
\Psi \equiv \frac {\Phi(x,y,\y1)}{\Upsilon} - y =
\frac {F_x + g'(x)} {2 C_1\ (\y1+p'(x))} + p(x) + \frac{C_2}{2 C_1}$$
From and $\Lambda$ and $\Psi$ are related by:
$$\label{case_1_subcase_part2}
\frac {\pa}{\pa x} \medl((\y1+p'(x))\ \Lambda\medr) +
\frac {\pa}{\pa \y1} \medl((\y1+p'(x))\ \Psi\medr)
= p(x) + \frac {C_2}{2 C_1}$$
where the only unknowns are $p(x)$, $C_1$, and $C_2$. By differentiating the equation above w.r.t 1 and $x$ we obtain two equations where the only unknown is $p'(x)$:
$$\begin{aligned}
& \Lambda_{y'} p''(x) + (\Lambda_{x y'} + \Psi_{y' y'}) (\y1+p'(x)) +
\Lambda_x + 2 \Psi_{y'} = 0 &
\label{case_1_subcase_part3}
\\*[.11 in]
& \Lambda\ p'''(x) + (\Lambda_{x x} + \Psi_{y' x}) (\y1 + p'(x))
+ (\Lambda_x + \Psi_{y'}) p''(x) + \Psi_x = p'(x) &
\label{case_1_subcase_part4}\end{aligned}$$
from where we obtain $p'(x)$ by solving a linear algebraic equation built by eliminating $p''(x)$ between and [^8]. Also, as a shortcut, if $(\Lambda_{x y'} + \Psi_{y' y'})
/ \Lambda_{y'}$ depends on 1, then we can build a linear algebraic equation for $p'(x)$ by solving for $p''(x)$ in and differentiating w.r.t. 1. $\triangle$
If depends neither on $p'(x)$ nor on $p''(x)$ this scheme will not succeed. However, in that case the original ODE is actually linear and given by . To see this, we set to zero the coefficients of $p'(x)$ and $p''(x)$ in , obtaining:
$$\label{case_1_subcase_part5}
\Lambda_{y'} = \Lambda_{x y'} + \Psi_{y' y'} = \Lambda_x + 2
\Psi_{y'}=0$$
which implies that $\Lambda$ is a function of $x$ only, and then
$$\label{case_1_subcase_part6}
\Psi_{y' y'} = 0$$
If we now rewrite $F(x,\y1)$ as
$$\label{case_1_subcase_part7}
F(x,\y1) = Z(x,\y1) - g(x) - \Lambda (\y1 + p')^2\,C_1$$
and introduce this expression in , we obtain $Z_{y'} =
0$; similarly, using this result, , and we obtain $Z_x=0$. Hence, $Z$ [*is a constant*]{}, and taking into account and , the ODE which led us to this case is just a non-homogeneous linear ODE of the form
$$\label{case_1_subcase_part9}
(y+p)''+ (\Lambda' (y+p)' - 2 (y + p) - C_2/C_1) / 2 \Lambda=0$$
whose homogeneous part does not depend on $p(x)$:
$$\yt+\frac{\Lambda'(x)}{2{\Lambda(x)}}\,\y1-{\frac {y}{\Lambda(x)}}=0$$
and as mentioned, it is the same as .
Kamke’s ODE 66.
This ODE is given by
$$\yt = a\left (c+bx+y\right )\left ({\y1}^{2}+1\right )^{3/2}$$
Proceeding as in Case A, we determine $\Upsilon$, $w(x,y)$, and $\H(y+p(x))$ as
$$\Upsilon = a\left (\y1^{2}+1\right )^{3/2};\ \ \ \ \ \ \
w(x,y) = 1;\ \ \ \ \ \ \
\H = 0$$
From the last equation we realize that we are in Case D. We determine $\Lambda$ and $\Psi$ (Eqs. (\[Lambda1\]), (\[Psi1\])) as:
$$\begin{aligned}
\Lambda & = & \frac{1}{\left ({\y1}^{2}+1\right )^{3/2}\, a}
\nonumber\\*[.07 in]
\Psi & = & c + b\, x\end{aligned}$$
We then build for this ODE:
$$\frac{p''(x)}{\left ({\y1}^{2}+1\right )^{3/2}\, a} +
c + b\, x = p(x) + \frac {C_2}{2 C_1}$$
Differentiating w.r.t. 1 leads to :
$$-3 \frac{p''(x)\, \y1}{\left ({\y1}^{2}+1\right )^{5/2}\, a} = 0$$
from which it follows that $p''(x) = 0$. Using this in we obtain:
$$p'(x) = b$$
after which becomes
$$\F(x,\y1)= \frac{\y1+b} {a \left ({\y1}^{2}+1\right )^{3/2}}$$
: $\H' = 0$ and $\H\neq 0$
In this case $\H(y+p(x)) = \G''/\G'= C_1$, so $\G'$ is an exponential function of its argument $(y+p(x))$ and hence from
$$G(x,y) = C_2 e^{(y+p(x))C_1} + (y+p(x))C_3 + g(x)$$
for some constants $C_2$, $C_3$ and some function $g(x)$. In this case one of the conditions to be satisfied is
$$\Upsilon_{y}= \mbox{constant} \neq 0$$
and $\Phi(x,y,\y1)$ will be of the form
$$\label{case_1_subcase_0}
\Phi(x,y,\y1) = - \frac {F_x(x,\y1) + g'(x)
+ \l(C_2 C_1 e^{ (y + p(x))C_1} + C_3\r)
\l(\y1+p'(x)\r)}{F_{y'}(x,\y1)}$$
Taking advantage of the fact that we explicitly know $C_1$, we build a first expression for $p'$ by dividing $C_1 e^{y C_1}$ by $\Upsilon$:
$$\label{Lambda2}
\Lambda \equiv - \frac {F_{y'}} {C_2 e^{p(x)C_1}\ (\y1+p'(x))}$$
We obtain a second expression for $p'$ by multiplying $\Phi$ by $\Lambda$ and subtracting $C_1 e^{C_1 y}$
$$\label{Psi2}
\Psi \equiv \frac {1}{C_2 e^{p(x)C_1}} \l(\frac {F_x + g'(x)} {\y1+p'(x)}
+ C_3\r)$$
As in Case D, $\Lambda$ and $\Psi$ are related by
$${\frac {\pa}{\pa x} \medl((\y1+p'(x))\,\Lambda\medr)
+ \l(\y1+p'(x)\r)\, p'(x) \Lambda C_1
+ \frac {\pa}{\pa \y1} \medl((\y1+p'(x))\,\Psi\medr)}
=\frac {C_3}{C_2 e^{p(x)C_1}}
\label{case_1_subcase_0_part2}$$
where the only unknowns are $C_2$, $C_3$ and $p(x)$. Differentiating with respect to $\y1$ we have
$$\begin{array}{rcl}
\lefteqn{
\label{case_1_subcase_0_part3}
\medl(p''(x) + {p'(x)}^2 C_1\medr) \Lambda_{y'}
+ p'(x) \medl(\y1 \Lambda_{y'} C_1 + \Lambda C_1
+ \Lambda_{x y'} + \Psi_{y' y'}\medr)
}
\\*[.15in]
&\mbox{\hspace{5.5cm}} &
+ 2 \Psi_{y'} + \Lambda_x + \y1 \Lambda_{x y'}
+ \y1 \Psi_{y' y'} = 0
\nonumber
\end{array}$$
The problem now is that, due to the exponential on the RHS of , differently from Case D, we are not able to obtain a second expression for $p'(x)$ by differentiating w.r.t $x$. The alternative we have found can be summarized as follows. We first note that if $\Lambda_{y'}=0$, is already a linear algebraic equation[^9] for $p'$, so that we are only worried with the case $\Lambda_{y'} \neq 0$. With this in mind, we divide by $\Lambda_{y'}$ and, [*if*]{} the resulting expression depends on 1, we directly obtain a linear algebraic equation in $p'(x)$ by just differentiating w.r.t 1. $\triangle$
$$\yt=\frac {\y1\,\left (x\y1+1\right )\left (-2+{e^y}\right )}{\y1 x^2+\y1-1}$$
We determine $\Upsilon$, $w(x,y)$, and $\H(y+p(x))$ as
$$\Upsilon = \frac {\y1 (x \y1 + 1) e^y}{\y1 x^2 + \y1 - 1};\ \ \ \ \ \ \
w(x,y) = e^y;\ \ \ \ \ \ \
\H = 1$$
From the last equation we know that we are in Case E. We then determine $\Lambda$ and $\Psi$ as in Eqs. (\[Lambda2\]) and (\[Psi2\]):
$$\begin{aligned}
\Lambda & = & \frac {\y1 x^2 + \y1 - 1}{\y1 (x \y1 + 1)}
\nonumber\\*[.07 in]
\Psi & = & -2\end{aligned}$$
Now, we build :
$$\frac {1}{x \y1+1}
\left(\left (p''+ {p'}^{2}+ {\y1}^2 \frac{(x p'-1)}{x \y1+1}\right)
\left (x^2 +1-\frac {1}{\y1} \right)+2\,x p'-2 \right)
=\frac {C_3}{C_2\, e^p}$$
and, differentiating w.r.t. 1, we obtain :
$$\frac {2\,x \y1+1-(x^3+x) {\y1}^2}{ {\y1}^2 (x \y1+1)^2}\left (p''+ {p'}^2
\right )+\frac {2\,\y1-1-2\,x+x \y1}{\left (x \y1+1\right )^3}(x p'-1)=0$$
Proceeding as explained, dividing by $\Lambda_{y'}$ and differentiating w.r.t. 1, we have
$$\frac {\pa}{\pa \y1} \left({\y1}^2 \frac {2\,\y1-1-2\,x+x \y1}{\left (x
\y1+1\right ) (2\,x \y1+1-(x^3+x) {\y1}^2 )}\right)(x p'-1)=0$$
Solving for $p'(x)$ gives $p'(x)=1/x$, from which becomes:
$$\F(x,\y1)= \l(\y1-\frac{1}{x}\r)\frac {\y1 x^2 + \y1 - 1}{\y1 (x \y1 + 1)}$$
The final branch occurs when divided by $\Lambda_{y'}$ does not depend on $\y1$ (so that we will not be able to differentiate w.r.t 1). In this case we can build a linear algebraic equation for $p'(x)$ as follows. Let us introduce the label $\beta(x,p',p'')$ for divided by $\Lambda_{y'}$, so that becomes:
$$\label{Lambda_beta}
\Lambda_{y'}(x,\y1)\ \beta(x,p',p'') = 0$$
Since we obtained by differentiating with respect to 1, can be written in terms of $\beta$ by integrating with respect to 1:
$$\Lambda(x,\y1) \beta(x,p',p'') + \gamma(x,p',p'') = \frac {C_3} {C_2
e^{p(x) C_1}}
\label{Lambda_beta_gamma}$$
where $\gamma(x,p',p'')$ is the constant of integration, and can be determined explicitly in terms of $x$, $p'$ and $p''$ by comparing with . Taking into account that ${\beta}(x,p',p'')=0$, reduces to:
$$\label{gamma}
\gamma(x,p',p'') = \frac{C_3}{C_2 e^{p(x)C_1}}$$
We can remove the unknowns $C_2$ and $C_3$ after multiplying by $e^{p(x)C_1}$, differentiating with respect to $x$, and then dividing once again by $e^{p(x)C_1}$. We now have our second equation for $p'$, which we can build explicitly in terms of $p'$, since we know $\gamma(x,p',p'')$ and $C_1$:
$$\label{gamma_p}
\frac {d \gamma}{dx}+C_1\ p' \gamma=0$$
Eliminating the derivatives of $p'$ between and leads to a linear algebraic equation in $p'$. Once we have $p'$, the determination of $\F(x,\y1)$ follows directly from . $\triangle$
Integrating factors of the form $\mu(y,\y1)$
--------------------------------------------
From , the ODE family admitting an integrating factor of the form $\mu(y,\y1)$ is given by
$$\label{reducible2}
\yt=-{\frac {\y1}{\mu}}
\l(G_{{y}}+\frac{\partial}{\partial y}\int \!\mu\;{d\y1}\r)-{\frac
{G_{{x}}}{\mu}}$$
where $\mu(y,y')$ and $G(x,y)$ are arbitrary functions of their arguments. For this ODE family, it would be possible to develop an analysis and split the problem into cases as done in the previous section for the case $\mu(x,\y1)$. However, it is straightforward to notice that under the transformation $y(x) \rightarrow x,\ x \rightarrow y(x)$, transforms into an ODE of the form with integrating factor $\mu(x,\y1^{-1})/\y1^2$. It follows that an integrating factor for any member of the ODE family above can be found by merely changing variables in the given ODE and calculating the corresponding integrating factor of the form $\mu(x,y')$.
$$\yt -{\frac {{\y1}^{2}}{y}} + \sin(x)\,\y1\,y +\cos(x)\,{y}^{2} = 0
\label{gon}$$
Changing variables $y(x) \rightarrow x,\ x \rightarrow y(x)$ we obtain
$$\yt+{\frac {\y1}{x}}-\sin(y)\,{\y1}^{2}x-\cos(y)\,{x}^{2}{\y1}^{3}=0
\label{gon_y_x}$$
Using the algorithm outlined in the previous section, an integrating factor of the form $\mu(x,\y1)$ for is given by
$${\frac {1}{{\y1}^{2}x}}$$
from where an integrating factor of the form $\mu(y,\y1)$ for is $1/y$, leading to the first integral
$$\sin(x)\;y+{\frac {\y1}{y}}+C_1=0,$$
which is a first order ODE of Bernoulli type. The solution to then follows directly. This example is interesting since from [@gonzales] has no point symmetries.
Integrating factors and symmetries {#mu_and_X}
==================================
Besides the formulas for integrating factors of the form $\mu(x,y)$, the main result presented in this paper is a systematic algorithm for the determination of integrating factors of the form $\mu(x,\y1)$ and $\mu(y,\y1)$ [*without solving any auxiliary differential equations or performing differential Groebner basis calculations*]{}, and these last two facts constitute the relevant point. Nonetheless, it is interesting to briefly compare the standard integrating factor ($\mu$) and symmetry approaches, so as to have an insight of how complementary these methods can be in practice.
To start with, both methods tackle an [*n$^{th}$*]{} order ODE by looking for solutions to a linear [*n$^{th}$ order determining PDE*]{} in $n+1$ variables. Any given ODE has infinitely many integrating factors and symmetries. When many solutions to these [*determining*]{} PDEs are found, both approaches can, in principle, give a multiple reduction of order.
In the case of integrating factors there is one unknown function, while for symmetries there is a pair of infinitesimals to be found. But symmetries are defined up to an arbitrary function, so that we can always take one of these infinitesimals equal to zero[^10]; hence we are facing approaches of equivalent levels of difficulty and actually of equivalent solving power.
Also valid for both approaches is the fact that, unless some [*restrictions*]{} are introduced on the functional dependence of $\mu$ or the infinitesimals, there is no hope that the corresponding determining PDEs will be easier to solve than the original ODE. In the case of symmetries, it is usual to restrict the problem to ODEs having [*point symmetries*]{}, that is, to consider infinitesimals depending only on $x$ and $y$. The restriction to the integrating factors here discussed is similar: we considered $\mu$’s depending on only two variables.
At this point it can be seen that the two approaches are complementary: the determining PDEs for $\mu$ and for the symmetries are different[^11], so that even using identical restrictions on the functional dependence of $\mu$ and the infinitesimals, problems which may be untractable using one approach may be easy or even trivial using the other one.
As an example of this, consider Kamke’s ODE 6.37
$$\yt + 2\,y\, \y1 + f(x)\left (\y1+{y}^{2}\right ) - g(x) = 0
\label{k37}$$
For [*arbitrary*]{} $f(x)$ and $g(x)$, this ODE has an integrating factor depending only on $x$, easily determined using the algorithms presented. Now, for [*non-constant*]{} $f(x)$ and $g(x)$, this ODE has no point symmetries, that is, no infinitesimals of the form $[\xi(x,y),\
\eta(x,y)]$, except for the particular case in which $g(x)$ can be expressed in terms of $f(x)$ as in[^12]
$$g(x)=
\fr{\it f''}{4}\,
+ \fr{3\,f\,{\it f'}}{8}
+ \fr{f^3}{16}
-{\frac {C_2\,{\exp{\l(-3/2\,\displaystyle\int \!f(x){dx}\r)}}}
{4\,\left (2\,C_1+ \displaystyle
\int \!{\exp\l({-1/2\,\int \!f(x){dx}}\r)}\,{dx}\right )^{3}}}$$
Furthermore, this ODE does not have non-trivial symmetries of the form $[\xi(x,\y1),\ \eta(x,\y1)]$ either, and for symmetries of the form $[\xi(y,\y1),\ \eta(y,\y1)]$ the determining PDE does not split into a system.
Another ODE example of this type is found in a paper by [@gonzales] (1988):
$$\yt - {\frac {{\y1}^{2}}{y}} - g(x)\,p\,{y}^{p}\y1 - {\it g'}\,{y}^{p+1} = 0
\label{gon2}$$
In that work it is shown that for constant $p$, the ODE above only has point symmetries for very restricted forms of $g(x)$. For instance, is a particular case of the ODE above and has no point symmetries. On the other hand, [*for arbitrary*]{} $g(x)$, has an obvious integrating factor depending on only one variable: $1/y$, leading to a first integral of Bernoulli type:
$${\frac {\y1}{y}}-g(x){y}^{p}+C_1=0$$
so that the whole family is integrable by quadratures.
We note that and are respectively particular cases of the general reducible ODEs having integrating factors of the form $\mu(x)$:
$$\yt=-{\frac {\left (\mu_{{x}}+G_{{y}}\right )}{\mu(x)}}\y1-{\frac {G_{
{x}}}{\mu(x)}}$$
where $\mu(x)$ and $G(x,y)$ are arbitrary; and $\mu(y)$:
$$\yt=-{\frac {( \mu_{{y}}\,\y1 + {G_{{y}}}) }{\mu(y)}} \,{\y1}
-{\frac {G_{{x}}}{\mu(y)}}$$
In turn, these are very simple cases if compared with the general ODE families and , respectively having integrating factors of the forms $\mu(x,\y1)$ and $\mu(y,\y1)$, and which can be systematically reduced in order using the algorithms here presented.
It is then natural to conclude that the integrating factor and the symmetry approaches are useful for solving different types of ODEs, and can be viewed as equivalently powerful and general, and in practice complementary. Moreover, if for a given ODE, an integrating factor and a symmetry are known, in principle one can combine this information to build two first integrals and reduce the order by two at once (see for instance [@stephani]).
Tests
=====
After plugging the reducible-ODE scheme here presented into the ODEtools package [@odetools2], we tested the scheme and routines using Kamke’s non-linear 246 second order ODE examples[^13]. The purpose was to confirm the correctness of the returned results and to determine which of these ODEs have integrating factors of the form $\mu(x,y)$, $\mu(x,\y1)$ or $\mu(y,\y1)$. The test consisted of determining $\mu$ and testing the exactness condition .
In addition, we ran a comparison of performances in solving a subset of Kamke’s examples having integrating factors of the forms $\mu(x,\y1)$ or $\mu(y,\y1)$, using different computer algebra ODE-solvers (Maple, Mathematica, MuPAD and the Reduce package Convode). The idea was to situate the new scheme in the framework of a sample of relevant packages presently available.
To run the comparison of performances, the first step was to classify Kamke’s ODEs into: [*missing x*]{}, [*missing y*]{}, [*exact*]{} and [*reducible*]{}, where the latter refers to ODEs having integrating factors of the forms $\mu(x,\y1)$ or $\mu(y,\y1)$. ODEs missing variables were not included in the test since they can be seen as first order ODEs in disguised form, and as such they are not the main target of the algorithm being presented. The classification we obtained for these 246 ODEs is as follows
[|p[1.24 in]{}|p[4.8 in]{}|]{} Classification & ODE numbers as in Kamke’s book\
99 ODEs are missing $x$ or missing $y$ & 1, 2, 4, 7, 10, 12, 14, 17, 21, 22, 23, 24, 25, 26, 28, 30, 31, 32, 40, 42, 43, 45, 46, 47, 48, 49, 50, 54, 56, 60, 61, 62, 63, 64, 65, 67, 71, 72, 81, 89, 104, 107, 109, 110, 111, 113, 117, 118, 119, 120, 124, 125, 126, 127, 128, 130, 132, 137, 138, 140, 141, 143, 146, 150, 151, 153, 154, 155, 157, 158, 159, 160, 162, 163, 164, 165, 168, 188, 191, 192, 197, 200, 201, 202, 209, 210, 213, 214, 218, 220, 222, 223, 224, 232, 234, 236, 237, 243, 246\
13 are in exact form & 36, 42, 78, 107, 108, 109, 133, 169, 170, 178, 226, 231, 235\
40 ODEs are [*reducible*]{} with integrating factor $\mu(x,\y1)$ or $\mu(y,\y1)$ [*and*]{} missing $x$ or $y$& 1, 2, 4, 7, 10, 12, 14, 17, 40, 42, 50, 56, 64, 65, 81, 89, 104, 107, 109, 110, 111, 125, 126, 137, 138, 150, 154, 155, 157, 164, 168, 188, 191, 192, 209, 210, 214, 218, 220, 222, 236\
28 ODEs are [*reducible*]{} and not missing $x$ or $y$ & 36, 37, 51, 66, 78, 97, 108, 123, 133, 134, 135, 136, 166, 169, 173, 174, 175, 176, 178, 179, 193, 196, 203, 204, 206, 215, 226, 235\
For our purposes, the interesting subset is the one comprised of the 28 ODEs not already missing variables. The results we obtained using the aforementioned computer algebra ODE-solvers[^14] are summarized as follows[^15]:
[|p[1.2cm]{}|p[2.2cm]{}|p[2.6cm]{}|p[2.6cm]{}|p[4.7cm]{}|]{}\
& Convode & Mathematica 3.0 & MuPAD 1.3 & ODEtools\
Solved: & 51, 166, 173, 174, 175, 176, 179. & 78, 97, 108, 166, 169, 173, 174, 175, 176, 178, 179, 206. & 78, 97, 108, 133, 166, 169, 173, 174, 175, 176, 179, & 51, 78, 97, 108, 133, 134, 135, 136, 166, 169, 173, 174, 175, 176, 178, 179, 193, 196, 203, 204, 206, 215.\
& 7 & 12 & 11 &
22
\
Reduced: & & & & 36, 37, 66, 123, 226, 235.\
& 0 & 0 & 0 &
6
\
\
As shown above, while the scheme here presented is finding first integrals in all the 28 ODE examples, opening the way to solve 22 of them to the end, the next scores are only 12 and 11 ODEs, respectively solved by Mathematica 3.0 and MuPAD 1.3.
Concerning the six reductions of order returned by [**odsolve**]{}, it must be said that neither MuPAD nor Mathematica provide a way to convey them, so that perhaps their ODE-solvers are obtaining first integrals for these cases but the routines are giving up when they cannot solve the problem to the end.
Conclusions
===========
In connection with second order ODEs, this paper presented a systematic method for determining the existence of integrating factors and their explicit form, when they have the forms $\mu(x,y)$, $\mu(x,\y1)$ and $\mu(y,\y1)$. The scheme is new, as far as we know, and its implementation in the framework of the computer algebra package ODEtools has proven to be a valuable tool. Actually, the implementation of the scheme solves ODEs not solved by using standard or symmetry methods (see ) or some other relevant and popular computer algebra ODE-solvers (see ).
Furthermore, the algorithms presented involve only very simple operations and do not require solving auxiliary differential equations, except in one branch of the $\mu(x,y)$ problem. So, even for examples where other methods also work, for instance by solving the related PDE system Eqs.(\[PDE\_A\]) and (\[PDE\_B\]) using ansatzes and differential Groebner basis techniques, the method here presented can return answers faster and avoiding potential explosions of memory[^16].
On the other hand, we have restricted the problem to the universe of second order ODEs having integrating factors depending only on two variables while packages as CONLAW (in REDUCE) can try and in some cases solve the PDE system Eqs.(\[PDE\_A\]) and (\[PDE\_B\]) by using more varied ansatzes for $\mu$.
A natural extension of this work would be to develop a scheme for building integrating factors of restricted but more general forms, now for higher order ODEs. We are presently working on these possible extensions[^17], and expect to succeed in obtaining reportable results in the near future.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the State University of Rio de Janeiro (UERJ), Brazil, and by the Symbolic Computation Group, Faculty of Mathematics, University of Waterloo, Ontario, Canada. The authors would like to thank K. von Bülow[^18] for a careful reading of this paper.
Kamke, E., [*Differentialgleichungen: L[ö]{}sungsmethoden und L[ö]{}sungen*]{}. Chelsea Publishing Co, New York (1959).
S.C. Anco and G.W. Bluman, [*Integrating Factors and First Integrals for Ordinary Differential Equations*]{}, Preprint, University of British Columbia (1997).
T. Wolf, [*Computer algebra algorithms and routines for the computation of conservation laws and fixing of gauge in differential expressions*]{}, J. Symb. Comp. 27 (1999), pp 221-238.
Murphy, G.M., [*Ordinary Differential Equations and their solutions*]{}. Van Nostrand, Princeton, 1960.
Sheftel, M.B., [*A Course in Group Analysis of Differential Equations, Part II: Ordinary Differential Equations*]{}. St. Petersburg Institute of Economics and Finance, 1997. In Russian.
F. Boulier, Lazard, Ollivier and Petitot, [*Representation for the radical of a finitely generated differential ideal*]{}, Proceedings of ISSAC95, pp. 158-166
G.J. Reid and A.D. Wittkopf, [*Long Guide to the Standard Form Package*]{}, Pre-print, Department of Mathematics, University of British Columbia, Vancouver, Canada, 1993.
A. González-López, [*Symmetry and integrability by quadratures of ordinary differential equations*]{}, Phys. Lett. [**A**]{}, V. 133, N 4.5 (1998) 190.
Stephani, H. [*Differential equations: their solution using symmetries*]{}, ed. M.A.H. MacCallum, Cambridge University Press, New York and London (1989).
E.S. Cheb-Terrab, L.G.S. Duarte and L.A.C.P. da Mota, “Computer Algebra Solving of Second Order ODEs Using Symmetry Methods". Computer Physics Communications 108 (1998) 90
Appendix A {#appendix-a .unnumbered}
==========
We display here both the integrating factors obtained for the 28 Kamke’s ODEs used in the tests (see ) and the “case" corresponding to each ODE when using just the algorithms for $\mu(x,\y1)$ or $\mu(y,y')$[^19]. As explained in , the algorithm presented is subdivided into different cases: A, B, C, D, E and F, and case B is always either A or C.
[|p[1 in]{}|p[3 in]{}|p[.35 in]{}|]{} Integrating factor & Kamke’s book ODE-number & Case\
$1 $ & 36 & D\
${e^{\int \!f(x){dx}}} $ & 37 & A\
${{y'}}^{-1} $ & 51, 166, 169, 173, 175, 176, 179, 196, 203, 204, 206, 215 & C\
${\frac {b+{y'}}{\left (1+{{y'}}^{2}\right )^{3/2}}} $ & 66 & D\
$x $ & 78 & D\
${x}^{-1} $ & 97 & A\
$y $ & 108 & D\
${y}^{-1} $ & 123 & A\
${\frac {1+{y'}}{\left ({y'}-1\right ){y'}}} $ & 133 & C\
${\frac {{y'}-1}{\left (1+{y'}\right ){y'}}} $ & 134 & C\
${\frac {{y'}-1}{\left (1+{y'}\right )\left (1+{{y'}}^{2}\right )}} $ & 135 & C\
${\frac {{y'}-1}{h({y'})}} $ & 136 & C\
${\frac {x}{2\,x{y'}-1}} $ & 174 & C\
$\left (1+{y'}\right )^{-1} $ & 178 & C\
${\frac {1}{{y'}\,\left (1+2\,y{y'}\right )}} $ & 193 & C\
${y'} $ & 226 & A\
$h({y'}) $ & 235 & C\
\[lastpage\]
[^1]: We say that a second order ODE is in explicit form when it appears as $\yt-\Phi(x,y,\y1)=0$. Also, we exclude from the discussion the case of a linear ODE and an integrating factor of the form $\mu(x)$, already known to be the solution to the adjoint ODE.
[^2]: In a recent work by [@bluman1], the authors arrive at and departing from the adjoint linearized system corresponding to a given ODE; the possible splitting of into an overdetermined system for $\mu$ is also mentioned. However, in that work, $\yt$ of above appears replaced by $\Phi(x,y,\y1)$, and the authors discuss possible alternatives to tackle Eqs.(\[PDE\_AB\]) and (\[PDE\_B\]) instead of Eqs.(\[PDE\_A\]) and (\[PDE\_B\]).
[^3]: The result for presented in this subsection is also presented as lemma 3.8 in [@sheftel].
[^4]: We recall that if $\mu$ is an integrating factor leading to a first integral $\psi$, then the product $\mu F(\psi)$ - where $F$ is an arbitrary function - is also an integrating factor, which however does not lead to a first integral independent of $\psi$.
[^5]: When the given ODE is linear, is just the corresponding adjoint equation.
[^6]: ODEs [*missing y*]{} may also have integrating factors of the form $\mu(x,\y1)$. Such an ODE however can always be reduced to first order by a change of variables, so that the determination of a $\mu(x,\y1)$ for it is equivalent to solving a first order ODE problem - not the focus of this work.
[^7]: We are only interested in the case $\mu_{y'}=F_{y'y'}\neq 0$.
[^8]: From , $\Lambda \neq 0$, so that always depends on $p'''(x)$, and solving for $p''(x)$ and substituting twice into will lead to the desired equation for $p'(x)$. If depends on $p'(x)$ but not on $p''(x)$, then itself is already a linear algebraic equation for $p'(x)$.
[^9]: We can see this by assuming that $\Lambda_{y'}=0$ and that does not contain $p'$, and then arriving at a contradiction as follows. We first set the coefficients of $p'$ in to zero, arriving at
$$\label{case_1_subcase_0_part3_b}
\mbox{\hspace{3.5cm}}
0 = C_1\ \Lambda + \Psi_{y' y'} = 2 \Psi_{y'} + \Lambda_x
+ \Psi_{y' y'}
\y1
\mbox{\hspace{3.5cm}}
{(A)}$$
Eliminating $\Psi_{y' y'}$ gives $$2 \Psi_{y'} = C_1\ \Lambda\ \y1 - \Lambda_x$$
Differentiating the expression above w.r.t 1 and since $\Lambda_{y'}=0$, we have
$$2 \Psi_{y' y'} = C_1\ \Lambda$$
Finally, using Eq.(A), $0 = \Lambda$, contradicting $F_{y'} \neq 0$.
[^10]: Symmetries $[\xi(x,y,..y^{(n-1)}),\eta(x,y,..y^{(n-1)})]$ of an order ODE can always be rewritten as $[G, (G - \xi) y' + \eta]$, where $G(x,y,..y^{(n-1)})$ is an arbitrary function (for first order ODEs, $y'$ must replaced by the right-hand-side of the ODE). Choosing $G$ = 0 the symmetry acquires the form $[0,\bar{\eta}]$
[^11]: We are considering here ODEs of order greater than one.
[^12]: To determine $g(x)$ in terms of $f(x)$ we used the [*standard form*]{} Maple package by Reid and Wittkopf complemented with some basic calculations.
[^13]: Kamke’s ODEs 6.247 to 6.249 cannot be made explicit and are then excluded from the tests.
[^14]: Maple R4 is not present in the table since it is not solving any of these 28 ODEs. This situation is being resolved in the upcoming Maple R5, where the ODEtools routines are included in the Maple library, and the previous ODE-solver was replaced by [**odsolve**]{}. However, the scheme here presented was not ready when the development library was closed; the [*reducible*]{} scheme implemented in Maple R5 is able to determine, when they exist, integrating factors only of the form $\mu(\y1)$.
[^15]: Some of these 28 ODEs are given in Kamke in exact form and hence they can be easily reduced after performing a check for exactness; before running the tests all these ODEs were rewritten in explicit form by isolating $\yt$.
[^16]: Explosions of memory may happen when calculating all the integrability conditions involved at each step in the differential Groebner basis approach.
[^17]: See http://lie.uwaterloo.ca/odetools.html
[^18]: Symbolic Computation Group of the Theoretical Physics Department at UERJ - Brazil.
[^19]: We note that for non-linear ODEs these two algorithms work as well when $\mu_{y'} = 0$, but in practice these very simple examples are covered by the algorithm for $\mu(x,y)$ presented in .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate an analytically tractable toy model for thermally induced polymorphic dynamics of cooperatively rearranging biofilaments - like microtubules. The proposed 4 -block model, which can be seen as a coarse-grained approximation of the full polymorphic tube model, permits a complete analytical treatment of all thermodynamic properties including correlation functions and angular fourier mode distributions. Due to its mathematical tractability the model straightforwardly leads to some physical insights in recently discussed phenomena like the “length dependent persistence length”. We show that a polymorphic filament can disguise itself as a classical wormlike chain on small and on large scales and yet display distinct anomalous tell-tale features indicating an inner switching dynamics on intermediate length scales.'
author:
- 'Hervé Mohrbach$^{1,2}$ and Igor M. Kulić$^{2}$'
title: A Solvable Model for Polymorphic Dynamics of Biofilaments
---
Introduction
============
Biological filaments of the cytoskeleton are large macromolecules formed by a hierarchical self-assembly of smaller yet often highly complex protein subunits. The monomer complexity can allow the subunits to undergo rearrangements between several conformational states. Once inserted into a macromolecular lattice these individual subunits can start to interact giving rise to cooperative phenomena which can affect equilibrium and dynamical properties of the whole assembly in unexpected manners. It is the goal of this paper to explore in detail the statistical mechanics of this type of cooperatively switching supramolecular assemblies, whose paradigm example might be found in the stiffest cytoskeletal protein filaments of eukaryotic cells, the microtubules [@Alberts]. Microtubules are hollow nanotubes whose walls are formed by lateral self-association of parallel protofilaments that themselves are built by a head-to-tail polymerization of $\alpha\beta$-tubulin heterodimer protein subunits. This very complex architecture confers to microtubules their high stiffness as well as a number of unique static and dynamic properties. In [@MTKulic1], guided by experimental findings, we built the case for a novel model of microtubules, based on two hypotheses: the elementary tubulin dimer units can fluctuate between a curved and a straight configuration and can interact cooperatively along a protofilament axis. This implies that the ground state of microtubules is not, as usually accepted a straight Euler beam, but instead a fluctuating several micron sized cooperative super-helix. The resulting polymorphic dynamics of the microtubule lattice seems to quantitatively explain several experimental puzzles including anomalous scaling of dynamic fluctuations of grafted microtubules [@Taute], their apparent length-stiffness relation [@Pampaloni] and their remarkably curved-helical appearance in general [@Venier]. These results rely on phenomenological modelling where the cylindrical symmetry of the microtubule lattice is approximated by a continuous symmetry [@MTKulic1]. This approximation seems reasonable as the number of protofilaments is typically large, ranging from $9$ to $17$ for taxol stabilized in-vitro microtubules with a predominant $14$ protofilament structure [@Wade; @Chretien]. In vivo, microtubules most commonly appear with $13$ protofilaments[@Bouchet], although many exceptions exist depending of the cell type. The approach developed in [@MTKulic1] revealed the existence of a unique and unusual zero mode dynamics which has strong consequences : microtubules are permanently coherently reshaping -i.e. changing their reference ground state configuration- by thermal fluctuations.
Going beyond the continuous phenomenological approach of [@MTKulic1], we explore here other important aspects of the polymorphic microtubules theory by considering a simplified minimal solvable model. We will adopt a coarse-grained approach where the microtubule lattice is considered as made up of only $4$ blocks of protofilaments that can fluctuate between a curved and a straight configuration. Although the coarse-grained model to some extent loses the (quasi)continuous zero mode dynamics, it captures a number of important features of the full model [@MTKulic1] and exact analytic computations of relevant observables becomes accessible in this simpler case. In particular the tubulin-tubulin state correlation function, the persistence length, the thermodynamic stability of conformational states as well as the tangent angle spectrum can be computed rather elegantly. This simplified approach should give more detailed analytical insights into microtubule’s static properties and make connection with the previous phenomenological model’s parameters introduced in [@MTKulic1]. While it is not our prime goal here to compare the developed toy model with experiments we hope that some of the derived results (like the tangent angle power spectrum) can become useful guides for future experimental quantification of polymorphic filament fluctuations. For a more detailed experiment-theory comparison and deeper motivation of the model we refer the reader to the articles [@MTKulic1].
The 4-block polymorphic tube model
==================================
The microtubule lattice (see Fig. 1) is modelled as a continuum tube material made of a variable number of protofilaments with $R_{i}\approx7.5nm,$ $R_{o}\approx11.5nm$ - the inner and outer microtubule radii respectively. The protofilaments are twisted around the microtubule’s longitudinal axis with the corresponding internal twist $q_{0}$ - or equivalently the pitch $\lambda=2\pi
q_{0}^{-1}$ - being a lattice type dependent constant that takes typical discrete values $\lambda=+3.4\mu m,+25\mu m,-6\mu m$ for $12$, $13$ and $14$ protofilament microtubules respectively [@Wade; @Chretien][@Ray][@ChretienFuller].
To describe the tube’s geometry we introduce two reference frames (cf. Fig. 1A). One is the material frame with base vectors $(\vec{u}_{1},\vec{u}_{2},\vec{u}_{3})$ attached to each microtubule cross-section. The other is an external fixed laboratory frame with base vectors $(\vec{u}_{x},\vec{u}_{y},\vec{u}_{z})$ with respect to which all deformations are measured. Putting the filament along the $\vec{u}_{z}$ axis direction and considering small angular deflections we have $\vec{u}_{z}\approx\vec{u}_{3}$. In this case the two frames are simply related by a rotation transformation $\hat{R}(s)$ given by internal microtubule lattice twist $q_{0}$, such that $(\vec{u}_{x},\vec{u}_{y})=\hat{R}(s)(\vec{u}_{1},\vec{u}_{2})$ with $$\hat{R}(s)=\left(
\begin{array}
[c]{cc}\cos(q_{0}s+\varphi) & -\sin(q_{0}s+\varphi)\\
\sin(q_{0}s+\varphi) & \cos(q_{0}s+\varphi)
\end{array}
\right) \label{R}$$ and $s\in\left[ 0,L\right] $ the longitudinal position variable along the microtubule centerline. The arbitrary rotation angle $\varphi$ corresponds to the angular deviation between the two frames at $s=0.$ As each protofilament is built by a self association of $N$ discrete GDP-tubulin dimers of length $b\approx8nm$ it is natural to introduce a discrete variable $i$ such that all microtubule’s cross-section positions are written as $s=ib$ with $i=1..N.$
The coarse-grained block approximation consists of grouping neighboring protofilaments into $n=4$ effective blocks (cf. Fig. 1B). The fluctuations of the block-dimers between 2 states - a straight and a curved state with intrinsic curvature $\kappa_{FP}$ and with an energy difference $\Delta G>0$ favoring the curved state are modelled by a two state variable $\sigma_{i}^{k}=0,1$ where $i=1..N$ is the longitudinal position and $k=1..4$ the block’s index (cf. Fig. 1C).
To complete the description of the polymorphic tube model two order parameters at each microtubule cross-section can be introduced [@MTKulic1]. The first is the *vectorial polymorphic order parameter* $\vec{p}_{i}=\vec{u}_{1}(\sigma_{i}^{1}-\sigma_{i}^{2})+\vec{u}_{2}(\sigma_{i}^{4}-\sigma_{i}^{3}),$ a 2D vector attached to each cross-section (cf Fig. 1D) describing the asymmetry of the curved state distribution - a kind of “conformational polarization vector” of the block states. For instance the “all-straight” or “all-curved” protofilament state correspond both to the same value $\vec{p}=0$, cf. Fig. 1D (as the curved state distribution is isotropic in both special cases). A second (scalar) quantity $m_{i}=\sum_{k=1}^{4}\sigma_{i}^{k}$ counts the total number of blocks in the curved state at each cross-section $i$. The tubulin cooperativity *along* each protofilament (block) axis is modelled by an Ising type nearest-neighbor cooperative interaction with an interaction energy $J>0$ favoring longitudinal nearest neighbors to be in the same state. The interaction energy at cross-section $i$ reads: $$e_{inter,i}=-J\sum\nolimits_{k=1}^{4}\left( 2\sigma_{i}^{k}-1\right) \left(
2\sigma_{i+1}^{k}-1\right) \label{eJ}$$ The total elastic + polymorphic energy can then be written as [@MTKulic1]: $$E_{MT}=\sum\nolimits_{i=1}^{N}\left( e_{i}+e_{inter,i}\right) .\label{EMT}$$ where $e_{i}$ is the combined elastic energy and the energy resulting from the switching of tubulin dimers at i-th cross-section:$$e_{i}=\frac{Bb}{2}\left[ \left( \vec{\kappa}_{i}-\vec{\kappa}_{pol,i}\right) ^{2}+\frac{\kappa_{1}^{2}}{2}\left( \frac{\pi\gamma}{2}m_{i}-p_{i}^{2}\right) \right] \label{e}$$ with the elastic bending modulus $B=\frac{Y\pi}{4}\left( R_{o}^{4}-R_{i}^{4}\right) $ and $Y$ the Young modulus. The effective lattice curvature $\kappa_{1}=\frac{\left( R_{0}-R_{1}\right) ^{2}}{\pi\left( R_{o}^{2}+R_{i}^{2}\right) }\kappa_{FP}$ results directly from the preferred curvature of the individual protofilament $\kappa_{PF}$ [@MTKulic1]. Here $\vec{\kappa}$ is the microtubule centerline curvature vector and $\vec
{\kappa}_{pol}$ the polymorphic curvature vector which in the external coordinate frame $\left( \vec{u}_{x},\vec{u}_{y}\right) $ is $$\vec{\kappa}_{pol,i}=2^{-1/2}\kappa_{1}\hat{R}(ib)\vec{p}_{i}\label{kapppol}$$ In Eq. \[e\] we introduced an important dimensionless parameter $\gamma=\frac{\kappa_{PF}}{\kappa_{1}}-\frac{8\Delta G}{bB\kappa_{1}^{2}}$ which measures the competition between block switching and elastic energy and ultimately determines the microtubule shape. It effectively acts as an external field that biases the curved lattice state for $\gamma<0$ and favors the straight state for $\gamma>0$ (cf. below). For small deflections $\theta_{x/y}<<1$ around the $z$-axis, the unit vector tangent to the microtubule’s centerline is approximately given by $\vec{t}\approx(\theta
_{x},\theta_{y},1)$ in the laboratory frame $\left( \vec{u}_{x},\vec{u}_{y},\vec{u}_{z}\right) $ where $\vec{\theta}=\left( \theta_{x},\theta
_{y}\right) $ are the centerline deflection angles in x/y direction. The global centerline curvature vector $\vec{\kappa}=d\vec{t}/ds$ can then be approximated as $\overrightarrow{\kappa}\approx\left( \theta_{x}^{\prime
},\theta_{y}^{\prime},0\right) .$ Writing the total curvature as $\overrightarrow{\kappa}=\overrightarrow{\kappa}_{pol}+\overrightarrow
{\theta^{\prime}}_{el}$ with $\overrightarrow{\theta^{\prime}}_{el}$ the purely elastic contribution we readily see from Eqs. \[eJ\],\[EMT\] that the partition function decomposes into a product of two independent elastic and polymorphic contributions $Z=Z_{el}Z_{pol},$ with $Z_{el}=\int D\theta
\exp\left( -\frac{l_{B}}{2}\int\nolimits_{0}^{L}\theta_{el}^{\prime
2}ds\right) $ (and $l_{B}=B/k_{B}T$ the bending persistence length with $k_{B}T$ the thermal energy). Therefore our main goal reduces to the computation of $Z_{pol}=\sum_{\sigma}\exp(-\frac{1}{k_{B}T}E_{\sigma})$ with $E_{\sigma}$ the polymorphic contribution of $E_{MT}$. In the general case (with large number of blocks) this appears to be a formidable task, that however can be exactly performed in the case of $4$ blocks as we will see in the following.
Reduction to a 2 $\times$ 2-block model
---------------------------------------
We first note that, at each cross-section $p_{i}^{2}=\left( \sigma_{i}^{4}-\sigma_{i}^{3}\right) ^{2}+(\sigma_{i}^{2}-\sigma_{i}^{1})^{2}$ decomposes into a sum over mutually facing blocks i.e. only blocks 1-2 and 3-4 couple directly. Consequently the partition function can be factorized into a product of two simpler ones $Z_{pol}=Z_{12}Z_{34}$ with $Z_{kl}$ the partition function of the Ising variables $\sigma^{k}$ and $\sigma^{l}$ of the blocks that face each other, cf. Fig. 1D. Therefore from now on we may consider the $2$ block model, say the block pair $1-2$. The polymorphic order parameters are in this case $m_{i}=\sigma_{i}^{1}+\sigma_{i}^{2}$ and $\overrightarrow
{p}_{i}=\left( \sigma_{i}^{1}-\sigma_{i}^{2}\right) \vec{u}_{1}.$ By introducing new more convenient Ising variables $S_{i}^{k}=\pm1$ defined as $S_{i}^{k}=2\sigma_{i}^{k}-1$ the energy Eq. \[eJ\] takes the familiar form of a ladder type Ising model (cf. Fig. 1E). In terms of these variables, $Z_{12}=Z_{0}\widetilde{Z}$ with $Z_{0}=\exp(\frac{1}{k_{B}T}B\kappa_{1}^{2}L\left( \pi\gamma-1\right) /4)$ and $$\widetilde{Z}=\sum_{\left\{ S^{1},S^{2}\right\} =-1,1}\exp(-\widetilde
{E})\label{Ztilde}$$ where from Eqs. \[eJ\]-\[e\] we obtain $$\widetilde{E}= {\textstyle\sum_{i=1}^{N}}
\left[ \widetilde{B}S_{i}^{1}S_{i}^{2}+\widetilde{\gamma}\left( S_{i}^{1}+S_{i}^{2}\right) -\widetilde{J}\left( S_{i}^{1}S_{i+1}^{1}+S_{i}^{2}S_{i+1}^{2}\right) \right] \label{E12}$$ with $\widetilde{B}=\frac{B\kappa_{1}^{2}b}{4k_{B}T},$ $\widetilde{J}=\frac
{J}{k_{B}T}$ and $\widetilde{\gamma}=\frac{\pi}{2}\gamma\widetilde{B}.$ With the two spins - on either side of the ladder (Fig. 1E)- being interchangeable we have $\left\langle S_{i}^{1}\right\rangle =\left\langle S_{i}^{2}\right\rangle $ (and thus $\left\langle \sigma_{i}^{1}\right\rangle
=\left\langle \sigma_{i}^{2}\right\rangle $) for all $i$ as the energy is translationally invariant. From Eq. \[kapppol\], we readily see that $\left\langle \vec{\kappa}_{pol,i}\right\rangle =0.$ And yet this does not mean that the microtubule is on average in a straight configuration. Just on the contrary, the microtubule can form a three dimensional super helix [@MTKulic1] that is “coherently” reshaping as the curvature alternately exchanges from one block to the other with a non-vanishing value of $\left\langle \kappa_{pol}^{2}\right\rangle $. Therefore the order parameter that characterizes the typical curvature of the lattice is the mean “polarization” $P=\sqrt{N^{-1}\sum_{i}\left\langle \overrightarrow{p}_{i}^{2}\right\rangle }$ of the curved states (as $\left\langle \kappa_{pol}^{2}\right\rangle =\kappa_{1}^{2}P^{2}$). The other important quantity is $M=\frac{1}{N}\sum\nolimits_{i}\left\langle
\sigma_{i}^{k}\right\rangle =\left\langle
\sigma_{i}^{k}\right\rangle $ - the mean concentration of curved states, or in the Ising terminology the mean “magnetization” (up to a trivial additive constant) . Within the same terminology the parameter $\widetilde {\gamma}$ takes the role of a “magnetic field” that, according to its sign, favors one or the other possible spin orientation (curved or straight block) -but the same orientation for blocks on both sides of the ladder in Fig. 1E. The “ferromagnetic coupling constant ” $\widetilde{J}$ promotes the longitudinal parallel alignments of spins along a given block axis. The “anti-ferromagnetic coupling” $\widetilde{B}$ on the other hand favors $S_{i}^{1}$ and $S_{i}^{2}$ to be antiparallel - a tendency which competes with $\widetilde{\gamma}$ and the lattice is frustrated. For large $\left\vert
\widetilde{\gamma}\right\vert >>\widetilde{B}$, we therefore expect the alignment tendency to win on both sides of the ladder so that $\left\langle S_{i}^{1/2}\right\rangle \approx\pm1,$ ($+1$ if $\widetilde{\gamma}<0$ and $-1$ if $\widetilde{\gamma}>0$) and thus $M\approx1$ or $0.$ In this situation the microtubule is in a straight state with $P\approx0$, which is either completely unstressed with $M\approx$ $0$ (for positive $\widetilde{\gamma}$) or maximally prestrained state with $M\approx1$ (for negative $\widetilde {\gamma}$).
For $\widetilde{\gamma}=0,$ the lattice is not frustrated and on average when on one ladder-side $S_{i}^{1}=1$ then $S_{i}^{2}=-1$ on the other (and vice versa). Consequently $\left\langle S_{i}^{k}\right\rangle =0$ and $M=\left\langle \sigma_{i}^{k}\right\rangle =1/2.$ In this case blocks $1$ and $2$ at the cross-section $i$ fluctuate alternately between the curved and straight state and the tube bends alternately in the directions $\vec{u}_{1}$ and $-\vec{u}_{1}.$ The cooperative ferromagnetic interaction $\widetilde{J}$ implies a correlation between the spins along the contour and the formation of domains with size of the order of the spin-spin correlation length $\xi$ computed below. It is these whole domains that alternately bend the tube in the $\vec{u}_{1}$ and $-\vec{u}_{1}$ direction that lead to $\left\langle
\overrightarrow{p}_{i}\right\rangle =0$ but $P\neq0.$ In fact $P$ will take its largest value $P\approx1$ in this non frustrated case. One remarks that for $\xi$ much smaller that the internal twist wavelength $\lambda$ the typical domain looks like a circular arc, whereas for $\xi\gg\lambda,$ a typical coherent domain is a super-helix in the 3 dimensional space with pitch $\lambda$ as the direction of bending $\pm\vec{u}_{1}$ is slowly rotating in the external frame. For $L>>\xi,$ the microtubule is made up of a juxtaposition of uncorrelated fluctuating helices or of uncorrelated circular arcs that bend independently in the $\pm\vec{u}_{1}$ direction. Consequently we expect a long distance behavior similar to an usual worm like chain. This is true despite the fact that elastic contributions where neglected in this qualitative discussion. Indeed there are no elastic fluctuations here (as for usual semiflexible filaments): the polymorphic transition from curved to straight states of short uncorrelated segments only mimic elastic fluctuations. For $\widetilde{\gamma}$ small but non zero the lattice is slightly frustrated and thus less curved (smaller value of $P$) displaying nevertheless a qualitatively similar physical behavior.
2-block model thermodynamics
============================
The partition function Eq. \[Ztilde\], can be exactly computed via the transfer matrix method. For this purpose, we impose periodic boundary conditions on both sides of the ladder $S_{1}^{1/2}=S_{N+1}^{1/2}$. This permits us to write the partition function as $\widetilde{Z}=$Tr$(T^{N})$ with the symmetric $\left( 4\times4\right) $ matrix $T$ defined by $$\begin{aligned}
& \left\langle S_{i}^{1},S_{i}^{2}\right\vert T\left\vert S_{i+1}^{1},S_{i+1}^{2}\right\rangle =\\
& e^{-\frac{\widetilde{\gamma}}{2}(S_{i}^{1}+S_{i}^{2}+S_{i+1}^{1}+S_{i+1}^{2})-\frac{\widetilde{B}}{2}\left( S_{i}^{1}S_{i}^{2}+S_{i+1}^{1}S_{i+1}^{2}\right) +\widetilde{J}S_{i}^{1}S_{i+1}^{1}+\widetilde{J}S_{i}^{2}S_{i+1}^{2}}$$ with explicit elements given by $$T=\left(
\begin{array}
[c]{cccc}e^{2\widetilde{J}-2\widetilde{\gamma}-\widetilde{B}} & e^{-\widetilde{\gamma}}
& e^{-\widetilde{\gamma}} & e^{-2\widetilde{J}-\widetilde{B}}\\
e^{-\widetilde{\gamma}} & e^{2\widetilde{J}+\widetilde{B}} & e^{-2\widetilde
{J}+\widetilde{B}} & e^{\widetilde{\gamma}}\\
e^{-\widetilde{\gamma}} & e^{-2\widetilde{J}+\widetilde{B}} & e^{2\widetilde
{J}+\widetilde{B}} & e^{\widetilde{\gamma}}\\
e^{-2\widetilde{J}-\widetilde{B}} & e^{\widetilde{\gamma}} & e^{\widetilde
{\gamma}} & e^{2\widetilde{J}+2\widetilde{\gamma}-\widetilde{B}}\end{array}
\right)$$ Denoting $U$ the matrix diagonalizing $T$, such that $UTU^{-1}=\Lambda$ is a diagonal matrix with eigenvalues $\lambda_{r}$ with $r=1,4$ , the partition function can be written as $Z=Tr\left( \Lambda^{N}\right) =\sum
_{r=1,4}\lambda_{r}^{N}$. Denoting $\lambda_{1}$ the largest eigenvalue the free energy per lattice site $f=F/N=-k_{B}T/N\ln Z$ reduces in the thermodynamic limit of $N\rightarrow\infty$ to the expression $f=-k_{B}T\ln\lambda_{1}.$ From the free energy, all other thermodynamic quantities can be derived. In particular the curved state density $M=\frac{1}{2}-\frac{1}{4}\frac{\partial}{\partial\widetilde{\gamma}}\ln\lambda_{1}$ and the polarization $P=\sqrt{\frac{1}{2}+\frac{1}{2}\frac{\partial}{\partial
\widetilde{B}}\ln\lambda_{1}}.$
The explicit expression for $\lambda_{1}$ is rather cumbersome and we omit it here. Instead we provide the plot of $M$ and $P$ in terms of $\widetilde
{\gamma}$ for different values of $\widetilde{J}$ with $\widetilde{B}=1$ (in [@MTKulic1], $\widetilde{B}$ was indeed found to be of order unity) in Fig. 2. We first remark that $P(-\gamma)=P(\gamma)$, so the microtubule’s mean curvature is symmetric with respect to the sign of $\gamma.$ Further we observe that there always exists a range of $\widetilde{\gamma}$ where $P\ $is close to unity which is quite spread for small coupling $J\ll k_{B}T$ and that for $\left\vert \widetilde{\gamma}\right\vert \gg\widetilde{B},$ the polarization $P$ is larger for smaller values of $J.$ Although in this regime the total energy density is minimal for straight unstressed or prestrained states, these states have a small entropy. In comparison, curved tube states have a higher energy density in this regime but also a higher entropy, as blocks of small size $\xi$ fluctuate independently. Consequently these curved -helical- states can have a smaller free energy. With growing coupling $J$, the size of coherent blocks increases (as $\xi$ grows with $J$) and the energy contribution becomes dominant over entropy so $P$ gets lower with growing $J.$ From Fig. 2, we observe that for $J=1.5k_{B}T,$ the entropy is already negligible as for $\left\vert \widetilde{\gamma}\right\vert >\widetilde{B}$ the straight states with $P\approx0$ are selected. On the contrary for $\left\vert \widetilde{\gamma}\right\vert <\widetilde{B},$ the microtubule adopts a curved conformation with a quasi constant value $P\approx1$- corresponding to maximum lattice curvature which is independent of $\widetilde{\gamma}$. This plateau region shows that in order to observe a curved (or helical) conformation microtubule does not require a very precise fine tuning of $\widetilde{\gamma}$ - as long it is smaller than $\widetilde{B}$. This approximate $\widetilde{\gamma}$ independence allows us to limit ourselves to the analytically most elegant case $\widetilde{\gamma
}=0$.
Simple case $\widetilde{\gamma}=0$
----------------------------------
In this case all quantities can be computed in compact form. In particular the largest eigenvalue of the transfer matrix simplifies to $\lambda_{1}=2\cosh(\widetilde{B})\cosh(2\widetilde{J})+D$ with $D=(3+\cosh(2\widetilde
{B})+2\cosh(4\widetilde{J})\sinh^{2}(\widetilde{B}))^{1/2}.$ As already mentioned the curved state density is constant $M=1/2$ (independent of $\widetilde{B}$ and $\widetilde{J}$), whereas the total “polarization” is $P^{2}=1-2\left\langle \sigma_{i}^{1}\sigma_{i}^{2}\right\rangle =\frac{1}{2}+\frac{\sinh(\widetilde{B})\cosh(2\widetilde{J})}{D}$ which tends to $1$ for $\widetilde{J}\gtrsim1$. This means that $\left\langle \sigma_{i}^{1}\sigma_{i}^{2}\right\rangle \rightarrow0$ which can be understood as both states (on opposing ladder sides) at the same site $i$ fluctuate alternately between $0$ and $1.$ The coupling $\widetilde{J}$ which is a measure of the cooperativity determines over which distance the curved state is coherent. The cooperativity can be measured by the spin-spin correlation function $g_{ij}^{kl}=\left\langle \sigma_{i}^{k}\sigma_{j}^{l}\right\rangle
-\left\langle \sigma_{i}^{k}\right\rangle \left\langle \sigma_{j}^{l}\right\rangle =\left\langle \sigma_{i}^{k}\sigma_{j}^{l}\right\rangle
-1/4$ with $k,l=1,2$ at two different sites $i$ and $j$. By symmetry $g_{ij}^{11}=g_{ij}^{22}$ and $g_{ij}^{12}=g_{ij}^{21}.$ Introducing the two matrices $\underline{V}^{1}$ and $\underline{V}^{2}$ defined as $$\underline{V}^{1}=\left(
\begin{array}
[c]{cccc}1 & 0 & 0 & 0\\
0 & 1 & 0 & 0\\
0 & 0 & -1 & 0\\
0 & 0 & 0 & -1
\end{array}
\right) \text{ \ and }\underline{V}^{2}=\left(
\begin{array}
[c]{cccc}1 & 0 & 0 & 0\\
0 & -1 & 0 & 0\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & -1
\end{array}
\right)$$ the correlation functions of the variable $S^{k}$ with $k=1,2$ can be written as $\left\langle S_{i}^{k}S_{j}^{l}\right\rangle =\frac{1}{Z}$Tr$(T^{i-1}\underline{V}^{k}T^{j-i-1}\underline{V}^{l}T^{N-j+i+1})$ for $i<j.$ With the matrix $U$ diagonalizing $T$, i.e. $U^{-1}TU=\Lambda$ and the following transformed matrices $V^{k}=U^{-1}\underline{V}^{k}U$, we have $\left\langle
S_{i}^{k}S_{j}^{l}\right\rangle =\frac{1}{Z}$Tr$(\Lambda^{i-1}V^{k}\Lambda^{j-i}V^{l}\Lambda^{N-j+1}).$ In the limit $N\rightarrow\infty,$ this expression can be evaluated : $$\begin{aligned}
\left\langle S_{i}^{k}S_{j}^{l}\right\rangle & =V_{11}^{k}V_{11}^{l}+\left(
\frac{\lambda_{2}}{\lambda_{1}}\right) ^{j-i}V_{21}^{k}V_{12}^{l}\nonumber\\
& +\left( \frac{\lambda_{3}}{\lambda_{1}}\right) ^{j-i}V_{31}^{k}V_{13}^{l}+\left( \frac{\lambda_{4}}{\lambda_{1}}\right) ^{j-i}V_{41}^{k}V_{41}^{l} \label{SpinSpin}$$ where the eigenvalues are ordered $\lambda_{1}>..>\lambda_{4}.$ From Eq. \[SpinSpin\] all $g_{ij}^{kl}$ can be deduced. It turns out that the most interesting quantity connected to the spatial fluctuations of the microtubule -its persistence length (see next section) - is connected to $\Delta
g_{ij}\equiv g_{ij}^{11}-g_{ij}^{12}=\left\langle \sigma_{i}^{1}\sigma_{j}^{1}\right\rangle -\left\langle \sigma_{i}^{1}\sigma_{j}^{2}\right\rangle .$ An explicit computation of the matrix elements of $V^{k}$ shows that $V_{14}^{k}=V_{11}^{k}=0$ and $V_{31}^{1}V_{13}^{1}=V_{13}^{2}V_{31}^{1}$ as well as $V_{21}^{1}V_{12}^{1}=-V_{12}^{2}V_{21}^{1}=P^{2}$. This leads to $$\Delta g_{ij}=\frac{1}{2}P^{2}\exp(-\left\vert i-j\right\vert /\widetilde{\xi
}) \label{CC}$$ where the correlation length is $\widetilde{\xi}=\left[ \ln\left(
\frac{\lambda_{1}}{\lambda_{2}}\right) \right] ^{-1}$ with $\lambda
_{2}=2\exp(\widetilde{B})\sinh(2\widetilde{J}).$ With the result for $2$ blocks (ladder model, Fig 1E) we are now able to compute the persistence length for the $4$-block polymorphic tube (Fig 1B).
Angular persistence length
==========================
The persistence length $l_{p}$ is the length scale characterizing the filament’s resistance to thermally induced bending moments. For usual biofilaments like DNA or actin filaments $l_{p}=l_{B}$ is a material constant equal to the bending persistence length $l_{B}\equiv B/k_{B}T$. However for microtubules $l_{p}$ is known to be length dependent [@Pampaloni]. Here we compute the angular persistence length for the $4$-block model from the usual definition in terms of the tangent-tangent correlation function $\left\langle
\overrightarrow{t}(s^{\prime})\cdot\overrightarrow{t}(s)\right\rangle
=\exp(-\left\vert s^{\prime}-s\right\vert /l_{p})$ where $\overrightarrow
{t}(s)$ is the unit-tangent vector at the position $s$ of the microtubule centerline. From $\overrightarrow{t}(s)=(\theta_{x}(s),\theta_{y}(s),\sqrt{1-\theta_{x}^{2}(s)-\theta_{y}^{2}(s)})$ in the external frame, we deduce to quadratic order in $\theta$ that $\left\langle \overrightarrow
{t}(s^{\prime})\cdot\overrightarrow{t}(s)\right\rangle \approx1-\frac{1}{2}V(s^{\prime}-s)$ with the angular variance $V(s^{\prime}-s)=\left\langle
(\theta_{x}(s^{\prime})-\theta_{x}(s))^{2}\right\rangle +\left\langle
(\theta_{y}(s^{\prime})-\theta_{y}(s))^{2}\right\rangle .$ Therefore one can write $\left\langle \overrightarrow{t}(s^{\prime})\cdot\overrightarrow
{t}(s)\right\rangle \approx\exp(-\left\vert s^{\prime}-s\right\vert /l_{p})$ with a persistence length $l_{p}(s^{\prime}-s)=2\frac{\left\vert s^{\prime
}-s\right\vert }{V(s^{\prime}-s)}$ that will be manifestly distance dependent in our case. Now by writing $\overrightarrow{\theta}=\overrightarrow{\theta
}_{pol}+\overrightarrow{\theta}_{el}$ and from the independence of polymorphic and elastic contributions $\left\langle \overrightarrow{\theta}_{pol}\cdot\overrightarrow{\theta}_{el}\right\rangle =\left\langle \overrightarrow
{\theta}_{pol}\right\rangle \cdot\left\langle \overrightarrow{\theta}_{el}\right\rangle =0,$ we deduce that the total persistence length can be decomposed as $l_{p}=(\frac{1}{l_{B}}+\frac{1}{l_{pol}})^{-1}$ with a polymorphic persistence length $$l_{pol}(s^{\prime}-s)=2\frac{\left\vert s^{\prime}-s\right\vert }{V_{pol}(s^{\prime}-s)}\label{PL}$$ At this level, for the geometric description of the microtubule shape in space it is more convenient to use a continuum description and replace the discrete index $i$ by the continuous variable $s,$ so that $\overrightarrow{\theta
}_{pol,i}\rightarrow\overrightarrow{\theta}_{pol}(s)$ which is obtained from the integration over $s$ of the curvature Eq. \[kapppol\] $$\overrightarrow{\theta}_{pol}(s)=\overrightarrow{\theta}_{pol}(0)+\int_{0}^{s}\vec{\kappa}_{pol}(s)ds\label{tethapol}$$ Now we can compute the polymorphic variance from the correlation function $G_{pol,x}(s^{\prime},s)=\left\langle (\theta_{pol,x}(s^{\prime})-\theta
_{x}(0)(\theta_{pol,x}(s)-\theta_{x}(0))\right\rangle .$ Note that by symmetry $G_{pol,y}(s^{\prime},s)=G_{pol,x}(s^{\prime},s).$ From Eqs. \[tethapol\],\[CC\] we readily obtain for the full $4$-block model$$\begin{aligned}
G_{pol,x}(s^{\prime},s) & =\frac{\kappa_{1}^{2}P^{2}}{2}\int_{0}^{s^{\prime}}\int_{0}^{s}ds_{1}ds_{2}\cos\left( q_{0}(s_{2}-s_{1}\right) )\nonumber\\
& \exp\left( -\frac{\left\vert s_{2}-s_{1}\right\vert }{\widetilde{\xi}b}\right) .\label{Gpol}$$ where for a free microtubule (not angularly constrained at the ends) we have to integrate over the arbitrary initial angle $\varphi\in\left[
0,2\pi\right] .$ The variance can now be computed and we obtain : $$\begin{aligned}
V_{pol}(s^{\prime}-s) & =\frac{2\kappa_{1}^{2}P^{2}\xi}{1+\xi^{2}q_{0}^{2}}\left\{ \left\vert s^{\prime}-s\right\vert -\frac{1-\xi^{2}q_{0}^{2}}{1+\xi^{2}q_{0}^{2}}\xi\right. \nonumber\\
& +\frac{\xi e^{-\left\vert s^{\prime}-s\right\vert /\xi}}{1+\xi^{2}q_{0}^{2}}\left( \left( 1-q_{0}^{2}\xi^{2}\right) \cos\left( q_{0}(s^{\prime
}-s)\right) \right. \nonumber\\
& \left. \left. -2q_{0}\xi\sin\left( q_{0}\left\vert s^{\prime
}-s\right\vert \right) \right) \right\} \label{VARI}$$ with $\xi=\widetilde{\xi}b.$ From Eq. \[VARI\] the persistence length $l_{pol}$ and thus $l_{p}$ can be deduced. The polymorphic variance Eq. \[VARI\] has the following generic asymptotic behaviors : At large distance $s>>\xi,$ the variance $V_{pol}(s)\approx2\kappa_{1}^{2}P^{2}\frac{\xi}{1+\xi^{2}q_{0}^{2}}s$ scales linearly with $s$ indicating that $\overrightarrow{\theta}_{pol}$ is performing a simple (angular) random walk. At such a scale the microtubule looses its “coherent nature” and is replaced by a collection of uncorrelated segments. So not surprisingly we recover the classical results of a semiflexible chain again. In this asymptotic regime the effective persistence length reaches saturation with a renormalized constant value $l_{p}\left( \infty\right) =1/\left( l_{pol}^{-1}+l_{B}^{-1}\right)
$ where $$l_{pol}\left( \infty\right) =\frac{1+\xi^{2}q_{0}^{2}}{\xi\kappa_{1}^{2}P^{2}}\label{lpinfinit}$$ At short distance such that $s<<$min($\xi,q_{0}^{-1}$) the variance has a quadratic behavior $V_{pol}(s)\approx\kappa_{1}^{2}P^{2}s^{2}$. In this regime, the polymorphic fluctuations are completely dominated by purely “classical” semiflexible chain fluctuations and $l_{p}\left( 0\right)
=l_{B}.$ Starting from this value of $l_{p}$ at $s=0$, polymorphic fluctuations begin to contribute reducing the persistence length $l_{p}\left(
s\right) \approx(\frac{1}{l_{B}}+\frac{\kappa_{1}^{2}P^{2}s}{2})^{-1}$. This behavior is universal and independent of $\xi$ and $q_{0}^{-1}.$
In the intermediate regime with $s$ of the order of $\xi$, the behavior of $V_{pol}(s)$ depends on the value of $\xi q_{0}$ - a kind of “helix coherence” parameter. When $\xi q_{0}\leq1$ (low helix coherence) oscillations in Eq. \[VARI\] are damped and the persistence length $l_{p}\left( s\right) $ is monotonously deceasing from $l_{B}$ to the constant $l_{p}\left(
\infty\right) $ with $l_{pol}\left( \infty\right) \approx(\xi\kappa_{1}^{2}P^{2})^{-1}$ obtained from Eq. \[lpinfinit\] (cf. Fig.4). This corresponds also to the situation of no internal twist $q_{0}=0.$
In the most interesting (high helical coherence) regime $\xi q_{0}>1,$ the behavior of $V_{pol}(s)$ is a combination of two effects: an oscillation with wave length $\lambda=2\pi q_{0}^{-1}$ originating from the helical nature of the microtubule, but which is now damped by the presence of thermally induced defects (due to a finite $\xi$) reducing the coherence of the helix and enhancing the linearly growing behavior (the random walk). As a consequence for high helical coherence $l_{p}\left( s\right) $ displays three different regimes (cf Fig.4):
I. In the limit of very short , $l_{p}\left( s\right) \approx(\frac{1}{l_{B}}+\frac{\kappa_{1}^{2}P^{2}s}{2})^{-1}$ that attains a global minimum at $s_{\min}\approx\pi q_{0}^{-1}$
II\. For intermediate length values $s_{\min}<s\lesssim$ $\xi$, the total persistence length displays a non-monotonic oscillatory behavior of period $\lambda$ with damped amplitude around a nearly linearly growing average reflecting the polymorphic fluctuation of the helix.
III\. For distances $s>>\xi,$ the oscillations in Eq. \[VARI\] are completely damped as the helix forgets its “coherent nature” on these scales and reaches the limiting value $l_{p}^{\ast}\left( \infty\right) =1/\left( l_{pol}^{-1}+l_{B}^{-1}\right) $ with $l_{pol}\left( \infty\right) $ given by Eq. \[lpinfinit\].
High cooperativity limit and uniform states
===========================================
In the previous paragraphs we considered the thermodynamic limit $L=Nb\rightarrow\infty$ for which the correlation length $\xi$ is always smaller that the total length $L$ [@Footnote1]. It also interesting to consider the opposite regime $\xi>>L$. In this case of large cooperativity we may take formally the limit $J\rightarrow\infty$ . Consequently the two states variables $\sigma^{i}$ become uniform all along their respective block axes $i=1..4$. The decoupling of the full MT lattice into 2 independent $2\times2$ sub-blocks implies the equality of the energy and entropy of the two independent 2-sub-blocks sets so that we can focus on only a single sub-lattice made of two blocks, say $1$ and $2.$ The corresponding polymorphic energy is $E=\frac{B\kappa_{1}^{2}}{2}L\left[ (\frac{\pi}{2}\gamma-1)\left(
\sigma^{1}+\sigma^{2}\right) +2\sigma^{1}\sigma^{2}\right] $ and its partition function reads now $Z=1+2\exp\left( -2(\widetilde{\gamma
}-\widetilde{B})L/b\right) +\exp\left( -4\widetilde{\gamma}L/b\right) .$ For two blocks there are then only 4 possible states with different probabilities of realization $\rho(\sigma^{1},\sigma^{2})=\exp(-\beta E)/Z$ that are: $\rho(0,0)=1/Z$ for the straight unstrained state, $\rho
(1,1)=\exp\left( -4\widetilde{\gamma}L/b\right) /Z$ for the straight prestrained state and finally $\rho(0,1)=\rho(1,0)=\exp\left( -2(\widetilde
{\gamma}-\widetilde{B})L/b\right) /Z$ for the two polymorphic helical states. The average value of the spin and the correlation functions are given respectively by $\left\langle \sigma^{k}\right\rangle =\rho(1,0)+\rho(1,1)$ for $k=1..4$ and $\left\langle \sigma^{1}\sigma^{2}\right\rangle =\left\langle
\sigma^{3}\sigma^{4}\right\rangle =\rho(1,1)$.
Going back to the full lattice model made of four blocks the polymorphic order parameter is then given $P^{2}(L)=2(\left\langle \sigma^{1}\right\rangle
-\left\langle \sigma^{1}\sigma^{2}\right\rangle +\left\langle \sigma
^{3}\right\rangle -\left\langle \sigma^{3}\sigma^{4}\right\rangle
)=4\rho(1,0)$. The angular variance at position $s$ and $s^{\prime}$ can be easily computed and we find $V(s^{\prime}-s)=\frac{\kappa_{1}^{2}P^{2}(L)}{q_{0}^{2}}\left( 1-\cos\left( q_{0}(s^{\prime}-s\right) )\right) $ which obviously corresponds to the limit of infinite $\xi$ of Eq. \[VARI\] but with the additional factor $P^{2}(L)$ responsible for the length dependence of the variance. The weight of the different possible configurations depends on $\widetilde{\gamma}$ and the discrimination between them is more pronounced with growing length. This can be easily understood from the polymorphic entropy $\Sigma_{\sigma}$ which remarkably is non extensive. Note that $\Sigma_{\sigma}$ is always twice the entropy of the individual 2-block-sub-lattices. For the asymptotic limiting case of short $L\rightarrow0$ all $4\times4$ configurations of the full lattice have the same probability and therefore $\Sigma_{\sigma}(0)=2k_{B}\ln4.$ For larger $L,$ the entropy will continuously change to reach a constant value depending on the dominant configurations that are selected by $\gamma.$ For values of $\left\vert \widetilde{\gamma}\right\vert >\widetilde{B}$, we see on Fig. 5 that $P(L)$ goes to zero with growing values of $L$. In this regime there is only one configuration (the same for the two sublattices) $\rho(0,0)$ or $\rho(1,1)$ depending on the sign of $\widetilde{\gamma}$ and $\Sigma_{\sigma
}(L)\rightarrow0.$ Interestingly for $\widetilde{\gamma}=\pm\widetilde{B},$ there is an additional configuration for a given sub-lattice with the same probability $\rho(0,1)=\rho(1,1)=\rho(0,0)=1/Z,$ so that for the 4-blocks we have $3\times3$ configurations and $\Sigma_{\sigma}(L)\rightarrow2k_{B}\ln3.$ This coexistence leads to a state which is the superposition between straight and curved states with different probability such that $P^{2}(L)=\frac
{4}{3+\exp\left( -4\widetilde{B}L/b\right) }\approx\frac{4}{3}$ that is (quasi) length independent. When $\left\vert \widetilde{\gamma}\right\vert $ is very close to $\widetilde{B}$ , the evolution of $\Sigma_{\sigma}(L)$ versus $L,$ can be slow and therefore there is a regime of lengths where the the entropy is non-extensive and where the straight states have comparable free energy with the helical one. In this regime the microtubule will fluctuate between its almost degenerate straight and (two) curved states.
For $\left\vert \widetilde{\gamma}\right\vert <\widetilde{B},$ there are for a given sub-lattice two dominant degenerate configurations $(0,1)$ and $(1,0),$ so that the number of configuration for the full lattice model is $2\times2$ and $\Sigma_{\sigma}(L)\rightarrow2k_{B}\ln2$ with growing $L.$ In this regime the mean curvature is built up progressively with the length $L,$ to reach -faster for smaller value of $\left\vert \widetilde{\gamma}\right\vert $ - its maximum value $P=\sqrt{2}.$ Here also when $\left\vert \widetilde{\gamma
}\right\vert $ is very close to $\widetilde{B}$ the realization of the curved state can be slow with growing $L.$
Angular power spectrum
======================
Other information on the equilibrium properties of the microtubule can be deduced from the analysis of its Fourier mode distribution. To do so we express the microtubule shape $\overrightarrow{\theta}(s)$ as a superposition of Fourier modes $\overrightarrow{\theta}(s)=\overrightarrow{\theta}(0)+\sqrt{2/L}\sum_{n=1}^{+\infty}\overrightarrow{a}_{n}\cos(q_{n}s)+\sqrt{2/L}\sum_{n=1}^{+\infty}\overrightarrow{b}_{n}\sin(q_{n}s)+\left(
\theta_{x}(L)-\theta_{x}(0)\right) s/L$ with the wave vector $q_{n}=\frac{2n\pi}{L}$ and where we add the “ramp term” $\left( \theta
_{x}(L)-\theta_{x}(0)\right) s/L,$ as $\overrightarrow{\theta}(s)$ is not a periodic function in general, i.e., $\overrightarrow{\theta}(L)\neq
\overrightarrow{\theta}(0)$ [@Footnote2]. Decomposing the shape as $\overrightarrow{\theta}(s)=\overrightarrow{\theta}_{pol}(s)+\overrightarrow
{\theta}_{el}(s)$ we readily see that the equilibrium square of the amplitude of the elastic Fourier modes $G_{el}(q_{n})\equiv\left\langle \overrightarrow
{a}_{el,n}^{2}\right\rangle =\left\langle \overrightarrow{b}_{el,n}^{2}\right\rangle =\frac{2}{l_{B}q_{n}^{2}}$ has the usual thermally driven elastic bending spectrum of a wormlike chain. On the other hand from Eq. \[Gpol\] the polymorphic mode distribution $G_{pol}(q)$ in the limit $L$ large (with $q_{n}\rightarrow q$ continuous) can be deduced : $$G_{pol}(q)=\frac{2\kappa_{1}^{2}P^{2}\xi}{q^{2}}\cdot\frac{1+\xi^{2}(q^{2}+q_{0}^{2})}{1+\xi^{4}(q^{2}-q_{0}^{2})^{2}+2\xi^{2}(q^{2}+q_{0}^{2})}\label{Gp}$$ As shown in Fig. 6, when the correlation length $\xi$ is of order of the helix pitch $\lambda=2\pi/q_{0}$, the function $G_{pol}(q)$ departs very much from the wormlike chain bending spectrum $G_{el}(q)$ and displays a non-monotonous shape with a peak at $q_{0}$. This peak is more pronounced and sharper for larger value of $\xi q_{0}$ (helix more coherent) and disappears when $\xi
q_{0}\leq1$. This is the main Fourier mode signature of the -polymorphic fluctuating- helicoidal nature of the microtubule. For long wavelength modes $q<<q_{0},$ the spectrum has a semiflexible like behavior $G_{pol}(p)\approx\frac{2}{l_{pol}\left( \infty\right) q^{2}}$ reflecting the incoherence of the helix at this scale, with the persistence length given by Eq. \[lpinfinit\]. For short wavelength modes $q>>q_{0},$ the behavior becomes quarticaly decreasing $G_{pol}(p)\approx\frac{2\kappa_{1}^{2}}{\xi
q^{4}}$ and polymorphic fluctuations are strongly damped at short distance in agreement with high cooperativity at this scale.
Quantitative comparison with experiments at this point seems difficult, because Fourier analysis of microtubules in-vitro was mainly performed on microtubules strongly confined into an almost 2-D geometry. This experimental necessity (originating from optical microscopy sample flatness requirements) is expected to induce artifacts as in this case, within our helical model, the microtubules would likely not fully equilibrate between their different equivalent curved states as large barriers will appear under confinement. In agreement with this argument the experimental Fourier modes of microtubules systematically reveal some type of “frozen-in” curvature which is much larger than the fluctuations around it [@GITTES] - in particular for longer wavelength modes. This is in so far striking as no clear argument for large systematic built-in curvature in microtubules (other than the here described) is evident, indicating that this curvature is in fact the very character of the microtubule elasticity. Similarly in vivo, it was found that growing microtubules have large scale frozen-in curvature [@Brangwyne]. This phenomenon was interpreted in terms of a random interaction of the growing microtubule tip with the surrounding cytoskeleton. Instead it could be that the curvature has an intrinsic origin (like in the present model) but is not allowed to fully equilibrate due to constraints present in the cytoplasm.
Conclusion and Outlook
======================
We have investigated the thermodynamics of a toy model for cooperatively switching (polymorphic) biofilaments, whose paradigm example is believed to be the microtubule. The polymorphism of the microtubule’s subunits and the short range cooperative interaction leads to properties that are very different from usual biofilaments. In particular the ground state itself is not a single fixed shape but is found to consist of a set of degenerate helicoidal configurations. At finite temperature, the fluctuations of the subunits create defects that smoothen the perfect helix and can even destroy it on very long scale where the filament retrieves a typical worm-like chain behavior. Another peculiar characteristic of the polymorphic tube is the length dependence of the persistence length. The Fourier spectrum turns out also to be markedly different from classical biofilaments. While it has a typical semiflexible chain behavior at long scale the Fourier modes amplitude are enhanced for modes around the characteristic scale of the helix. At very short scales the modes are strongly dampened due to the strong cooperative dimer interaction. We believe that the Fourier spectrum will be of some value to discriminate between different models in the analysis of future experiments (as in [@Pampaloni]) where confinement is absent and microtubules can thus rearrange and equilibrate.
The model developed here was inspired and tailored to microtubules. However it can be adapted to other polymorphic filaments with similar cross-sectional symmetry (circular rod or tube). Polymorphic sheets, rods and tubes are not uncommon in nature with the best documented example the bacterial flagellum [@Flagellum]. In the latter case the polymorphic states are likely very rigid (or coherent in our terminology) on thermal energy scales with correlation lengths $\xi$ larger than filament’s length $L$. Rearrangements between equivalent lattice states could then face high barriers in the flagellum and suppress the here described polymorphic fluctuations on measurement time scales. However other filaments might be even better candidates for soft polymorphic tubes (or rods) and deserve a closer investigation. Observations of unusual flaring-up of peculiar helical modes in bacterial pili (cf. Fig. 1 in [@Pili]) as well as in actin filaments cooperatively interacting with drugs like cofilin (cf. Fig. 1c in [@ActinCofilin]) visually suggest the presence of soft polymorphic mode dynamics.
[99]{}
Alberts, B., A. Johnson, J. Lewis, M. Raff, K. Roberts, and P. Walter. 2005. Molecular Biology of the Cell.
Mohrbach, H., A. Johner, and I. M. Kulic. 2010. Tubulin bistability and polymorphic dynamics of microtubules. *Phys. Rev. Lett.* 105:268102; Cooperative Lattice Dynamics and Anomalous Fluctuations of Microtubules. arXiv:1108.4800 cond-mat.soft. 2011, Submitted.
Taute, K. M., F. Pampaloni2, E. Frey, and E-L. Florin, 2008. Microtubule dynamics depart from the wormlike chain model. *Phys. Rev. Lett.* 100, 028102.
Pampaloni, F., G. Lattanzi, A. Jonas, T. Surrey, E. Frey, and E. L. Florin. 2006. Thermal fluctuations of grafted microtubules provide evidence of a length-dependent persistence length. *Proc. Natl. Acad. USA*. 103:10248-10253.
Venier, P., A. C. Maggs, M. F. Carlier, and D. Pantaloni. 1994. Analysis of microtubule rigidity using hydrodynamic flow and thermal fluctuations. *J. Biol. Chem.* 269: 13353.
Wade, R. H., D. Chrétien, and D. Job. 1990. Characterization of microtubule protofilament numbers. How does the surface lattice accommodate?. *J. Mol. Bio*. 212:775-786.
Chretien, D., and R. H. Wade. 1991. New data on the microtubule surface lattice. *Bio. Cell. 71*:161-174.
Bouchet-Marquis, C., B. Zuber, A-M. Glynn, M. Eltsov, M. Grabenbauer, K. N. Goldie, D. Thomas, A. S. Frangakis, J. Dubochet, and D. Chrétien. 2007. Visualization of cell microtubules in their native state. *Bio. Cell*. 99:45-53.
Ray, S., E. Meyhofer, R. A. Milligan, and J. Howard. 1993. Kinesin follows the microtubule’s protofilament axis. *J. Cell Biol*. 121:1083-1093.
Chretien, D., and S. D. Fuller. 2000. Microtubules switch occasionally into unfavorable configurations during elongation. *J. Mol. Biol.* 298:663-676.
The periodic boundary conditions on the spin inherent to the tranfer matrix method implies that Eq. \[CC\] is valid for $L>>\xi.$ For $L$ of order $\xi,$ we need to consider free boundary conditions for which we have computed the spin correlations numerically for growing values of $N.$ A difference between numerics and the analytical - asymptotic $N\rightarrow
\infty$ expression Eq. \[CC\] comes from the correlation function that decreases suddently slightly faster than $\exp(-\left\vert i-j\right\vert
/\widetilde{\xi})$ for $i$ or $j$ near the ends. This boundary effect is rapidly negligeable and already for $N=100$ a very good correspondance between the numerics and \[CC\] was found.
We have tried the more conventional decomposition $\overrightarrow{\theta}(s)=a_{0}+\sqrt{2/L}\sum_{n=1}^{+\infty}\overrightarrow{a}_{n}\cos(\frac{n\pi}{L}s)$ as in [@GITTES] and, from it, computed the variance $V_{pol}(s)$ and found that is does not give Eq. \[VARI\]. Fourier decomposition around a non-trivial ground state is a complex issue with mathematical “underwater mines” that we will discuss elsewhere.
Gittes, F., B. Mickey, J. Nettleton, and J. Howard. 1993. Flexural rigidity of microtubules and actin filaments measured from thermal fluctuations in shape. *J. Cell Biol.* 120:923-934; Brangwynne, C., G. Koenderink, E. Barry, Z. Dogic, F. MacKintosh, and D. Weitz. 2007. Bending dynamics of fluctuating biopolymers probed by automated high-resolution filament tracking. *Biophys. J.* 93:346-359; Janson, M. E. and M. Dogterom. 2004. A bending mode analysis for growing microtubules: evidence for a velocity-dependent rigidity. *Biophys. J*. 87:2723-2736.
Brangwyne C., F. C. MacKintosh, D. A. Weitz. 2007. Force Fluctuations and Polymerization Dynamics of Intracellular Microtubules. *PNAS*. 104:16128-16133.
S. V. Srigiriraju & T. R. Powers Phys. Rev. Lett. 94, 248101 (2005); H. Wada and R. R. Netz, Europhys. Lett, 82, 28001 (2008)
J.M. Skerker and H.C.Berg, Direct observation of extension and retraction of type IV pili. PNAS 98, 126901 (2001)
B.R. McCullough, L.B., J-L M. , E. M. De La Cruz , Cofilin Increases the Bending Flexibility of Actin Filaments: Implications for Severing and Cell Mechanics, J. Mol. Biol. (2008) 381, 550– 558
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The first lecture provides an introduction to the physics of color superconductivity in cold dense quark matter. The main color superconducting phases are briefly described and their properties are listed. The second lecture covers recent developments in studies of color superconducting phases in neutral and $\beta$-equilibrated matter. The properties of gapless color superconducting phases are discussed.'
author:
- |
Igor A. Shovkovy[^1]\
Frankfurt Institute for Advanced Studies\
and\
Institut für Theoretische Physik,\
Johann Wolfgang Goethe-Universität,\
D-60054 Frankurt am Main, Germany
title: 'Two lectures on color superconductivity[^2]'
---
Introduction into color superconductivity
=========================================
Dense baryonic matter
---------------------
Almost all matter around us is made of several dozen chemical elements and their isotopes. Each element (atom) has a compact nucleus, made of protons and neutrons, and a cloud of electrons surrounding the nucleus. The proton and the neutron are baryons, i.e., “heavy” particles. They are about two thousand times heavier than the electron. (The actual values of the masses are $m_{\rm p} \approx 938.3\,\mbox{MeV}/c^2$, $m_{\rm n} \approx
939.6\,\mbox{MeV}/c^2$, and $m_{\rm e} \approx 0.511\,\mbox{MeV}/c^2$, see Ref. [@HDG].) Thus, the mass of an atom comes mostly from its heavy nucleus. A typical size of a nucleus is about five orders of magnitude smaller than a typical atomic size. It is clear, therefore, that the density of matter inside a nucleus is much larger than the density of matter made of chemical elements. In fact, typical densities inside nuclei are of order $10^{14}\,\mbox{g}/
\mbox{cm}^{3}$. To get an impression how large this is, it suffices to say that a tablespoon of a material with this density would weigh about a billion tonnes. Creating even a tiny macroscopic sample of such matter in laboratory is impossible. On the other hand, baryonic matter at even higher densities exists in the Universe in the interior of compact stars. Therefore, it is of fundamental importance to study such very dense matter.
Protons and neutrons, as well as all mesons and baryons that may appear in dense matter, are strongly interacting particles. Their properties and the properties of dense baryonic matter are, in principle, described by the microscopic theory of strong interactions, Quantum Chromodynamics (QCD). The only problem is that QCD is a rather complicated non-Abelian gauge theory. The Lagrangian density of the theory is written in terms of quark fields which carry color charges. Because of the property of confinement, however, quarks cannot exist in vacuum as free particles. Instead, they combine with other quarks and antiquarks to form color neutral hadrons (i.e., mesons and baryons) whose description is not easy in the framework of QCD.
It is known that baryons are not point-like particles. They have a typical size of about $1\,\mbox{fm}=10^{-13}\,\mbox{cm}$. In sufficiently dense matter, therefore, baryons can be forced to stay so close to one another that they would overlap. At such densities, constituent quarks are shared by neighboring baryons and, with increasing the density further, quarks should eventually become mobile over large distances. This means that quarks become deconfined. When this happens, it does not make sense to talk about hadronic matter any more. It is transformed into quark matter. The properties of such matter is the topic of these lectures.
It was suggested long time ago that quark matter may exist inside central regions of compact stars [@quark-star]. By making use of the property of asymptotic freedom in QCD [@GWP], it was argued that quarks interact weakly, and that realistic calculations taking full account of strong interactions are possible for sufficiently dense matter [@ColPer]. The argument of Ref. [@ColPer] consisted of the two main points: (i) the long-range QCD interactions are screened in dense medium causing no infrared problems, and (ii) at short distances, the interaction is weak enough to allow the use of the perturbation theory. As we shall see below, the real situation in dense quark matter is slightly more subtle.
### Weakly interacting quark matter
By assuming that very dense matter is indeed made of weakly interacting quarks, one could try to understand the thermodynamic properties of the corresponding ground state by first completely neglecting the interaction between quarks. In order to construct the ground state, it is important to keep in mind that quarks are fermions, i.e., particles with a half-integer spin, $s=1/2$. Therefore, quarks should obey the Pauli exclusion principle which says that no two identical quarks can simultaneously occupy the same quantum state.
In the ground state of zero temperature non-interacting quark matter, quarks occupy all available quantum states with the lowest possible energies. This is described formally by the following quark distribution function: $$f_F(\mathbf{k}) =\theta\left(\mu-E_\mathbf{k}\right), \quad \mbox{at}
\quad T=0,
\label{f_F_T=0}$$ where $\mu$ is the quark chemical potential, and $E_\mathbf{k}\equiv
\sqrt{k^2+m^2}$ is the energy of a free quark (with mass $m$) in the quantum state with the momentum $\mathbf{k}$ (by definition, $k\equiv
|\mathbf{k}|$). As one can see, $f_F(\mathbf{k})=1$ for the states with $k<k_F\equiv \sqrt{\mu^2-m^2}$, indicating that all states with the momenta less than the Fermi momentum $k_F$ are occupied. The states with the momenta greater than the Fermi momentum $k_F$ are empty, i.e., $f_F(\mathbf{k})=0$ for $k>k_F$.
In the chiral limit, i.e., when the quark mass is vanishing, the pressure of zero temperature free quark matter is given by [@Kapusta] $$P^{(0)} = 2 N_f N_c \int \frac{d^3\mathbf{k}}{(2\pi)^3}
\left(\mu-E_\mathbf{k}\right) \theta\left(\mu-E_\mathbf{k}\right)
-B =\frac{N_f N_c}{12 \pi^2} \mu^4 - B,
\label{pressure_0}$$ where the overall factor $2 N_f N_c$ counts the total number of degenerate quark states, namely 2 spin states (i.e., $s=\pm 1/2$), $N_f$ flavor states (e.g., up, down and strange), and $N_c$ color states (e.g., red, green and blue). The extra term $B$ in the expression for the pressure (\[pressure\_0\]), called the bag constant, was added by hand. This term effectively assigns a nonzero contribution to the vacuum pressure and, in this way, provides the simplest modelling of the quark confinement in QCD [@bag-model]. \[Note that the vacuum pressure is higher than the pressure of quark matter when $\mu<\left(4\pi^2 B/N_f\right)^{1/4}$.\] Here, we ignore the requirements of the charge neutrality and the $\beta$ equilibrium in quark matter. These will be addressed in detail in the second lecture. The energy density of quark matter reads $$\epsilon^{(0)} \equiv \mu \frac{\partial P^{(0)}}{\partial \mu}
- P^{(0)} =\frac{N_f N_c}{4 \pi^2} \mu^4 +B.
\label{e_0}$$ Thus, one arrives at the following equation of state: $$P^{(0)}=\frac{1}{3} \left(\epsilon^{(0)} -4B\right).
\label{EoS_0}$$ This equation of state could be easily generalized to the case of nonzero quark masses and/or to the case of non-equal chemical potentials for different quarks. Also, it could be further improved by adding the lowest order corrections due to the interaction. For example, to leading order in coupling, the interaction results in the following correction to the pressure [@FreMcL]: $$\delta P^{(0)} =-\frac{\alpha_s N_f (N_c^2-1)}{16 \pi^3} \mu^4
+O\left(\alpha_s^2\right),
\label{pressure_cor}$$ where $\alpha_s\equiv g^2/4\pi$ is the value of the running coupling constant of strong interactions defined at the scale of the quark chemical potential. The next to leading order corrections can also be calculated [@FreMcL].
It is generally believed that the equation of state in Eq. (\[pressure\_0\]), or its generalization, provides a good approximation for the description of weakly interacting quark matter. Below we argue, however, that some underlying assumptions regarding the ground state of such quark matter are not correct.
Cooper instability and color superconductivity
----------------------------------------------
It appears that the perturbative ground state of quark matter, characterized by the distribution function in Eq. (\[f\_F\_T=0\]), is unstable when there is an attractive (even arbitrarily weak in magnitude!) interaction between quarks. This is because of the famous Cooper instability [@Cooper] that develops as a result of the formation of Cooper pairs $\langle q_{\mathbf{k}}\,
q_{-\mathbf{k}}\rangle$ made of quarks from around the highly degenerate Fermi surface, i.e., quarks with the absolute value of momenta $k\simeq k_F$. Cooper pairs are bosons, and they occupy the same lowest energy quantum state at zero temperature, producing a Bose condensate. In the presence of such a condensate of Cooper pairs, the ground state of quark matter becomes a (color) superconductor. This is very similar to the ground state of the electron system in the Bardeen-Cooper-Schrieffer theory of low-temperature superconductivity [@BCS]. The only real difference comes from the fact that quarks, unlike electrons, come in various flavors (e.g., up, down and strange) and carry non-Abelian color charges. To emphasize this difference, superconductivity in quark matter is called [*color superconductivity*]{}. For recent reviews on color superconductivity, see Ref. [@reviews].
It was known for a long time that dense quark matter should be a color superconductor [@BarFra; @Bail]. In many studies, however, this fact was commonly ignored. Only recently, the potential importance of this phenomenon was appreciated. To large extent, this was triggered by the observation in Ref. [@cs] that the value of the color superconducting gap could be as large as $100\,\mbox{MeV}$ at baryon densities existing in the central regions of compact stars, i.e., at densities which are a few times larger than the normal nuclear density, $n_0\simeq 0.15 \mbox{~fm}^{-3}$. This very natural estimate for the value of the gap in QCD, in which a typical energy scale itself is $200\,\mbox{MeV}$, opened a wide range of new theoretical possibilities, and the subject bursted with numerous studies. The main reason is that the presence of such a large energy gap in the quark spectrum may allow to extract clear signatures of color superconducting states of matter in observational data from compact stars.
As in low-temperature superconductors, one of the main consequences of color superconductivity in dense quark matter is the appearance of a nonzero energy gap in the one-particle spectrum, $${\cal E}_{\mathbf{k}}
= \sqrt{\left(E_{\mathbf{k}}-\mu\right)^2+\Delta^2},
\label{E_qp}$$ where $\Delta$ is the gap. The presence of the gap in the energy spectrum should affect transport properties (e.g., conductivities and viscosities) of quark matter. Thus, if quark matter exists in the interior of compact stars, this will be reflected, for example, in the cooling rates and in the rotational slowing down of such stars. Also, a nonzero gap modifies thermodynamic properties, e.g., the specific heat and the equation of state. In application to stars, this could modify theoretical predictions for the mass-radius relations, or even suggest the existence of a new family of compact stars.
Color superconductivity may affect directly as well as indirectly many other observed properties of stars. In some cases, for example, superconductivity may be accompanied by the baryon superfluidity and/or the electromagnetic Meissner effect. If matter is superfluid, rotational vortices would be formed in the stellar core, and they would carry a portion of the angular momentum of the star. Because of the Meissner effect, the star interior could become threaded with magnetic flux tubes. In either case, the star evolution may be affected.
In general, it is of great phenomenological interest to perform a systematic study of all possible effects of color superconductivity in compact stars. Before this can be done, however, one needs to know the structure of the QCD phase diagram and properties of various color superconducting phases in detail. Despite the recent progress in the field, such knowledge still remains patchy. While many different phases of quark matter have been proposed, there is no certainty that all possibilities have already been exhausted. This is especially so when additional requirements of charge neutrality and $\beta$ equilibrium are imposed.
In the rest of this lecture, the main color superconducting phases, that involve different number of pairing quark flavors, will be reviewed briefly. The role of charge neutrality and $\beta$ equilibrium in quark matter will be discussed in the second lecture.
Two-flavor color superconductivity ($N_f=2$)
--------------------------------------------
It is instructive to start this lecture from the description of the simplest color superconducting phase, namely the two-flavor color superconductor (2SC). This is a color superconducting phase with the spin-0 Cooper pairing in quark matter made of up and down quarks.
As we discussed in the previous subsection, color superconductivity comes as a result of the Cooper instability driven by an attractive interaction between quarks. To study this instability in detail, one needs to specify the origin and the strength of the interaction. At very large densities, when quark matter is weakly interacting, the dominant interaction between quarks is provided by the one-gluon exchange. Then, the Cooper pairing dynamics, including the screening effects due to dense medium as well as the effects due to higher order corrections, can be studied from first principles within the framework of QCD. Some key details of the analysis in this regime are given below.
We consider quark fields $\psi_i^a$ in the fundamental representation of the SU(3)$_c$ color gauge group. These fields carry flavor ($i=1,2$) and color ($a=1,2,3$) indices. The QCD Lagrangian density reads $${\cal L}_{\rm QCD} = \bar\psi_i^a\left(i\gamma^\mu \partial_\mu
+\gamma^0\mu -m^{(0)}_i\right) \psi_i^a
+g A^{A}_{\mu} \bar\psi_i^a \gamma^\mu T^{A}_{ab} \psi_i^b
-\frac{1}{4}G^{A}_{\mu\nu}G^{A,\mu\nu},
\label{QCD-action}$$ where $A^{A}_{\mu}$ is the vector gauge field in the adjoint representation of SU(3)$_c$, $G^{A}_{\mu\nu}= \partial_\mu A^{A}_{\nu}
-\partial_\nu A^{A}_{\mu}+g f^{ABC}A^{B}_{\mu}A^{C}_{\nu}$ is the field strength, and the generators of color transformations are defined as $T^{A}_{ab} \equiv \frac{1}{2}\left(\lambda^{A} \right)_{ab}$ where $\lambda^{A}$ are the Gell-Mann matrices. The current quark masses and the quark chemical potential are denoted by $m^{(0)}_i$ and $\mu$, respectively. At high densities, one can neglect small current masses of quarks. Then, the QCD Lagrangian density (\[QCD-action\]) becomes invariant under SU(2)$_L\times$SU(2)$_R$ global chiral transformations.
In QCD, the constituent masses of quarks $m_i$ are generated dynamically, and they can be very different from the current masses $m^{(0)}_i$, appearing in the Lagrangian density. At zero density, for example, typical values of the constituent quark masses are of order $\frac{1}{3}
m_{\rm n}\approx 313\,\mbox{MeV}$ even in the chiral limit. At high densities, on the other hand, the masses of the up and down quarks become small. This is because they are proportional to the value of the chiral condensate $\langle \bar\psi_L \psi_R\rangle$ which melts in dense matter. Thus, it is often also justified to neglect the constituent quark masses in studies of color superconducting phases.
The interaction vertex in the Lagrangian density (\[QCD-action\]) has a non-trivial color structure given by the color generators $T^{A}_{ab}$. As a result, the quark-quark scattering amplitude in the one-gluon exchange approximation is proportional to the following color tensor: $$\sum_{A=1}^{N_c^2-1} T^{A}_{a a^\prime} T^{A}_{b^\prime b} =
-\frac{N_c+1}{4N_c}\left(\delta_{a a^\prime}\delta_{b^\prime b}
-\delta_{ab^\prime}\delta_{a^\prime b}\right)
+\frac{N_c-1}{4 N_c}\left(\delta_{a a^\prime}\delta_{b^\prime b}
+\delta_{a b^\prime}\delta_{a^\prime b}\right).
\label{T-T}$$ The first antisymmetric term corresponds to the [*attractive*]{} antitriplet channel, while the second symmetric term corresponds to the repulsive sextet channel, see Fig. \[fig-1-glu\]. It is the first antisymmetric antitriplet channel that plays the crucial role in Cooper pairing.
![The diagrammatic representation of the one-gluon exchange interaction between two quarks in QCD. The color structure of the corresponding amplitude contains an antisymmetric antitriplet and a symmetric sextet channel.[]{data-label="fig-1-glu"}](one-gluon1.eps){width="40.00000%"}
Here, it is appropriate to note that quark matter is unlikely to be truly weakly interacting at moderate densities existing in the central regions of compact stars, i.e., $n{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}10n_0$. In this case, the use of the microscopic theory of strong interactions is very limited, and one has to rely on various effective models of QCD. A very simple type of such a model is the Nambu-Jona-Lasinio (NJL) model with a local four-fermion interaction [@NJL]. Originally, this model was introduced for the description of nucleons with dynamically generated masses. Nowadays, it is also commonly used for the description of quarks. For example, the NJL type models were used in Refs. [@cs; @4fermi; @4fermi-rev; @huang_2sc] for the description of color superconducting phases. One of the simplest NJL models, that respects the SU(2)$_L\times$SU(2)$_R$ global chiral symmetry (in the limit $m^{(0)}_i\to 0$), is defined by the following Lagrangian density [@SKP]: $$\begin{aligned}
{\cal L}_{\rm NJL} &=& \bar\psi_i^a\left(i\gamma^\mu \partial_\mu
+\gamma^0\mu -m^{(0)}_i\right) \psi_i^a
+ G_S\left[(\bar\psi \psi)^2
+ (i\bar\psi \gamma_5\vec{\tau}\psi)^2\right]
\nonumber\\
&+&G_D (i \bar{\psi}^C \varepsilon \epsilon^a \gamma_5 \psi )
(i \bar{\psi} \varepsilon \epsilon^a \gamma_5 \psi^C),
\label{NJL-action}\end{aligned}$$ where $\psi^C=C \bar{\psi}^T$ is the charge-conjugate spinor and $C=i\gamma^2\gamma^0$ is the charge conjugation matrix. The matrix $C$ is defined so that $C\gamma_{\mu} C^{-1}=-\gamma_{\mu}^{T}$. Regarding the other notation, $\vec{\tau}=(\tau^1,\tau^2,\tau^3)$ are the Pauli matrices in the flavor space, while $(\varepsilon)^{ik} \equiv
\varepsilon^{ik}$ and $(\epsilon^a)^{bc} \equiv \epsilon^{abc}$ are the antisymmetric tensors in the flavor and in the color spaces, respectively. The dimensionful coupling constant $G_S=5.01\,\mbox{GeV}^{-2}$ and the momentum integration cutoff parameter $\Lambda=0.65\,\mbox{GeV}$ (which appears only in loop calculations) are adjusted so that the values of the pion decay constant and the value of the chiral condensate take their standard values in vacuum QCD: $F_\pi = 93$ MeV and $\langle\bar{u}u\rangle=\langle\bar{d}d\rangle =(-250\,\mbox{MeV})^3$ [@SKP]. The strength of the coupling constant $G_D$ is taken to be proportional to the value of $G_S$ as follows: $G_D=\eta G_S$ where $\eta$ is a dimensionless parameter of order 1. Note that the value of $\eta$ is positive which corresponds to having the antisymmetric diquark channel [*attractive*]{}. This is motivated by the microscopic QCD interaction, as well as by the instanton motivated models [@cs].
### Color and flavor structure of the condensate
Now let us discuss the color and flavor structure of the condensate of Cooper pairs in the 2SC ground state. This can be determined from the nature of the diquark interaction and the Pauli exclusion principle.
The color structure of the one-gluon exchange interaction in Eq. (\[T-T\]) shows that the color antisymmetric antitriplet channel is attractive. The antisymmetric color channel is also attractive in the instanton motivated models of Ref. [@cs]. This is reflected in the color structure of the condensate of spin-0 Cooper pairs, $$\left\langle \left(\bar{\psi}^C\right)_i^a \gamma^5 \psi_j^b
\right\rangle
\sim \varepsilon_{ij} \epsilon^{abc} ,
\label{2sc-cond}$$ which is antisymmetric in the color indices of the constituent quarks. As one can check (not shown explicitly), this spin-0 condensate is also antisymmetric in Dirac indices. Then, in accordance with the Pauli exclusion principle, it should be antisymmetric in flavor indices as well. The presence of the $\gamma^5$ matrix on the left hand side of Eq. (\[2sc-cond\]) makes the condensate a scalar. A similar condensate without $\gamma^5$ is a pseudoscalar which is disfavored in QCD because of the instanton effects.
The color antitriplet condensate (\[2sc-cond\]) may have an arbitrary orientation in the color space. By making use of the global color transformations, this orientation can be changed as convenient. It is conventional to choose the condensate to point in the third (blue) color direction, $\left\langle
\left(\bar{\psi}^C\right)_i^a \gamma^5 \psi_j^b \right\rangle
\sim \varepsilon_{ij} \epsilon^{ab3}$. In this case, the Cooper pairs in the 2SC phase are made of the red and green quarks only, and the blue quarks do not participate in the pairing. These unpaired blue quarks give rise to gapless quasiparticles in the low-energy effective theory.
The flavor antisymmetric structure in Eq. (\[2sc-cond\]) corresponds to a singlet representation of the global SU(2)$_L\times$SU(2)$_R$ chiral group. This means that the chiral symmetry is not broken in the 2SC ground state. In fact, there are no other global continuous symmetries which are broken in the 2SC phase. There exist, however, several approximate symmetries which are broken. One of them is the approximate U(1)$_A$ symmetry which is a good symmetry at high density when the instantons are screened [@scr_inst]. Its breaking in the 2SC phase results in a pseudo-Nambu-Goldstone boson [@low-e-2sc]. Additional four exotic pseudo-Nambu-Goldstone states may appear as a result of a less obvious approximate axial color symmetry discussed in Ref. [@pseudoNG].
In the 2SC ground state, the vector-like SU(3)$_c$ color gauge group of QCD is broken down to the SU(2)$_c$ subgroup. Therefore, five out of total eight gluons of SU(3)$_c$ become massive due to the Anderson-Higgs mechanism. The other three gluons, which correspond to the unbroken SU(2)$_c$, do not interact with the gapless blue quasiparticles. They give rise to a pure SU(2)$_c$ gluodynamics. The red and green quasiparticles decouple from this low-energy SU(2)$_c$ gluodynamics because they are gapped [@RSS].
### Quark propagator in the Nambu-Gorkov basis
To simplify the study of diquark condensates in dense quark matter, instead of using the Dirac spinors for quark fields, it is more convenient to introduce the following eight-component Nambu-Gorkov spinors: $$\Psi \equiv \left(\begin{array}{l}\psi \\ \psi^C \end{array}\right),$$ where $\psi^C$ is the charge-conjugate spinor, defined earlier. The structure of the inverse quark propagator in this basis reads $$S_0^{-1} = \left(
\begin{array}{cc}
\left[ G_0^+ \right]^{-1} & 0 \\
0 & \left[ G_0^- \right]^{-1}
\end{array}
\right),$$ where $$\left[ G_0^\pm \right]^{-1} = \gamma^\mu K_\mu \pm \mu \gamma_0$$ are the inverse Dirac propagators for massless quarks ($G_0^{+}$) and charge-conjugate quarks ($G_0^{-}$).
In the ground state with Cooper pairing, see Eq. (\[2sc-cond\]), the quark propagator in the Nambu-Gorkov basis also has nonzero off-diagonal components, $$S^{-1} = \left(
\begin{array}{cc}
\left[ G_0^+ \right]^{-1} & \Delta^{-} \\
\Delta^{+} & \left[ G_0^- \right]^{-1}
\end{array}
\right),
\label{inv_prop}$$ where $\Delta^{-} = -i \epsilon^{3}\varepsilon\gamma^5 \Delta$ and $\Delta^{+} \equiv \gamma^0 \left(\Delta^{-}\right)^{\dagger} \gamma^0
= -i \epsilon^{3}\varepsilon\gamma^5\Delta^{*}$ are the matrices in the Dirac space, and $\Delta$ is the diquark gap parameter. This is the simplest ansatz for the quark propagator in the color superconducting ground state. It can be made more general, for example, by adding a regular (diagonal) part of the quark self-energy. Such a self-energy plays an essential role in the dynamics of chiral symmetry breaking. In dense quark matter, on the other hand, it is not so important. Thus, for simplicity, it is neglected here (the effects of the diagonal part of the self-energy were studied in Ref. [@self-e]).
From the inverse propagator in Eq. (\[inv\_prop\]), one can derive the following expression for the quark propagator: $$S = \left(\begin{array}{cc}
G^{+} & \Xi^{-} \\
\Xi^{+} & G^{-}
\end{array}\right),
\label{quarkpropagator}$$ where $$\begin{aligned}
G^{\pm} &=&
\left[\left(G_0^{\pm}\right)^{-1}
-\Delta^{\mp}G_0^{\mp}\Delta^{\pm}\right]^{-1}, \\
\Xi^{\pm} &=& -G_0^{\mp} \Delta^{\pm} G^{\pm} .\end{aligned}$$ This structure of the quark propagator carries basic information about the properties of the 2SC ground state. For instance, the location of its poles in the energy plane gives the quasiparticle dispersion relations.
The only quantity that remains unknown in the ansatz (\[inv\_prop\]) is the value of the gap parameter $\Delta$. This should be determined from a self-consistent gap equation that takes into account the pairing dynamics between quarks. This is discussed in more detail in the next three subsections.
### One-gluon exchange interaction
In dense QCD, in which quarks are weakly interacting, the one-gluon exchange is the dominant interaction between quarks. It appears, however, that the one-gluon interaction is partially screened by surrounding dense medium. Thus, before studying the pairing dynamics between quarks, we first address the modification of gluon properties due to dense medium.
The inverse propagator of the medium modified gluon can be written in the following form: $$\left({\cal D}^{-1}\right)_{\mu\nu}^{AB} =
\left(D_0^{-1}\right)_{\mu\nu}^{AB}
+\Pi_{\mu\nu}^{AB},
\label{glu_prop}$$ where $\Pi_{\mu\nu}^{AB}$ is the gluon self-energy, or the polarization tensor. The gluons with soft momenta, $p{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}g_s\mu$, play the key role in Cooper pairing of quarks. The main contribution to the corresponding soft gluon polarization tensor comes from the quark loop with hard internal momenta of order $\mu$. This contribution is large compared to all other contributions in QCD (e.g., those from the ghost and the gluon loops) because it is proportional to the density of quark states at the Fermi surface (as well as to the coupling constant $\alpha_s$), i.e., $\Pi_{\mu\nu}^{AB}\sim
\alpha_s\mu^2$. The corresponding approximation, in which the sub-leading terms suppressed by powers of the quark chemical potential $\mu$ are neglected, is called the hard dense loop (HDL) approximation.
The calculation of the polarization tensor in the HDL approximation was performed in Refs. [@Vija; @Manuel] (in the case of dense QCD in $2+1$ dimensions, similar calculations were performed in Ref. [@qcd2+1]). Here, we present the final result and briefly comment on its main features.
The HDL polarization tensor can be written as follows: $\Pi_{\mu\nu}^{AB}\equiv \delta^{AB}\Pi_{\mu\nu}$, where the explicit form of the components of $\Pi_{\mu\nu}$ are given by $$\begin{aligned}
\Pi^{00}(p_0, \mathbf{p}) & = & \Pi_{l}(p_0, \mathbf{p}) ,
\label{Pi-00}\\
\Pi^{0i}(p_0, \mathbf{p}) & = &
p_0 \frac{p^i}{p^2} \Pi_{l} (p_0, \mathbf{p}) ,
\label{Pi-0i}\\
\Pi^{ij}(p_0, \mathbf{p}) & = & \left
( \delta^{ij}- \frac{p^i p^j}{p^2} \right)
\Pi_{t} (p_0,\mathbf{p})+ \frac{p^i p^j} {p^2}
\frac{p_0^2}{p^2} \Pi_{l} (p_0, \mathbf{p}),
\label{Pi-ij}\end{aligned}$$ with $$\begin{aligned}
\Pi_{l}(p_0,\mathbf{p})&=&m_D^2\left(\frac{p_0}{2p}
\ln\left|\frac{p_0+p}{p_0-p}\right|-1
-i\pi\frac{p_0}{2p}\theta(-p^2) \right),\\
\Pi_{t}(p_0,\mathbf{p})&=&\frac{1}{2}m_D^2-\frac{P^2}{2p^2}
\Pi_{l}(p_0,\mathbf{p}).\end{aligned}$$ Here, we use the notation $p^2\equiv \mathbf{p}^2$, $P^2=p_0^2-p^2$, and $m_D^2\equiv 2N_f\alpha_{s}\mu^2/\pi$. It is remarkable that this result coincides with the polarization tensor derived in the framework of the [*classical*]{} transport theory of dense Yang-Mills plasma [@Heinz]. One can check that the HDL polarization tensor is transverse, i.e., $$P^{\mu} \Pi_{\mu\nu}(p_0, \mathbf{p})=0.$$ It is convenient to write the tensor $\Pi_{\mu\nu}$ in terms of two transverse projection operators, $$\Pi_{\mu\nu}=-O^{(1)}_{\mu\nu} \Pi_{t}
+O^{(2)}_{\mu\nu} \left(2\Pi_{t}-m_D^2\right).$$ where $$\begin{aligned}
O^{(1)}_{\mu\nu}&=& g_{\mu\nu}-u_{\mu} u_{\nu}
+\frac{\mathbf{p}_{\mu}\mathbf{p}_{\nu}}{p^{2}},
\label{def-O1} \\
O^{(2)}_{\mu\nu}&=& u_{\mu} u_{\nu}
-\frac{\mathbf{p}_{\mu}\mathbf{p}_{\nu}}{p^{2}}
-\frac{P_{\mu}P_{\nu}}{P^{2}},
\label{def-O2} \end{aligned}$$ and $u_{\mu}=(1,0,0,0)$ is the time-like four-vector that defines the rest frame of dense quark medium. In order to have a complete set of projectors such that $\sum_I O^{(I)}_{\mu\nu}
=g_{\mu\nu}$, one has to include also the following longitudinal projector: $$O^{(3)}_{\mu\nu}= \frac{P_{\mu}P_{\nu}}{P^{2}}.
\label{def-O3}$$ By making use of this set of three projectors, inverse gluon propagator in Eq. (\[glu\_prop\]) can be represented as follows: $$\left({\cal D}^{-1}\right)_{\mu\nu}^{AB}
=i\delta^{AB} \left[\left(P^2-\Pi_{t}\right) O^{(1)}_{\mu\nu}
+\left(P^2+2\Pi_{t}-m_D^2\right) O^{(2)}_{\mu\nu}
+\frac{P^2}{\lambda} O^{(3)}_{\mu\nu}\right],
\label{inv-D}$$ where $\lambda$ is the gauge fixing parameter. This expression can be easily inverted, leading to the following result for the gluon propagator: $${\cal D}_{\mu\nu}^{AB} =-i\delta^{AB}\left(
\frac{1}{P^2-\Pi_{t}} O^{(1)}_{\mu\nu}
+\frac{1}{P^2+2\Pi_{t}-m_D^2} O^{(2)}_{\mu\nu}
+\frac{\lambda}{P^2} O^{(3)}_{\mu\nu}\right).
\label{D}$$ The gluon modes which are transverse and longitudinal with respect to the [*three-momentum*]{} $\mathbf{p}$ are called the magnetic and electric modes, respectively. These two types of physical gluon modes are described by the following fields: $$\begin{aligned}
a^{\rm (mag)}_{\mu}&=& O^{(1)}_{\mu\nu}A^{\nu},
\label{def-mag} \\
a^{\rm (el)}_{\mu}&=&O^{(2)}_{\mu\nu}A^{\nu} .
\label{def-el} \end{aligned}$$ From Eq. (\[D\]) one can see that the propagators of these modes are naturally separated from each other, as well as from the unphysical longitudinal mode, described by the field $a^{\rm (\parallel)}_{\mu}
=O^{(3)}_{\mu\nu}A^{\nu}$.
In the study of the Cooper pairing dynamics, the most relevant part of the one-gluon exchange interaction is the interaction with the space-like momenta such that $p_0\ll p$. This should be clear from the kinematics of the quark scattering around the Fermi surface: the energy exchange between pairing quarks is typically much smaller than the change of their momenta. It is justified, therefore, to use the approximate form of $\Pi_{t}$, obtained in the region $p_0\ll p$, $$\Pi_{t}\simeq m_D^2\left(i\pi \frac{p_0}{4p}
\theta(P^2)+\frac{p_0^2}{p^2}+\dots\right).
\label{asym-2}$$ By substituting this asymptote into the gluon propagator (\[D\]), one can check that the magnetic modes produce a long-range (dynamically screened) interaction, while the electric mode works only at short distances, $r {\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}1/m_D$.
In the next subsection, the HDL gluon propagator is used in the study of the Cooper pairing dynamics. As we shall see there, the long-range interaction mediated by the magnetic modes will play a particularly important role in the dynamics.
### Gap equation in QCD
Color superconductivity is an essentially non-perturbative phenomenon which cannot be addressed with perturbative techniques even if the theory is weakly interacting. The standard method for studying superconductivity is the method of the Schwinger-Dyson equations. In general, the complete set of Schwinger-Dyson equations contains an infinite number of coupled equations for Green’s functions. There exist no systematic ways of solving such equations exactly. However, there exist various standard approximations that can be justified if the dynamics is controlled by a small parameter (e.g., a weak coupling constant, or a small $1/N$ parameter where $N$ is the number of flavors or colors in the model).
In the case of QCD at asymptotic densities, the main features of its dynamics are well described by the so-called improved rainbow approximation. In this approximation, one uses the bare quark-gluon vertices in the Schwinger-Dyson equation for the quark propagator. One also uses the gluon propagator with the screening effects taken into account in the HDL approximation. (Note that, in the simple rainbow approximation, the screening is usually neglected.) The graphical representation of the Schwinger-Dyson equation is shown in Fig. \[fig-sd-eq\].
![The graphical representation of the Schwinger-Dyson equation in the improved rainbow approximation. The quark propagator in the Nambu-Gorkov basis and the gluon propagator are denoted by the solid lines and the wavy lines, respectively.[]{data-label="fig-sd-eq"}](SD11-equations.eps)
The analytical form of this equation reads $$S^{-1}(K) = S_0^{-1}(K) +4\pi\alpha_s \int\frac{d^4 p}{(2\pi)^4}
\Gamma^{A}_{\mu} S(P) \Gamma^{B}_{\nu} {\cal D}_{AB}^{\mu\nu}(K-P),
\label{sd-eq}$$ where the quark-gluon vertices in the Nambu-Gorkov basis are given by $$\Gamma^{A}_{\mu} = \gamma_\mu \left(
\begin{array}{cc}
T^A & 0 \\
0 & -\left(T^A\right)^T
\end{array}
\right).$$ In the case of the simplest ansatz for the quark propagator, as in Eq. (\[inv\_prop\]), the Schwinger-Dyson equation (\[sd-eq\]) reduces to an equation for the gap parameter $\Delta$. For this reason, it is also called the gap equation. After neglecting the dependence of the gap on the three-momentum and after performing the momentum integration, one arrives at the following approximate form of the gap equation in Euclidean space ($k_4\equiv ik_0$) [@weak-son; @weak; @weak-sp1]: $$\Delta(k_4) \simeq \frac{\alpha_s}{9\pi} \int
\frac{dp_4 \Delta(p_4)}
{\sqrt{p_4^2+\Delta^2}}\ln\frac{\Lambda}{|k_4-p_4|},
\label{gap_eq-qcd}$$ where $\Lambda = 2 (4\pi)^{3/2}\mu \alpha_s^{-5/2}$. If the quark wave function renormalization from the diagonal part of the self-energy were taken into account, the extra factor $\exp\left(-\frac{4+\pi^2}{8}\right)$ would appear in the expression for $\Lambda$ [@self-e]. The appearance of the logarithm in the integrand of the gap equation (\[gap\_eq-qcd\]) is an artifact of the long-range force in QCD, mediated by the magnetic gluon modes. An approximate solution to the gap equation reads $$\Delta(0) \simeq \Lambda \exp\left(-\frac{3\pi^{3/2}}{2^{3/2}
\sqrt{\alpha_s}}\right).
\label{gap-sol}$$ The result is plotted in Fig. \[fig-gap-sol\]. Here, the value of the coupling constant $\alpha_s\equiv \alpha_s(\mu)$ is taken at the scale of the quark chemical potential.
![The value of the gap from Eq. (\[gap-sol\]) as a function of the quark chemical potential. The coupling constant is defined at the scale of the quark chemical potential. The solid (dashed) line gives the result with (without) taking the quark wave function renormalization into account.[]{data-label="fig-gap-sol"}](gap_vs_mu1.eps){width="50.00000%"}
It appears that, at densities that exist inside stars, i.e., $n{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}10n_0$, the value of the quark chemical potential cannot be much larger than $500\,\mbox{MeV}$. The corresponding value of the QCD coupling constant $\alpha_s\equiv \alpha_s(\mu)$ is not small, and, in contrast to the situation in asymptotically dense QCD, the Cooper pairing dynamics is not weakly interacting. This suggests that the above analysis of the Schwinger-Dyson equation may not be suitable for the description of matter inside stars. In spite of this, the solution in Eq. (\[gap-sol\]) is of fundamental importance in theory. This is one of very few non-perturbative solutions in QCD which is obtained from first principles, and which can be systematically improved by studying higher order corrections. Also, this solution shows that the phase diagram of QCD in the plane of temperature and chemical potential contains color superconducting phases at least at asymptotic densities.
### Gap equation in the NJL model
If deconfined quark matter exists in the central regions of compact stars, it is likely to be color superconducting even when it is not weakly interacting. In order to study such a strongly interacting regime of quark matter with a typical value of the quark chemical potential $\mu\simeq 500\,\mbox{MeV}$, one can use various effective models of QCD. The simplest type of such models is the NJL model with a local four-fermion interaction in Eq. (\[NJL-action\]).
The NJL model in Eq. (\[NJL-action\]) lacks gluons, and, as reflection of this, it possesses the global instead of the local (gauge) SU(3)$_{c}$ color symmetry. This model can be viewed as a result of integrating out heavy gluons from the QCD action. This is possible if the gluons get nonzero masses from non-perturbative effects. One arrives at an effective NJL model similar to that in Eq. (\[NJL-action\]) when the gauge fixing in QCD, needed to perform the integration, is consistent with the global color symmetry.
The global color symmetry of the NJL model is broken in the 2SC phase. Then, because of the Goldstone theorem, five massless Nambu-Goldstone bosons should appear in the low-energy spectrum of such a theory. In QCD, however, there is no room for such Nambu-Goldstone bosons. The seeming contradiction is removed by noting that these Nambu-Goldstone bosons are not physical. Their appearance is an artifact of the gauge fixing. In particular, there exist a gauge choice in QCD, namely the unitary gauge, in which these bosons can be completely eliminated.
The gap equation in the NJL model in the mean field approximation looks as follows: $$\Delta \simeq \frac{4 G_D}{\pi^2} \int_0^{\Lambda}
\left(\frac{\Delta}{\sqrt{(p-\mu)^2+\Delta^2}}
+\frac{\Delta}{\sqrt{(p+\mu)^2+\Delta^2}}\right)
p^2 d p .
\label{gap-eq-NJL}$$ This gap equation can be obtained from a Schwinger-Dyson equation similar to that in QCD in Eq. (\[sd-eq\]) after the gluon long-range interaction is replaced by a local interaction proportional to $\delta(K-P)$.
The approximate solution to the gap equation in Eq. (\[gap-eq-NJL\]) reads $$\Delta \simeq 2\sqrt{\Lambda^2-\mu^2}
\exp\left(-\frac{\pi^2}{8G_D\mu^2}+\frac{\Lambda^2-3\mu^2}{2\mu^2}\right).$$ This is very similar to the Bardeen-Cooper-Schrieffer solution in the case of low temperature superconductivity in solid state physics [@BCS]. As in the Bardeen-Cooper-Schrieffer theory, it has the same type non-analytic dependence on the coupling constant and the same type dependence on the density of quasiparticle states at the Fermi surface.
When the quark chemical potential $\mu$ takes a value in the range between $400\,\mbox{MeV}$ and $500\,\mbox{MeV}$, and the strength of the diquark pairing is $G_D=\eta G_S$ with $\eta$ between $0.7$ and $1$, the value of the gap appears to be of order $100\,\mbox{MeV}$. In essence, this is the result that was obtained in Ref. [@cs].
### Properties of quark matter in the 2SC phase
Now let us briefly summarize the main physical properties of the 2SC phase of dense quark matter. To large extent, these follow directly from the symmetry of the ground state, determined by the structure of the condensate in Eq. (\[2sc-cond\]), and the magnitude of the diquark gap $\Delta$.
As we argued earlier, the blue up and blue down quarks do not participate in Cooper pairing in the 2SC phase. They give rise to gapless quasiparticles in the low energy spectrum of the theory. The density of states of such quasiparticles is proportional to $\mu^2$ and, therefore, is very large. This simple fact has important implications for the quark matter properties. At small temperatures, these gapless quasiparticles give dominant contributions to the specific heat, as well as to the electrical and heat conductivities of the 2SC phase. Also, the presence of the ungapped blue up and blue down quarks should result in a large neutrino emissivity due to the $\beta$-processes, $d_b \to u_b +e^{-}+\bar{\nu}_{e}$ and $u_b +e^{-}\to d_b +\nu_{e}$ [@Iwamoto].
The other four quark quasiparticles, originating from the red and green paired quarks, are gapped. Their dispersion relations look like that in Eq. (\[E\_qp\]). At small temperatures, $T\ll \Delta$, the contributions of these quasiparticles to all transport and many thermodynamic quantities are suppressed by the exponentially small factor $\exp\left(-\Delta/T\right)$. Gluons do not play a very important role either. They are bosons and their thermal number densities are small at small temperatures. In addition, five out of total eight gluons are gapped because of the color Meissner effect.
The presence of the unpaired blue quarks is also connected with the absence of baryon superfluidity in the 2SC phase. The generator of the baryon number conservation symmetry is defined as follows: $$\tilde{B}=B-\frac{2}{\sqrt{3}}T_8
=\mbox{diag}_{\rm color}\left(0,0,1\right),
\label{U1_B}$$ where $B$ is the generator of the U(1)$_B$ symmetry in vacuum. This vacuum generator $B$ mixes with the color generator $T_8$ to produce the generator $\tilde{B}$ of the $\tilde{\mbox{U}}(1)_B$ symmetry in medium. From Eq. (\[U1\_B\]) it is clear that only (anti-)blue quasiparticles carry a nonzero baryon number in the 2SC phase.
In the 2SC phase, there is also an unbroken $\tilde{\mbox{U}}(1)_{\rm em}$ gauge symmetry. The corresponding generator reads $$\tilde{Q}=Q-\frac{1}{\sqrt{3}}T_8 ,
\label{U1_em_2sc}$$ where $Q=\mbox{diag}_{\rm flavor}(\frac{2}{3},-\frac{1}{3})$ is the generator of the U(1)$_{\rm em}$ symmetry in vacuum. The gauge boson of $\tilde{\mbox{U}}(1)_{\rm em}$ is nothing else but the medium photon. It should be clear, therefore, that the 2SC ground state is not subject to the electromagnetic Meissner effect. Thus, if there is an external magnetic field, it is not expelled from such a medium.
One should notice that the quark quasiparticles carry the following $\tilde{Q}$-charges: $$\begin{aligned}
\tilde{Q}(u_r)=\tilde{Q}(u_g)=\frac{1}{2}, && \tilde{Q}(u_b)=1,\\
\tilde{Q}(d_r)=\tilde{Q}(d_g)=-\frac{1}{2}, && \tilde{Q}(d_b)=0.\end{aligned}$$ It is interesting that the blue up quark, which remains unpaired in the 2SC phase, is charged (this is not true for the blue down quark which is also unpaired). It means that this phase of matter is a $\tilde{Q}$-conductor, and that the value of the electrical conductivity is very high.
The pressure of cold dense quark matter is dominated by the Pauli pressure. The partial contribution of each quark is $\mu^4/12\pi^2$, see Eq. (\[pressure\_0\]). The correction due to color superconductivity is parametrically suppressed by the factor $\left(\Delta/\mu\right)^2$. The reason is that the Cooper pairing affects only those quark states that are close to the Fermi surface. In contrast, it is the whole Fermi sphere that contributes to the Pauli pressure. At weak coupling, the contribution due to diquark pairing can be easily calculated. It is given by $\left(\mu\Delta/2\pi\right)^2$ per each gapped quark quasiparticle [@eff-pot]. In the 2SC phase, there are six (two flavors times three colors) quarks in total, and four of them give rise to quasiparticles with the gap $\Delta$ in their energy spectra. Thus, the pressure of such matter is approximately given by [@eff-pot; @kappa] $$P_{\rm (2SC)}\simeq \frac{N_c N_f \mu^4}{12\pi^2} -B
+ 4 \left(\frac{\mu\Delta}{2\pi}\right)^{2}
=\frac{\mu^4}{2\pi^2} -B
+ \frac{\mu^2\Delta^2}{\pi^2}.
\label{pres-2sc}$$ Note that the bag pressure was also added on the right hand side. If present, this term partially cancels the leading $\mu^4$ term. Because of this, color superconductivity may have a large effect on the hadron-quark phase transition [@LH02] and even on the properties of compact stars [@AlRe03]. By making use of thermodynamic identities, one can derive the corresponding energy density, $$\epsilon_{\rm (2SC)}\simeq \frac{3\mu^4}{2\pi^2} +B
+ \frac{\mu^2\Delta^2}{\pi^2}
\left(1+\frac{2\mu}{\Delta}\frac{\partial\Delta }{\partial\mu}\right).
\label{ener-2sc}$$ The expressions in Eqs. (\[pres-2sc\]) and (\[ener-2sc\]) give a parametric representation of the equation of state of dense quark matter in the 2SC phase. Such an equation of state is the key input in the Tolman-Oppenheimer-Volkoff equations [@Tol] which determine the interior structure of compact stars. In short, such equations describe hydrostatic equilibrium of matter inside a star when the pressure gradient in each radial layer of the star is balanced by the gravitational weight of the layer itself. Several constructions of compact stars with quark matter in their interior were presented in Refs. [@LH02; @AlRe03; @BB02; @BBBNOS02; @SHH; @BGAYT; @BNOS].
Color-flavor locked phase ($N_f=3$)
-----------------------------------
It may happen that dense baryonic matter is made not only of the lightest up and down quarks, but of strange quarks as well. In fact, because of a possible reduction in the free energy from converting non-strange quarks into strange quarks, one may even speculate that strange quark matter is the true ground state of baryonic matter [@Bod].
The constituent strange quark mass in vacuum QCD is estimated to be of order $500\,\mbox{MeV}$. Its current mass is about $100\,\mbox{MeV}$. In dense baryonic matter (say, with $\mu\simeq 500\,\mbox{MeV}$), the value of the strange quark mass should be in the range between $100\,\mbox{MeV}$ and $500\,\mbox{MeV}$. It is plausible then that strange quarks also participate in Cooper pairing.
Here, we consider an idealized version of three-flavor quark matter, in which all quarks are assumed to be massless. The more realistic case of a nonzero strange quark mass will be briefly discussed in the second lecture. In the massless case, the quark model possesses the global SU(3)$_L\times$SU(3)$_R$ chiral symmetry and the global U(1)$_B$ symmetry connected with the baryon number conservation. This is in addition to SU(3)$_c$ color gauge symmetry. Note that the generator $Q=\mbox{diag}_{\rm flavor}(\frac{2}{3},-\frac{1}{3},
-\frac{1}{3})$ of the U(1)$_{\rm em}$ symmetry of electromagnetism is traceless, and therefore it coincides with one of the vector-like generators of the SU(3)$_L\times$SU(3)$_R$ chiral group.
To large extent, the color and flavor structure of the diquark condensate of Cooper pairs in the three-flavor quark matter is again determined by the symmetry of the attractive diquark channel (i.e., color-antisymmetric antitriplet) and the Pauli exclusion principle. The spin-0 condensate corresponds to the following ground state expectation value [@cfl]: $$\left\langle \left(\bar{\psi}^C\right)_i^a \gamma^5 \psi_j^b
\right\rangle
\sim \sum_{I,J=1}^{3} c^{I}_{J}\varepsilon_{ijI} \epsilon^{abJ}
+\cdots,
\label{cfl-cond}$$ which is antisymmetric in the color and flavor indices of the constituent quarks, cf. Eq. (\[2sc-cond\]). The $3\times 3$ matrix $c^{I}_{J}$ is determined from the global minimum of the free energy. It appears that $c^{I}_{J}=\delta^{I}_{J}$. The ellipsis on the right hand side stand for a contribution which is symmetric in color and flavor. A small contribution of this type is always induced in the ground state, despite the fact that it corresponds to a repulsive diquark channel [@cfl; @weak-cfl]. This is not surprising after noting that the symmetric condensate $\left\langle
\left(\bar{\psi}^C\right)_i^a \gamma^5 \psi_j^b \right\rangle
\sim \delta_{i}^{a}\delta_{j}^{b}+\delta_{j}^{a}\delta_{i}^{b}$ does not break any additional symmetries [@cfl].
In the ground state, determined by the condensate (\[cfl-cond\]), the chiral symmetry is broken down to its vector-like subgroup. However, the mechanism of this symmetry breaking is very unusual. To see this, let us rewrite the condensate in Eq. (\[cfl-cond\]) as follows: $$\left\langle \psi_{L,i}^{a,\alpha} \epsilon_{\alpha\beta}
\psi_{L,j}^{b,\beta}\right\rangle
=-\left\langle \psi_{R,i}^{a,\dot\alpha} \epsilon_{\dot\alpha\dot\beta}
\psi_{R,j}^{b,\dot\beta}\right\rangle
\sim \sum_{I=1}^{3} \varepsilon_{ijI} \epsilon^{abI}
+\cdots,
\label{LL-RR-cond}$$ where $\alpha,\beta,\dot\alpha,\dot\beta=1,2$ are the spinor indices. The condensate of left-handed fields in Eq. (\[LL-RR-cond\]) breaks the SU(3)$_c$ color symmetry and the SU(3)$_L$ chiral symmetry, but leaves the diagonal SU(3)$_{L+c}$ subgroup unbroken. Indeed, as one can check, this condensate remains invariant under the simultaneous flavor transformation $g_L$ and the compensating color transformation $g_c=(g_L)^{-1}$. Similarly, the condensate of right-handed fields in Eq. (\[LL-RR-cond\]) leaves the SU(3)$_{R+c}$ subgroup unbroken.
When both condensates are present, the symmetry of the ground state is SU(3)$_{L+R+c}$. At the level of global symmetries, the original SU(3)$_{L}\times$SU(3)$_{R}$ is broken down to the vector-like SU(3)$_{L+R}$, just like in vacuum. Unlike in vacuum, however, this breaking does [*not*]{} result from any condensates mixing left- and right-handed fields. Instead, it results from two separate condensates, made of left-handed fields only and of right-handed fields only. The color-flavor orientations of the two condensates are “locked” to each other by color transformations. This mechanism is called locking, and the corresponding phase of matter is called color-flavor-locked (CFL) phase [@cfl].
To large extent, the gap equation in the three-flavor quark matter is the same as in the two-flavor case. The differences come only from a slightly more complicated color-flavor structure of the off-diagonal part of the inverse quark propagator (gap matrix) [@cfl; @weak-cfl], $$\Delta^{ij}_{ab}=i\gamma^5\left[
\frac{1}{3}\left(\Delta_1+\Delta_2\right)\delta^{i}_{a}\delta^{j}_{b}
-\Delta_2\delta^{i}_{b}\delta^{j}_{a}\right],
\label{delta_cfl}$$ where two parameters $\Delta_1$ and $\Delta_2$ determine the values of the gaps in the quasiparticles spectra. In the ground state that respects the SU(3)$_{L+R+c}$ symmetry, the original nine quark states give rise to a singlet and an octet of quasiparticles. The singlet has gap $\Delta_1$, and the octet has gap $\Delta_2$.
When the small symmetric diquark condensate is neglected, one finds that $\Delta_1=2\Delta_2$, i.e., the gap of the singlet is twice as large as the gap of the octet. In real case, however, this relation holds only approximately. The value of the octet gap $\Delta_2$ in asymptotically dense QCD is given by an expression similar to that in Eq. (\[gap-sol\]), but the overall coefficient is multiplied by an extra factor $2^{-1/3}\left(2/3\right)^{5/2}\approx 0.288$ [@weak-cfl; @loops].
### Properties of quark matter in the CFL phase
Let us briefly review physical properties of the CFL phase. In contrast to the 2SC phase, there are no gapless quark quasiparticles in the low energy spectrum of the CFL phase. This implies that, at small temperatures, $T\ll \Delta$, the contributions of quark quasiparticles to all transport and many thermodynamic quantities are suppressed by the exponentially small factor $\exp\left(-\Delta/T\right)$. Gluons do not play any important role either. All of them are gapped because of the color Meissner effect.
Unlike the 2SC phase, the CFL phase is superfluid. This is because the U(1)$_B$ baryon number symmetry is broken in the ground state. If such a phase appears in a core of a rotating star, it will be threaded with rotational vortices. This may be related to the existence of such phenomena as glitches. These are sudden changes of the rotational frequency observed in some pulsars. The glitches may be caused by occasional releasing of the angular momentum of rotational vortices, that remained pinned to the stellar crust [@gl-super].
Because of the Goldstone theorem, the breaking of the U(1)$_B$ baryon number symmetry should result in the appearance of a Nambu-Goldstone boson in the low-energy theory. In absence of gapless quark quasiparticles, this Nambu-Goldstone boson turns out to play an important role in many transport properties of cold CFL matter [@kappa; @opaque].
The CFL phase, like the 2SC phase, is not an electromagnetic superconductor. Therefore, it does not expel a magnetic flux from its interior. This is the result of having an unbroken $\tilde{\mbox{U}}(1)_{\rm em}$ gauge symmetry in the ground state. The corresponding generator reads $$\tilde{Q}=Q-T_3-\frac{1}{\sqrt{3}}T_8 .
\label{U1_em_cfl}$$ The quark quasiparticles carry the following $\tilde{Q}$-charges: $$\begin{aligned}
\tilde{Q}(u_r)=0, && \tilde{Q}(u_g)=\tilde{Q}(u_b)=1,\\
\tilde{Q}(d_r)=-1, && \tilde{Q}(d_g)=\tilde{Q}(d_b)=0,\\
\tilde{Q}(s_r)=-1, && \tilde{Q}(s_g)=\tilde{Q}(s_b)=0.\end{aligned}$$ Since all quasiparticles are gapped, the corresponding phase of matter is a $\tilde{Q}$-insulator. Its electrical conductivity at small temperatures is dominated by thermally excited electrons and positrons [@opaque]. At sufficiently low temperatures, when the thermal density is low, the CFL phase becomes transparent to light [@opaque; @opaque2].
In order to write down the expression for the pressure of three-flavor quark matter in the CFL phase, we take the Pauli pressure contributions of nine (three flavors times three colors) quarks and add the correction due to color superconductivity from one quasiparticle with gap $\Delta_1$ and eight quasiparticles with gap $\Delta_2$. Thus, we derive $$\begin{aligned}
P_{\rm (CFL)} &\simeq &\frac{N_c N_f \mu^4}{12\pi^2} -B
+ \left(\frac{\mu\Delta_1}{2\pi}\right)^{2}
+ 8\left(\frac{\mu\Delta_2}{2\pi}\right)^{2}
\nonumber \\
&\simeq& \frac{3\mu^4}{4\pi^2} -B + 3\frac{\mu^2\Delta^2}{\pi^2},
\label{pres-cfl}\end{aligned}$$ where we used the approximate relation between the singlet and octet gaps, $\Delta_1=2\Delta_2\equiv 2\Delta$. By making use of thermodynamic identities, one can derive the corresponding energy density, $$\epsilon_{\rm (CFL)}\simeq \frac{9\mu^4}{4\pi^2} +B
+ 3\frac{\mu^2\Delta^2}{\pi^2}
\left(1+\frac{2\mu}{\Delta}\frac{\partial\Delta }{\partial\mu}\right).
\label{ener-cfl}$$ These two expressions give a parametric representation of the equation of state of dense quark matter in the CFL phase.
### Low-energy effective action
As in the case of chiral perturbation theory in vacuum QCD [@GasLeut], one could write down an effective low-energy theory in the CFL phase. From the symmetry breaking pattern, it is known that there are nine Nambu-Goldstone bosons and one pseudo-Nambu-Goldstone boson in the low-energy spectrum of the theory [@HRZ; @CasGat]. Eight of the Nambu-Goldstone bosons are similar to those in vacuum QCD: three pions ($\pi^0$ and $\pi^\pm$), four kaons ($K^0$, $\bar{K}^0$ and $K^\pm$) and the eta-meson ($\eta$). The additional Nambu-Goldstone boson ($\phi$) comes from breaking the U(1)$_B$ baryon symmetry. Finally, the pseudo-Nambu-Goldstone boson ($\eta^\prime$) results from breaking the approximate axial U(1)$_A$ symmetry.
The low energy action for the Nambu-Goldstone bosons in the CFL phase was derived in Refs. [@HRZ; @CasGat; @SonSt]. In the chiral limit, the result reads $$\begin{aligned}
{\cal L}_{eff} &=& \frac{f_\pi^2}{4} {\rm Tr}\left[
\partial_0\Sigma\partial_0\Sigma^\dagger - v^2
\partial_i\Sigma\partial_i\Sigma^\dagger \right]
+ \frac{1}{2} \left[ (\partial_0 \phi)^{2}
- v^2 (\partial_i\phi)^{2} \right]
\nonumber \\
&+&\frac{1}{2} \left[ (\partial_0 \eta^{\prime})^{2}
- v^2 (\partial_i\eta^{\prime})^{2} \right],
\label{cov-der}\end{aligned}$$ where $f^2_\pi=(21-8\ln 2)(\mu/6\pi)^2$ and $v^2=1/3$ were calculated in asymptotically dense QCD in Refs. [@SonSt; @fpi-more]. (A nonzero mass of the $\eta^\prime$-meson was neglected here [@SSZh].) By definition, $\Sigma\equiv \exp \left(i\lambda^A \pi^A/f_\pi\right)$ is a unitary matrix field which describes the octet of the Nambu-Goldstone bosons, transforming under the chiral $SU(3)_{L} \times SU(3)_{R}$ group as follows: $$\Sigma \to U_{L} \Sigma U_{R}^{\dagger},$$ where $(U_{L},U_{R}) \in SU(3)_{L} \times SU(3)_{R}$.
The extra terms in the low-energy action should be added when quark masses are nonzero [@SonSt; @mmm]. Since the quark masses break the chiral symmetry of dense QCD explicitly, these type of corrections produce non-vanishing masses for all Nambu-Goldstone bosons, except one. The single boson left massless is the Nambu-Goldstone boson related to the baryon number breaking. Of course, this symmetry is not affected by the quark masses.
Within the framework of the high density effective theory [@loops; @hdet], it was shown that an additional effect of nonzero quark masses may appear when the strange quark mass exceeds a critical value $m_s\sim m_u^{1/3}\Delta^{2/3}$. In this case, the CFL phase is expected to undergo a phase transition to a phase with a meson (e.g., kaon or/and eta) condensate [@BS; @KR; @Kry2004cw]. In the low-energy action of the corresponding phase, additional Nambu-Goldstone bosons with very unusual properties can appear [@abnormal].
Spin-1 color superconductivity ($N_f=1$)
----------------------------------------
In the case of neutral matter in $\beta$ equilibrium, as we shall discuss in the second lecture in more detail, it may happen that different quark flavors cannot create Cooper pairs because of a large mismatch between their Fermi momenta. In this case, one could consider the possibility of a much weaker spin-1 Cooper pairing [@Bail; @weak-sp1; @sp1995; @spin-1; @spin-1-Meissner; @andreas]. It is the Pauli principle that does not allow to construct spin-0 Cooper pairs from quarks of the same flavor. The color antisymmetric wave function of a pair can only be symmetric in spin indices. This corresponds to a spin-1 state.
The pairing in a spin-1 triplet state and in a color antisymmetric antitriplet state can lead to a rather complicated structure of the diquark condensate. The general structure of the gap matrix can be written in the following form [@andreas]: $$\Delta^{ab}=i\Delta_0 \sum_{c,i=1}^{3} \epsilon^{abc} {\cal C}_{ci}
\left[\hat{k}^i\cos\theta +\gamma^i_\perp\sin\theta\right],
\label{delta_1sc}$$ where $\hat{\mathbf{k}}\equiv \mathbf{k}/k$, $\gamma^i_\perp\equiv
\gamma^i-\hat{k}^i (\mbox{\boldmath $\gamma$}\cdot\hat{\mathbf{k}})$, and the explicit form of the $3\times 3$ matrix ${\cal C}_{ci}$ as well as the angular parameter $\theta$ are determined by the minimization of the quark matter free energy. In the special cases with $\theta=0$ and $\theta=\pi/2$, the corresponding gaps are called longitudinal and transverse, respectively.
By choosing various specific matrices ${\cal C}$, one could construct many different spin-1 phases. The four most popular of them are the so-called A-phase, the color-spin-locked (CSL) phase, the polar and the planar phases. These are determined by the following matrices [@andreas]: $$\begin{aligned}
{\cal C}^{\rm (A-phase)}
=\frac{1}{\sqrt{2}}\left(\begin{array}{lll}
0 & 0 & 0 \\
0 & 0 & 0 \\
1 & i & 0
\end{array}\right), &\quad &
{\cal C}^{\rm (CSL)}
=\frac{1}{\sqrt{3}}\left(\begin{array}{lll}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}\right), \\
{\cal C}^{\rm (polar)}
=\left(\begin{array}{lll}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 1
\end{array}\right),&\quad &
{\cal C}^{\rm (planar)}
=\frac{1}{\sqrt{2}}\left(\begin{array}{lll}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 0
\end{array}\right).\end{aligned}$$ The corresponding four phases are characterized by different symmetries of their ground state. In particular, the original group SU(3)$_c\times
$SO(3)$_J\times $U(1)$_{\rm em}$ of one-flavor quark matter breaks down as follows [@spin-1; @spin-1-Meissner; @andreas]: $$\begin{aligned}
\mbox{A-phase:} & &
SU(2)_c\times \widetilde{SO}(2)_J \times \tilde{U}(1)_{\rm em},\\
\mbox{CSL:} & & \widetilde{SO}(3)_{J},\label{csl}\\
\mbox{Polar:} & &
SU(2)_c\times SO(2)_J \times \tilde{U}(1)_{\rm em},\\
\mbox{Planar:} & &
\widetilde{SO}(2)_J \times \tilde{U}(1)_{\rm em}.\end{aligned}$$ It was argued recently that the phase with the lowest free energy is the transverse A-phase [@andreas].
The value of the spin-1 gap is estimated to be about two or three orders of magnitude smaller than a typical spin-0 gap. At best, it can be about $1\,\mbox{MeV}$, but more realistically it is expected to be about $0.1\,\mbox{MeV}$, or less. The presence of such a small gap is unlikely to modify considerably any thermodynamic, or even transport properties of dense quark matter. It is fair to mention, however, that the cooling history of stars might be sensitive enough to feel the effect of such small gaps [@B_cool]. It was also suggested that the electromagnetic properties of spin-1 phases can be of phenomenological importance [@spin-1-Meissner]. In contrast to the 2SC and CFL phases, discussed earlier, spin-1 phases reveal the electromagnetic Meissner effect. As one can see from Eq. (\[csl\]), this clearly applies to the CSL phase. It turns out that the electromagnetism is also subject to the Meissner effect when there appears a mixture of any two independent spin-1 condensates made of quarks with different charges (e.g., one is made of up quarks and the other is made of down quarks) [@spin-1-Meissner]. The type-I superconductivity in such a system could have observable effects on the magnetic field relaxation in pulsars.
Summary of the first lecture
----------------------------
In this lecture, only the main three color superconducting phases of dense quark matter have been discussed. These are the simplest possible phases that can be realized in one-, two-, and three-flavor quark matter. It should be emphasized, however, that the conditions in dense matter were idealized by assuming that the chemical potentials of pairing quarks were equal. In real situation which, for example, is realized in the interior of compact stars, matter happens to be very different from ideal. In particular, the charge neutrality and the $\beta$ equilibrium are two very important conditions that may modify the quark chemical potentials and, therefore, affect the properties of quark matter. These issues are discussed in the second lecture.
The aim of this lecture was to introduce the general idea of color superconductivity in cold dense baryonic matter. It was argued that such matter is expected to be deconfined and, because of the Cooper instability, it should be color superconducting. To large extent, this phenomenon is the same as low-temperature superconductivity in condensed matter physics [@BCS]. One feature in dense QCD is very remarkable. The attractive interaction between quarks comes from the gauge boson exchange. This is in contrast to the Bardeen-Cooper-Schrieffer theory in solid state physics, where the gauge boson (photon) exchange is repulsive, while an attractive interaction comes from the phonon exchange.
One of the key points of this lecture is the observation that QCD at asymptotic densities is a weakly interacting (although non-perturbative) theory. Moreover, the properties of its color superconducting ground state can be studied analytically from first principles, providing a rare example of an essentially solvable limit in a non-Abelian theory. By itself, this has a fundamental theoretical importance. Also, this result may provide valuable insights in the theory of strong interactions. One of the examples might be the idea of duality between the hadronic and quark description of QCD [@cont].
Although the use of weakly interacting QCD for the description of dense quark matter is very instructive by itself, it does not seem to be very useful for the quantitative description of the ground state at densities existing inside stars. The reason is that the density of matter in stars is bounded from above by the condition of the hydrostatic equilibrium, $n{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}10n_0$. At such not so large densities, one cannot treat QCD as a weakly interacting theory. In this situation, the use of various effective models of QCD has proved to be very useful.
Color superconductivity in neutral matter
=========================================
Dense matter inside stars
-------------------------
As we discussed in the first lecture, it is natural to expect that some color superconducting phases may exist in the interior of compact stars. The estimated central densities of such stars might be sufficiently large for producing deconfined quark matter. Such matter should develop the Cooper instability and become color superconducting. It should also be noted that typical temperatures inside compact stars are so low that the diquark condensate, if produced, would not melt.
In the first lecture, we discussed the idealized version of dense matter, in which the Fermi momenta of pairing quarks were assumed to be equal. This does not describe the real situation that should occur inside compact stars. The reason is that matter in the bulk of a compact star should be neutral (at least, on average) with respect to electric as well as color charges. Otherwise, the star would not be bound by gravity which is much weaker than electromagnetism. Matter should also remain in $\beta$ equilibrium, i.e., all $\beta$ processes (e.g., such as $d \to u + e^{-} + \bar\nu_{e}$, $u + e^{-} \to d + \nu_{e}$, $s \to u + e^{-} + \bar\nu_{e}$, and $u + e^{-} \to s + \nu_{e}$) should go with equal rates in both directions.
As we shall see below, after the charge neutrality and the $\beta$ equilibrium in quark matter are enforced, the chemical potentials of different quarks would satisfy some relations that may interfere with the dynamics of Cooper pairing. If this happens, some color superconducting phases may become less favored than others. For example, it was argued in Ref. [@no2sc], that a mixture of unpaired strange quarks and the non-strange 2SC phase, made of up and down quarks, is less favorable than the CFL phase after the charge neutrality condition is enforced. A similar conclusion was also reached in Ref. [@n_steiner].
Assuming that the constituent medium modified mass of the strange quark is large (i.e., larger than the corresponding strange quark chemical potential), in Ref. [@SH] it was shown that neutral two-flavor quark matter in $\beta$-equilibrium can have another rather unusual ground state called the gapless two-flavor color superconductor (g2SC). While the symmetry in the g2SC ground state is the same as that in the conventional 2SC phase, the spectrum of the fermionic quasiparticles is different.
The existence of gapless color superconducting phases was confirmed in Refs. [@GLW; @var-appr; @rusterR], and generalized to nonzero temperatures in Refs. [@HS; @LZ]. In addition, it was also shown that a gapless CFL (gCFL) phase could appear in neutral strange quark matter [@gCFL; @gCFL-long]. At nonzero temperature, the gCFL phase and several other new phases (e.g., the so-called dSC and uSC phases) were studied in Refs. [@dSC; @RSR; @FKR]. If the surface tension is sufficiently small, as suggested in Ref. [@RR], the mixed phase composed of the 2SC phase and the normal quark phase [@SHH] may be more favored.
Here, it may be appropriate to mention that a non-relativistic analogue of gapless superconducting phases could appear in a trapped gas of cold fermionic atoms [@WilLiu; @Deb; @LWZ; @FGLW; @GWM]. (Note that an alternative ground state for the atomic system, similar to the quark mixed phases in Refs. [@RR; @SHH; @mixed], was proposed in Ref. [@Bed].)
Gapless two-flavor color superconductivity
------------------------------------------
### Neutrality vs. color superconductivity
Let us discuss the effect of charge neutrality on color superconductivity. It is instructive to start with the case of two-flavor quark matter first. In this case, the effect is most prominent and, thus, it is easiest to understand. Later, the results for three-flavor quark matter will be also discussed.
In order to get an impression about the importance of the charge neutrality condition in a large macroscopic chunk of matter, such as a core of a compact star, let us estimate the corresponding Coulomb energy. A simple calculation leads to the following result: $$E_{\rm Coulomb} \sim n_Q^2 R^5 \sim M_\odot c^2
\left(\frac{n_Q}{10^{-15} e/\mbox{fm}^3}\right)^2
\left(\frac{R}{1\,\mbox{km}}\right)^5,
\label{E_Coulomb}$$ where $R$ is the radius of the quark matter core, whose charge density is denoted by $n_Q$. It is easy to see that this energy is not an extensive quantity: the value of the corresponding [*energy density*]{} increases with the size of the system as $R^2$. By taking a typical value of the charge density in the ideal 2SC phase, $n_Q\sim 10^{-2}
e/\mbox{fm}^3$, the energy in Eq. (\[E\_Coulomb\]) becomes a factor of $10^{26}$ larger than the rest mass energy of the Sun! To avoid such an incredibly large energy price, the charge neutrality $n_Q=0$ should be satisfied with a very high precision.
In the case of two-flavor quark matter, one can argue that the neutrality is achieved if the number density of down quarks is approximately twice as large as number density of up quarks, $n_d \approx 2n_u$. This follows from the fact that the negative charge of the down quark ($Q_d=-1/3$) is twice as small as the positive charge of the up quark ($Q_u=2/3$). When $n_d \approx 2n_u$, the total electric charge density is vanishing in absence of electrons, $n_Q\approx Q_d n_d
+Q_u n_u\approx 0$. It turns out that even a nonzero density of electrons, required by the $\beta$ equilibrium condition, could not change this relation much.
The argument goes as follows. Let us consider noninteracting massless quarks. In $\beta$ equilibrium, the chemical potentials of the up quark and the down quark, $\mu_u$ and $\mu_d$, should satisfy the relation $\mu_d=\mu_u+\mu_e$ where $\mu_e$ is the chemical potential of electrons (i.e., up to a sign, the chemical potential of the electric charge). By assuming that $\mu_d\approx 2^{1/3}\mu_u$, i.e., $n_d\approx2n_u$ as required by the neutrality in absence of electrons, one obtains the following result for the electron chemical potential: $\mu_e=\mu_d-\mu_u \approx \frac{1}{4}\mu_u$. The corresponding density of electrons is $n_e\approx 5\cdot 10^{-3}n_u$. We see that $n_e\ll n_u$ which is in agreement with the original assumption that $n_d\approx 2n_u$ in neutral matter.
While the approximate relation $n_d\approx 2n_u$ may be slightly modified in an interacting system, the main conclusion would remain qualitatively the same. The Fermi momenta of up and down quarks, whose pairing is responsible for color superconductivity, are generally non-equal after the neutrality and $\beta$ equilibrium conditions are imposed. In this case, the Cooper pairing may be substantially modified, or even prevented. The study of this issue is the topic of the rest of this lecture.
### Model
In view of strongly coupled dynamics and various non-perturbative (e.g., instanton) effects in dense QCD at moderate densities $n{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}10n_0$, existing in central regions of compact stars, there is no real advantage in using microscopic theory of QCD. For purposes of the current presentation, it suffices to use an effective model such, for example, as the NJL model with the Lagrangian density in Eq. (\[NJL-action\]), in which the common quark chemical potential $\mu$ is replaced by a properly chosen color-flavor matrix $\hat\mu$ of chemical potentials [@no2sc; @n_steiner; @huang_2sc].
In $\beta$ (chemical) equilibrium, the matrix $\hat\mu$ is determined only by a few independent chemical potentials which can be introduced in the corresponding partition function as Lagrange multipliers in front of conserved charges $Q_i$ of the model, $$Z=\mbox{Tr}\,\exp\left(-\frac{H+\sum_i \mu_{i} Q_i}{T}\right),$$ In two-flavor quark matter, we consider only three relevant conserved charges: the baryon number $n_B$, the electric charge $n_Q$, and the color charge $n_{8}$. Then, the matrix of quark chemical potentials is given in terms of the baryon chemical potential (by definition, $\mu_B\equiv 3\mu$), the electron chemical potential ($\mu_e$) and the color chemical potential ($\mu_8$) [@SH], $$\hat\mu_{ij, \alpha\beta}= (\mu \delta_{ij}- \mu_e Q_{ij})
\delta_{\alpha\beta} + \frac{2}{\sqrt{3}}\mu_{8} \delta_{ij}
(T_{8})_{\alpha \beta},$$ where $Q$ and $T_8$ are the generators of U(1)$_{\rm em}$ of electromagnetism and the U(1)$_{8}$ subgroup of SU(3)$_{c}$ color gauge group. The explicit expressions for chemical potentials of different quarks read $$\begin{aligned}
\mu_{ur} =\mu_{ug} =\mu -\frac{2}{3}\mu_{e} +\frac{1}{3}\mu_{8},
\label{mu-ur-ug} \\
\mu_{dr} =\mu_{dg} =\mu +\frac{1}{3}\mu_{e} +\frac{1}{3}\mu_{8},
\label{mu-dr-dg} \\
\mu_{ub} =\mu -\frac{2}{3}\mu_{e} -\frac{2}{3}\mu_{8},
\label{mu-ub} \\
\mu_{db} =\mu +\frac{1}{3}\mu_{e} -\frac{2}{3}\mu_{8}.
\label{mu-db} \end{aligned}$$ Note that there exist two mutually commuting color charges in the model. They are related to the generators $T_3$ and $T_8$ of the SU(3)$_{c}$ gauge group. Then, generally speaking, one could introduce two independent chemical potentials for the corresponding two color charges. Because of the invariance of the 2SC ground state under the transformations of the SU(2)$_{c}$ color gauge subgroup, the introduction of the second nontrivial color chemical potential $\mu_{3}$ is not necessary. This additional color chemical potential, however, is generally needed in three-flavor quark matter.
Here, several comments are in order regarding the introduction of color quark chemical potential $\mu_8$. This is required for enforcing color charge neutrality only in the color superconducting ground state. In the NJL model, it is added by hand in the partition function, and then adjusted to neutralize the ground state. In QCD, on the other hand, the charge neutrality is realized dynamically due to the generation of the gluon condensate $\langle A_0^8\rangle\neq 0$ [@DD]. The appearance of such a condensate is equivalent to having a nonzero value of the chemical potential $\mu_8\sim g_s \langle
A_0^8\rangle$. A similar relation exists between the chemical potential $\mu_3$ and the gluon condensate $\langle A_0^3\rangle$ in three-flavor quark matter. It should be clear, therefore, that nonzero gluon condensates, such as $\langle A_0^8\rangle$ or $\langle A_0^3\rangle$, do not break explicitly any symmetries in QCD. They appear spontaneously only in the color superconducting ground state, and only after gauge fixing.
### Effective potential
In order to derive the effective potential in the NJL model (\[NJL-action\]) in the mean field approximation, it is convenient to use the standard Hubbard-Stratonovich method [@Hub-Stra]. Then, the original NJL Lagrangian density is replaced by $$\begin{aligned}
\tilde{\cal L}_{\rm NJL} &=& \bar\psi\left[i\gamma^\mu \partial_\mu
+ \gamma^0 \hat{\mu} - m^{(0)} -\sigma
-i\gamma_5 (\vec{\mbox{\boldmath $\pi$}} \cdot\vec{\tau})\right]\psi
-\frac{\sigma^2+\vec{\mbox{\boldmath $\pi$}}^2}{4G_S}
\nonumber\\
&-& \frac{\phi_a^{*} \phi^a}{4G_D}
-\frac{i}{2} \bar{\psi}^C \gamma_5 \varepsilon \epsilon^{a}
\phi^a \psi
-\frac{i}{2} \bar{\psi} \varepsilon \epsilon^{a} \phi_a^{*}
\gamma_5 \psi^C ,
\label{HS-action}\end{aligned}$$ where $\sigma$, $\vec{\mbox{\boldmath $\pi$}}$, and $\phi_a$ are auxiliary fields. After taking into account the equations of motion for these fields, $$\begin{aligned}
\sigma &=& - 2 G_S \left(\bar\psi \psi\right) ,\\
\vec{\mbox{\boldmath $\pi$}} &=& - 2 G_S\left(\bar\psi i\gamma_5
\vec{\tau} \psi\right),\\
\phi^a &=& - 2G_D \left( i \bar{\psi} \varepsilon
\epsilon^{a} \gamma_5 \psi^C \right) ,\\
\phi_a^{*} &=& - 2G_D \left(i \bar{\psi}^C \gamma_5 \varepsilon
\epsilon^{a} \psi\right) ,\end{aligned}$$ one can check that the Lagrangian density in (\[HS-action\]) is equivalent to the original NJL model. In the mean field approximation, the auxiliary fields are replaced by their vacuum expectation values, $\langle\sigma\rangle = m-m^{(0)}$, $ \langle \vec{\mbox{\boldmath $\pi$}} \rangle=0$ and $\langle \phi^a \rangle=\Delta\delta^{a}_{3}$, and the quantum fluctuations are neglected. Then, the effective potential of quark matter in $\beta$-equilibrium (with massless electrons) takes the form [@HS]: $$\begin{aligned}
\Omega &=& \Omega_{0}
-\frac{1}{12\pi^2}\left(\mu_{e}^{4}+2\pi^{2}T^{2}\mu_{e}^{2}
+\frac{7\pi^{4}}{15} T^{4} \right)
+\frac{\left(m-m^{(0)}\right)^2}{4G_S} \nonumber\\
&+&\frac{\Delta^2}{4G_D}
-\sum_{I} \int\frac{d^3 p}{(2\pi)^3} \left[E_{I}
+2 T\ln\left(1+e^{-E_{I}/T}\right)\right],
\label{pot}\end{aligned}$$ where $\Omega_{0}$ is a constant added to make the pressure of the vacuum vanishing (the bag constant can also be included on the right hand side if necessary). Here, while including electrons, we neglected the contribution of neutrinos. This properly describes the situation inside compact stars after the deleptonization had occurred, i.e., after the temperature of matter dropped so low that the neutrino mean free path became larger than a typical stellar size. In protoneutron stars this happens in less than a minute after their birth.
The sum in the second line of Eq. (\[pot\]) runs over all (6 quark and 6 antiquark) quasiparticles. The dispersion relations and the degeneracy factors of the quasiparticles read $$\begin{aligned}
E_{ub}^{\pm} &=& E(p) \pm \mu_{ub} , \hspace{26.6mm} [\times 1]
\label{disp-ub} \\
E_{db}^{\pm} &=& E(p) \pm \mu_{db} , \hspace{26.8mm} [\times 1]
\label{disp-db}\\
E_{\Delta^{\pm}}^{\pm} &=& E_{\Delta}^{\pm}(p) \pm \delta \mu .
\hspace{26.5mm} [\times 2]
\label{2-degenerate}\end{aligned}$$ These can be obtained in a rather straightforward way from the quark propagator. Here, the following shorthand notation was introduced: $$\begin{aligned}
E(p) &\equiv& \sqrt{{\bf p}^2+m^2}, \\
E_{\Delta}^{\pm}(p) &\equiv&
\sqrt{[E(p) \pm \bar{\mu}]^2 +\Delta^2},\\
\bar{\mu} &\equiv&
\frac{\mu_{ur} +\mu_{dg}}{2}
=\frac{\mu_{ug}+\mu_{dr}}{2}
=\mu-\frac{\mu_{e}}{6}+\frac{\mu_{8}}{3}, \label{mu-bar}\\
\delta\mu &\equiv&
\frac{\mu_{dg}-\mu_{ur}}{2}
=\frac{\mu_{dr}-\mu_{ug}}{2}
=\frac{\mu_{e}}{2}. \label{delta-mu}\end{aligned}$$ The thermodynamic potential that determines the pressure of quark matter, $\Omega_{\rm phys} =-P$, is obtained from the effective potential $\Omega$ in Eq. (\[pot\]) after substituting the values of $\mu_{8}$, $\mu_{e}$, $m$ and $\Delta$ which solve the color and electric charge neutrality conditions, i.e., $$n_{8}\equiv
-\frac{\partial \Omega}{\partial \mu_{8}}=0, \quad \mbox{and} \quad
n_{Q}\equiv
-\frac{\partial \Omega}{\partial \mu_{e}}=0,
\label{Q=0}$$ as well as the gap equations, i.e., $$\frac{\partial \Omega}{\partial m}=0, \quad \mbox{and} \quad
\frac{\partial \Omega}{\partial \Delta}=0.
\label{gap-eqs}$$ Only a solution that satisfies Eqs. (\[Q=0\]) and (\[gap-eqs\]) can correspond to a neutral ground state of quark matter. If there exist several such solutions, it is the solution which gives the highest value of the pressure that gives the ground state.
### Three regimes in neutral matter
In studying neutral two-flavor quark matter, it was found that there exist three qualitatively different dynamical regimes which differ by the strength of diquark coupling, as well as by the properties of the ground state [@SH; @HS].
The first regime corresponds to weak diquark coupling. In this case, a weak Cooper pairing is completely suppressed by the mismatch between the Fermi momenta of the up and down quarks. In other words, the appearance of the 2SC phase is in conflict with the constraint of charge neutrality. The ground state of neutral matter corresponds to the normal phase in this regime. One should note, though, that a much weaker spin-1 pairing between quarks of same flavor is not forbidden in such neutral matter. In fact, if the temperature is sufficiently low, two independent spin-1 condensates, one made of up quarks and the other made of down quarks, are inevitable.
The other limiting case is the strongly coupled regime. It is clear that, if the value of the diquark coupling is sufficiently large, the color condensation could be made as strong as needed to overcome a finite mismatch between the Fermi surfaces of pairing quarks. In this argument, it is taken into account that the charge neutrality constraint is only slightly affected by the value of the diquark coupling. This is the case in all models that have been studied so far. In this regime, the ground state is in the 2SC phase, and the neutrality has little effect. Here, it might be appropriate to mention that the unpaired blue up and blue down quarks of the 2SC phase cannot produce any additional spin-1 condensate. The main reason is that their chemical potentials are badly mismatched.
In Fig. \[2sc-quasi\] the dispersion relations of quasiparticles in the 2SC phase (i.e., in the strongly coupled regime) and in the normal quark matter phase (i.e., in the weakly coupled regime) are shown by the solid and dashed lines, respectively. The dispersion relations of unpaired blue quasiparticles are not shown. It should be clear, however, that they are similar to those in the normal phase, albeit slightly shifted in the horizontal direction.
![The dispersion relations of quark quasiparticles in the 2SC phase (solid lines) and the normal quark matter phase (dashed lines). The low-energy part of the dispersion relations is shown in the right panel.[]{data-label="2sc-quasi"}](2sc.eps){width="99.00000%"}
![The dispersion relations of quark quasiparticles in the 2SC phase (solid lines) and the normal quark matter phase (dashed lines). The low-energy part of the dispersion relations is shown in the right panel.[]{data-label="2sc-quasi"}](disp-relation-large-gap.eps){width="99.00000%"}
We see from Fig. \[2sc-quasi\] that there are two types of gapped quasiparticles in the spectrum of the 2SC phase. In fact, each of them is a doublet with respect to the unbroken SU(2)$_c$ gauge group. One of the doublets has the gap $\Delta+\delta\mu$, and the other has the gap $\Delta-\delta\mu$, where $\delta\mu\equiv \mu_e/2$ and $\Delta$ are the solutions of Eqs. (\[Q=0\]) and (\[gap-eqs\]) that describe the ground state. When the mismatch parameter $\delta\mu$ is vanishing, two types of quasiparticles become degenerate. This corresponds to the ideal 2SC phase, considered in the first lecture.
The actual values of the coupling constants, which define the strongly and weakly coupled dynamical regimes, can differ very much from one effective model of QCD to another. Thus, it is convenient to use the value of the following dimensionless ratio $\Delta/\delta\mu$ instead of the value of the diquark coupling constant in a quantitative definition of such regimes. In this ratio, only the value of the gap $\Delta$ is strongly sensitive to the coupling constant.
From our classification above, it should be clear that $\Delta/
\delta\mu=0$ corresponds to the weakly coupled regime. This should not be confused with the $G_D=0$ case. In fact, in any model, there should exist a critical value $G_D^{(1)}$ such that $\Delta/\delta\mu=0$ when $G_D<G_D^{(1)}$. As we shall see below, the strongly coupled regime is given by $\Delta/\delta\mu>1$. This makes sense after noting that the values of the two gaps in the 2SC phase, $\Delta+\delta\mu$ and $\Delta-\delta\mu$, are positive when $\Delta/\delta\mu>1$ (see Fig. \[2sc-quasi\]).
It turns out, that there also exists an intermediate regime, in which the diquark coupling is neither too weak nor too strong. In this third regime, the ground state is given by the so-called gapless 2SC phase [@SH; @HS]. This phase corresponds to the ratio $\Delta/\delta\mu$ in the range from $0$ to $1$ (i.e., $0<\Delta/\delta\mu<1$). This regime will be discussed in more detail in the next subsection.
### Gapless 2SC phase
In the case of the NJL model used in Refs. [@SH; @HS], the intermediate dynamical regime with $0<\Delta/\delta\mu<1$ is realized when the strength of the diquark coupling is given by $0.7{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}\eta {\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}}0.8$ where $\eta\equiv G_D/G_S$. To be specific, below we use $\eta=0.75$.
Let us start from a brief discussion of the neutrality conditions in Eq. (\[Q=0\]). At zero temperature, they are satisfied approximately when $\mu_8\approx 0$ and $$\begin{aligned}
\mu_e & \approx & \frac{3}{5} \mu, \quad \mbox{for} \quad
\Delta > \delta\mu,
\label{Q=0_1}\\
\Delta & \approx & \sqrt{\frac{\mu_e^2}{4}-
\left(\frac{(\mu-2\mu_e/3)^3-\mu_e^3}
{6(\mu-\mu_e/6)^2}\right)^2} ,
\quad \mbox{for} \quad \Delta \leq \delta\mu.
\label{Q=0_2}\end{aligned}$$ It should be noted that the solutions in the two regions of parameters, $\Delta> \delta\mu$ and $\Delta \leq \delta\mu$, are very different. One could check that this reflects a qualitatively different nature of the quasiparticle spectra in the two regions. While all the modes described by the dispersion relations in Eq. (\[2-degenerate\]) are gapped when $\Delta > \delta\mu$, there appear a gapless doublet mode when $\Delta \leq \delta\mu$.
After imposing neutrality, the chemical potentials $\mu_e$ and $\mu_8$ are not free parameters, but (implicit) functions of $\Delta$. By making use of these functions $\mu_e=\mu_e(\Delta)$ and $\mu_8=\mu_8(\Delta)$ in Eq. (\[pot\]), one can define the effective potential for neutral matter as a function of the gap parameter $\Delta$. This potential is shown graphically in Fig. \[V2D\] (solid line).
![The effective potential as a function of the diquark gap $\Delta$ calculated at a fixed value of the electron chemical potential $\mu_e\approx 148$ MeV (dashed line), and the effective potential defined along the neutrality line (solid line). The results are plotted for $\mu=400$ MeV and $G_{D}=\eta G_{S}$ with $\eta=0.75$.[]{data-label="V2D"}](eff-pot-Omega-T0.eps){width="8cm"}
To emphasize the role of the neutrality condition, in Fig. \[V2D\], the result for the effective potential of non-neutral quark matter at a fixed value of the electron chemical potential $\mu_e\approx 148$ MeV is also shown (dashed line). The chemical potential $\mu_e$ is chosen so that the electric charge density vanishes when the value of the gap corresponds to the maximum of the potential. This effective potential (labelled “$\mu_e=148$ MeV”) describes negatively charged matter to the left from the maximum, and positively charged matter to the right from the maximum. On both sides, it is unphysical because of a large (infinite in an infinite volume) contribution of the Coulomb energy, see Eq. (\[E\_Coulomb\]), that was neglected in Eq. (\[pot\]). It is fair to note that the charged minima of the dashed line may become physically important if mixed phases are allowed [@RR; @SHH; @mixed].
From the location of the minimum of the effective potential for neutral quark matter (solid line in Fig. \[V2D\]), one determines the value of the gap parameter in the ground state, $\Delta\approx
68$ MeV. It appears that $\Delta<\delta\mu\equiv \mu_e/2\approx
74\,\mbox{MeV}$ in such a ground state, i.e., it corresponds to the intermediate regime in our classification. The low energy quasiparticle spectrum in this phase contains additional (as compared to the 2SC phase) gapless modes [@SH; @HS]. This is also seen clearly from the analytical expressions in Eqs. (\[disp-ub\])–(\[2-degenerate\]). Graphically, the dispersion relations are shown in Fig. \[disp-rel\]. These should be compared with the corresponding quasiparticle dispersion relations in the 2SC phase in Fig. \[2sc-quasi\].
![The dispersion relations of quark quasiparticles in the g2SC phase (solid lines) and the normal quark matter phase (dashed lines). The low-energy part of the dispersion relations is shown in the right panel.[]{data-label="disp-rel"}](g2sc.eps){width="99.00000%"}
![The dispersion relations of quark quasiparticles in the g2SC phase (solid lines) and the normal quark matter phase (dashed lines). The low-energy part of the dispersion relations is shown in the right panel.[]{data-label="disp-rel"}](disp-relation-small-gap.eps){width="99.00000%"}
The most remarkable property of the quasiparticle spectrum in the g2SC phase is that the low energy excitations ($E\ll \delta\mu-\Delta$) are very similar to those in the normal phase. The only difference is that the values of the chemical potentials of the free up and down quarks are replaced by $\mu^{\pm}\equiv \bar{\mu}\pm
\sqrt{(\delta\mu)^2-\Delta^2}$. This simple observation may suggest that the low energy (large distance scale) properties of the g2SC phase are similar to those in the normal phase [@HS]. For example, the Debye screening mass of the gluons of the SU$(2)_c$ subgroup should be nonzero. The latter, in fact, should be proportional to the density of the gapless modes [@pi], $$m_{1,D}^2\simeq \frac{2\alpha_s}{\pi}\left(
\frac{(\mu^{-})^2\delta\mu}{\sqrt{(\delta\mu)^2-\Delta^2}}
+\frac{(\mu^{+})^2\delta\mu}{\sqrt{(\delta\mu)^2-\Delta^2}}
\right)\theta(\delta\mu-\Delta).
\label{m_D_1}$$ Note that the corresponding value for the Debye screening mass in the ordinary 2SC phase is vanishing [@R-meissner]. The overall coefficient in Eq. (\[m\_D\_1\]) can be fixed by matching the value of the Debye screening mass with the known result in the normal phase. The Meissner screening mass of the gluons of the unbroken SU$(2)_c$ subgroup is vanishing, $m_{1,M}^2=0$, as it should be [@pi].
The magnetic properties of the g2SC phase are nothing like those in the normal phase. The values of the Meissner screening masses of five gluons are [*imaginary*]{}, indicating a chromomagnetic instability in dense two-flavor quark matter [@pi]. This instability may lead to a gluon condensation, and possibly to breaking the rotational or the translational symmetry in the true ground state of quark matter. This could be a completely new state, or this could be one of the states proposed earlier. For example, it may be a state with deformed quark Fermi surfaces [@MutherSed], or it may turn into the Larkin-Ovchinnikov-Fulde-Ferrell state [@LOFF], which was discussed in Ref. [@loff-cs] in the context of quark matter. To make a certain statement, further studies in this direction are needed.
### Finite temperature properties
The finite temperature properties of neutral quark matter in the intermediate coupling regime are also very unusual [@HS; @LZ]. The most striking are the results for the temperature dependence of the gap parameter for various values of the diquark coupling strengths. This is demonstrated by the plot in Fig. \[gap-eta\]. As one can see, the results for not very strong coupling are very unusual. The gap function may take a finite value at nonzero temperatures even though it is exactly zero at zero temperature. This possibility comes about only because of the neutrality condition which introduces a mismatch between the Fermi momenta of pairing quarks. The thermal excitations smear the Fermi surfaces of the up and down quarks, and this opens the possibility of Cooper pairing which is forbidden at zero temperature. With increasing the temperature, the thermal fluctuations will eventually become too strong, and the pairing will be destroyed.
![The temperature dependence of the diquark gap in neutral quark matter calculated for several values of the diquark coupling strength $\eta=G_D/G_S$ (left panel), and the ratio of the critical temperature to the zero temperature gap in neutral quark matter as a function of the coupling strength $\eta$ (right panel).[]{data-label="gap-eta"}](gap-T-eta.eps){width="99.00000%"}
![The temperature dependence of the diquark gap in neutral quark matter calculated for several values of the diquark coupling strength $\eta=G_D/G_S$ (left panel), and the ratio of the critical temperature to the zero temperature gap in neutral quark matter as a function of the coupling strength $\eta$ (right panel).[]{data-label="gap-eta"}](Tc-gap-ratio.eps){width="99.00000%"}
By looking at the results in the left panel of Fig. \[gap-eta\], one should immediately realize that the value of the ratio of the critical temperature $T_c$ to the value of the gap at zero temperature $\Delta_0$ is not a universal number. In contrast to the Bardeen-Copper-Schrieffer theory, the ratio $T_c/\Delta_0$ in the neutral phase depends on the diquark coupling constant. This is shown explicitly in the right panel in Fig. \[gap-eta\]. As we see, the ratio $T_c/\Delta_0$ can be arbitrarily large, and even remain strictly infinite for a range of coupling strengths.
Gapless color-flavor locked phase
---------------------------------
While considering a realistic model of strange quark matter, one should take into account that the strange quark mass is much larger than the masses of the light up and down quarks. Most likely, the actual value of strange quark mass in dense medium should be in the range between about $100\,\mbox{MeV}$ and $500\,\mbox{MeV}$. This is not negligible in comparison with a typical value of the quark chemical potential $\mu\simeq 500\,\mbox{MeV}$ at densities existing in compact stars.
In application to the Cooper pairing dynamics responsible for color superconductivity, the main effect of a non-vanishing strange quark mass is a reduction of the strange quark Fermi momentum, $$k_F^{(s)} = \sqrt{\mu^2-m_s^2} \simeq \mu-\frac{m_s^2}{2\mu}.
\label{k_F_shift}$$ The magnitude of the reduction is approximately given by the value of $m_s^2/2\mu$. This quantity plays the role of a mismatch parameter in three-flavor quark matter, which is similar to $\delta\mu\equiv
\mu_e/2$ in two-flavor quark matter. For obvious reasons, this mismatch should interfere with Cooper pairing between strange and non-strange quarks.
Because of a nonzero strange quark mass, the diquark condensate of the ideal CFL phase (\[LL-RR-cond\]) may get distorted as follows [@n_steiner; @gCFL]: $$\left\langle \psi_{L,i}^{a,\alpha} \epsilon_{\alpha\beta}
\psi_{L,j}^{b,\beta}\right\rangle
=-\left\langle \psi_{R,i}^{a,\dot\alpha} \epsilon_{\dot\alpha\dot\beta}
\psi_{R,j}^{b,\dot\beta}\right\rangle
\sim \Delta_1 \varepsilon_{ij1} \epsilon^{ab1}
+ \Delta_2 \varepsilon_{ij2} \epsilon^{ab2}
+ \Delta_3 \varepsilon_{ij3} \epsilon^{ab3}
+\cdots,
\label{LL-RR-gCFL}$$ where the ellipsis denote terms symmetric in color and flavor. As in the ideal CFL phase, they are small and not crucial for the qualitative understanding of strange quark matter [@RSR].
The three gap parameters $\Delta_1$, $\Delta_2$ and $\Delta_3$ in Eq. (\[LL-RR-gCFL\]) describe down-strange, up-strange and up-down condensates of Cooper pairs, respectively. A nonzero value of $m_s^2/2\mu$ affects first and foremost the pairing between the strange and non-strange quarks, i.e., the gap parameters $\Delta_1$ and $\Delta_2$. Because of the color-flavor locking, preserved in the diquark condensate (\[LL-RR-gCFL\]), this translates into a special role played by the blue quarks.
By starting from the massless case and gradually increasing the value of strange quark mass, one finds that the CFL phase stays robust until a critical value of the control parameter $m_s^2/2\mu \simeq \Delta$ is reached [@gCFL]. Here, $\Delta$ is the gap parameter in the CFL phase. (Strictly speaking, $\Delta\equiv \Delta_1=\Delta_2\approx\Delta_3$ in the CFL phase when $m_s\neq 0$.) Above the critical value, the charge neutrality cannot be accommodated in the CFL phase, and a new gapless phase appears [@gCFL].
A nice feature of the CFL phase is that it stays almost automatically electrically neutral [@enforce_n]. The reason is that Cooper pairing in the CFL phase helps to enforce the equal number densities of all three quark flavors, $n_u=n_d=n_s$. Since the sum of the charges of the up, down and strange quarks add up to zero, this insures that the electric charge density is vanishing, $n_Q=\frac{2}{3}n_u-\frac{1}{3}
n_d-\frac{1}{3}n_s=0$. This is exactly what happens in the CFL phase even at nonzero, but sub-critical values of the strange quark mass.
It is the color rather than the electric charge neutrality that plays an important role in three-flavor quark matter at a nonzero value of $m_s^2/2\mu$. It has been mentioned above that, because of the color-flavor locking, the blue quarks have a special status in the Cooper pairing dynamics. In order to avoid the violation of the color neutrality, then, a nonzero compensating value of the color chemical potential $\mu_8\simeq -m_s^2/2\mu$ is needed [@no2sc]. (Note that, in QCD with dynamical gluons, a nonzero value of the gluon condensate $\langle A_0^{8} \rangle$ will be generated instead [@DD].)
It was shown in Ref. [@gCFL] that the CFL phase becomes gapless when $m_s^2/2\mu > \Delta$. In essence, the mechanism is the same as in two-flavor quark matter, discussed earlier in this lecture. In order to get a slightly better understanding, it is instructive to consider how the strange quark mass interferes with the pairing between blue down quarks and green strange quarks [@gCFL]. After taking into account the shift of the strange quark Fermi momentum in Eq. (\[k\_F\_shift\]), the effective values of the chemical potentials of blue down and green strange quarks become $$\begin{aligned}
\mu_{db} & \simeq & \mu - \frac{2}{3} \mu_8
\simeq \mu + \frac{m_s^2}{3\mu},\\
\mu^{\rm eff}_{sg} & \simeq & \mu + \frac{1}{3} \mu_8
- \frac{m_s^2}{2\mu}
\simeq \mu - \frac{2 m_s^2}{3\mu} ,\end{aligned}$$ where we took into account that $\mu_8\simeq -m_s^2/2\mu$. Up to an effective shift of the strange quark chemical potential, these expressions for the quark chemical potentials are similar to those in Eqs. (\[mu-ur-ug\])–(\[mu-db\]) where $\mu_e=0$. In analogy with the two-flavor case, see Eq. (\[delta-mu\]), we define $$\delta\mu =\frac{\mu_{db}-\mu^{\rm eff}_{sg}}{2} = \frac{m_s^2}{2\mu},$$ as a formal value of the mismatch parameter. As soon as $\delta\mu
>\Delta$, one should get the gapless modes in the quasiparticle spectrum. Their dispersion relations are given by the same expression (\[2-degenerate\]), except that $\bar\mu \to
(\mu_{db}+\mu^{\rm eff}_{sg})/2$.
![The quasiparticle dispersion relations for electric and color neutral color superconducting quark matter at $T=0$, $\mu=500\,\mbox{MeV}$, and $m_s=400\,\mbox{MeV}$, see Ref. [@RSR].[]{data-label="disp-numer"}](epsilon.eps){width="90.00000%"}
A typical result for the low-energy dispersion relations of all nine quasiparticle are shown in Fig. \[disp-numer\] which is taken from Ref. [@RSR]. It should be noted that the two dispersion relations in panel (d) are qualitatively the same as those in Fig. \[disp-rel\] in the gapless 2SC case.
Physical properties of the gCFL phase are very different from those of the CFL phase. The presence of gapless quasiparticle modes with a density of states proportional to $\mu^2$ has a large effect on the thermodynamics as well as on the transport. In contrast to the CFL phase which is an insulator, the gCFL phase is a metal with a nonzero number density of electrons. Also, the neutrino emissivity rate from the gCFL phase should be rather high. It is dominated by the $\beta$-processes: $\tilde{d}_b \to \tilde{u}_b + e^{-} + \bar{\nu}_{e}$ and $\tilde{u}_b + e^{-} \to \tilde{d}_b+\nu_{e}$, where $\tilde{u}_b$ and $\tilde{d}_b$ denote the gapless quasiparticles whose dispersion relations are represented by the solid lines in panels (c) and (d) in Fig. \[disp-numer\]. Relatively large contributions to the neutrino emissivity may also come from the processes with additional gluons: $\tilde{s}_{r,g} \to \tilde{u}_b + e^{-} + \bar{\nu}_{e}+g_{4,5,6,7}$ and $\tilde{u}_b + e^{-} \to \tilde{s}_{r,g} + \nu_{e} + g_{4,5,6,7}$, where $g_{4,5,6,7}$ denotes one of the gluons described by the gauge fields $A^{4}_{\mu}$, $A^{5}_{\mu}$, $A^{6}_{\mu}$ or $A^{7}_{\mu}$.
Mixed phases with color superconducting components
--------------------------------------------------
In addition to homogeneous (one-component) phases, one could also study various mixed phases of dense quark matter. The neutrality in such phases is imposed not locally, but [*globally*]{} (i.e., on average). For example, one can construct a phase which is made of alternating layers of different coexisting components. While each of the component may have a nonzero charge density, the average charge density of the whole construction is vanishing. The Coulomb energy in a mixed phase is nonzero, but it does not grow with the volume of the system. There is also a contribution to the total energy that comes from surfaces separating components of the mixed phase [@RR; @SHH; @mixed]. If the corresponding surface tension is sufficiently large, the mixed phase is disfavored energetically, and therefore it cannot appear. If, on the other hand, the surface tension is small, the mixed phase can be the ground state of matter, replacing a homogeneous gapless phase.
### Gibbs construction
Here is a brief introduction to the general method of constructing mixed phases by imposing the Gibbs conditions of equilibrium [@glen92; @Weber]. From the physical point of view, the Gibbs conditions enforce the mechanical as well as chemical equilibrium between different components in a mixed phase. This is achieved by requiring that the pressure of different components inside the mixed phase are equal, and that the chemical potentials ($\mu$ and $\mu_{e}$) are the same across the whole mixed phase. For example, in application to the mixed phase made of the normal and the 2SC quark components [@SHH], these conditions read $$\begin{aligned}
P^{(NQ)}(\mu,\mu_{e}) &=& P^{(2SC)}(\mu,\mu_{e}),
\label{P=P}\\
\mu &=& \mu^{(NQ)}=\mu^{(2SC)},
\label{mu=mu}\\
\mu_{e} &=& \mu^{(NQ)}_{e}=\mu^{(2SC)}_{e}.
\label{mue=mue}\end{aligned}$$ It is easy to visualize these conditions by plotting the pressure surfaces as functions of chemical potentials ($\mu$ and $\mu_{e}$) for the components of the mixed phase. Graphically, this is shown in Fig. \[fig-front-quark\]. The Gibbs conditions are automatically satisfied along the intersection line of the pressure surfaces (dark solid line in Fig. \[fig-front-quark\]).
![\[fig-front-quark\] Pressure as a function of $\mu\equiv\mu_B/3$ and $\mu_e$ for the hadronic phase (at the bottom), for the 2SC phase (on the right in front) and the normal quark phase (on the left). The dark thick line follows the neutrality line in the hadronic phase, and in two mixed phases: (i) the mixed phase of hadronic and normal quark matter; and (ii) the mixed phase of normal and color superconducting quark matter.](gibbs_const.ps){width="60.00000%"}
Different components of the mixed phase occupy different volume fractions in space. To describe this quantitatively, it is convenient to introduce the volume fraction of each component. As an example, let us consider the quark mixed phase made of the normal and the 2SC component [@SHH]. We denote the volume fraction of the normal phase by $\chi^{NQ}_{2SC}\equiv V_{NQ}/V$ (in general, the notation $\chi^{A}_{B}$ means “the volume fraction of phase A in a mixture with phase B"). Then, the volume fraction of the 2SC phase is given by $\chi^{2SC}_{NQ}=(1-\chi^{NQ}_{2SC})$. From the definition, it is clear that $0\leq \chi^{A}_{B} \leq 1$.
The average electric charge density of the mixed phase is determined by the charge densities of its components taken in the proportion of the corresponding volume fractions. Thus, $$n^{(MP)}_{e}(\mu,\mu_e) = \chi^{NQ}_{2SC} n^{(NQ)}_{e}(\mu,\mu_e)
+(1-\chi^{NQ}_{2SC}) n^{(2SC)}_{e}(\mu,\mu_e).$$ If the charge densities of the two components have opposite signs, one can impose the global charge neutrality condition, $n^{(MP)}_{e}=0$. Otherwise, a neutral mixed phase could not exist. In the case of quark matter, the charge density of the normal quark phase is negative, while the charge density of the 2SC phase is positive along the line of the Gibbs construction (dark solid line in Fig. \[fig-front-quark\]). Therefore, a neutral mixed phase exists. The volume fractions of its components are $$\begin{aligned}
\chi^{NQ}_{2SC} &=& \frac{n^{(2SC)}_{e}}{n^{(2SC)}_{e}-n^{(NQ)}_{e}}, \\
\chi^{2SC}_{NQ} &\equiv& 1-\chi^{NQ}_{2SC}=
\frac{n^{(NQ)}_{e}}{n^{(NQ)}_{e}-n^{(2SC)}_{e}}.\end{aligned}$$
After the volume fractions have been determined from the condition of the global charge neutrality, the energy density of the corresponding mixed phase can also be calculated, $$\varepsilon^{(MP)}(\mu,\mu_e) = \chi^{NQ}_{2SC} \varepsilon^{(NQ)}(\mu,\mu_e)
+(1-\chi^{NQ}_{2SC}) \varepsilon^{(2SC)}(\mu,\mu_e).$$ This is essentially all what one needs in order to construct the equation of state of the mixed phase.
### Surface tension and Coulomb forces
In the above construction, the effects of the Coulomb forces and the surface tension between different components of the mixed phase were neglected. In reality, these might be important. In particular, the balance between the Coulomb forces and the surface tension determines the geometries of different components inside the mixed phase.
In the mixed phase made of the normal and 2SC components, for example, nearly equal volume fractions of the two quark phases are likely to form alternating layers (slabs) of matter. The energy cost per unit volume to produce such layers scales as $\sigma^{2/3}(n_{e}^{(2SC)}
-n_{e}^{(NQ)})^{2/3}$ where $\sigma$ is the surface tension [@geometry]. Therefore, the quark mixed phase is a favorable phase of matter only if the surface tension is not too large. The estimates of Ref. [@RR] show that the value of $\sigma$ at the boundary between the normal and the 2SC phases is only about $5$ to $10\,\mbox{MeV/fm}^{2}$. The maximum value of $\sigma$ allowed in the mixed phase, on the other hand, is of order $10$ to $15\,\mbox{MeV/fm}^{2}$ when the value of the quark chemical potential is in the range between $400$ and $500\,\mbox{MeV}$. Therefore, one should conclude that the mixed phase is more favorable than the gapless 2SC phase.
Here, it is fair to mention that the conclusion of Ref. [@RR] may not be final yet. The reason is that the thickness of the layers, which scales as $\sigma^{1/3} (n_{e}^{(2SC)}-n_{e}^{(NQ)})^{-2/3}$ [@geometry], is estimated to be of order $5$ to $10\,\mbox{fm}$ in the mixed phase [@RR; @SHH]. This is comparable to the value of the Debye screening length in each of the two quark phases. Therefore, the effects of charge screening, which have been neglected so far and which are known to increase the surface energy [@screening], may still change the conclusion.
Summary
-------
In recent years there was a lot of progress in understanding the mechanism of color superconductivity and in studying physical properties of various possible color superconducting phases of dense quark matter. At present, one can say that the QCD theory of strong interactions at asymptotic densities is most likely to have a color superconducting ground state. Unfortunately, however, one cannot say with certainty what is the critical value of the baryon density at which color superconductivity first appears. Because of this, one does not know for sure if color superconducting phases could appear at highest densities existing in the Universe.
In nature, very dense baryonic matter exists in compact stars. Thus, it is possible that color superconductivity exists in the central regions of the heaviest compact stars. This conjecture has not yet been (dis-)proved, and it should come under scrutiny in the near future. In this connection, it might be appropriate to mention that there were recent observations on the cooling [@quark1] of one neutron star and the radius [@quark2] of another that led the authors to suggest the presence of quark matter. These suggestions have been disputed [@no-quark1; @no-quark2]. However, the current situation, where there is no unambiguous evidence that deconfined quarks play a role in compact stars, may well change as further observations are made.
One of the main consequences of color superconductivity in dense matter is the appearance of an energy gap in the quasiparticle spectrum. In QCD, the typical value of the gap is $100\,\mbox{MeV}$. The presence of such a large gap is likely to affect many physical properties of matter, which in their turn should be reflected in the observational properties of stars. It is of great interest, therefore, to perform a systematic study of the properties of dense baryonic matter under conditions typical for the interior of stars. The goal of such studies should be a clear theoretical prediction of the most favorable phases of matter inside stars. Also, one should work out the key physical properties of different phases that could have an unambiguous imprint on the observational data from stars. Such a theoretical basis would be crucial in an attempt to discover new (e.g., quark) states of matter in nature.
The current understanding of color superconducting phases of QCD at large baryon chemical potential $\mu$ may also turn out to be useful for developing alternative methods to treat QCD at nonzero $\mu$ on the lattice [@HH; @Hands]. As is known, there are not so many methods to study QCD from first principles. One of them is the method of lattice calculations [@lattice]. As of now, the lattice calculations have produced many results for QCD at finite temperatures and vanishing baryon chemical potential $\mu=0$ [@lattice-res]. There are also some results at nonzero but small values of the baryon chemical potential [@lattice-mu]. The lattice calculations at large nonzero $\mu$ are problematic because of the famous sign problem. At $\mu\neq 0$, the fermion determinant in the functional integral becomes complex, and the standard numerical techniques, which rely on a probability interpretation of the integrand, cannot be easily applied. Thus, new approaches and new ideas are needed.
Building on the knowledge gained from lattice calculations at nonzero temperature as well as from the effective models of QCD at nonzero chemical potential, one would like to get eventually a complete QCD phase diagram in the plane of temperature and quark chemical potential. The knowledge of such a diagram is not only of a theoretical importance. Different regions of the phase diagram describe physical conditions that can be realized, for example, during the evolution of the early Universe or in the heavy ion collisions. There is also a region at nonzero baryon chemical potentials which is responsible for the physics of stars. In order to understand the corresponding physics, one first has to know the locations of the critical lines and the types of the phase transitions. Some basic features of the QCD diagram have already started to emerge [@RSR; @FKR; @Halasz; @Rajpd]. It will probably take many years before the complete phase diagram is understood in detail.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank the organizers of the IARD 2004 conference and the organizers of the Helmholtz International Summer School and Workshop on [*Hot points in Astrophysics and Cosmology*]{} for their invitation to present these lectures. Also, I thank Michael Buballa, Valery Gusynin, Paul Ellis, Matthias Hanauske, Deog-Ki Hong, Mei Huang, Volodya Miransky, Micaela Oertel, Dirk Rischke, Stefan Rüster, Gordon Semenoff, and Rohana Wijewardhana for very fruitful collaboration on various topics in color superconductivity. This work was supported by Gesellschaft für Schwerionenforschung (GSI) and by Bundesministerium für Bildung und Forschung (BMBF).
[99]{}
K. Hagiwara [*et al.*]{} \[Particle Data Group Collaboration\], Phys. Rev. D [**66**]{}, 010001 (2002). D. Ivanenko and D.F. Kurdgelaidze, Astrofiz. [**1**]{}, 479 (1965); Lett. Nuovo Cim. [**2**]{}, 13 (1969); N. Itoh, Prog. Theor. Phys. [**44**]{}, 291 (1970); F. Iachello, W.D. Langer, and A. Lande, Nucl. Phys. A [**219**]{}, 612 (1974). H.D. Politzer, Phys. Rev. Lett. [**30**]{}, 1346 (1973); D.J. Gross and F. Wilczek, Phys. Rev. D [**8**]{}, 3633 (1973); Phys. Rev. D [**9**]{}, 980 (1974). J.C. Collins and M.J. Perry, Phys. Rev. Lett. [**34**]{}, 1353 (1975). J. I. Kapusta, [*Finite-temperature field theory*]{}, (University Press, Cambridge, 1989).
A. Chodos, R. L. Jaffe, K. Johnson, C. B. Thorn, and V. F. Weisskopf, Phys. Rev. D [**9**]{}, 3471 (1974); A. Chodos, R. L. Jaffe, K. Johnson, and C. B. Thorn, Phys. Rev. D [**10**]{}, 2599 (1974). B. A. Freedman and L. D. McLerran, Phys. Rev. D [**16**]{}, 1169 (1977). L.N. Cooper, Phys. Rev. [**104**]{}, 1189 (1956). J. Bardeen, L.N. Cooper, and J.R. Schrieffer, Phys. Rev. [**106**]{}, 162 (1957); Phys. Rev. [**108**]{}, 1175 (1957). K. Rajagopal and F. Wilczek, hep-ph/0011333; D. K. Hong, Acta Phys. Polon. B [**32**]{}, 1253 (2001); M. Alford, Ann. Rev. Nucl. Part. Sci. [**51**]{}, 131 (2001); T. Schäfer, hep-ph/0304281; D. H. Rischke, Prog. Part. Nucl. Phys. [**52**]{}, 197 (2004); H.-c. Ren, hep-ph/0404074; M. Huang, hep-ph/0409167. B.C. Barrois, Nucl. Phys. [**B129**]{}, 390 (1977); S.C. Frautschi, in “Hadronic matter at extreme energy density", edited by N. Cabibbo and L. Sertorio (Plenum Press, 1980).
D. Bailin and A. Love, Nucl. Phys. [**B190**]{}, 175 (1981); Nucl. Phys. [**B205**]{}, 119 (1982); Phys. Rep. [**107**]{}, 325 (1984). M. Alford, K. Rajagopal, and F. Wilczek, Phys. Lett. B [**422**]{}, 247 (1998); R. Rapp, T. Schäfer, E. V. Shuryak, and M. Velkovsky, Phys. Rev. Lett. [**81**]{}, 53 (1998). Y. Nambu and G. Jona-Lasinio, Phys. Rev. [**122**]{}, 345 (1961); Phys. Rev. [**124**]{}, 246 (1961). G. W. Carter and D. Diakonov, Phys. Rev. D [**60**]{}, 016004 (1999); J. Berges and K. Rajagopal, Nucl. Phys. B [**538**]{}, 215 (1999); M. Buballa and M. Oertel, Nucl. Phys. A [**703**]{}, 770 (2002); F. Gastineau, R. Nebauer, and J. Aichelin, Phys. Rev. C [**65**]{}, 045204 (2002); M. Huang, P. F. Zhuang, and W. Q. Chao, Phys. Rev. D [**65**]{}, 076012 (2002). M. Buballa, hep-ph/0402234. M. Huang, P. F. Zhuang, and W. Q. Chao, Phys. Rev. D [**67**]{}, 065015 (2003). T. M. Schwarz, S. P. Klevansky, and G. Papp, Phys. Rev. C [**60**]{}, 055205 (1999). D. J. Gross, R. D. Pisarski, and L. G. Yaffe, Rev. Mod. Phys. [**53**]{}, 43 (1981); E. V. Shuryak, Nucl. Phys. B [**203**]{}, 140 (1982); D. T. Son, M. A. Stephanov, and A. R. Zhitnitsky, Phys. Lett. B [**510**]{}, 167 (2001). R. Casalbuoni, Z. Y. Duan, and F. Sannino, Phys. Rev. D [**62**]{}, 094004 (2000). V. A. Miransky, I. A. Shovkovy, and L. C. R. Wijewardhana, Phys. Rev. D [**62**]{}, 085025 (2000); Phys. Rev. D [**64**]{}, 096002 (2001). D. H. Rischke, D. T. Son, and M. A. Stephanov, Phys. Rev. Lett. [**87**]{}, 062001 (2001). W. E. Brown, J. T. Liu, and H.-c. Ren, Phys. Rev. D [**61**]{}, 114012 (2000); Q. Wang and D. H. Rischke, Phys. Rev. D [**65**]{}, 054005 (2002). H. Vija and M.H. Thoma, Phys. Lett. B [**342**]{}, 212 (1995). C. Manuel, Phys. Rev. D [**53**]{}, 5866 (1996). V. A. Miransky, G. W. Semenoff, I. A. Shovkovy, and L. C. R. Wijewardhana, Phys. Rev. D [**64**]{}, 025005 (2001); P. Jaikumar and I. Zahed, Phys. Rev. D [**64**]{}, 014035 (2001). U. W. Heinz, Annals Phys. [**168**]{}, 148 (1986). D. T. Son, Phys. Rev. D [**59**]{}, 094019 (1999). D. K. Hong, V. A. Miransky, I. A. Shovkovy, and L. C. R. Wijewardhana, Phys. Rev. D [**61**]{}, 056001 (2000); T. Schäfer and F. Wilczek, Phys. Rev. D [**60**]{}, 114033 (1999); S. D. H. Hsu and M. Schwetz, Nucl. Phys. [**B572**]{}, 211 (2000). R. D. Pisarski and D. H. Rischke, Phys. Rev. D [**61**]{}, 051501 (2000); Phys. Rev. D [**61**]{}, 074017 (2000). N. Iwamoto, Phys. Rev. Lett. [**44**]{}, 1637 (1980). The pressure correction due to color superconductivity is determined, up to a sign, by the value of the effective potential at its minimum in V. A. Miransky, I. A. Shovkovy, and L. C. R. Wijewardhana, Phys. Lett. B [**468**]{}, 270 (1999). I. A. Shovkovy and P. J. Ellis, Phys. Rev. C [**66**]{}, 015802 (2002); C. Manuel, A. Dobado, and F. J. Llanes-Estrada, hep-ph/0406058. G. Lugones and J. E. Horvath, Phys. Rev. D [**66**]{}, 074017 (2002). M. Alford and S. Reddy, Phys. Rev. D [**67**]{} (2003) 074024. R. C. Tolman, Phys. Rev. [**55**]{}, 364 (1939); J. R. Oppenheimer and G. M. Volkoff, Phys. Rev. [**55**]{}, 374 (1939). S. Banik and D. Bandyopadhyay, Phys. Rev. D [**67**]{} (2003) 123003. M. Baldo, M. Buballa, F. Burgio, F. Neumann, M. Oertel, and H.J. Schulze, Phys. Lett. B [**562**]{} (2003) 153. I. Shovkovy, M. Hanauske, and M. Huang, Phys. Rev. D [**67**]{}, 103004 (2003). D. Blaschke, H. Grigorian, D.N. Aguilera, S. Yasui, and H. Toki, AIP Conf. Proc. [**660**]{} (2003) 209. M. Buballa, F. Neumann, M. Oertel, and I. Shovkovy, Phys. Lett. B [**595**]{}, 36 (2004). A. R. Bodmer, Phys. Rev. D [**4**]{}, 1601 (1971); E. Witten, Phys. Rev. D [**30**]{}, 272 (1984); C. Alcock, E. Farhi, and A. Olinto, Astrophys. J. [**310**]{}, 261 (1986). M. G. Alford, K. Rajagopal, and F. Wilczek, Nucl. Phys. [**B537**]{}, 443 (1999). I. A. Shovkovy and L. C. R. Wijewardhana, Phys. Lett. B [**470**]{}, 189 (1999); T. Schäfer, Nucl. Phys. [**B575**]{}, 269 (2000). T. Schäfer, Nucl. Phys. A [**728**]{}, 251 (2003). P. W. Anderson and N. Itoh Nature [**256**]{}, 25 (1975). I. A. Shovkovy and P. J. Ellis, Phys. Rev. C [**67**]{}, 048801 (2003). C. Vogt, R. Rapp, and R. Ouyed, Nucl. Phys. A [**735**]{}, 543 (2004). J. Gasser and H. Leutwyler, Nucl. Phys. B [**250**]{}, 465 (1985); Annals Phys. [**158**]{}, 142 (1984); G. Ecker, Prog. Part. Nucl. Phys. [**35**]{}, 1 (1995); A. Pich, Rept. Prog. Phys. [**58**]{}, 563 (1995). D. K. Hong, M. Rho, and I. Zahed, Phys. Lett. B [**468**]{}, 261 (1999). R. Casalbuoni and R. Gatto, Phys. Lett. B [**464**]{}, 111 (1999). D. T. Son and M. A. Stephanov, Phys. Rev. D [**61**]{}, 074012 (2000); Erratum: Phys. Rev. D [**62**]{}, 059902 (2000). K. Zarembo, Phys. Rev. D [**62**]{}, 054003 (2000); S. R. Beane, P. F. Bedaque, and M. J. Savage, Phys. Lett. B [**483**]{}, 131 (2000); V. A. Miransky, I. A. Shovkovy, and L. C. R. Wijewardhana, Phys. Rev. D [**63**]{}, 056005 (2001). D. T. Son, M. A. Stephanov, and A. R. Zhitnitsky, Phys. Rev. Lett. [**86**]{}, 3955 (2001). C. Manuel and M. H. G. Tytgat, Phys. Lett. B [**479**]{}, 190 (2000); T. Schäfer, Phys. Rev. D [**65**]{}, 074006 (2002). D. K. Hong, Phys. Lett. B [**473**]{}, 118 (2000); Nucl. Phys. B [**582**]{}, 451 (2000); G. Nardulli, Riv. Nuovo Cim. [**25N3**]{}, 1 (2002); P. T. Reuter, Q. Wang, and D. H. Rischke, nucl-th/0405079. P. F. Bedaque and T. Schäfer, Nucl. Phys. A [**697**]{}, 802 (2002). D. B. Kaplan and S. Reddy, Phys. Rev. D [**65**]{}, 054042 (2002). A. Kryjevski, D. B. Kaplan, and T. Schäfer, hep-ph/0404290. V. A. Miransky and I. A. Shovkovy, Phys. Rev. Lett. [**88**]{}, 111601 (2002); T. Schäfer, D. T. Son, M. A. Stephanov, D. Toublan, and J. J. M. Verbaarschot, Phys. Lett. B [**522**]{}, 67 (2001). M. Iwasaki and T. Iwado, Phys. Lett. B [**350**]{}, 163 (1995). T. Schäfer, Phys. Rev. D [**62**]{}, 094007 (2000); M. G. Alford, J. A. Bowers, J. M. Cheyne, and G. A. Cowan, Phys. Rev. D [**67**]{}, 054018 (2003); M. Buballa, J. Hošek, and M. Oertel, Phys. Rev. Lett. [**90**]{}, 182002 (2003). A. Schmitt, Q. Wang, and D. H. Rischke, Phys. Rev. Lett. [**91**]{}, 242301 (2003); Phys. Rev. D [**69**]{}, 094017 (2004)
A. Schmitt, nucl-th/0405076. D. Blaschke, D. N. Voskresensky, and H. Grigorian, astro-ph/0403171. T. Schäfer and F. Wilczek, Phys. Rev. Lett. [**82**]{}, 3956 (1999); Phys. Rev. D [**60**]{}, 074014 (1999). M. Alford and K. Rajagopal, JHEP [**0206**]{}, 031 (2002). A.W. Steiner, S. Reddy, and M. Prakash, Phys. Rev. D [**66**]{}, 094007 (2002). I. Shovkovy and M. Huang, Phys. Lett. B [**564**]{}, 205 (2003). E. Gubankova, W.V. Liu, and F. Wilczek, Phys. Rev. Lett. [**91**]{}, 032001 (2003). A. Mishra and H. Mishra, Phys. Rev. D [**69**]{}, 014014 (2004). S. B. Rüster and D. H. Rischke, Phys. Rev. D [**69**]{}, 045011 (2004). M. Huang and I. Shovkovy, Nucl. Phys. A [**729**]{} (2003) 835. J.F. Liao and P.F. Zhuang, Phys. Rev. D [**68**]{}, 114016 (2003). M. Alford, C. Kouvaris, and K. Rajagopal, Phys. Rev. Lett. [**92**]{}, 222001 (2004). M. Alford, C. Kouvaris, and K. Rajagopal, hep-ph/0406137. K. Iida, T. Matsuura, M. Tachibana, and T. Hatsuda, Phys. Rev. Lett. [**93**]{}, 132001 (2004). S. B. Rüster, I. A. Shovkovy, and D. H. Rischke, Nucl. Phys. [**A743**]{}, 127 (2004). K. Fukushima, C. Kouvaris and K. Rajagopal, hep-ph/0408322. S. Reddy and G. Rupak, nucl-th/0405054. W.V. Liu and F. Wilczek, Phys. Rev. Lett. [**90**]{}, 047002 (2003). B. Deb, A. Mishra, H. Mishra, and P.K. Panigrahi, Phys. Rev. A [**70**]{}, 011604(R) (2004). W. V. Liu, F. Wilczek, and P. Zoller, cond-mat/0404478. M. M. Forbes, E. Gubankova, W. V. Liu, and F. Wilczek, hep-ph/0405059. E. Gubankova, F. Wilczek, and E. G. Mishchenko, cond-mat/0409088. F. Neumann, M. Buballa, and M. Oertel, [*Nucl. Phys.*]{} [**A714**]{}, 481 (2003). P.F. Bedaque, H. Caldas, and G. Rupak, Phys. Rev. Lett. [**91**]{}, 247002 (2003); H. Caldas, Phys. Rev. A [**69**]{}, 063602 (2004). A. Gerhold and A. Rebhan, Phys. Rev. D [**68**]{}, 011502 (2003); A. Kryjevski, Phys. Rev. D [**68**]{}, 074008 (2003); D. D. Dietrich and D. H. Rischke, Prog. Part. Nucl. Phys. [**53**]{}, 305 (2004). R. L. Stratonovich, Doklady Akad. Nauk S.S.S.R. [**115**]{}, 1097 (1957); \[translation: Sov. Phys. Doklady [**2**]{}, 416 (1958)\]; J. Hubbard, Phys. Rev. Lett. [**3**]{}, 77 (1959). M. Huang and I. A. Shovkovy, Phys. Rev. D [**70**]{}, 051501(R) (2004); hep-ph/0408268. D. H. Rischke, Phys. Rev. D [**62**]{}, 034007 (2000); D. H. Rischke and I. A. Shovkovy, Phys. Rev. D [**66**]{}, 054019 (2002). H. Muther and A. Sedrakian, Phys. Rev. D [**67**]{}, 085024 (2003). A. I. Larkin and Yu. N. Ovchinnikov, Sov. Phys. JETP [**20**]{}, 762 (1965); P. Fulde and R. A. Ferrell, Phys. Rev. [**135**]{}, A550 (1964). M. G. Alford, J. A. Bowers, and K. Rajagopal, Phys. Rev. D [**63**]{}, 074016 (2001); J. A. Bowers, J. Kundu, K. Rajagopal, and E. Shuster, Phys. Rev. D [**64**]{}, 014024 (2001); R. Casalbuoni, R. Gatto, M. Mannarelli, and G. Nardulli, Phys. Rev. D [**66**]{}, 014006 (2002); I. Giannakis, J. T. Liu, and H.-c. Ren, Phys. Rev. D [**66**]{}, 031501 (2002); J. A. Bowers and K. Rajagopal, Phys. Rev. D [**66**]{}, 065002 (2002). K. Rajagopal and F. Wilczek, Phys. Rev. Lett. [**86**]{}, 3492 (2001). N. K. Glendenning, Phys. Rev. D [**46**]{}, 1274 (1992). F. Weber, [*Pulsars as Astrophysical Laboratories for Nuclear and Particle Physics*]{} (Institute of Physics, Bristol, 1999).
H. Heiselberg, C. J. Pethick, and E. F. Staubo, Phys. Rev. Lett. [**70**]{}, 1355 (1993); N. K. Glendenning and S. Pei, Phys. Rev. C [**52**]{}, 2250 (1995); N. K. Glendenning and J. Schaffner-Bielich, Phys. Rev. Lett. [**81**]{}, 4564 (1998). T. Norsen and S. Reddy, Phys. Rev. C [**63**]{}, 065804 (2001); D. N. Voskresensky, M. Yasuhira, and T. Tatsumi, Nucl. Phys. A [**723**]{}, 291 (2003). P. Slane, D.J. Helfand, and S.S. Murray, [*Astrophys. J.*]{} [**571**]{}, L45 (2002). J.J. Drake, et al., [*Astrophys. J.*]{} [**572**]{}, 996 (2002). D.G. Yakovlev, A.D. Kaminker, P. Haensel, and O.Y. Gnedin, [*Astron. & Astrophys.*]{} [**389**]{}, L24 (2002). F.M. Walter and J.M. Lattimer, [*Astrophys. J.*]{} [**576**]{}, L145 (2002). D. K. Hong and S. D. H. Hsu, Phys. Rev. D [**68**]{}, 034011 (2003). S. Hands, I. Montvay, S. Morrison, M. Oevers, L. Scorzato, and J. Skullerud, Eur. Phys. J. C [**17**]{}, 285 (2000); J. B. Kogut, D. K. Sinclair, S. J. Hands, and S. E. Morrison, Phys. Rev. D [**64**]{}, 094505 (2001). K. G. Wilson, Phys. Rev. D [**10**]{}, 2445 (1974); M. Creutz, Phys. Rev. D [**21**]{}, 2308 (1980). F. Karsch, E. Laermann, and A. Peikert, Nucl. Phys. B [**605**]{}, 579 (2001); M. Asakawa, T. Hatsuda, and Y. Nakahara, Prog. Part. Nucl. Phys. [**46**]{}, 459 (2001); F. Csikor, Z. Fodor, S. D. Katz, and T. G. Kovacs, JHEP [**0311**]{}, 070 (2003). F. Karsch, Nucl. Phys. Proc. Suppl. [**83**]{}, 14 (2000); Z. Fodor and S. D. Katz, Phys. Lett. B [**534**]{}, 87 (2002); JHEP [**0203**]{}, 014 (2002); C. R. Allton [*et al.*]{}, Phys. Rev. D [**66**]{}, 074507 (2002). M. A. Halasz, A. D. Jackson, R. E. Shrock, M. A. Stephanov, and J. J. M. Verbaarschot, Phys. Rev. D [**58**]{}, 096007 (1998); M. A. Stephanov, K. Rajagopal, and E. V. Shuryak, Phys. Rev. Lett. [**81**]{}, 4816 (1998); P. de Forcrand and O. Philipsen, Nucl. Phys. B [**642**]{}, 290 (2002). K. Rajagopal, Nucl. Phys. A [**661**]{}, 150 (1999).
[^1]: On leave from Bogolyubov Institute for Theoretical Physics, 03143, Kiev, Ukraine.
[^2]: Lectures delivered at the IARD 2004 conference, Saas Fee, Switzerland, June 12 - 19, 2004, and at the Helmholtz International Summer School and Workshop on [*Hot points in Astrophysics and Cosmology*]{}, JINR, Dubna, Russia, August 2 - 13, 2004.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The one-dimensional repulsive SU$(n)$ Hubbard model is investigated analytically by bosonization approach and numerically using the density-matrix renormalization-group (DMRG) method for $n=3,4$, and $5$ for commensurate fillings $f=p/q$ where $p$ and $q$ are relatively prime. It is shown that the behavior of the system is drastically different depending on whether $q>n$, $q=n$, or $q<n$. When $q>n$, the umklapp processes are irrelevant, the model is equivalent to an $n$-component Luttinger liquid with central charge $c=n$. When $q=n$, the charge and spin modes are decoupled, the umklapp processes open a charge gap for finite $U>0$, whereas the spin modes remain gapless and the central charge $c=n-1$. The translational symmetry is not broken in the ground state for any $n$. On the other hand, when $q<n$, the charge and spin modes are coupled, the umklapp processes open gaps in all excitation branches, and a spatially nonuniform ground state develops. Bond-ordered dimerized, trimerized or tetramerized phases are found depending on the filling.'
author:
- 'E. Szirmai, Ö. Legeza, and J. S[ó]{}lyom'
title: |
Spatially nonuniform phases in the one-dimensional SU$(n)$ Hubbard model\
for commensurate fillings
---
Introduction
============
Recently, the SU$(n)$-symmetric generalization of the standard SU(2) Hubbard model[@Hubb1-4] has been intensively studied theoretically.[@marston; @assaraf; @assaraf_half; @honer; @assaad; @szirmai01; @szirmai02; @buchta_sun] Apart from its theoretical interest this model may mimic strongly correlated electron systems where the orbital degrees of freedom of $d$ and $f$ electrons play important role and these extra degrees of freedom are taken into account by considering $n$-component fermions. On the other hand, ultracold gases in optical lattices may also be simulated by such multi-component models.
The Hamiltonian of the model is usually written in the form $$\begin{split}
\label{eq:ham}
{\mathcal H} & = - t\sum_{i=1}^N\sum_{\sigma=1}^n (c_{i,\sigma}^\dagger
c_{i+1,\sigma}^{\phantom \dagger} + c_{i+1,\sigma}^\dagger
c_{i,\sigma}^{\phantom\dagger}) \\
& \phantom{=\,} + \frac{U}{2}\sum_{i=1}^N
\sum_{\substack{\sigma,\sigma'=1 \\ \sigma \neq \sigma'}}^n n_{i, \sigma}n_{i,\sigma'} ,
\end{split}$$ where $N$ is the number of sites in the chain. The operator $c_{i,\sigma}^\dagger$ ($c_{i,\sigma}^{\phantom \dagger}$) creates (annihilates) an electron at site $i$ with spin $\sigma$, where the spin index is allowed to take $n$ different values. $n_{i,\sigma}$ is the particle-number operator, $t$ is the hopping integral between nearest-neighbor sites, and $U$ is the strength of the on-site Coulomb repulsion. In what follows $t$ will be taken as the unit of energy.
The model behaves as an $n$-component Tomonaga–Luttinger liquid at generic fillings. Other type of behavior may appear at commensurate fillings due to umklapp processes. The possible phases, their nature and the critical coupling where they appear have been studied in detail for two special commensurate fillings of the band, namely for half filling and $1/n$ filling.[@marston; @assaraf; @assaraf_half; @honer; @assaad; @szirmai01; @szirmai02; @buchta_sun] It is well established by now that the ground state is a fully gapped bond-ordered dimerized state in the half-filled case for any $n>2$. Contrary to this, the ground state remains translationally invariant in the $1/n$-filled case, and only the charge mode acquires a gap for $U>U_{\rm
c}$. While Assaraf [*et al.*]{} [@assaraf] argued that $U_{\rm c}$ is finite, our recent numerical work[@buchta_sun] has suggested a much less, perhaps $U_{\rm c}=0$ critical value above which multiparticle umklapp processes become relevant.
It is worth mentioning that the SU($n$) Hubbard model has a rich phase diagram in the attractive case, too.[@rapp; @zhao] The one-third-filled SU(3) model has two distinct phases in the high-dimensional limit. In one of them the fermions form trions while in the other phase a color superfluid state emerges. The existence of these phases is not yet settled in one dimension.
In the present paper, the role of multiparticle umklapp processes will be further analyzed for general commensurate fillings $f=p/q$, where $p$ and $q$ are relatively prime. We try to establish under what conditions the umklapp processes can generate gaps in the charge or spin sectors, and when and how the translational symmetry is broken. To this end, partly analytic, partly numerical procedures will be applied. We will generalize the method used in Ref. \[\] to the one-third-filled SU($n$) model to show analytically that the ground state cannot be spatially uniform. It is trimerized at least in the large-$n$ limit. The numerical work will show that in fact this trimerized state with gapped excitations exists for $n > 3$ already.
In the numerical part, the length-dependence of the entropy of finite blocks of a long chain is studied. Recently, it has been shown that quantum phase transitions can be conveniently studied by calculating some measure of entanglement.[@osborne; @osterloh; @zanardi; @gu; @vidal; @yang; @gu2; @wu; @legeza_qpt] This can either be a local quantity, e.g., the concurrence[@wootters], a global quantity, e.g., the fidelity,[@zanardi2] or the entropy of a block of several sites.[@vidal10] As has been demonstrated recently,[@legeza_incomm] the oscillatory behavior of the block entropy can reveal the position of soft modes in the excitation spectrum of critical systems or the spatial inhomogeneity of gapped models. This will allow us to demonstrate that at commensurate fillings $f=p/q$ the type of ground state of the one-dimensional SU($n$) models depends on whether $q=n$, $q<n$, or $q>n$.
The paper is organized as follows. The oscillatory behavior of the block entropy, the corresponding peaks in its Fourier spectrum, and their relationship to the known properties of the half-filled and $1/n$-filled models are recalled in Sec. II, where some new results necessitating further studies are also given. An analytical investigation of the role of umklapp processes at commensurate fillings is presented in Sec. III and the possibility of spatial inhomogeneity of the ground state is discussed. The numerical results for various fillings are presented in Sec. IV. Finally our findings and conclusions are summarized in Sec. V.
Oscillatory length dependence of the block entropy
==================================================
If a finite block of length $l$ of a long chain of $N$ sites is considered, it is in a mixed state, even if the long chain is in its ground state. The mixed state can be described by a density matrix $\rho_N(l)$ and the corresponding von Neumann entropy is $$s_N(l) = -{\text{Tr}}\big[ \rho_N(l) \ln \rho_N(l) \big] \,.
\label{eq:neumann}$$ It is well known[@vidal10; @korepin1] that this entropy as a function of the block size grows logarithmically if the system is critical and the spectrum is gapless. In addition, the central charge $c$ can be derived[@holzhey; @cardy] from the initial slope of the length dependence of $s_N(l)$, $$s_N(l) = \frac{c}{6}\ln \left[ \frac{2N}{\pi} \sin \left( \frac{\pi l}{N}
\right) \right] + g \,,
\label{eq:cardy}$$ where $g$ is a shift due to the open boundary. It contains a constant term which depends on the ground-state degeneracy and an alternating term decaying with a power of the distance from the boundary.[@affleck; @laflorence] On the other hand for noncritical, gapped models, $s_N(l)$ saturates to a finite value when $l$ is far from the boundaries.
Recently it has been pointed out by some of us [@legeza_incomm] that a wider variety of behavior may be found for the length-dependence of the block entropy. Namely, we have shown that in some cases oscillations may appear in $s_N(l)$. This can be best analyzed by considering the Fourier transform $$\tilde{s}(k) = \frac{1}{N}\sum_{l=0}^N e^{- i k l}s_N(l)
\label{eq:sq}$$ for discrete wave numbers, $k=2\pi j/N$ lying in the range $(-\pi,\pi)$. Since $s_N(l) = s_N(N-l)$, $\tilde{s}(k)$ is real. It has a large peak at $k=0$, all other Fourier components are negative. Peaks in $|\tilde{s}(k)|$ carry information about the position of soft modes or the spatial inhomogeneity of the ground state. More precisely if the amplitude of a peak at a nonzero wave number $k^*$ remains finite in the thermodynamic limit, this indicates a periodic spatial modulation of the ground state with wavelength $\lambda=2\pi/k^{\ast}$. On the other hand, if a marked peak appears in $|\tilde{s}(k)|$ but its amplitude vanishes as $N\rightarrow\infty$, this allows to identify the wave vector of soft modes in critical models.
In a recent work [@buchta_sun] we have shown that such oscillations appear in $s_N(l)$ for the SU($n$) Hubbard model as well. The periodicity depends on both $n$ and the band filling $f=p/q$. This is shown for the $n=3$ and $n=4$ models for the $1/n$-filled and half-filled cases in Fig. \[fig:s\_l\_fig1\] for a large value of $U$ $(U=10)$.
![Block entropy $s_N(l)$ of finite chains with $N=18$ and $N=16$ site, respectively, for $n=3$ and 4 at fillings $f=1/n$ and $f=1/2$ for $U=10$. The solid line is our fit using Eq. (\[eq:cardy\]).[]{data-label="fig:s_l_fig1"}](sun_sl_fig1.eps)
In the $1/n$-filled cases, $s_N(l)$ increases logarithmically with the block length (and then goes down as $l$ approaches $N$). When every third or fourth values are taken, depending on the periodicity, these selected values can be fitted to (\[eq:cardy\]) as shown by the solid lines in panels (a) and (c). This indicates gapless behavior and gives $c=n-1$. This is in agreement with the theoretical expectation, since the charge mode becomes gapped due to multiparticle umklapp processes and only the $n-1$ spin modes are gapless. A distinct behavior is found in half-filled systems, as seen in panels (b) and (d). The quantity $s_N(l)$ oscillates with period two, and if only every second point is taken, it seems to saturate beyond some block length, before decreasing again, indicating that the corresponding models are fully gapped.
The finite-size dependence of the peaks of $|\tilde{s}(k)|$ appearing at $k^{\ast}= 2k_{\text{F}} = 2 \pi f$ characterizing the oscillation is shown in Fig. \[fig:sq\_n\]. It is seen that in the $1/n$-filled case the Fourier components at $k^{\ast}=2\pi/n$ vanish in the thermodynamic limit, while a finite value is obtained at $k^{\ast}=\pi$ for half-filled models. This corroborates our finding that the $1/n$-filled SU($n$) models are critical with a spatially uniform ground state while a gapped bond-ordered dimerized phase appears at half filling.
![Finite-size dependence of $|\tilde s(k^*)|$ for various $n$ and fillings for $U=10$. The solid line is the finite-size-scaling fit.[]{data-label="fig:sq_n"}](sun_sq_n.eps)
We have done similar calculations for more general commensurate fillings of the band. Figure \[fig:s\_l\_fig2\] shows the results obtained for $n=3$, $f=2/5$ as well as for $n=4$, $f=1/3$. When every fifth points are taken for the $f=2/5$ filled SU(3) model, they can be fitted to yielding $c=3$ for the central charge. This indicates that all modes, including the charge mode, are gapless.
![Same as Fig. \[fig:s\_l\_fig1\] but for (a) $n=3, f=2/5$, $N=20$ and (b) $n=4, f=1/3$, $N=24$.[]{data-label="fig:s_l_fig2"}](sun_sl_fig2.eps "fig:") $l$ $l$\
Such a fit does not work for the one-third-filled SU(4) model. To better see the difference $|\tilde{s}(k)|$ is considered again. The amplitude of the Fourier component at $k^{\ast}= 4\pi/5$, also displayed in Fig. \[fig:sq\_n\] for the $n=3$ model, vanishes in the $N\rightarrow\infty$ limit. On the other hand $|\tilde{s}(k)|$ remains finite at $k^{\ast} = 2\pi/3$ in the one-third-filled $n=4$ model.
When the same calculations are repeated for the $n=5$ model at $f=1/2, 1/3,
1/4$, and $1/5$, peaks appear in $|\tilde s(k)|$ at $k^*=\pi, 2\pi/3$, $\pi/2$, and $2\pi/5$, respectively. As is seen in Fig. \[fig:su5\_sq\_n\], the amplitude of the peaks remains finite even when $N\rightarrow\infty$ in the first three cases, while it vanishes in the last case.
![Same as Fig. \[fig:sq\_n\] but for $n=5$. The solid line is the finite-size-scaling fit.[]{data-label="fig:su5_sq_n"}](su5_sq_n.eps)
These results indicate that the role of umklapp processes depends on the relationship between the number of components $n$ and the relative primes $p$ and $q$ characterizing the commensurate filling. In what follows this problem will be studied first analytically in a bosonization approach and large-$n$ expansion technique, and then numerically using the DMRG method.
Analytical considerations
=========================
The role of umklapp processes: a bosonization approach
------------------------------------------------------
Following the usual procedure we write the Hamiltonian in momentum space and linearize the free-particle spectrum around the two Fermi points $(\pm k_{\text{F}})$. The underlying assumption is that the low-lying excitations determine the physics of the system. Depending on whether the momentum of the fermions is close to $+k_{\text{F}}$ or $-k_{\text{F}}$, one can distinguish left- and right-moving particles, and the interaction processes also can be classified on the basis of whether the incoming and scattered particles are right or left movers and the momentum transfer is small (forward scattering) or large, of the order of $2k_{\text{F}}$ (backward scattering). In a generic model the strength of the various scattering processes may be different. For the sake of simplicity we neglect chiral processes in which both particles move in the same direction before and after the interaction, since they lead to the renormalization of the Fermi velocity only.
One can recognize that at generic fillings, where umklapp processes do not play a role, the forward and backward scattering processes can be interpreted as current–current interactions and their contribution to the Hamiltonian density can be conveniently rewritten using Dirac fermions[@zinn-justin] in the following short form (automatic summation for the repeated indices is understood): $$\label{eq:intham}
H_{\textrm{int}}(x) = {\textstyle{\frac{1}{2}}}
g_{\sigma_1 \sigma_2 \sigma_3 \sigma_4}\bar\psi_{\sigma_1}(x)\gamma_\mu
\psi_{\sigma_2}(x)\bar\psi_{\sigma_3}(x)\gamma_\mu\psi_{\sigma_4}(x) .$$ Here $\sigma_i$ denote the spin indices that can take the values $1,\ldots, n$, $\gamma_\mu$ with $\mu=1,2$ are the Dirac matrices, in our case the standard Pauli matrices ($\sigma_x$, $\sigma_y$), and $\bar\psi(x)=\psi^\dagger(x)\gamma_1$. While the Hubbard model contains a single interaction parameter $U$, the couplings $g_{\sigma_1 \sigma_2 \sigma_3 \sigma_4}$ may be different for physically different processes in more realistic models. In the renormalization-group treatment we will assume this to be the case. It is assumed, however, that the spin of the fermions does not change in the scattering process and the couplings are symmetric under the exchange $(\sigma_1, \sigma_2) \leftrightarrow (\sigma_3, \sigma_4)$. If the fermion field $\psi_{\sigma}(x)$ is decomposed into left- and right-moving components according to $$\psi_{\sigma}(x)=\begin{pmatrix}
R_{\sigma}(x) \\ L_{\sigma}(x)
\end{pmatrix}, \hskip 0.4cm
\hskip 0.4cm
\bar\psi_{\sigma}(x)=\left( L^\dagger_{\sigma}(x), \, R^\dagger_{\sigma}(x) \right),$$ the usual backward- and forward-scattering processes are in fact recovered. In the standard $g$-ology[@solyom] notation $g_{\sigma \sigma' \sigma' \sigma}$ is denoted by $-g_1$, and $g_{\sigma \sigma \sigma' \sigma'}$ by $g_2$.
The well-known renormalization-group (RG) equations, the $\beta$ function can be written for these scattering processes in a short form:[@zinn-justin] $$\begin{split}
\frac{\partial \ln g_{\sigma_1 \sigma_2 \sigma_3 \sigma_4}}{\partial \ln \Lambda'/\Lambda} \equiv
\beta_{\sigma_1 \sigma_2 \sigma_3 \sigma_4 } & =
g_{\sigma_1 \sigma_i \sigma_3 \sigma_j}g_{ \sigma_i \sigma_2 \sigma_j \sigma_4} \\
& \phantom{=\, }-
g_{\sigma_1 \sigma_i \sigma_j \sigma_4}g_{\sigma_i \sigma_2 \sigma_3 \sigma_j},
\end{split}$$ where $\Lambda$ is the cut-off parameter. These RG equations have been analyzed earlier[@marston; @szirmai01] and it was found that the backward-scattering processes scale out at generic fillings in the SU($n$) Hubbard model and for this reason this model is equivalent to an $n$-component Luttinger liquid in this case. The Hamiltonian can be diagonalized[@szirmai02] and the excitation spectrum can be determined exactly in bosonic phase-field representation.[@bozonizacio] There is one symmetric combination of the phase fields with different spin indices, this is the so-called charge mode: $$\label{eq:modec}
\phi_\mathrm{c}(x) = \frac{1}{\sqrt{n}}\sum_{\sigma=1}^n\phi_{\sigma}(x),$$ while the $n-1$ antisymmetric combinations give the spin modes: $$\label{eq:modes}
\phi_{m\mathrm{s}}(x) = \frac{1}{\sqrt{m(m+1)}}\left[ \sum_{\sigma=1}^m
\phi_{\sigma}(x) - m\phi_{m+1}(x) \right]$$ with $m=1,\ldots ,n-1$.
Similarly to the spin-charge separation in the two-component Luttinger model, one finds complete mode separation. The Hamiltonian density is the sum of the contributions of the individual modes, $$H(x) = \sum_j H_j(x) \,,$$ where $j={\text{c}},1{\text{s}},2{\text{s}},\ldots ,(n-1){\text{s}}$. Each term has the usual bosonic form: $$\label{eq:newham}
H_j(x) = \frac{\hbar u_j}{2}\left\{ \pi K_j \left[\Pi_j(x)\right]^2
+ \frac{1}{\pi K_j} \left[\partial_x\phi_j(x)\right]^2 \right\} ,$$ where $\Pi_j(x)$ is the momentum canonically conjugated to $\phi_j(x)$. The renormalized velocities and the Luttinger parameters can be given in terms of the new couplings $g_{2;j}$ appearing after diagonalization[@szirmai02] in the spin indices:
u\_j = & v\_,\
K\_j = & .
In a finite system, where the momentum is quantized in units of $2\pi/L$, the excitation spectrum of the Luttinger model can be written as: $$\label{eq:Lutt-Ham-1}
E = \sum_j \hbar u_j \frac{2\pi}{L} \left( n_+^j + n_-^j + \Delta_+^j +
\Delta_-^j \right),$$ where $n_\pm ^j$ are integers describing the bosonic excitations and $$\label{eq:Lutt-Ham-2}
\Delta_\pm^j = \frac{1}{16} \Big( \sqrt{K_j} J_j \pm
\frac{1}{\sqrt{K_j}} \delta N_j \Big)^2 \,,$$ where $\delta N_j$ is the change in the number of particles in the $j$th channel, and similarly $J_j$ describes the current in the $j$th channel created by adding particles to or removing them from the branches of the dispersion relation.
The total momentum is given by $$P=\hbar k_\textrm{F} J_c + \sum_j \hbar \frac{2 \pi}{L}\left(n_+^j -
n_-^j + \Delta_+^j - \Delta_-^j \right).$$ Thus soft modes appear not only at zero momentum but at integer multiples of $2k_\textrm{F}$, too, since the charge current $J_c$ is an even number if the total charge is conserved. Since $$2 k_\textrm{F}=\frac{N_{\text{c}}^0}{n}\frac{2 \pi}{L} \,,$$ where $N_{\text{c}}^0$ is the number of particles in the system, and the filling of the band is $f = N_{\text{c}}^0/nN$, the position of these soft modes depends on the filling only, $2k_{\text{F}}= 2 \pi f$.
We know that the usual umklapp processes, scattering of two right movers into left-moving states or vice versa, that were neglected so far, are relevant in a half-filled system for any $n$.[@marston; @szirmai01] Multiparticle umklapp processes may become relevant at other commensurate fillings. To see what kind of processes are allowed we have to take into account that the total quasimomentum transferred in an umklapp process has to be an integer multiple of $2\pi$. If the band is $f=p/q$ filled and consequently $k_{\text{F}} = p\pi/q$, $q$ particles have to be scattered from one Fermi point to the opposite one to satisfy this condition. This is illustrated in Fig. \[fig:umklappok\]. Since there are $n$ different types of fermion in the SU($n$) model, and the Hubbard interaction is local, only such umklapp processes are allowed by the Pauli principle in which the number of scattered particles is less than or equal to $n$. Thus $q$-particle umklapp processes are forbidden in our model when $q > n$.
![Umklapp processes in the SU(3) symmetric Hubbard model. (a) Any two of the three types of fermion can be scattered in the half-filled case. (b) Fermions with the three possible spin orientations participate in the scattering process in the one-third-filled case.[]{data-label="fig:umklappok"}](umklappok.eps)
The role of multiparticle umklapp processes in the conductivity of the SU(2) Hubbard model has been studied in Ref. \[\]. A different aspect, whether the charge and spin modes are coupled or not by the possible multiparticle umklapp processes has been considered for the SU($n$) model.[@buchta_sun] The $q$-particle umklapp processes can be described in the bosonic phase-field representation by $$\label{eq:umkl}
H_\textrm{U} (x) = g_3 \int \textrm{d}x \sum_{\{\sigma_i\}'}
\cos\big[2\left(\phi_{\sigma_1}(x)+\ldots +\phi_{\sigma_q}(x)\right)\big] \,.$$ The phase fields appearing here are the phases of the bosonic representation of the particles participating in the scattering process, and $\{\sigma_i\}'$ indicates that all spin indices have to be different.
This Hamiltonian can be expressed in terms of the phase fields corresponding to the charge and spin modes, making use of the inverse of (\[eq:modec\]) and (\[eq:modes\]). It is clear that if $q=n$, only the symmetric combination of the phase fields, that is the charge mode defined in (\[eq:modec\]) appears in . This means that in this case the $q$-particle umklapp processes couple to the charge mode only. The Hamiltonian density of the charge mode is identical to that of the well-known sine-Gordon model which has a fully gapped excitation spectrum. Thus the charge mode becomes gapped for finite $U$. The $n-1$ spin modes are not influenced by the umklapp processes in the $1/n$-filled case. They remain gapless and the central charge is $c=n-1$.
When $q<n$, the sum of the $q$ phase fields corresponding to the $q$ particles required for umklapp processes, will contain various combinations of the $n$ boson fields, leading to a mixing (coupling) of the charge and spin modes, thus opening gaps in all modes. The model becomes noncritical for $U >
U_{\text{c}}$.
When on the other hand $q>n$, umklapp processes are forbidden by Pauli’s exclusion principle when the interaction is local. The charge and spin modes remain gapless, thus $c=n$ for $U>0$. The expected behavior for different cases is summarized in Table \[tab:analitikus\].
$n$ $c$ phase $k^*$
------- -- ----------- -- ------- -- ------------ -- ------------
$q=n$ any $n$ $n-1$ C0S$(n-1)$ $2\pi p/n$
$q<n$ $n\neq 2$ – C0S0 $2\pi p/q$
$q>n$ any $n$ $n$ C1S$(n-1)$ $2\pi p/q$
: Central charge $c$ and the type of phase characterized by the number of soft modes in the charge and spin sectors (CxSy) for the $p/q$-filled SU($n$) Hubbard model.[]{data-label="tab:analitikus"}
Spatial inhomogeneity in the large-$n$ limit of the SU($n$) Hubbard-Heisenberg model at one-third filling
---------------------------------------------------------------------------------------------------------
We have seen that the spectrum is fully gapped when $q<n$. The question naturally arises whether the opening of a gap at multiples of $k^*=2k_{\text{F}}$ is related to a breaking of the translational symmetry, an instability against the formation of a spatially inhomogeneous state with the corresponding wave number.
To analyze the stability of the homogeneous state we generalize the procedure used by Marston and Affleck[@marston] to the one-third-filled case in the large-$U$ limit. When $n$ is an integer multiple of 3, the number of fermions sitting on each site is an integer in a homogeneous sample, and a finite energy is needed to add an extra particle. This energy gap at $k_{\text{F}}$ may imply a tripling of the spatial period in the ground state. To search for this spatial inhomogeneity a more general model, the SU($n$) symmetric generalization of the Hubbard-Heisenberg model will be considerd. Its Hamiltonian is $$\begin{aligned}
\label{eq:hub-hei-ham}
H &=&\sum_{i=1}^N \bigg[ - \frac{J}{n} \sum_{\sigma,\sigma'=1}^n (c_{i,\sigma}^\dagger
c^{\phantom \dagger}_{i+1,\sigma})(c_{i+1,\sigma'}^\dagger c^{\phantom \dagger}_{i,\sigma'}) \\
&- & t\sum_{\sigma=1}^n (c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i+1,\sigma} +
\text{h.c.} ) + \frac{U}{n} \bigg( \sum_{\sigma=1}^n c_{i,\sigma}^\dagger
c^{\phantom \dagger}_{i,\sigma} - \frac{n}{3} \bigg)^2 \bigg]. \nonumber\end{aligned}$$ The chemical potential is shifted to zero at one-third filling, and the Hubbard coupling $U$ and Heisenberg coupling $J$ are rescaled by $2/n$ so that the spacing of the energy levels remain the same as $n$ increases. We note that the usual $J\sum_{<i,j>}\mathbf{S}_i\mathbf{S}_j$ nearest-neighbor Heisenberg interaction breaks up into three terms in fermionic representation of the spin operators. One of them only shifts the chemical potential, another corresponds to nearest-neighbor Coulomb interaction (that is unimportant in the large $U$ limit), therefore they are neglected.
The equilibrium state will be determined from the minimum of the free energy that can be derived from the partition function of the system. In functional integral formalism the partition function can be expressed with the Lagrangian of the model. At finite temperatures the imaginary time Lagrangian is $L[c,c^\dagger]=\sum_{i,\sigma} \big(c_{i,\sigma}^\dagger (\textrm{d}/\textrm{d}\tau)
c^{\phantom \dagger}_{i,\sigma} + H\big)$ and the partition function is $$\label{eq:Z}
Z=\int [\textrm{d}c][\textrm{d}c^\dagger ]\textrm{exp}\bigg( -\int_0^\beta
\textrm{d}\tau \, L[c,c^\dagger] \bigg).$$ Here $\beta$ is the inverse temperature. The integral occurring in the above expression cannot be calculated in a simple way due to the quartic terms in the Lagrangian. However, these quartic terms can be eliminated by a Hubbard-Stratonovich transformation based on the integral identity $$\label{eq:hub-strat}
\textrm{exp}(VX^2) \propto \int \textrm{d}Y \textrm{exp} (-Y^2/4V + XY).$$ In our case the quantities corresponding to $X$ are $\sum_{\sigma}
c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i,\sigma}$ and $\sum_{\sigma} c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i+1,\sigma}$. Therefore we have to introduce $2N$ bosonic fields: $\phi_i$ and $\chi_{i,i+1}$. We note that the fields $\phi_i$ and $\chi_{i,i+1}$ correspond[@marston] to site- and bond-centered densities, respectively.
Adding the appropriate terms to the Lagrangian gives $$\begin{gathered}
\label{eq:lagr-def}
L[c,c^\dagger,\phi,\chi] \stackrel{\textrm{def}}{=} L[c,c^\dagger] \\
+ \frac{n}{U} \sum_i \bigg[ \frac{\phi_i}2 + i\frac{U}{n} \Big(\sum_{\sigma}
c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i,\sigma} - \frac{n}{3}\Big) \bigg]^2 \\ +
\frac{n}{J} \sum_i \bigg| \frac{1}{2}\chi_{i,i+1} + \frac{J}{n} \sum_{\sigma}
c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i+1,\sigma} \bigg|^2.\end{gathered}$$ The new Lagrangian is quadratic in the fields, however the fermionic fields are coupled to the bosonic ones. The explicit form of the Lagrangian is $$\begin{gathered}
\label{eq:lagr}
L [ c,c^\dagger ,\phi ,\chi ] = \sum_{i,\sigma} \bigg\{ c_{i,\sigma}^\dagger
\Big(\frac{\textrm{d}}{\textrm{d}\tau} + i \phi_i \Big) c^{\phantom \dagger}_{i,\sigma}
+\frac{1}{4J} | \chi_{i,i+1} |^2 \\
+ \big[ ( \chi_{i,i+1} - t ) c_{i,\sigma}^\dagger c^{\phantom \dagger}_{i+1,\sigma} + h.c. \big]
+ \frac{1}{4U} \phi_i^2 - i \frac{1}{3} \phi_i \bigg\}.\end{gathered}$$ The partition function is obtained by integrating over $c$ and $c^\dagger$: $$\begin{gathered}
Z[\phi,\chi] = \int [\textrm{d}c][\textrm{d}c^\dagger ]\textrm{exp}\bigg(
-\int_0^\beta \textrm{d}\tau \, L[c,c^\dagger,\phi,\chi] \bigg) \,.\end{gathered}$$ Writing it in the form $Z[\phi,\chi]\equiv
\textrm{exp}(-S_{\textrm{eff}}[\phi,\chi])$, this defines the effective action. The free energy can be expressed in a usual way via $Z[\phi,\chi]$ as $$F[\phi,\chi]=-1/\beta \, \, \textrm{ln}\big( Z[\phi,\chi] \big).$$
As mentioned in the previous subsection, if spatial oscillations occur in the system, they are expected to appear with wave number $k^*=2\pi f$. Thus we may expect spontaneous trimerization at one-third filling. Therefore we suppose that the boson field $\phi_i$ takes three different values depending on whether $i=3l$, $i=3l+1$ or $i=3l+2$ with integer $l$. They will be denoted as $\phi_1$, $\phi_2$, and $\phi_3$. Similar assumption holds for the fields $\chi_{i,i+1}$, too. The three values are $\chi_1$, $\chi_2$, and $\chi_3$. The lattice is thus decomposed into three sublattices.
To get a real free energy the fields $\phi_\alpha$ ($\alpha=1,2,3$) are redefined by continuing to the complex plane ($i\phi_\alpha\rightarrow
\phi_\alpha$). The free energy can then be written as $$\begin{gathered}
\label{eq:free-en}
F(\{\phi_\alpha \},\{\chi_\alpha \})= \frac13 \sum_{\alpha=1}^3 \left(
\frac{Nn}{4J}\chi_\alpha^2 + \frac{Nn}{4U} \phi_\alpha^2 -
\frac{Nn}{3}\phi_\alpha \right) \\ + \frac n3 \sum_k \Big( E(k) - 1/3 \Big),\end{gathered}$$ and the summation for $k$ has to be performed over the reduced Brillouin zone which is now one-third of the original one ($k$ runs from $-\pi/3$ to $\pi/3$) and $E(k)$ is the energy spectrum of a single fermion coupled to the boson fields. It is the eigenvalues of the Hamiltonian $$\begin{gathered}
\label{eq:sp-Ham}
H = \sum_i\Big[ \Big( (\chi_1-t)a_i^\dagger b^{\phantom \dagger}_{i+1} +
(\chi_2-t)b_{i+1}^\dagger c^{\phantom \dagger}_{i+2} \\ + (\chi_3-t)c_{i+2}^\dagger
a^{\phantom \dagger}_{i+3} + {\text{h.c.}} \Big) \\
+ \phi_1 a_i^\dagger a^{\phantom \dagger}_i + \phi_2 b_{i+1}^\dagger
b^{\phantom \dagger}_{i+1} + \phi_3 c_{i+2}^\dagger c^{\phantom \dagger}_{i+2} \Big] ,\end{gathered}$$ where the operators $a$, $b$ and $c$ belong to different sublattices. In order to determine the one-particle spectrum one has to diagonalize Hamiltonian in momentum space. Therefore, we are looking for the eigenvalues of the matrix $$\label{eq:matrix}
\begin{pmatrix}
\phi_1 & (\chi_1 - t){\text{e}}^{-ik} & (\chi_3 - t){\text{e}}^{ik} \\
(\chi_1 - t){\text{e}}^{ik} & \phi_2 & (\chi_2 - t){\text{e}}^{-ik} \\
(\chi_3 - t){\text{e}}^{-ik} & (\chi_2 - t){\text{e}}^{ik} & \phi_3
\end{pmatrix}.$$ The energy spectrum $E(k)$ has three branches corresponding to the three (real) solutions of the third-order eigenvalue equation:
\[eq:spectrum\] & E\_1(k) - (\_1 + \_2 + \_3)/3 =\
& -(Q) ,\
& E\_[2,3]{}(k) - (\_1 + \_2 + \_3)/3 =\
& -(Q) .
Here $P$ and $Q$ are the parameters in the eigenvalue equation when transformed to the form $\tilde{E}^3(k) + P \tilde{E}(k) + Q =0$ with $\tilde{E}(k)=E(k)-(\phi_1 + \phi_2 + \phi_3)/3$.
The minimum of the free energy of a system with such a spectrum cannot be evaluated quite generally. Fortunately, we are only interested whether the three $\phi_{\alpha}$ and $\chi_{\alpha}$ are different or not. This analysis can be carried out easier in terms of the linear combinations:
:= & (\_1 + \_2 + \_3)/3,\
\_[1]{} := & (\_1 - \_2 )/2,\
\_[2]{} := & (\_1 +\_2 - 2\_3)/6,
and similar definitions for $\chi$, $\Delta\chi_1$ and $\Delta\chi_2$. One finds that although the free energy has an extremum at $\Delta\phi_{1}=\Delta\phi_{2}=0$ and $\Delta\chi_{1}=\Delta\chi_{2}=0$, the free energy of the uniform state is not a local minimum. Thus a density-wave has to appear in the system. We can conclude that the SU($n$) Hubbard model is unstable against the Heisenberg coupling in the large-$n$ limit and exhibits inhomogeneous spatial ordering for $U>0$.
>From this analysis alone – due to the rather complicated one-particle spectrum – we cannot decide whether the system is dimerized, trimerized, or some other periodicity occurs, and whether the density wave is site-centered or bond-centered. It is natural to relate the nonuniform phase to the fully gaped excitation spectrum. Thus when our previous considerations are taken into account, trimerized phase is expected in one-third-filled models. This will be supported by the numerical calculation. We will also see, that the trimerized phase is not a special feature of the systems with integer number of electrons per site. It occurs in one-third-filled system for arbitrary $n>3$.
Numerical study of the spatial inhomogeneity
============================================
In this section we present our numerical results obtained by the DMRG[@white] method for the length dependence of the block entropy $s_N(l)$ and its Fourier transform $\tilde{s}(k)$ to relate them to the number and position of soft modes when the model is critical, or to spatial inhomogeneity of the ground state for gapped models.
The spatial modulation of the ground state can be a site- or a bond-centered density wave. A site-centered density wave would manifest itself in an oscillation of the von Neumann entropy of single sites, $s_i$, with $i=1,\dots, N$ or in the local electron density defined by $$\langle n_{i} \rangle = \sum_{a=1}^n \langle \Psi_{\rm GS}| n_{i,a} |
\Psi_{\rm GS} \rangle\,,
\label{eq:n_i}$$ where $|\Psi_{\rm GS} \rangle$ is the ground-state wavefunction. The wave number of the charge oscillation can again be determined from peaks in the Fourier transform of $s_i$ or $\langle n_i \rangle$ denoted as $s_1(k)$ and $n(k)$, respectively.
The existence of a bond-centered density wave can be demonstrated by studying the variation of the bond energy or the two-site entropy along the chain. To avoid boundary effects we have calculated the difference of two-site entropies in the middle of the chain, between first, second, third and so on neighbor bonds:
D\_s(N) = & s\_[N/2,N/2+1]{}-s\_[N/2+1,N/2+2]{} , \[eq:d\_s\]\
T\_s(N) = & s\_[N/2,N/2+1]{}-s\_[N/2+2,N/2+3]{} , \[eq:t\_s\]\
Q\_s(N) = & s\_[N/2,N/2+1]{}-s\_[N/2+3,N/2+4]{} , \[eq:q\_s\]\
P\_s(N) = & s\_[N/2,N/2+1]{}-s\_[N/2+4,N/2+5]{} . \[eq:p\_s\]
For convenience the number of sites in the chain was always even. Moreover, since we expect dimerized, trimerized or tetramerized phases depending on the commensurate filling $p/q$, the number of sites $N$ was always taken to be an integer multiple of $q$.
When a doubling of the lattice periodicity of the ground state is indicated by a finite peak in $|\tilde s(k)|$ at $k^*=\pi$, a truly dimerized phase gives equal finite values for $D_s$ and $Q_s$ and vanishing $T_s$ and $P_s$ in the $N\rightarrow\infty$ limit. Stronger and a weaker bonds alternate along the chain. When the peak in $|\tilde s(k)|$ appears at $k^*=2\pi/3$ and a trimerized phase is expected, $T_s$ should be finite and $Q_s$ should vanish. Symmetry considerations imply that two equally strong bonds are followed by a weaker or stronger bond in this case. In a tetramerized phase, the peak in $|\tilde s(k)|$ appears at $k^*=\pi/2$, $D_s$, $T_s$, and $Q_s$ may be finite in the $N\rightarrow\infty$ limit and only $P_s$ vanishes necessarily.
The numerical procedure
-----------------------
The numerical calculations presented in this paper have been performed on finite chains with open boundary condition (OBC) using the DMRG technique and the dynamic block-state selection (DBSS) approach.[@legeza_dbss; @legeza_qdc] We have set the threshold value of the quantum information loss $\chi$ to $10^{-5}$ for $n=3,4$ and to $10^{-4}$ for $n=5$ and the minimum number of block states $M_{\rm min}$ to $256$. In spite of the large number of degrees of freedom per site in the $n=5$ case the entropy analysis allows one to study this problem as well. The ground state has been targeted using four to eight DMRG sweeps until the entropy sum rule has been satisfied. The accuracy of the Davidson diagonalization routine has been set to $10^{-7}$ and the largest dimension of the superblock Hamiltonian was around three millions. As an indication of the computational resources used in the present work we note that the maximum number of block states was around 1600 for $n=3$ and 900 for $n=4$ and $5$.
The large-$N$ limit of the entropies and amplitudes of the peaks in the Fourier spectrum can be obtained if appropriate scaling functions are used. In a critical, gapless model, in leading order, these are expected to scale to zero as $1/N$ while in a noncritical model the scaling function depends on the boundary condition. Therefore for any quantity $A$ the finite-size scaling ansatz $$A(N) = A_0 + a/{N^\beta}
\label{eq:scale_obc}$$ is used to evaluate the data obtained with OBC, where $A_0$, $a$, $\beta$ are free parameters to be determined by the fit.
The numerical results
---------------------
### Models with $q=n$
The $1/n$-filled case ($q=n$) has already been considered in Ref. and some of the results were listed in Sec. II. As has been shown in Fig. \[fig:s\_l\_fig1\], $s_N(l)$ oscillates with period $n$ for finite systems. These oscillations are due to the soft modes located at wave numbers $k^{\ast} = 2\pi/n$. Taking every $n$th value only, $s_N(l)$ can be fitted accurately using (\[eq:cardy\]) for relatively short chains already if $U$ is large. After a proper finite-size scaling, the fit gives $c=n-1$, as expected.
Taking the Fourier transform of $s_N(l)$, besides the large positive peak at $k=0$, additional negative peaks are found at the positions of the soft modes, at $k^*=\pi$, $2\pi/3$, $\pi/4$, and $2\pi/5$ for $n=2,3,4,5$, respectively. Their amplitude vanishes, however, in the $N \rightarrow \infty$ limit. As a further check that there are neither site- nor bond-centered oscillations in the ground state we have analyzed $s_i$, $\langle n_i \rangle$ and $s_{i,i+1}$. All Fourier components of these quantities scale to zero in the thermodynamic limit. As an example, the finite-size dependence of $n(k^*)$ for $n=3$ and $n=4$ is shown in Fig. \[fig:nq\_n\] at $U=10$.
![Finite-size dependence of $n(k^*)$ for various $n$ and fillings for $U=10$. The solid line is the finite-size-scaling fit.[]{data-label="fig:nq_n"}](sun_nq_n.eps)
When the finite-size scaling of $D_s$, $T_s$, and $Q_s$ is analyzed one finds that they all vanish in the thermodynamic limit as shown in Figs. \[fig:ds\_n\], \[fig:ts\_n\], and \[fig:qs\_n\] for $U=10$. All this shows that the ground state of the $1/n$-filled SU($n$) Hubbard model is spatially homogeneous, the translational symmetry is not broken.
![Finite-size dependence of $D_s$ for various $n$ and fillings for $U=10$. The solid line is the finite-size-scaling fit.[]{data-label="fig:ds_n"}](sun_ds_n.eps)
![Finite-size dependence of $T_s$ for various $n$ and fillings for $U=10$. The solid line is the finite-size-scaling fit.[]{data-label="fig:ts_n"}](sun_ts_n.eps)
![Finite-size dependence of $Q_s$ for various $n$ and fillings for $U=10$. The solid line is the finite-size-scaling fit.[]{data-label="fig:qs_n"}](sun_qs_n.eps)
### Models with $q>n$
We have chosen as an example $n=3$ and $f=2/5$. As seen in Fig. \[fig:s\_l\_fig2\](a), the block entropy $s_N(l)$ oscillates with period $5$. When every fifth data points are fitted to (\[eq:cardy\]), $c=3$ is obtained. This indicates that the model remains critical for finite $U$ as well. The finite peak in $|\tilde s(k)|$ at $k^*=4\pi/5$ is due to soft modes. The amplitude of the peak disappears in the $N \rightarrow \infty$ limit, the ground state of the system is uniform. This is confirmed by the calculation of $\tilde s(k^*)$, $n(k^*)$, $D_s$, $T_s$, and $Q_s$ shown in Figs. \[fig:sq\_n\], \[fig:nq\_n\], \[fig:ds\_n\], \[fig:ts\_n\], and \[fig:qs\_n\], respectively. There is neither a site- nor a bond-centered oscillation in the occupation number or bond strength.
### Models with $q<n$
One realization of this condition, the half-filled case for $n>2$ has been studied by us earlier.[@buchta_sun] It was found, as shown in Fig. \[fig:s\_l\_fig1\], that the block entropy oscillates with period $2$ for any $n > 2$. The peak in $|\tilde s(k)|$ at $k^*=\pi$ does not vanish in the thermodynamic limit (see Fig. \[fig:sq\_n\]). In agreement with this $D_s$ and $Q_s$ are finite and converge to the same value as shown in Figs. \[fig:ds\_n\] and \[fig:qs\_n\], while $T_{\rm s}$ vanishes (see Fig. \[fig:ts\_n\]). The same behavior is found in the $n=5$ model at half filling, as seen in the upper panel of Fig. \[fig:su5\_ds\].
![Finite-size dependence of $D_s, T_s$, $Q_s$, and $P_s$ for the half-, one-third-, and quarter-filled one-dimensional SU(5) Hubbard model at $U=10$.[]{data-label="fig:su5_ds"}](su5_dimer_fig1.eps)
On the other hand, $s_1(k^*)$ and $n(k^*)$ vanish in the thermodynamic limit (see Fig. \[fig:nq\_n\]). The translational symmetry of the Hamiltonian is broken, the ground state of the half-filled model is dimerized. Stronger and weaker bonds alternate along the chain.
In the one-third-filled case of the $n=4$ and $n=5$ models ($q<n$) $s_N(l)$ oscillates with period $3$ and the amplitude of the Fourier component $\tilde s(k^*=2\pi/3)$ remains finite even for $N\rightarrow\infty$ (see Figs. \[fig:sq\_n\] and \[fig:su5\_sq\_n\]). This indicates that the spatial periodicity is tripled in the ground state. This is corroborated by our results shown in Figs. \[fig:ds\_n\], \[fig:ts\_n\], \[fig:qs\_n\], and the middle panel of \[fig:su5\_ds\]. $D_s$ and $T_s$ scale to the same finite value while $Q_s$, $s_1(k^*)$, and $n(k^*)$ vanish. In the ground state, two bonds of equal strength are followed by a weaker or stronger bond.
As a last example we have studied the quarter-filled SU(5) Hubbard chain. We found that $|\tilde s(k)|$ scales to a finite value at $k^*=\pi/4$ as shown in Fig. \[fig:su5\_sq\_n\]. All Fourier components of the site entropy and local charge density vanish for long chains, while $D_s$, $T_s$, and $Q_s$ scale to finite values. Only $P_s$ scales to zero, as shown in the lower panel of Fig. \[fig:su5\_ds\]. The ground state is a bond-ordered tetramerized state. Fig. \[fig:su5\_bond\] shows schematically the periodic modulation of the bond strength along the chain for half-, one-third, and quarter-filled models.
![Schematic plot of the local bond strength in half-, one-third, and quarter-filled SU(5) Hubbard chains.[]{data-label="fig:su5_bond"}](su5_dimer_fig2.eps)
Conclusion
==========
To study the role of multiparticle umklapp processes, we have treated the one-dimensional SU$(n)$ Hubbard model analytically by bosonization approach and numerically using the DMRG method for $n=3,4$, and $5$ for commensurate fillings $f=p/q$ where $p$ and $q$ are relative primes.
Our results confirm that umklapp processes play essentially different role depending on the relationship between $q$ and $n$. When $q=n$ (this is the case in the $1/n$-filled case) the charge and spin modes are not coupled, the umklapp processes open gap only in the spectrum of charge modes. The system remains critical with $n-1$ gapless spin modes, the central charge is $c=n-1$, and the translational symmetry of the Hamiltonian is not broken in the ground state.
When $q>n$, the leading-order umklapp processes are forbidden in the model with local interaction by Pauli’s exclusion principle. The model is equivalent to an $n$-component Luttinger liquid with $c=n$ and the ground state is spatially uniform for $U\geq 0$.
When, however, $q<n$ the charge and spin modes are coupled by the umklapp processes and gap opens in the spectrum of all modes. Even more interestingly, a spatially nonuniform ground state emerges whose periodicity depend on the filling. Half-filled models develop a dimerized ground state, trimerized state appears in one-third-filled models, the ground state is tetramerized in quarter-filled models. Other periodicities would probably be found at other fillings. Our findings are summarized in Table \[tab:sun\_numerical\] which can be compared to the analytical results given in Table \[tab:analitikus\].
$n$ $p/q$ $c$ periodicity $k^*$
------- -- ----- -- ------- -- ----- -- -------------- -- ----------
$q=n$ 2 1/2 1 unform $\pi$
3 1/3 2 uniform $2\pi/3$
4 1/4 3 uniform $\pi/2$
5 1/5 4 uniform $2\pi/5$
$q<n$ 3 1/2 - dimerized $\pi$
4 1/2 - dimerized $\pi$
4 1/3 - trimerized $2\pi/3$
5 1/2 - dimerized $\pi$
5 1/3 - trimerized $2\pi/3$
5 1/4 - tetramerized $\pi/2$
$q>n$ 3 2/5 3 uniform $4\pi/5$
: Central charge and spatial inhomogeneity for the $p/q$-filled SU($n$) Hubbard chain. $k^{\ast}$ in the last column gives the wave number of soft modes when the model is critical while it gives the wave number of the nonuniform ground state when the model is gapped.[]{data-label="tab:sun_numerical"}
We emphasize that our calculations were performed at a relatively large value of $U$, where the nonuniformity of the ground state is well developed, and the finite value of the dimer, trimer or tetramer order parameter can easily be detected. We conjecture, based on our earlier calculations,[@buchta_sun] that the critical value $U_c$ above which the nonuniform phase appears, is $U_c=0$.
This research was supported in part by the Hungarian Research Fund (OTKA) Grants No. K 68340, F 046356 and NF 61726 and the János Bolyai Research Fund. The authors acknowledge computational support from Dynaflex Ltd. under Grant No. IgB-32.
[99]{}
J. Hubbard, Proc. Roy. Soc. A [**276**]{}, 238 (1963); [ *ibid.*]{} [**277**]{}, 237 (1964); [*ibid.*]{} [**281**]{}, 401 (1964); [ *ibid.*]{} [**285**]{}, 542 (1965).
J. B. Marston and I. Affleck, Phys. Rev. B [**39**]{}, 11538 (1989).
R. Assaraf, P. Azaria, M. Caffarel, and P. Lecheminant, Phys. Rev. B [**60**]{}, 2299 (1999).
R. Assaraf, P. Azaria, E. Boulat, M. Caffarel, and P. Lecheminant, Phys. Rev. Lett. [**93**]{}, 016407 (2004).
C. Honerkamp and W. Hofstetter, Phys. Rev. Lett. [**92**]{}, 170403 (2004).
F. F. Assaad, Phys. Rev. B [**71**]{}, 075103 (2005).
E. Szirmai and J. Sólyom, Phys. Rev. B [**71**]{}, 205108 (2005).
E. Szirmai and J. Sólyom, Phys. Rev. B [**74**]{}, 155110 (2006).
K. Buchta, Ö. Legeza, E. Szirmai, J. Sólyom, Phys. Rev. B [**75**]{}, 155108 (2007).
Á. Rapp, G. Zaránd, C. Honerkamp, and W. Hofstetter, Phys. Rev. Lett. [**98**]{}, 160405 (2007).
J. Zhao, K. Ueda, and X. Wang, arXiv:cond-mat/0702582.
T. J. Osborne and M. A. Nielsen, Phys. Rev. A [**66**]{}, 32110 (2002).
A. Osterloh, L. Amico, G. Falci, and R. Fazio, Nature (London) [**416**]{}, 608 (2002).
P. Zanardi, Phys. Rev. A [**65**]{}, 42101 (2002).
S.-J. Gu, H.-Q. Lin, and Y.-Q. Li, Phys. Rev. A [**68**]{}, 42330 (2003).
J. Vidal, G. Palacios, and R. Mosseri, Phys. Rev. A [**69**]{}, 022107 (2004); J. Vidal, R. Mosseri, and J. Dukelsky, [*ibid.*]{} [**69**]{}, 054101 (2004).
M.-F. Yang, Phys. Rev. A. [**71**]{}, 030302(R) (2005).
S.-J. Gu, S.-S. Deng, Y.-Q. Li, and H.-Q. Lin, Phys. Rev. Lett. [**93**]{}, 86402 (2004).
L.-A. Wu, M. S. Sarandy, and D. A. Lidar, Phys. Rev. Lett. [**93**]{}, 250404 (2004).
. Legeza and J. Sólyom, Phys. Rev. Lett. [**96**]{}, 116401 (2006).
W. K. Wootters, Phys. Rev. Lett. [**80**]{}, 2245 (1998).
P. Zanardi and N. Paunković, Phys. Rev. E [**74**]{}, 031123 (2006); P. Zanardi, M. Cozzini, and P. Giorda, arXiv:quant-ph/0606130(unpublished).
G. Vidal, J. I. Latorre, E. Rico, and A. Kitaev, Phys. Rev. Lett. [**90**]{}, 227902 (2003).
. Legeza, J. Sólyom, L. Tincani and R. M. Noack, Phys. Rev. Lett. [**99**]{}, 087203 (2007).
V. E. Korepin, Phys. Rev. Lett. [**92**]{}, 096402 (2004).
C. Holzhey, F. Larsen, and F. Wilczek, Nucl. Phys. [**B424**]{}, 443 (1994).
P. Calabrese and J. Cardy, J. Stat. Mech.: Theor. Exp. P06002 (2004).
I. Affleck and A. W. W. Ludwig, Phys. Rev. Lett. [**67**]{}, 161 (1991).
N. Laflorencie, E. S. S[ø]{}rensen, M.-S. Chang and I. Affleck, Phys. Rev. Lett. [**96**]{}, 100603 (2006).
J. Zinn-Justin, [*Quantum Field Theory and Critical Phenomena*]{}, Clarendon Press, Oxford (2001).
J. Sólyom, Adv. Phys. [**28**]{}, 201 (1979).
For a review, see A. O. Gogolin, A. A. Nersesyan, A. M. Tsvelik, *Bosonization and Strongly Correlated Systems*, Cambridge University Press, Cambridge (1998).
T. Giamarchi and A. J. Millis, Phys. Rev. B [**46**]{}, 9325 (1992).
S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992); Phys. Rev. B [**48**]{}, 10345 (1993).
. Legeza, J. R[ö]{}der, and B. A. Hess, Phys. Rev. B [**67**]{}, 125114 (2003).
. Legeza and J. Sólyom, Phys. Rev. B [**70**]{}, 205118 (2004).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In contrast to gravity in the weak-field regime, which has been subject to numerous experimental tests, gravity in the strong-field regime is largely unconstrained by observations. We show that gravity theories that cannot be rejected by solar system tests but that diverge from general relativity in the strong-field regime predict neutron stars with significantly different properties than their general relativistic counterparts. In particular, the range of redshifts of surface atomic lines predicted by such gravity theories is significantly larger than the uncertainty introduced by our lack of knowledge of the equation of state of ultra-dense matter. Measurements of such redshifted lines with current X-ray observatories such as Chandra and XMM-Newton can thus provide interesting new constraints on strong-field gravity.'
author:
- Simon DeDeo
- Dimitrios Psaltis
title: 'Towards New Tests of Strong-field Gravity with Measurements of Surface Atomic Line Redshifts from Neutron Stars'
---
The rapid variability and high luminosities of accreting compact objects strongly suggest that their high-energy emission originates very deep in their gravitational potentials. As a result, the properties of their X-ray and $\gamma$-ray emission can, in principle, be used to test directly the strong-field regime of a gravity theory. Unfortunately, there are many astrophysical considerations that complicate the modeling of the emission of accreting neutron stars and black holes and it is difficult to disentangle the effects of gravity from those due to astrophysical processes. For this reason tests of gravity involving high-energy phenomena around compact objects have received little attention in the past. This is despite the fact that the behavior of gravity in the strong-field regime is largely unconstrained by observations.
The properties of a neutron star and its external spacetime, in contrast to the high-energy phenomena above its surface, are determined solely by the strong-field behavior of gravity and the equation of state. To date, studies of neutron stars have focused on better determining the equation of state, whose details are still debated. We will argue, however, that the uncertainty in the behavior of strong-field gravity introduces much greater differences in neutron-star properties than do current uncertainties in the neutron-star equation of state. Furthermore, new observational developments allow for the measurement of the gravitational redshift of surface atomic lines, which is the cleanest source of information about the mass, radius, and spin of neutron stars. The first such observations, with Chandra and XMM-Newton, have already been reported [@cpm02; @spzt02].
Testing rigorously General Relativity (GR) in the strong-field regime requires a general framework of which GR is a subset; this would be equivalent to the Parametrized Post-Newtonian formulation in the weak-field. Such a framework is yet to be constructed. Thus, in order to provide a proof of principle for the efficacy of the proposed tests, we will investigate here the constraints that can be imposed on a parametrized subclass of scalar-tensor theories (see, e.g., Ref. [@de92]). As an additional example, we will also consider the limits that can be placed on Rosen’s bimetric theory [@r73]; although binary-pulsar experiments have already excluded the bimetric theory [@we77], our results demonstrate the potential of the proposed tests for constraining gravity theories.
We consider first *scalar-tensor theories*. The general class of scalar-tensor theories, where the gravitational force felt by matter is mediated by both a rank-two tensor, $g_{\mu\nu}$, and a scalar field, $\phi$, is one of the most natural extensions of Einstein’s theory. The action is [@de92] $$S = \frac{1}{16\pi G_{*}}\int d^4x \sqrt{-g_{*}}(R_{*}-2g^{\mu\nu}_{*}\phi_{,\mu}\phi_{,\nu})$$$$\label{lagrangian}
+ S_m[\Psi_m;A^2(\phi)g_{*\mu\nu}].$$ Here $\Psi_m$ refers collectively to all matter fields other than $\phi$, $G_*$ is a dimensional constant, and $A(\phi)$ is a function of $\phi$. The tensor $g_{*\mu\nu}$ is the metric in the Einstein frame and the Brans-Dicke frame tensor is $g_{\mu\nu}=A^2(\phi)g_{*\mu\nu}$. For $A(\phi)$ unity, we recover GR.
Solar system experiments constrain the coupling function $A$ to be very flat at the cosmological value of the scalar field $\phi$. If, however, the scalar field inside a compact object fluctuates far enough away from its cosmological value to discover a steep part of $A$, the system will be able to reach the more energetically favorable configuration with large non-zero $\phi$ near the center of the object. This non-perturbative effect leads to neutron stars that are significantly more massive and larger than their GR counterparts, but it is invisible to solar system experiments that probe the weak-field regime. This is the phenomenon of “spontaneous scalarization” discovered by Damour & Esposito-Farèse [@de93], who used the Hulse-Taylor pulsar to put constraints on the strong-field regime of such theories.
Following Damour & Esposito-Farèse [@de93], we choose $$A(\phi)=e^{\frac{1}{2}\beta\phi^2},$$ where $\beta$ is a real number. Spontaneous scalarization occurs only when $\beta\leq-4.85$. Throughout this paper we will be interested only in solutions where $\phi_0$, the cosmological value of the scalar field, is zero and $A(\phi_0)=1$.
Following Ref. [@de93], we write the relativistic equations of stellar structure for a [**non-rotating**]{} neutron star in scalar-tensor gravity. We write the metric as $$\begin{aligned}
ds_*^2 &=& -e^{\nu(r)}dt^2+\left[1-\frac{2\mu(r)}{r}\right]^{-1}dr^2\nonumber\\
& &+ r^2(d\theta^2+\sin^2\theta d\phi^2),
\label{metric}\end{aligned}$$ and describe the matter fields as a perfect fluid in the physical frame, i.e., $$T_{\mu\nu}=(\rho+p)u_{\mu}u_{\nu}+pg_{\mu\nu}.$$ Here $p$ is the pressure, $\rho$ is the density, and $u_{\mu}$ is the four-velocity of the fluid. The differential equations to be integrated are then [@h98] $$\label{eqs}
\mu^{\prime} = 4\pi G_*r^2A^4\rho+\frac{1}{2}r(r-2\mu)\psi^2,$$$$\nu^{\prime} = 8\pi g_*\frac{r^2A^4p}{r-2\mu}+r\psi^2+\frac{2\mu}{r(r-2\mu)},$$$$\phi^{\prime} = \psi,$$ $$\begin{aligned}
\psi^{\prime} &=& 4\pi G_* \frac{rA^4}{r-2\mu}[\alpha(\rho-3p)+r(\rho-p)\psi]
\nonumber\\
&&+ \frac{\mu}{r(r-2\mu)}\psi,
\label{harry}\end{aligned}$$ $$N^{\prime} = 4\pi nA^3r^2\left(1-\frac{2\mu}{r}\right)^{-1/2},$$ $$p^{\prime} = -(\rho+p)\left[4\pi G_*\frac{r^2A^4p}{r-2\mu}+\frac{1}{2}r\psi^2\right]$$ $$\label{eqs2}
-(\rho+p)\left[\frac{\mu}{r(r-2\mu)}+\alpha\psi\right].$$ Here $N$ is the baryon number, $n$ is the number density, and primes denote derivatives with respect to $r$. We supplement these equations with a (zero temperature) equation of state, $p=p(\rho)$ and $n=n(\rho)$.
We choose two commonly used equations of state, which cover a broad subset of the wide range discussed in Cook et al. [@cst94]. In order of increasing stiffness, these are EOS A [@eosa] and EOS UU [@eosuu]. Neutron star models computed with these two equations of state bracket the uncertainty introduced by our inability to calculate from first principles the properties of ultra-dense matter, when condensates or unconfined u-d-s quark matter is not taken into account.
Having chosen an equation of state, we then integrate the system of equations (\[eqs\])–(\[eqs2\]) from the center of the star, where we specify $\mu(0)=\nu(0)=N(0)=0,\phi(0)=\phi_c,$ and $\rho(0)=\rho_c$, to the surface, where $p(R)=0$. We then integrate equations (\[eqs\])–(\[harry\]) from the surface of the star to infinity to determine the form of the metric exterior to the star. We note that the Arnowitt-Deser-Misner (ADM) mass felt by an observer far away is in general different to the mass that contributes to the redshift. In addition to considering the scalar-tensor theory, we also investigate the bimetric theory, proposed by Rosen [@r73]. This involves, in addition to the dynamical metric $g_{\mu\nu}$, a nondynamical flat metric $\eta_{\mu\nu}$. The metric $g_{\mu\nu}$ in Rosen’s theory may be written as $$ds^2 = -e^{-2m_{\Phi}(r)/r} dt^2$$$$+ e^{2m_{\Lambda}(r)/r} [dr^2+r^2(d\theta^2+\sin^2 \theta d\phi^2)],$$where $m_{\Phi}(r)$ and $m_{\Lambda}(r)$ are given, interior to the star, by $$m_{\Phi}(r) = 4\pi \int^r_0 e^{\Phi+3\Lambda} (\rho+3p)r^2 dr,$$ $$m_{\Lambda}(r) = 4\pi \int^r_0 e^{\Phi+\Lambda} (\rho-p)r^2 dr,$$and $\Phi$ and $\Lambda$ are given by the field equations $$\Phi^{\prime} = Gm_{\Phi}(r)/r^2,$$ $$\Lambda^{\prime} = -Gm_{\Lambda}(r)/r^2.$$ The equation of hydrostatic equilibrium is $$p^{\prime} = -(\rho+p)\Phi^{\prime}.$$ Outside the star, $m_{\Phi}(r)=m_{\Phi}(R)\equiv M_{\Phi}$ and $m_{\Lambda}(r)=m_{\Lambda}(R)\equiv M_{\Lambda}$ are constants. $M_{\Phi}$ is equal to the Kepler-measured mass at large distances. We specify an equation of state, $p=p(\rho)$, and the boundary conditions $\rho(0)=\rho_c, \Phi(\infty)=\Lambda(\infty)=0.$ Figure \[1\] shows the relationship between ADM mass and neutron-star radius for the three theories under consideration. For the case of the scalar-tensor theory, we show, as an example, the case $\beta=-8$, which is comparable to the most negative value of $\beta$ not yet ruled out by binary pulsar timing [@de96]. For each of the three theories, we plot a shaded region corresponding to the allowed values of mass and radius for equations of state with stiffness between EOS A and EOS UU. We include only the stars stable to radial perturbation and, for configurations in the scalar-tensor theory we show only the one energetically preferred.
As can be seen immediately, all three theories allow neutron stars with masses in the currently observed range of $1.35-1.8 M_{\odot}$ to exist [@tc99]. As a result, measurement of neutron star masses alone cannot distinguish between them. This degeneracy is broken, however, by considering the predicted radii. In the astrophysically relevant range $M_{\text{ADM}}\geq 1.3 M_{\odot}$, the three theories occupy mutually exclusive regions in the $(M_{\text{ADM}},R)$ space. Observations that can put limits on both the mass and radius of a neutron star can thus constrain, without ambiguity, the permissible set of gravity theories.
One such observation is of the surface redshift factor, $z$, relative to infinity. This is defined by $E_{\infty}=[1/(1+z)]E_{\text{surf}}$, where $E_{\text{surf}}$ is the energy of a photon emitted from the stellar surface and measured at infinity with energy $E_{\infty}$. In the GR Schwartzschild metric, this provides a direct measurement of only the ratio $M_{\text{ADM}}/R$. In the case of the scalar-tensor and bimetric theories the relationship between $M_{\text{ADM}}$, $R$, and $z$ is more complicated. In all cases, however, the formula, $$E_{\infty}=\left(\frac{g_{00,s}}{g_{00,\infty}}\right)^{1/2}E_{\text{surf}},$$holds, where $g_{00,s}$ is evaluated at the surface of the star and $g_{00,\infty}$ is evaluated at infinity.
Figures \[3\] and \[4\] show the redshift factor $z$ for the scalar-tensor and GR theories as contours on a plot of the mass $M_{\text{ADM}}$ versus the parameter $\beta$, for EOS A and EOS UU, respectively. As discussed above, the phenomenon of spontaneous scalarization occurs only for $\beta\lesssim -4.35$, and so above this value the contours are parallel to the $\beta$ axis and equal to the GR value. The heavy lines on the plot show the upper mass limits for neutron stars as well as the bounding region, in the $(M_{\text{ADM}},\beta)$ space, where spontaneously scalarized stars are produced. The dashed line shows the value of $M_{\text{ADM}}$ as a function of $\beta$ for which the baryonic mass is equal to $1.4
M_{\odot}$. Note the $z=0.35$ contour, the value recently measured by Cottam et al. [@cpm02] in the neutron star source EXO 0748–676.
The redshifts found for stars in the scalar-tensor theories differ greatly from their general relativistic values for both equations of state. The trend is for the redshift values to increase (nearly always) monotonically as $\beta$ becomes more negative. If we can place constraints on the neutron star mass, a redshift measurement will put an upper bound on the value of $-\beta$.
Figure \[2\] shows the redshift factor $z$ as a function of $M_{\text{ADM}}$ for neutron stars produced in the GR and bimetric theories. We plot two shaded regions, one for GR and one for the bimetric theory, that cover the range between EOS A and EOS UU. The dashed line shows the recent $z=0.35$ measurement [@cpm02]. Even when a GR and a bimetric star have the same redshift, the mass serves to break the degeneracy, providing a test of strong field gravity despite the uncertainty in the equation of state.
The recent measured redshift of $z=0.35$ [@cpm02] cannot by itself constrain strong-field gravity in the theories we consider here. However, strong limits can be placed if the mass of the neutron star is measured directly from binary dynamics or constrained by arguments from formation mechanisms (e.g., restricting the baryonic mass to be above the Chandrasekhar limit for a degenerate core.) The neutron star EXO 0748–676, for which this redshift was measured, is a member of an eclipsing binary system with an orbital period of 3.8 h [@pwg86] and hence is a prime candidate for producing a mass measurement that can provide such a constraint. Simply requiring the baryonic mass of the star to be greater than $1.4 M_{\odot}$ (the Chandrasekhar limit for its degenerate progenitor) places an upper limit on $-\beta$ of 7 (EOS A) to 9 (EOS UU). As expected, the constraint on strong-field gravity depends only weakly on the equation of state.
In addition to the redshift of atomic lines, a number of other observational characteristics of neutron stars depend – in more complicated ways – on the of nature strong-field gravity. These are the maximum allowed spin frequency, the Eddington limit for accreting stars, bursting behavior, cooling rates, and the frequencies of quasi-periodic oscillations. The constraints imposed by quasi-periodic oscillations will be reported in a forthcoming paper.
The constraints on scalar-tensor discussed here have a wider applicability than the structure of compact objects because varieties of scalar-tensor theories naturally arise as low-energy limits of gravitational theories in higher dimensions [@ow98]. They are also under consideration in cosmology as explanations of evidence for an accelerating universe. To date, little attention has been given to the implications of such theories for the properties of compact objects that are now observable. The uncertainties in the equation of state at ultra-high densities were thought to make such studies fruitless. Our investigation has shown that this is not the case and that measurements of the redshifts of surface atomic lines from neutron stars can provide new tests of strong-field gravity.
We thank A. Loeb, J. Maldacena, C. Miller, and F. Özel for useful discussions. SD acknowledges the support of an NSF Graduate Research Fellowship. DP acknowledges the support of NSF grant PHY-0070928.
[19]{} 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: |
During the several months of the outburst of Mkn 501 in 1997 the source has been monitored in TeV [$\gamma$-rays ]{}with the HEGRA stereoscopic system of imaging atmospheric Cherenkov telescopes. Quite remarkably it turned out that the shapes of the daily [$\gamma$-ray ]{}energy spectra remained essentially stable throughout the entire state of high activity despite dramatic flux variations during this period. The derivation of a long term time-averaged energy spectrum, based on more than 38,000 detected TeV photons, is therefore physically meaningful. The unprecedented [$\gamma$-ray ]{}statistics combined with the 20% energy resolution of the instrument resulted in the first detection of $\gamma$-rays from an extragalactic source well beyond 10 TeV, and the first high accuracy measurement of an exponential cutoff in the energy region above 5 TeV deeply into the exponential regime. From 500 GeV to 24 TeV the differential photon spectrum is well approximated by a power-law with an exponential cutoff: ${\rm d} N/{\rm d} E=N_0 \, (E/1\,{\rm TeV})^{-\alpha} \, \exp{(-E/E_0)}$, with $N_0=(10.8 ~\pm0.2_{\rm stat}~\pm2.1_{\rm sys})
\cdot 10^{-11} \, \rm cm^{-2} s^{-1} TeV^{-1}$, $\alpha=1.92 ~\pm0.03_{\rm stat} ~\pm0.20_{\rm sys}$, and $E_0=(6.2 ~\pm0.4_{\rm stat} ~(-1.5 ~+2.9)_{\rm sys}) \, \rm TeV$. We summarize the methods for the evaluation of the energy spectrum in a broad dynamical range which covers almost two energy decades, and study in detail the principal sources of systematic errors. We also discuss several important astrophysical implications of the observed result concerning the production and absorption mechanisms of [$\gamma$-rays ]{}in the emitting jet and the modifications of the initial spectrum of TeV radiation due to its interaction with the diffuse extragalactic background radiation.
author:
- 'F.A. Aharonian, A.G. Akhperjanian, J.A. Barrio, K. Bernlöhr$^*$, H. Bojahr, I. Calle, J.L. Contreras, J. Cortina, A. Daum, T. Deckers, S. Denninghoff, V. Fonseca, J.C. Gonzalez, G. Heinzelmann, M. Hemberger, G. Hermann$^\dag$, M. He[ß]{}, A. Heusler, W. Hofmann, H. Hohl, D. Horns, A. Ibarra, R. Kankanyan, J. Kettler, C. Köhler, A. Konopelko$^\S$, H. Kornmeyer, M. Kestel, D. Kranich, H. Krawczynski, H. Lampeitl, A. Lindner, E. Lorenz, N. Magnussen, H. Meyer, R. Mirzoyan, A. Moralejo, L. Padilla, M. Panter, D. Petry$^\ddag$, R. Plaga, A. Plyasheshnikov$^\S$, J. Prahl, G. Pühlhofer, G. Rauterberg, C. Renault$^\#$, W. Rhode, A. Röhring, V. Sahakian, M. Samorski, D. Schmele, F. Schröder, W. Stamm, H.J. Völk, B. Wiebel-Sooth, C. Wiedner, M. Willmer, W. Wittek'
date: 'Received 16 March 1999; accepted 28 June 1999'
title: The time averaged TeV energy spectrum of Mkn 501 of the extraordinary 1997 outburst as measured with the stereoscopic Cherenkov telescope system of HEGRA
---
Introduction
============
Mkn 501, an active galactic nucleus (AGN) at a redshift $z = 0.034$, was discovered several years ago as a faint source of TeV $\gamma$-radiation (Quinn et al. [@quin:96] ; Bradbury et al. [@brad:97]). In 1997 it turned into a state of high activity, unique in both its strength and duration. The TeV emission of the source from March to September 1997 was characterized by a strongly variable flux. It was on average more than three times larger than the flux of the Crab Nebula, the strongest known persistent TeV source in the sky. Fortunately, the time period of the outburst coincided with the source visibility windows of several ground-based imaging atmospheric Cherenkov telescopes (IACTs) designed for the detection of very high energy (VHE) cosmic $\gamma$-rays. Thus almost continuous monitoring of Mkn 501 in TeV [$\gamma$-rays ]{}with several IACTs (CAT, HEGRA, TACTIC, Telescope Array, Whipple) located in the Northern Hemisphere was possible (e.g. Protheroe et al. [@prot:98]).
The observations of Mkn 501 by the HEGRA stereoscopic IACT system during this long outburst made a detailed study of the temporal and spectral characteristics of the source possible, based on an unprecedented statistics of more than 38,000 TeV photons (Aharonian et al. [@ahar:99a]; hereafter Paper 1). The “background-free” detection of [$\gamma$-rays ]{}, with an average rate of several hundred [$\gamma$-rays ]{}per hour (against $\approx 20$ background events caused by charged cosmic rays), allowed us to determine statistically significant signals for $\leq 5$ minute intervals during much of the 110 h observation time, spread over 6 months. Moreover, it was possible to monitor the energy spectrum of the source on a daily basis. Within the errors the energy spectrum maintained a constant form over the range from 1 TeV to 10 TeV. This was the case even though the flux varied strongly on time scales $\leq 1$ day. We believe that this is an important result, and it was to some extent unexpected.
The diurnal spectra exhibit a power-law shape at low energies (between 1 TeV and several TeV), with a gradual steepening towards higher energies (Paper 1). Such a spectral form could not be unequivocally ensured in the first analysis which was performed during the period of activity of the source, since the systematic errors of the recently commissioned stereoscopic system of HEGRA were not well studied at this time. As a consequence, the energy spectrum could not be determined more precisely than implied by a power law fit (Aharonian et al. [@ahar:97a]), even though the tendency for a gradual steepening of the observed spectra was noticed (Aharonian et al. [@ahar:97b]). The results of Paper 1 and the new results in a broader energy interval presented below are based on detailed systematic studies (see Paper 1 and Konopelko et al. [@kon:99]), and extend and supersede these previous results. This allowed us to come to the definite conclusion that the spectrum determined in the energy region from 1 to 10 TeV steepens significantly (Paper 1). A similar tendency has been found also by the Whipple (Samuelson et al. [@samuel:98]), Telescope Array (Kajino et al. 1999, private communication), and CAT (Djannati-Atai et al. [@CATcorrelation]) groups. Independent spectral measurements by the HEGRA telescopes CT1 and CT2 will be published elsewhere.
Apart from its astrophysical significance, the constancy of the spectral shape has the important practical consequence that it allows to measure the spectrum with small statistical errors also in the energy regions below 1 TeV and above 10 TeV. Indeed, the low photon statistics of the detector in both “extreme” energy bands (towards low energies basically due to the decrease of the detector’s collection area; towards high energies due to the steep photon spectrum) can be drastically increased by using the data accumulated over the whole period of observations.
In Sect. 2 we describe the HEGRA stereoscopic system and the specific form of the data analysis, based on Monte Carlo simulations of both, the air showers and the detection system. The data sample is the same as in Paper 1 and is described in Sect. 3. A detailed study of the systematic errors in the spectrum derivation is contained in Sect. 4; most of this methodology was actually developed in the context of the Mkn 501 data analysis. We believe that this is the first study of this kind. The experimental results are then presented in Sect. 5, whereas Sect. 6 attempts a first discussion. This discussion concentrates on the $\gamma$-ray results of Sect. 5 and what one can learn from them alone. A multiwavelength analysis is clearly outside the scope of this paper. Our conclusions are contained in Sect. 7.
Readers only interested in the astrophysical results, should skip Sect. 2-4 and proceed to Sect. 5.
The HEGRA system of imaging atmospheric Cherenkov telescopes
============================================================
The HEGRA Cherenkov telescope system
------------------------------------
The VHE $\gamma$-ray observatory of the HEGRA collaboration consists of six imaging atmospheric Cherenkov telescopes (IACTs) located on the Roque de los Muchachos on the Canary island of La Palma, at 2200 m above sea level. A prototype telescope (CT1) started operation in 1992 and has undergone significant hardware upgrades since then. This telescope continues to operate as an independent instrument. The stereoscopic system of Cherenkov telescopes consists of five telescopes (CT2 - CT6), and has been taking data since 1996, initially with three and four telescopes, and since 1998 as a complete five-telescope system. Four of the telescopes (CT2, CT4, CT5, CT6) are arranged in the corners of a square with roughly 100 m side length, and one telescope (CT3) is located in the center of the square. During 1997, when the data discussed in this paper were taken, CT2 was still used as stand alone detector.
The telescopes have an 8.5 m$^2$ tessellated reflector, focusing the Cherenkov light onto a camera with 271 photomultipliers (PMTs), covering a field of view of $4.3^\circ$ in diameter. A telescope is triggered when the signal in at least two adjacent PMTs exceeds an amplitude of 10 (before June 1997) or 8 (after June 1997) photoelectrons; in order to trigger the CT system and to initiate the readout of data, at least two telescopes have to trigger simultaneously. Typical trigger rates are in the 10-16 Hz range. The PMT signals are digitized and recorded by 120 MHz Flash-ADCs. The telescope hardware is described by Hermann ([@her:95]); the trigger system and its performance are reviewed by Bulian et al. ([@bul:98]).
Reconstruction of air showers with the HEGRA IACT system
--------------------------------------------------------
The routine data analysis (see Paper 1 for details) includes a screening of data to exclude data sets taken at poor weather conditions or with hardware problems. In particular, the mean system trigger rate proved to be a sensitive diagnostic tool. Reconstruction of data involves the deconvolution of Flash-ADC data (Hess et al. [@hess:98]), the calibration and flat-fielding of the cameras, the determination of Hillas image parameters, and the reconstruction of geometrical shower parameters based on the stereoscopic views of the air shower obtained with the different telescopes (Daum et al. [@daum:97], Aharonian et al.[@ahar:97a]). The characteristic angular resolution for individual $\gamma$-rays is $0.1^\circ$; by sophisticated procedures $\gamma$-ray sources can be located with sub-arcminute precision (Pühlhofer et al. [@pul:97]).
The separation of hadronic and electromagnetic showers is based on the shape of the Cherenkov images, in particular using the [*width*]{} parameter. The width of each image is normalized to the average [*width*]{} of a $\gamma$-ray image for a given impact parameter of the shower relative to the telescope, and a given image intensity. Here, impact distances are obtained from the stereoscopic reconstruction of the shower geometry. Cuts are then applied to the [*mean scaled width*]{} obtained by averaging the scaled width values over telescopes. For a point source both the [*pointing*]{} information and the image [*shape*]{} information are used. Each of them provides a cosmic-ray background rejection of up to 100.
The reconstruction of the energy of air showers is based on the relation between the shower energy and the image intensity ([*size*]{}) at a given distance from the shower axis (Aharonian et al. [@cell]). This relation is tabulated based on Monte Carlo simulations, with the zenith angle of the shower as an additional parameter. The distance between a given telescope and the shower core is known from the stereoscopic reconstruction of the shower, with a typical precision of 10 m or less, for not too distant showers. The energy estimates from the different telescopes are then averaged, taking into account the slightly different sensitivities of the telescopes. These sensitivities are calibrated to 1% by comparing the light yield in two telescopes for events with cores halfway between the two telescopes (Hofmann [@hof:97]). According to Monte Carlo simulations, this energy reconstruction provides an energy resolution of 15% to 20%, depending on the selection of the event sample.
Monte Carlo simulations of air showers and of the telescope response
--------------------------------------------------------------------
Any quantitative analysis of IACT data has to rely on detailed Monte Carlo simulations to evaluate the detection characteristics of the instrument.
The simulation of air showers and of Cherenkov light emission (Konopelko et al. [@kon:99]) includes all relevant elementary processes. On their trajectory to the detector, photons may be lost by ozone absorption, Mie scattering, and Rayleigh scattering. Atmospheric density profiles, ozone profiles and aerosol densities have been checked against local experimental data where available (e.g. Hemberger [@mark:98]).
On the detector side, the simulations include the wavelength dependence of the mirror reflectivity, of the light collection system, and of the PMT quantum efficiency. The point spread function of the mirror system is modeled after measurements of images of bright stars. The readout electronics is simulated in significant detail. PMT output waveforms are modeled by superimposing the response to single photoelectrons, with their relative timing and amplitude smearing. These signals are then sampled, quantitized, and fed into the same analysis path as regular Flash-ADC data. The simulation includes the measured saturation effects both in the PMT/preamplifier and in the Flash-ADC. Details concerning the Monte Carlo simulation used here, the performance of the system, and the comparison with experimental data are described by Konopelko et al. ([@kon:99]).
The Mrk 501 data sample
=======================
The analysis is based on the same data sample as used in Paper 1, corresponding to a total exposure time of 110 h, between March and October of 1997. Mrk 501 was observed in the so-called wobble mode, with the source positioned $\pm 0.5^\circ$ in declination away from the optical axis of the telescopes, alternating every 20 min. For background subtraction, an equivalent region displaced by the same amount in the opposite direction is used. The separation of $1^\circ$ of these on-source and off-source regions is large compared to the angular resolution of the telescope system. In total, the sample comprises about 38,000 $\gamma$-ray events.
To select $\gamma$-ray candidates, the same loose cuts were applied as in Paper 1. In particular, the reconstructed shower direction had to be within $0.22^\circ$ from the source, and the [*mean scaled width*]{} parameter had to be less than 1.2. Events were accepted up to a maximum impact parameter of 200 m from the central telescope CT3. Events with larger impact parameters frequently suffer from truncated images in the cameras and, due to small angles between the stereo views, have larger uncertainties on the shower parameters. Since the analysis presented in the following emphasizes the control of systematic errors, it was felt that in this case the advantage of having clean events and a well-defined effective area at large energies – basically all events within the 200 m radius trigger above a few TeV – outweighs the gain in statistics which could have been achieved by accepting all events [^1]. For the following analysis a software threshold of two or more triggered telescopes, each with at least 40 recorded photoelectrons was used.
Properties of this Mrk 501 data sample were examined in detail in Paper 1.
Determination of energy spectra and sources of systematic errors
================================================================
Compared to single IACTs, stereoscopic IACT systems permit drastic reduction of systematic errors, in particular for tasks like the precision determination of energy spectra. Using the redundant information provided by the multiple views, essentially all relevant characteristics, such as the radial distribution of Cherenkov light or the trigger probabilities of the telescopes can be verified experimentally (see also Hofmann [@hof:97]). Simultaneous sampling of the intensity of the Cherenkov light front in different locations emphasizes the calorimetric nature of the energy determination, and reduces the effect of local fluctuations. Finally, given the unambiguous reconstruction of the shower geometry and the fact that at energies of one TeV virtually all events within 100 m from the central telescope trigger the system, and that above a few TeV almost all events within 200 m trigger, the effective detection area above 1 TeV can be basically defined by pure geometry, without relying on simulations. Only the threshold region requires a critical consideration.
Under ideal conditions, the differential energy spectrum of the incident radiation is determined as $$\phi(E) = {r(E) \over \eta(E) A(E)} \, ,
\label{eqa}$$ where $r(E)$ is the measured rate of $\gamma$-rays of energy $E$ after background subtraction, $A(E)$ is the effective detection area and $\eta(E)$ is the efficiency of the cuts applied to isolate the signal and to suppress the background. The effective area and the cut efficiencies are usually derived from Monte Carlo simulations.
A complication arises from the finite energy resolution, described by the response function ${\cal R}(E',E)$, the probability that a $\gamma$-ray of energy $E$ is reconstructed at an energy $E'$. The measured rate is hence given by the convolution $$r(E') = \int dE~{\cal R}(E',E)~\eta(E) A(E) \phi(E) \, .
\label{eqb}$$ Eq. \[eqb\] can no longer be trivially inverted to yield $\phi(E)$. Options to find $\phi(E)$ include the explicit deconvolution using a suitable algorithm, which will usually make some assumption concerning the smoothness of the spectrum. Another approach is to assume a certain functional form for the shape of the spectrum, and to determine free parameters such as the flux and the spectral index from a fit of Eq. \[eqb\] to the data. Finally, one can absorb the effect of the energy smearing into a modified effective area $A$, defined such that Eq. \[eqa\] holds. The latter approach is the simplest, but has the disadvantage that now $A$ depends on the assumed shape of the spectrum. However, with the $<20\%$ energy resolution provided by the HEGRA CT system, shape-dependent corrections are negligible for most practical purposes, and results are stable after one iteration. Therefore, while both other techniques were pursued, the final results are based on this third method.
The typical energy dependence of the effective area is shown in Fig. \[fig\_area\]. Below 1 TeV, the effective detection area rises steeply with energy, and then saturates at around $10^5$ m$^2$. The saturation reflects the cut on a maximum distance from the central telescope of 200 m.
The technical implementation of the energy reconstruction and flux determination is described in detail in Paper 1. Compared to the brief discussion given above, a main complication for real data arises from the dependence of $A$ on the zenith angle $\theta$, which varies during runs and from run to run. In order to be able to interpolate between Monte Carlo-generated effective areas at certain discrete zenith angles, a semi-empirical scaling law is exploited, which relates the effective areas at different zenith angles. The variation of zenith angles with time is accounted for through replacing Eq. \[eqa\] by the sum over all events recorded within the observation time $T$ $$\phi(E) = {1 \over T} \sum_{events} {1 \over \eta(E,\theta) A(E,\theta)}
\label{eqc}$$ where each event is weighted with the appropriate effective area, given its energy and zenith angle. Note that for each period of a certain hardware configuration a set of effective areas is used which has been determined from the Monte Carlo simulations which model in detail the specific hardware performance, i.e. which take into account the trigger configuration and the mirror point spread function (see Paper 1).
The key aspect in a precise and reliable determination of $\gamma$-ray spectra is the control of systematic errors. Sources of systematic errors include, e.g.,
- Systematic errors in the determination of the absolute energy scale.
- Deviations from the linearity of the energy reconstruction, caused e.g. by threshold effects at very low energies, and possible saturation effects in the PMTs or the electronics at very high energies.
- Systematic errors in the determination of the effective area $A$; particularly critical is the threshold region, where $A$ is a very steep function of $E$.
- Systematic errors in the determination of the (energy-dependent) efficiency of the angular and image shape cuts.
As discussed in detail in Appendix A, a non-accurate modeling of the detector response or the atmospheric transmission possibly results in 1) a shift in the energy scale of the reconstructed $\gamma$-ray spectrum, and 2) a distortion of the shape of the spectrum. Most systematic uncertainties, e.g. mirror reflectivities and PMT quantum efficiencies, exclusively contribute to an error of the energy scale. The calculations in Appendix A show that the curvature of the spectrum is reconstructed correctly provided that the Monte Carlo simulations accurately model the correlation between the detector threshold and the reconstructed energies including the fluctuations involved. If the Monte Carlo description of this correlation is incorrect, the energies reconstructed in data and the effective areas computed from Monte Carlo simulations do not match. A shift of the reconstructed energies by a factor $(1+\epsilon)$ relative to the effective area $A$ used for the evaluation of the spectrum results in a flux error of $${\Delta \phi \over \phi} \approx {\Delta A(E) \over A(E)}
\approx \epsilon~{\mbox{d}\ln{A(E)} \over \mbox{d}\ln{E}}~~~.
\label{eqd}$$ The flux error is potentially large in the threshold region, where $A$ varies quickly with $E$, $A \propto E^\beta$ with $\beta \simeq 6$ and $\Delta \phi / \phi
\approx 6 \epsilon$, but is negligible at high energies, where $A$ is constant, and is simply governed by the geometrical cuts (see Fig. 1).
The remainder of this section is dedicated to a discussion of the various sources of systematic errors.
Reliability of the determination of the shower energy
-----------------------------------------------------
A crucial input for the determination of shower energies is of course the expected light yield as a function of core distance. Since the average core distance varies significantly with energy, inadequacies in the assumed relation will not only worsen the energy resolution, but will systematically distort the spectrum. With the redundant information provided by a system of Cherenkov telescopes, it is possible to actually measure the light yield as a function of the distance from the shower core and to verify the simulations (Aharonian et al. [@ahar:98b]). Briefly, the idea is to select showers with a fixed impact parameter relative to a telescope A, and with a fixed light yield in this telescope. This provides a sample of showers of constant energy, and now the light yield in other telescopes can be measured as a function of core distance. Fig. \[fig\_yield\] shows the characteristic shape of the light pool for $\gamma$-ray showers, which is well reproduced by the simulation; this holds also for the variation of the shape with shower energy and zenith angle (Aharonian et al. [@ahar:98b]).
A second key ingredient in the energy determination is the reconstruction of the core location. Unlike in the case of the angular resolution, where it is easy to show that the simulation reproduces the data by comparing the distribution of shower directions relative to the source with the simulations, a direct check of the precision of the core reconstruction is not possible. However, noting that two telescopes suffice for a stereoscopic reconstruction, one can split up the four-telescope system into two systems of two telescopes and compare the results (Hofmann [@hof:97]). Figure \[fig\_res\]a illustrates the difference in core coordinates between the two subsystems for data and for the Monte Carlo simulations. As can be recognized the Monte Carlo simulations accurately predict the distribution of the distances between the two cores. Under the assumption that the two measurements are uncorrelated and that the reconstruction accuracy achieved with two telescopes is the same for two telescope and four telescope events the width of the difference distribution should be $\sqrt{2}$ times the resolution of a two-telescope system. By this means, the 2 telescope resolution is determined to be $14~{\rm m}~/~\sqrt{2}~\approx~10$ m. The Monte Carlo simulations show, that this reconstruction accuracy does indeed agree nicely with the true reconstruction accuracy for 2 telescope events.
\
The same technique can be used to study the energy resolution (Fig. \[fig\_res\]b; see also Hofmann [@hof:97]). Also here the data is in excellent agreement with the simulations. In this case, the Monte Carlo studies predict, that the energies measured with the two subsystems are considerably correlated: assuming uncorrelated estimates, an energy resolution of $25\%~/~\sqrt{2}~=~ 18\%$ is inferred; the Monte Carlo simulations predict a true energy resolution for two telescope events of 23%. On the basis of simulations, fluctuations in the shower height can be identified as the origin of this correlation. Work is ongoing to use the stereoscopic determination of the shower height to improve the energy resolution. The excellent agreement between data and Monte Carlo simulations confirms that the experimental effects entering the energy determination are well under control, and that consequently the estimate of the energy resolution based on the simulations is reliable.
The absolute energy scale
-------------------------
The absolute energy calibration of IACTs is a significant challenge, lacking a suitable monoenergetic test beam. Factors entering the absolute energy calibration are the production of Cherenkov light in the shower, the properties and the transparency of the atmosphere, and the response of the detection system involving mirror reflectivities, PMT quantum efficiencies, electronics calibration factors etc. So far, mainly three techniques were used to calibrate the HEGRA telescopes:
- The comparison between predicted and measured cosmic-ray detection rates. Given the integral spectral index of 1.7 for cosmic rays, an error $\epsilon$ in the energy scale results in an error of $1.7 \epsilon$ in the rate above a given threshold. Apart from the slightly different longitudinal evolution of hadronic and of $\gamma$-ray induced showers, this test checks all factors entering the calibration. The Monte Carlo simulations reproduce the measured cosmic-ray trigger rates within 10% (Konopelko et al. [@kon:99]). The IACT system was furthermore used to determine the flux of cosmic-ray protons in the 1.3 to 10 TeV energy range (Aharonian et al. [@ahar:98c]). The measured flux of protons $${\rm d} N_{\rm p}/{\rm d} E=(0.11 \pm 0.02_{\rm stat} \pm 0.05_{\rm sys})
\times$$ $$\hspace*{0.5cm}
E^{-2.72 \pm 0.02_{\rm stat} \pm 0.15_{\rm sys}}~~~
\rm s^{-1}sr^{-1} m^{-2} TeV^{-1}$$ excellently agrees with a fit to the combined data of all other experiments, indicating that systematics are well under control and that the assigned systematic errors – partially related to the energy scale – are rather conservative.
- Given the measured characteristics of the telescope components such as mirrors or PMTs, the sensitivity of the telescopes can be determined with an overall error of 22%.
- Using a distant, calibrated, pulsed light source, an overall calibration of the response of the telescope and its readout electronics could be achieved, with a precision of 10% (Fra[ß]{} et al. [@frass:97]).
The last three techniques have to rely on the Monte Carlo simulations of the shower and of the atmospheric transparency. While Monte Carlo simulations have converged and different codes produce consistent results, details of the atmospheric model and assumptions concerning aerosol densities can change the Cherenkov light yield on the ground by about 8% (Hemberger 1998).
Within their errors, all calibration techniques are consistent. Overall, we believe that a 15% systematic error on the energy scale is conservative, given the state of simulations and understanding of the instrument. Based on the comparison with cosmic-ray rates, one would conclude that the actual calibration uncertainty is below 10%.
The threshold region and associated uncertainties
-------------------------------------------------
As discussed earlier in detail, the threshold region is very susceptible to systematic errors. In the sub-threshold regime events trigger only because of upward fluctuations in the light yield, and energy estimates tend to be biased towards larger values. Fig. \[fig\_ereco\] shows the mean reconstructed energy as a function of the true energy for a sample of simulated events at typical zenith angles. Note that for the 1997 data set we have a noticeable number of detected $\gamma$-rays events below 500 GeV. However in this energy region the bias is so strong that a reliable correction is no longer possible. Therefore spectra will only be quoted above this energy.
\
The first major source of systematic errors at low energies is the detailed modeling of the triggering process. In the simulation great emphasis was placed on the correct description of the pixel trigger probabilities as a function of signal amplitude. For each trigger configuration and for each telescope this dependence has been derived from air shower data using the recorded information about which pixels of a triggered telescope surpassed the discriminator threshold and which pixels did not. As discussed above and in Appendix A, the onset of the [*size*]{}-distribution provides a sensitive test of the quality of the simulation. Fig. \[fig\_size\] shows the measured and simulated [*size*]{}-distribution for two telescopes for a certain trigger configuration (“Data-period I” of Paper 1). For our current best simulations, fits to the rising edges of the [*size*]{}-distributions indicate for the individual telescopes and the different trigger configurations deviations of the [*size*]{} scale between data and Monte Carlo on the 5% level. For conservatively estimating the systematic error on the shape of the Mkn 501 spectrum, we allow for a $\pm$5% correlated shift of the thresholds of all four telescopes. The resulting uncertainties in the effective detection area $A$ are indicated in Fig. \[fig\_area\]; as expected, $A$ significantly changes in the threshold regime, but remains constant well above threshold.
Random threshold variations (Gaussian-distributed with a width of 15%) between individual pixels were found to be of relatively small influence compared to the systematic threshold shifts.
As another source of systematic errors of special importance in the threshold region, the accuracy of the energy reconstruction for zenith angles between the discrete simulated zenith angles ($\theta=0^\circ$, $20^\circ$, $30^\circ$, $45^\circ$) has been considered. Imperfections in the scaling law used to relate the Cherenkov light yield at different zenith angles $\theta$ could result in a systematic shift of the reconstructed energy at intermediate values of $\theta$. To test the description, the energy of the Monte Carlo showers with $\theta = 30^\circ$ was determined with the light yield tables of the $0^\circ$- and $45^\circ$-showers and the result was compared to the result based on using the $30^\circ$ light yield table. The systematic shift of the reconstructed energies was about 5% below 1 TeV and 2% above 1 TeV; based on this and other studies we believe that using the full set of simulations, systematic shifts in the reconstructed energy are below 5% and 2%, below and above 1 TeV respectively. The modified effective area used for the determination of the spectrum slightly depends on the assumed source spectrum. At the lowest and at the highest energies the source spectrum can not be determined with high statistical accuracy. For energies below 1 TeV we conservatively estimate the corresponding uncertainty in the modified effective area, by varying the spectral index of an assumed source spectrum dN/dE$\propto$E$^{-\alpha}$ from $\alpha=1.5$ to $\alpha=3$. At the highest energies above 15 TeV we vary the assumed source spectrum from a broken power law to a power law with an exponential cutoff, both specified by fits to the data. In addition we explore the statistical significance of the $\gamma$-ray excess at the highest energies by a dedicated $\chi^2$-analysis (see Sect. \[RES\]).
In the threshold region of the detector the total systematic error is dominated by the systematic shift of the telescope thresholds and by the possible systematic shift in the reconstructed energy due to the zenith angle interpolation. The total systematic error is computed by summing up the individual relative contributions in quadrature.
Saturation effects at high energies
-----------------------------------
The PMTs and readout chains of the HEGRA system telescopes provide a linear response up to amplitudes of about 200 photoelectrons. At higher intensities, nonlinearities of the PMT become noticeable, and also the 8-bit Flash-ADC saturates. With typical signals in the highest pixels of 25 photoelectrons per TeV $\gamma$-ray energy, the effects become important for energies around 10 TeV and above.
Saturation of the Flash-ADC can be partly recovered by using the recorded length of the pulse to estimate its amplitude, effectively providing a logarithmic characteristic (Hess et al. [@hess:98]). The saturation characteristics of the PMT/preamplifier assembly have been measured, and a correction is applied. These nonlinearities and the Flash-ADC saturation are included in the simulations. As a very sensitive quantity to test the handling of saturation characteristics, the dependence of the pulse height in the peak pixel on the image [*size*]{} emerged (Fig. \[fig\_ratio\]). Data and simulations are in very good agreement up to pixel amplitudes exceeding 10$^3$ photoelectrons, equivalent to $>$50 TeV $\gamma$-ray showers. Older versions of the simulations, which did not properly account for PMT/preamp nonlinearities, showed marked deviations for [*size*]{}-values above 10$^3$.
Independent tests of saturation and saturation corrections were provided by omitting, both in the data and in the simulation, the highest pixels in the image, and by comparing event samples in different ranges of core distance and zenith angle. The saturation effect is non-negligible only at very high energies, namely $E \geq 15 \, \rm TeV$. However, all tests show that given the quality of our current simulations, systematic errors induced by saturation effects even in this energy region are yet small compared to the statistical errors.
Efficiency of cuts
------------------
Among the sources of systematic errors, the influence of cuts
\
is less critical. The large flux of $\gamma$-rays from Mrk 501 combined with the excellent background rejection of the IACT system allows to detect the signal essentially without cuts. In the analysis, one can afford to apply only rather loose cuts, which keep over $80\%$ of all $\gamma$-rays; only at the lowest energies, a slightly larger fraction of events is rejected. Since only a small fraction of the signal is cut, the uncertainty in the cut efficiency is a priori small; in addition, the efficiency can be verified experimentally by comparing with the signal before cuts, see, e.g., Fig. \[fig\_cuteff\]. The cut efficiencies in Fig. \[fig\_cuteff\] deviate from the results shown in Paper 1 on the 5%-level due to slightly improved Monte Carlo simulations. We conservatively estimate the systematic error the cut efficiencies, to be 10% at 500 GeV decreasing to a constant value of 5% at 2 TeV and rising above 10 TeV to 15% at 30 TeV. At low energies the acceptances are corrected according to the measured efficiencies.
Other systematic errors and tests
---------------------------------
The precision in the determination of the effective areas is partly limited by Monte Carlo statistics. The universal scaling law used to relate Monte Carlo generated effective areas at different zenith angles involves a rescaling of shower energies. Statistical fluctuations in a Monte Carlo sample at a given energy and angle will hence influence the area over a range of energies. Because of the resulting slight correlation, Monte Carlo statistics is in the following included in the systematic errors.
To test for systematic errors, the data sample was split up into subsamples with complementary systematic effects, and spectra obtained for these subsamples were compared. Typical subsamples include
- Events where 2, 3 or 4 telescopes are used in the reconstruction.
- Events passing a higher ‘software trigger threshold’, requiring e.g. a minimum [*size*]{} of 100 photoelectrons, or a signal in the two peak pixels of at least 30 photoelectrons.
- Events with showers in a certain distance range from the center of the system (CT3), e.g. 0-120 m compared to 120-200 m. This comparison tests systematics in the light-distance relation as well as the correction of nonlinearities in the telescope response.
- Events where all pixels are below the threshold for nonlinearities.
- Events in different zenith angle ranges. The comparison of these spectra provides a sensitive test of the entire machinery, and also of nonlinearities, where the data at larger angles should be less susceptible because of the smaller [*size*]{} at a given energy.
For each of the subsamples, the effective area and cut efficiencies were determined, and a flux was calculated. In all cases, deviations between subsample spectra were insignificant, or well within the range of systematic errors. Among the variables studied, the most significant indication of remaining systematic effects is seen in the comparison of different ranges in shower impact parameter relative to the central telescope. The ratio of the spectra determined with the data of small ($<$120 m) and large (between 120 m and 200 m) impact distances is shown in Fig. \[fig\_subsamplesb\]. A fit to a constant gives a mean ratio of $1.04 \pm 0.02$ with a $\chi^2$-value of 23.1 for 12 degrees of freedom, corresponding to a chance probability for larger deviations of 5%.
Experimental results {#RES}
====================
In this section we first present the 1997 Mkn 501 time averaged energy spectrum. As discussed already in the introduction the derivation of a time averaged spectrum is meaningful since the changes in the spectral shape during the HEGRA observations were rather small, i.e. they were too small to be assessed with an accuracy of typically between 0.1 and 0.3 in the diurnal spectral indices. Moreover, as described in Paper 1, dividing the data into groups according to the absolute flux level or according to the rising or falling behavior of the source activity yielded mean spectra which did not differ significantly from each other in the one to ten TeV energy range. The weakness of the correlation between the absolute flux and the spectral shape will further be substantiated below over the energy region from 500 GeV to 15 TeV. Nevertheless, the importance of the spectral constancy should not be overestimated. If the spectral variability is not tightly correlated with the absolute flux, diurnal spectral variability characterized by a change of the spectral index at several TeV by approximately $\pm$0.1 is surely consistent with the HEGRA data. The time-averaged energy spectrum is shown in Fig. \[fig\_spectrum\]. For the determination of the spectrum also at energies below 800 GeV, only the data from zenith angles smaller 30$^\circ$ have been used (80 h observation time). The measurements extend from 500 GeV to 24 TeV. The hatched region in Fig. \[fig\_spectrum\] ff.gives our estimate of the systematic errors on the shape of the spectrum, except the 15% uncertainty on the absolute energy scale.
The spectrum shows a gradual steepening over the entire energy range. A fit of the data from 500 GeV to 24 TeV with a power law model with an exponential cut off gives: $$\label{pexp}
dN/dE \, = \, N_0 \,(E/1\, {\rm TeV})^{-\alpha}\,\exp{(-E/E_0)}~~,$$
$N_0=(10.8 ~\pm0.2_{\rm stat}~\pm2.1_{\rm sys})
\cdot 10^{-11} \, \rm cm^{-2} s^{-1} TeV^{-1}$, $\alpha=1.92 ~\pm0.03_{\rm stat} ~\pm0.20_{\rm sys}$, and $E_0=(6.2 ~\pm0.4_{\rm stat} ~(-1.5 ~+2.9)_{\rm sys})$ TeV. The systematic errors on the fit parameters result from worst case assumptions concerning the systematic errors of the data points, and their correlations and include the error caused by the 15% uncertainty in the energy scale. The errors on the fit parameters, especially on $\alpha$ and $E_0$, are strongly correlated. The variation of only one of the parameters within the quoted error range yields spectra which are inconsistent with the measured spectrum. The data points and their errors are summarized in Table 1.
[rrrr]{}\
[E$^a$]{} & [${\rm d}N/{\rm d}E$ $^b$]{} & [$\sigma_{\rm stat}~~^c$]{} & [$\sigma_{\rm sys}~~~^d$]{}\
\
\
0.56 & 3.29 $10^{-10}$ & 1.68 $10^{-11}$ & (+1.82 -1.07) $10^{-10}$\
0.70 & 2.01 $10^{-10}$ & 7.45 $10^{-12}$ & (+6.33 -4.75) $10^{-11}$\
0.88 & 1.15 $10^{-10}$ & 3.45 $10^{-12}$ & (+1.88 -1.68) $10^{-11}$\
1.11 & 7.33 $10^{-11}$ & 1.88 $10^{-12}$ & (+7.42 -7.08) $10^{-12}$\
1.39 & 4.40 $10^{-11}$ & 1.10 $10^{-12}$ & (+3.75 -3.45) $10^{-12}$\
1.75 & 2.87 $10^{-11}$ & 7.25 $10^{-13}$ & (+2.12 -1.97) $10^{-12}$\
2.20 & 1.64 $10^{-11}$ & 4.64 $10^{-13}$ & (+1.03 -0.97) $10^{-12}$\
2.76 & 1.02 $10^{-11}$ & 3.14 $10^{-13}$ & (+6.71 -6.30) $10^{-13}$\
3.46 & 5.64 $10^{-12}$ & 1.98 $10^{-13}$ & (+3.50 -3.30) $10^{-13}$\
4.35 & 3.12 $10^{-12}$ & 1.27 $10^{-13}$ & (+1.94 -1.83) $10^{-13}$\
5.46 & 1.90 $10^{-12}$ & 8.73 $10^{-14}$ & (+1.24 -1.16) $10^{-13}$\
6.86 & 9.29 $10^{-13}$ & 5.24 $10^{-14}$ & (+6.40 -5.99) $10^{-14}$\
8.62 & 4.71 $10^{-13}$ & 3.28 $10^{-14}$ & (+3.42 -3.19) $10^{-14}$\
10.83 & 1.95 $10^{-13}$ & 1.82 $10^{-14}$ & (+1.69 -1.56) $10^{-14}$\
13.60 & 6.24 $10^{-14}$ & 1.02 $10^{-14}$ & (+9.32 -8.11) $10^{-15}$\
17.08 & 2.39 $10^{-14}$ & 5.85 $10^{-15}$ & (+5.87 -4.71) $10^{-15}$\
21.45 & 1.24 $10^{-14}$ & 3.35 $10^{-15}$ & (+5.93 -4.01) $10^{-15}$\
26.95 & $<$ 1.0 $10^{-14}$ $\,^e$&&\
33.85 & $<$ 7.2 $10^{-15}$ $\,^e$&&\
42.51 & $<$ 3.1 $10^{-15}$ $\,^e$&&\
\
\
\
\
\
\
\
\
\[tab\_results\]
In the highest energy bin (19 TeV to 24 TeV) 40 excess events are found above a background of 13 events, corresponding to a nominal significance of S = (N$_{\rm on}$ - N$_{\rm off}$) / $\sqrt{\rm N_{on} + N_{off}}$ of 3.7 $\sigma$. However, due to the steep spectrum in this energy range, a part of these events may represent a spill-over from lower energies. To provide an absolutely reliable lower limit on the highest energies in the sample, the spectrum was fit to the form of Eq. \[pexp\], but with a sharp cutoff at $E = E_{cut}$: $dN/dE = N_0$ $(E/1 ~\rm TeV)^{-\alpha}$ $\exp{(-E/E_0)}$ $\Theta(E_{cut}-E)$. The best fit is achieved with $E_{\rm cut} = 28$ TeV; the $2 \sigma$ lower limit is $E_{cut} = 16$ TeV.
Fig. \[fig\_seng2\] illustrates the spectral energy distribution, $E^2 dN/dE(E)$ as determined from the small zenith angle data ($<$30$^\circ$, energy threshold 500 GeV) and the large zenith angle data (30$^\circ$ to 45$^\circ$, 32 h observation time, energy threshold 1 TeV). Note that the large zenith angle data has mainly been acquired during the second half of the 1997 data taking period. Nevertheless the shape of both spectra agrees within the statistical and systematic errors. The combined small and large zenith angle data set yields the same lower limit on $E_{cut}$ of 16 TeV as derived from the small zenith angle data alone. It can be recognized that the spectral energy distribution is essentially flat from 500 GeV up to $\simeq$ 2 TeV.
Figure \[fig\_seng\] (upper panel) shows the spectral energy distribution for the overall data sample and for periods of high and low flux separately: dN/dE(2 TeV) determined on diurnal basis above 3 and below 1.6 $10^{-11} {\rm cm^{-1} s^{-1} TeV^{-1}}$, with a ratio of the mean fluxes close to 5. The high and low flux spectra agree within statistical errors, as shown by the ratio of both spectra, presented in Fig. \[fig\_seng\] (lower panel). The systematic error is to good approximation the same for both data samples and cancels out in the ratio. The result thus confirms our previous conclusion about the flux-independence of the spectrum of Mkn 501 in 1997 between 1 and 10 TeV (Paper 1). Now the statement is extended to the broader energy region, from 500 GeV to 15 TeV. From 1 TeV to several TeV the slope of the spectrum is determined with high statistical accuracy, e.g. a power law fit in the energy region from 1 TeV to 5 TeV gives a differential index of -2.23 $\pm0.04_{\rm stat}$ and -2.26 $\pm$0.06$_{\rm stat}$ for the high and the low flux spectrum respectively. In the narrow energy range from 500 GeV to 1 TeV the statistical uncertainty on the spectral index is considerably larger, we compute 0.2 for the high flux sample and 0.4 for the low flux sample. Therefore, our 1997 Mkn 501 data would not contradict a correlation of emission strength and spectral shape below 1 TeV as tentatively reported by the CAT-group (Djannati-Atai et al. [@CATcorrelation]). For completeness, the HEGRA IACT system data are plotted in Fig. \[fig\_comp\] jointly with the HEGRA CT1 (Aharonian et al. [@Mkn501PartII]), the CAT (Barrau et al. [@barr:97]), the Telescope Array (Hayashida et al. [@hay:98]), and the Whipple (Samuelson et al. [@samuel:98]) results concerning the Mrk 501 energy spectrum during the 1997 outburst. Generally a good agreement can be recognized in the overlapping energy regions, except for a steeper Telescope Array spectrum.
Discussion
==========
The observations of Mkn 501 by the HEGRA IACT system during the long outburst in 1997 convincingly demonstrate for the first time that the energy spectrum of the source extends well beyond 10 TeV. We believe that this very fact, together with the discovery of a time- and flux- independent stable spectral shape of the TeV radiation will have considerable impact on our understanding of the nonthermal processes in AGN jets.
This discussion is not an attempt at detailed modeling of the result presented in the previous section; this will be done elsewhere. Neither shall we systematically invoke multi-wavelength observations. Our purpose is rather to point out the multiple facets of the [$\gamma$-ray ]{}phenomenon on its own. They stem from the fact that the emission probably originates from a population of accelerated particles within a spatially confined relativistic jet, specifically oriented towards the observer, and that subsequently the radiation must propagate through the diffuse extragalactic background radiation field (DEBRA) before it reaches us. Each of these circumstances can influence the characteristics of the emitted spectrum, and we shall address them in turn below.
Production and absorption of TeV photons in the jet
---------------------------------------------------
The enormous apparent VHE [$\gamma$-ray ]{}luminosity of the source, reaching $\sim
10^{45} \, \rm erg/s$ during the strongest flares which typically last $\Delta t = \, 1$ day or less, implies that the [$\gamma$-rays ]{}are most probably produced in a relativistic, small-scale (sub-parsec) jet which is directed along the observer’s line of sight. The determining quantity is the so-called Doppler factor $\delta_{\rm j}= [(1+z) \Gamma_{\rm
j}(1-\beta_{\rm j}\cos{\Theta})]^{-1}$, where z is the redshift of a source moving with velocity $\beta_{\rm j}c$ and Lorentz factor $\Gamma_{\rm
j}=(1-\beta_{\rm j}^2)^{-1/2}$ along the jet axis that makes an angle $\Theta$ with the direction to the observer. Moreover, the assumption of relativistic bulk motion appears to be unavoidable in order to overcome the problem of severe $\gamma -\gamma$ absorption by pair production on low-frequency photons inside the source (see e.g. Dermer & Schlickeiser [@Der:94]). Indeed, assuming that the $\gamma$-radiation is emitted isotropically in the frame of a relativistically moving source, the optical depth at the [*observed*]{} [$\gamma$-ray ]{}energy is easily estimated as $$\tau_{\gamma \gamma}
\simeq \frac{f_{\rm r} d^2 \sigma_{\rm T} \delta_{\rm j}^{-6}\,E}{8m_{\rm
e}^2 c^6 \Delta t} \simeq 0.065 f_{-10} \Delta t_{\rm day}^{-1}
\delta_{10}^{-6} H_{60}^{-2} E_{\rm TeV}
\label{tau}$$ Here $\delta_{10}=\delta_{\rm j}/10$, $\Delta t_{\rm day}=\Delta t /1 \,
\rm day$, $E_{\rm TeV} = E/1 \, {\rm TeV}$, with $E$ the energy of the gamma-ray in the laboratory frame, and $f_{-10} = f_{\rm r}/10^{-10}\, \rm erg/cm^2 s$ is the observed energy flux at $h \nu \simeq 100 \delta_{10}^2 \, E_{\rm
TeV}^{-1} \, \rm eV$ which for an order of magnitude estimate is assumed to be constant at the optical to UV wavelengths that predominantly contribute to the [$\gamma$-ray ]{}absorption. Furthermore $d = cz/H_0 = 170 H_{60}^{-1}$ Mpc is the distance to the source with redshift $z = 0.034$, normalized to the value of the Hubble constant $H_{60} = H_0/60 \, {\rm km/s Mpc}$. Assuming now that the observed optical/UV flux of Mkn 501, $f_{-10} \simeq 0.5$ (see e.g. Pian et al. [@Pian:98]), is produced in the jet, the absorption of 20 TeV [$\gamma$-rays ]{}becomes negligible only when $\delta_{\rm j} \geq 10$; already for $\delta_{\rm j}=8$ internal absorption would be catastrophic for $\Delta t_{\rm day} = 1$, with $e^{-\tau_{\gamma \gamma}}
\leq 10^{-2}$. One may interpret the detected VHE spectrum of Mkn 501 given by Eq. \[pexp\] as a power-law [*production*]{} spectrum, modified by internal $\gamma - \gamma$ extinction with optical depth $\tau = E_{\rm TeV}/6.2$. This would give an accurate determination of the jet’s Doppler factor taking into account the very weak dependence of $\delta_{\rm j}$ on all relevant parameters, namely $\delta_{\rm j}=8.5 \, f_{-10}^{1/6}
\Delta t_{\rm day}^{-1/6}\, H_{60}^{-1/3}$. Since there could be a number of other reasons for the steepening of the TeV spectrum, that estimate can only be considered as a robust lower limit on $\delta_{\rm j}$.
The strong steepening of the observed spectrum of Mkn 501 above several TeV could also be attributed, for example, to an exponential cutoff in the spectrum of accelerated particles. These could either be protons producing [$\gamma$-rays ]{} through inelastic $p-p$ interactions and subsequent $\pi^0$-decay , or electrons producing [$\gamma$-rays ]{}via inverse Compton (IC) scattering. In IC models an additional steepening of the [$\gamma$-ray ]{}spectrum is naturally expected due to the production of TeV [$\gamma$-rays ]{}in the Klein-Nishina regime if the energy losses of electrons are dominated by synchrotron radiation in the jet’s magnetic field. And finally, the exponential cutoff in the observed TeV spectrum could be caused by $\gamma$-$\gamma$ absorption of TeV photons in a possible dust torus surrounding the AGN (Protheroe & Biermann [@PrBr:96]) and in the extragalactic diffuse background radiation (Nikishov [@Nik:62], Gould & Schreder [@Gould:65], Jelly [@Jelly:65], Stecker et al. [@Steck:92]).
Our current poor knowledge about the distortion of the source spectrum caused by internal and intergalactic absorption does not allow us to distinguish between hadronic and leptonic source models on the basis of their predictions concerning the TeV energy spectra. Fortunately the temporal characteristics are to a large extent free from these uncertainties. Thus we believe that real progress in this area can only be achieved by the analysis of both the [*spectral and temporal*]{} characteristics of X-ray and TeV [$\gamma$-ray ]{}emissions obtained during [*multiwavelength*]{} campaigns that investigate several X-ray selected BL Lac objects in different states of activity, and located at different distances within several 100 Mpc. Notwithstanding this belief, we show here that the high-quality HEGRA spectrum of Mkn 501 [*alone*]{} allows us to make quite a few interesting inferences about the [$\gamma$-ray ]{}production and absorption mechanisms.
### IC models of the gamma ray emission.
Currently it is believed (e.g. Ulrich et al. [@umu:97]) that the correlated X-ray/TeV flares discovered by multiwavelength observations of Mkn 421 (Takahashi et al. [@tak:96], Buckley et al. [@buk:96]) and Mkn 501 (Catanese et al. [@cat:97], Pian et al. [@Pian:98], Paper 1), support the hypothesis of both emission components originating in relativistic jets due to synchrotron/IC radiation of the same population of directly accelerated electrons (Ghisellini et al. [@GhisMD:1996]; Bloom & Marscher [@BlMar:96]; Inoue & Takahara [@IT:1996]; Mastichiadis & Kirk [@Mast:97]; Bednarek & Protheroe [@BedPro:97]). One of the distinctive features of leptonic models is that they allow significant temporal and spectral variations of TeV radiation. The stable shape of the TeV spectrum of Mkn 501 during the 1997 outburst does not contradict these models. It rather requires them to have two important features.
First of all the form of the spectrum of accelerated electrons should be essentially stable in time and be independent of the strength of the flare up to electron energies responsible for the production of the highest observed [$\gamma$-ray ]{}energies $E \geq 16 {\rm TeV}$. For a typical Doppler factor $\delta_{\rm j} \sim 10$ this implies relatively modest electron energies $E_{\rm e} \leq 10 \, {\rm TeV}$ in the jet frame, assuming that the Compton scattering at the highest energies takes place in the Klein-Nishina limit. In this limit a significant fraction of the electron energy goes to the upscattered photon, i. e. $E \sim \delta_{\rm j} \, E_{\rm e}$.
For low values of the magnetic field in the emitting plasma, i.e. B [ ]{} 0.01 G, the X-ray spectrum could be more sensitive to accelerated electrons with energies above 10 TeV. The typical observed energy of X-rays produced by electrons of energy $E_{\rm e}$ in the jet frame is $$E_{\rm X} \simeq 20 (B/0.1 \, {\rm G}) \, (E_{\rm e}/1 \, {\rm TeV})^2 \,
\delta_{10} \, {\rm keV} .
\label{Ex}$$ The BeppoSAX observations of Mkn 501 in April 1997 showed that the X-ray spectrum becomes very hard during strong flares. This is interpreted as a shift of the synchrotron peak to energies in excess of 100 keV (Pian et al. [@Pian:98]). Formally this effect could be explained by a significant increase in each of the three parameters which determine the position of the synchrotron peak, i.e. the maximum electron energy $E_{{\rm e}, max}$, the magnetic field $B$, and the jet Doppler factor $\delta_{\rm j}$.
The rather stable energy spectrum of TeV radiation implies that the spectrum of the parent electrons does not significantly vary during the HEGRA observations, the latter performed with typical integration times between one and two hours. The condition of a constant acceleration spectrum does not yet guarantee a stable energy spectrum of TeV radiation. Therefore we need a second condition, namely to assume very effective radiative (synchrotron and IC) cooling of electrons, sufficiently fast to establish an equilibrium electron spectrum within $\Delta t^{\ast}=10 \, \delta_{10} \, \rm h$. The radiative cooling time is $$t_{\rm rad}= (\frac{4}{3} \, \sigma_{\rm T} \,
c \, w_0 \, E_{\rm e}/m_{\rm e} c^2)^{-1} \simeq
\label{time}$$ $$\hspace*{0.8cm}
\simeq
15.5 \, (w_0/1 \, {\rm erg/cm^3)^{-1}}
\, E_{\rm TeV} \, \rm s,$$ where $w_0=B^2/8 \pi + w_{\rm r}$ is the total energy density of magnetic and photon fields. Thus, for a jet magnetic field of about $0.1 \, \rm G$ and a comparable low-frequency photon density ($\approx 4 \cdot 10^{-4} \, \rm erg/cm^3$) a radiative cooling time of less than 5 hours (in the jet frame) could be easily achieved.
### $\pi^0$ origin of gamma rays
The lack of correlation between spectral shape and absolute flux, as well as the very fact that [$\gamma$-rays ]{}with energy $\geq 16 \, \rm TeV$ are observed, could also be explained, perhaps even in a more natural way, by the assumption of a ‘$\pi^0$-decay’ origin of the $\gamma$ -radiation. Yet the efficiency of this mechanism in the jet appears to be too low to explain the observed time variability and the high fluxes of the TeV radiation. This is due to the low density $n_{\rm H}$ of the thermal electron-proton plasma in the jet. The problem of variability could be at least in principle overcome by invoking adiabatic losses caused by relativistic expansion of the emitting “blob”. However, this assumption implies very inefficient [$\gamma$-ray ]{}production with a luminosity $L_{\gamma}=
L_{p}t_{\rm ad}/t_{\rm pp}^{\pi^0}$, where $L_{\rm p}$ is the luminosity in relativistic protons, $t_{\rm ad}\geq R/c \sim \Delta t \delta{\rm j}$ is the adiabatic cooling time, and $t_{\rm pp}^{\pi^0}
\sim 5 \cdot 10^{15}(n_{\rm H}/1
{\rm cm}^{-3})^{-1} \, s$ denotes the characteristic emission time scale of $\pi^0$-decay $\gamma$-rays. We shall assume here that the proton luminosity $L_{\rm p}$ should not exceed the total power of the central engine, roughly the Eddington luminosity $L_{\rm E} = 1.3 \cdot 10^{45} \, (M/M_{\odot})
\, \rm erg/s$ of a supermassive Black Hole of mass $M=10^7 M_{\odot}$, or more empirically, the apparent ($4 \pi$) total luminosity of the source which is $L_{\rm tot} = 4\cdot {\pi}d^2\cdot f_{\rm tot} \sim 10^{45}$ erg/s, where $f_{\rm tot} \sim {\rm few} \times
10^{-10} \, \rm erg /cm^2 s$ is the total observed radiative flux (see e.g. Pian et al. [@Pian:98]). Then the observed TeV-flux of about $2.5 \cdot 10^{-10}$ erg/s requires a lower limit $n_{\rm H}
\geq 10^6 {\rm cm}^{-3}$ on the density of the thermal plasma in the jet. This makes the relativistically moving ‘blob’ very heavy ($M \sim 0.05 M_{\odot}$) with an unacceptably large kinetic energy $E_{\rm kin} = M c^2 \Gamma_{\rm j}
\simeq 10^{54} \Delta t_{\rm day}^3 \delta_{10}^4 (n/10^{6} \, \rm cm^{-3})$ erg.
We would like to emphasize that these arguments hold against the $\pi^0$-origin of [$\gamma$-rays ]{}produced in a small-scale jet; they do not in general exclude hadronic models. In particular, scenarios like the one assuming $\gamma$-radiation produced by gas clouds that move across the jet (Bednarek & Protheroe [@BedPro:97]), Dar & Laor [@dar:97]) remain an attractive possibility for hadronic models. They do not exclude either a “proton blazar” model (Mannheim [@Man:93]). It implies a secondary origin of the relativistic electrons that are the result of electromagnetic cascade, triggered by photo-meson processes involving extremely high energy protons in a hadronic jet (see Mannheim [@Man:98] and references therein).
Intergalactic extinction
------------------------
The effect of intergalactic extinction of VHE [$\gamma$-rays ]{}in diffuse extragalactic background radiation fields became astrophysically significant (see e.g. Stanev & Franceschini [@Stanev:98], Funk et al. [@Magn:97], Biller et al. [@Bill:98], Stecker & de Jager [@Stec:98], Biller [@vbill:98], Primack et al. [@vprim:98], Stecker [@vsteck:98]) after the discovery of TeV radiation from Mkn 421 and Mkn 501 up to energies of 10 TeV, as reported by the Whipple (Zweerink et al. [@Zv:97]), and HEGRA (Aharonian et al. 1997a), CAT (Djannati-Atai et al. [@CATcorrelation]), and Telescope Array (Hayashida et al. [@hay:98]) groups.
### Gamma ray absorption
If we ignore the appearance of second generation $\gamma$-rays (see Sect. 6.2.2), then extinction is reduced to a simple absorption effect which can be described by a single absorption optical depth $\tau$.
The optical depth $\tau$ of the intergalactic medium for a $\gamma$-ray photon of energy $E$, emitted from a source at the distance $d=c z/H_0$, can be expressed for small redshifts z$\ll$ 1 in a convenient approximate form using a quantity $\tau=\tau^{\prime} \xi$, where $$\tau^{\prime}(E) = \frac{\sigma_{\rm T}}{4} \,
\epsilon_{\rm m} n(\epsilon_{\rm m}) \, d \simeq
\label{taup}$$ $$\hspace*{5mm}
0.08 \left(\frac{\epsilon_{\rm m}^2 n(\epsilon_{\rm m})}{10^{-3} \,
\rm eV/cm^3)}\right) \,
\left(\frac{z}{0.034}\right)
E_{\rm TeV} \, H_{\rm 60}^{-1}$$ with $H_{\rm 60}=H_0/60 \, \rm km/s \, Mpc$, $E_{\rm TeV}=E/1 \, \rm TeV$, $\epsilon_{\rm m}=4 m^2_{\rm e} c^4 /E \simeq 1 E_{\rm TeV}^{-1} \,
\rm eV$, and $\xi$ being a correction factor which accounts for the specific form of the differential DEBRA photon number density $n(\epsilon)$; the background photon energy is denoted by $\epsilon$. This expression is based on the narrowness of the $\gamma \gamma \rightarrow e^+ e^-$ cross-section $\sigma_{\gamma
\gamma}$ as a function of $(\epsilon /\epsilon_{\rm m})$ which peaks at $\epsilon / \epsilon_{\rm m} \simeq 1$ in an isotropic field of background photons (Herterich [@her:74]). Thus for a large class of broad DEBRA spectra $n(\epsilon)$ the optical depth is essentially caused by background photons with energy centered around $\epsilon_{\rm m}$. For (broad) power-law spectra, $n(\epsilon)=n_0 \epsilon^{-\gamma}$, the optical depth can be calculated analytically as $\tau(E)=\eta(\gamma) \cdot 4^\gamma (\sigma_{\rm T}/4)
\epsilon_{\rm m} n(\epsilon_{\rm m}) \, d$, where $\eta(\gamma)=7/6 \gamma^{-5/3} (1+\gamma)^{-1}$ (Svensson [@sven:87]). Thus for relatively flat power-law spectra with $1 < \gamma < 2.5 $ we obtain $\xi=4^\gamma \eta(\gamma) \simeq 2$, i.e. approximately half of $\tau_{\gamma \gamma}$ is contributed by background photons with $\epsilon$ in the interval given by $\epsilon_{\rm m} \pm 1/2
\epsilon_{\rm m}$.
We note that $\tau(E) \propto E^{\gamma
-1}$ for a power-law spectrum of DEBRA . For example, within the ‘valley’ of the energy density at mid infrared wavelengths from several $\mu \rm m$ to several tens of $\mu \rm m$, where the energy density is expected to be more or less constant (i.e. $\epsilon^2
n(\epsilon)= {\rm const}$), the intergalactic extinction of [$\gamma$-rays ]{}is largest at the highest observed energies. In particular, according to Eq. \[taup\], even at a very low and probably unrealistic level of the DEBRA intensity of $\epsilon^2 n(\epsilon)=10^{-4} \, \rm eV/cm^3$ at $\lambda \sim 30 \,
\rm \mu m$, approximately $35 \, \rm percent$ of the 25 TeV $\gamma$-rays emitted by Mkn 501 are extinguished before they reach the observer. This implies that the observed TeV spectrum of Mkn 501 contains important information about the DEBRA, at least at wavelengths $\lambda \geq 10 \mu \rm m$. Moreover, it is quite possible that a non-negligible intergalactic extinction takes place also at [*low*]{} [$\gamma$-ray ]{}energies due to interactions with near infrared (NIR) background photons. Interpreting the power-law shape of the spectrum at $2 \, \rm TeV$ as an indication for weak extinction by, say, a factor less than or equal to 2, and ignoring the contributions from all other wavelengths beyond the interval $0.5 \pm 0.25 \, \rm eV$ ($\lambda \simeq 2.5^{\rm +2.5}_{\rm -0.85} \, \, \mu \rm m$) from Eq. \[taup\], one obtains $\epsilon^2 n(\epsilon) \leq 4
\times 10^{-3} \, H_{\rm 60} \; \rm eV/cm^3$, not far from the flux of DEBRA experimentally inferred (Dwek et al. [@dwek:98], de Jager & Dwek [@gt:98]) and theoretically expected (Malkan & Stecker [@malk:98], Primack et al. [@vprim:98]) at these wavelengths.
However, to the extent that the source spectrum of the [$\gamma$-rays ]{}is unknown, this cannot be considered as a model-independent upper limit (e.g. Weekes et al. [@GRO4rev:97]). Obviously for any quantitative estimate of the DEBRA one needs to know the intrinsic [$\gamma$-ray ]{}spectrum, and this can not be obtained from [$\gamma$-ray ]{} observations alone. As already emphasized above, a multi-wavelength approach is indispensable, for example in the form of detailed modeling of the [*entire*]{} or at least a large wave-length range of the nonthermal spectrum. To be specific, one avenue would be modeling the nonthermal X-ray and [$\gamma$-ray ]{}spectra in the framework of synchrotron-inverse Compton models, based on simultaneous multi-wavelength observations of X-ray selected BL Lac objects (Coppi & Aharonian [@CoAh:98]).
An illustration for the need of more than [$\gamma$-ray ]{}data alone is given by a DEBRA spectrum $n(\epsilon) \propto \epsilon^{-1}$. It does not change the spectral shape of [$\gamma$-rays ]{}at all (grey opacity), although formally the extinction could be arbitrarily large. The curves marked as “1” in Fig. \[fig1\_dis\] and Fig. \[fig2\_dis\] may serve as an example. In the case of Mkn 501 this ambiguity can be significantly reduced by rather general arguments regarding $\gamma$-ray [*energetics*]{}. Requiring again that the [$\gamma$-ray ]{}luminosity should not exceed the total apparent luminosity of the source, $L_{tot} \sim 10^{45} \, \rm erg/s$, we must have $\rm e^{\tau} \leq 10^4 (\delta_{\rm j}/10)^{4}$. For [$\gamma$-ray ]{}production in the jet with $\delta_{\rm j} \sim 10$, this implies $\tau \leq 10$. In fact, already a value of $\tau \sim 3$ (corresponding to curve 1 of Fig. \[fig1\_dis\]) creates uncomfortable conditions for the majority of realistic models of high energy radiation from Mrk 501, assuming that the nonthermal emission is produced in the jet due to synchrotron and inverse Compton processes. Indeed, $\tau_{\gamma \gamma} \sim 3$ implies that the $\gamma$-ray luminosity of the source corresponding to the “reconstructed spectrum” (curve 1 in Fig. \[fig2\_dis\]) exceeds the luminosity of the source in all other wavelengths by an order of magnitude which hardly could be accepted for any realistic combination of parameters characterizing the jet.
Accepting the current lack of reliable knowledge of the [$\gamma$-ray ]{}source spectrum, it is nevertheless worthwhile to derive upper limits on DEBRA by formulating different [*a priori*]{}, but [*astrophysically*]{} meaningful requirements on the spectrum and the [$\gamma$-ray ]{}luminosity of the source. A possible criterion, for example, could be that within any reasonable source model the intrinsic spectrum of $\gamma$-rays, $J_0(E)=J_{\rm obs} \, \exp(\tau)$, should not contain a feature which exponentially rises with energy at any [*observed*]{} [$\gamma$-ray ]{}energy. In practice this implies that the ‘reconstruction’ factor $\exp(\tau)$ should not significantly exceed the exponential term of the observed [$\gamma$-ray ]{}spectrum from Eq. 6. This condition is most directly fulfilled by the power law-spectrum with $\gamma = 2$ that has equal DEBRA power per unit logarithmic bandwidth in energy. It results in $\tau (E)\propto E$ and then yields an upper limit for the DEBRA density close to $\epsilon^2 n(\epsilon)=10^{-3} \rm eV/cm^3$. This limit corresponds to the borderline on which the source spectrum becomes a pure power-law Figs. \[fig1\_dis\] and \[fig2\_dis\], curves 2. A slight increase of the DEBRA density by as little as a factor of 1.5 leads to a dramatic (exponential) deviation of the reconstructed spectrum at the highest observed [$\gamma$-ray ]{} energies around 16 TeV from the $E^{-1.9}$ power-law extrapolation.
A power law with $\gamma = 3$ would give similar and complementary results. The resulting source spectra (Fig. \[fig2\_dis\]) and the upper limits on the DEBRA density (Fig. \[fig1\_dis\]) obtained in this way assuming power-law spectra for DEBRA with $\gamma = 2$ and 3, are given by the curves 2 and 3, respectively. The power law $\gamma = 1$ corresponding to $\tau_{\gamma \gamma}=3$ complements Fig. \[fig1\_dis\].
It should be noted however that any realistically expected spectrum of the DEBRA in a broad range of wavelengths deviates from a simple power-law. In fact, all models of the DEBRA, independently of the details, predict two pronounced peaks in the spectrum at 1 $\mu $m and 100 $\mu $m contributed by the emission of the stars and of the interstellar dust, respectively, and a relatively flat ‘valley’ at mid IR wavelengths around 10 $\mu$m (see e.g. Dwek et al. [@dwek:98]). The strong impact of the DEBRA spectrum on calculations of the opacity of the intergalactic medium has been emphasized by Dwek & Slavin ([@DwSl:1994]) and Macminn and Primack ([@Pri:96]).
Note that the power law with $\gamma \sim 1$ characterizes the shape of the spectrum of DEBRA at near IR wavelengths, typically between 1 and several microns, and the power-law with $\gamma \geq 3$ characterizes the DEBRA between 10 and 100 microns. Interestingly, the sum of these two power law spectra “reproduces” a reasonable shape of the ‘valley’. This is seen in Fig. \[fig1\_dis\] and Fig. \[fig3\_dis\] where the measured fluxes or estimated upper and lower limits obtained directly at different wavelengths of the DEBRA are shown.
Finally, an interesting numerical criterion for the derivation of upper limits on the DEBRA was suggested by Biller et al. ([@Bill:98]). It relies on independent $\gamma$-ray observations. A variant of this approach, where the additional restriction forbids a differential source spectrum harder than $\propto E^{-1.5}$ within the observed energy range, is shown by the horizontal bars in Fig. \[fig1\_dis\].
The results described above could be “improved” assuming a more realistic, $\tau_{\gamma \gamma}=1$ optical depth for the $\gamma = 1$ power law branch at short wavelengths, and a steeper, $\gamma = 4$, power law branch at long wavelengths (see Fig. \[fig3\_dis\]). The latter choice for $\gamma$ is due to the rapid rise of the data points towards far infrared wavelengths $\lambda$ in order to fit the recent measurements of the flux at $140 \, \mu \rm m$ by DIRBE aboard the COBE satellite. For the far infrared (FIR) branch the absolute flux of the power-law with $\gamma = 4$ is chosen again from the condition that the differential $\gamma$-ray source spectrum should not exponentially rise at energies up to $E \sim 16 \, \rm TeV$ (see Fig. \[fig4\_dis\]). Note also that the criterion of $\tau_{\gamma \gamma}=1$ for the $\gamma = 1$ branch at short wavelengths is pretty close to the level of the flux of the recent tentative detection of DEBRA at $3.5 \, \mu \rm m$ (Dwek & Arendt [@DwAr:98]). The sum of the NIR and FIR power-law branches with the above indices and absolute fluxes results in a deeper mid IR “valley” and predicts a very steep spectrum of DEBRA between 30 and 100 $\mu \rm m$. In Figs. \[fig3\_dis\] and \[fig4\_dis\] this spectrum has been truncated at $\lambda \simeq 80 \mu \rm m$ that corresponds to the kinematic threshold of pair production at interactions with the maximum observed energy of [$\gamma$-rays ]{}of about 20 TeV. The spectrum is also truncated at $\lambda=1 \mu \rm m$ in order to avoid significant excess compared with the fluxes at optical/UV wavelengths recently derived from the Hubble Deep Field analysis (Pozzetti et al. [@Poz:98]).
The effect of “reconstruction” of the [$\gamma$-ray ]{}spectra of Mkn 501 corresponding to this “best estimate” of DEBRA between 1 and 80 $\mu$m is illustrated in Fig. \[fig4\_dis\]. It shows, in the simple absorption picture (using two truncated power-laws) that even a conservative choice for the DEBRA field implies intergalactic extinction at all observed energies. Especially the reconstructed spectrum around 1 TeV could be considerably harder than the observed spectrum, with a maximum of $E^2 \, {\rm d}N/{\rm d}E$ at 2 TeV (see Fig. \[fig\_spectrum\]). This also demonstrates that it would be dangerous to interpret the observed spectral slope at low energies in terms of a power law extending from still lower energies.
### The effect of cascading in the DEBRA
The discussion of intergalactic $\gamma - \gamma$ absorption effects is in principle incomplete without considering secondary radiations. Briefly, when a [$\gamma$-ray ]{}is absorbed by pair production, its energy is not lost. The secondary electron-positron pairs create new [$\gamma$-rays ]{}via inverse Compton scattering on the 2.7 K MBR. The new [$\gamma$-rays ]{}produce more pairs, and thus an electromagnetic cascade develops (Berezinsky et al. [@Berez:90], Protheroe & Stanev [@Prostan:93], Aharonian et al. [@AhCo:94]). In fact, in our discussion of absorption we have neglected any secondary [$\gamma$-rays ]{}in the field of view of the detector and we finally turn now to these.
For a primary [$\gamma$-ray ]{}spectrum ${\rm d}N/{\rm d}E$ harder than $E^{-2}$, extending to energies $E \gg 1 {\rm TeV}$, the cascade spectrum at TeV energies could strongly dominate over the primary [$\gamma$-ray ]{}spectrum. In addition, the spectrum of the cascade [$\gamma$-rays ]{}that reach the observer has a standard form independent of the primary source spectrum with a characteristic photon index of $1.8-2.0$ at energies between $\sim 100~\rm GeV$ and an exponential cutoff determined by the condition $\tau(d,E)=1$. Thus, somewhat surprisingly, the measured time-averaged spectrum of Mkn 501 can in principle be fitted by a cascade [$\gamma$-ray ]{}spectrum for a reasonable DEBRA flux level of about $10^{-3} \, \rm eV/cm^3$, provided that the [*invisible*]{} source spectrum extends well beyond 25 TeV, say up to 100 TeV.
For an intergalactic magnetic field (IGMF) $B > 10^{-12} \,
\rm G$ the cascade [$\gamma$-rays ]{}could be observed in the form of an extended emission from a giant pair halo with a radius up to several degrees formed around the central source (Aharonian et al. [@AhCo:94]). Although the possible extinction of [$\gamma$-rays ]{}from Mkn 501, at least above 10 TeV, unavoidably implies the formation of a pair halo, the TeV radiation of Mkn 501 cannot be attributed to such a halo simply due to arguments based on the detected angular size and the time variability of the radiation.
However, the speculative assumption of an extremely low IGMF still allows an interpretation of the observed TeV [$\gamma$-rays ]{}of Mkn 501 within the hypothesis of a cascade origin. Instead of extended and persistent halo radiation, we expect in this case that the cascade [$\gamma$-rays ]{}penetrate from the source to the observer almost on a straight line. Yet at cosmological distances to the source even very small deflections of the cascade electrons by the IGMF lead to significant time delays of the arriving $\gamma$-rays: $\Delta t_{\rm B} \simeq 2.4 (d/1 \,
{\rm
Gpc})(E/1 \, {\rm TeV})^{-2} (B/10^{-18} \, {\rm G})^2 \, \rm days$ (Plaga [@Plaga:95]). This implies that in order to see more or less synchronous activity (within several days or less) of Mkn 501 at different wavelengths, as was observed during multiwavelength observations of the source (see e.g. Pian et al. [@Pian:98]), we would have to require $B \leq 10^{-18} \, \rm G$. Although indeed quite speculative, such small magnetic fields on spatial scales large compared to 1 Mpc cannot be [*a priori*]{} excluded (e.g. Kronberg [@Kron:96]).
Conclusions
===========
In this paper we have presented the 1997 Mkn 501 time averaged spectrum as measured with the HEGRA IACT system in the energy range from 500 GeV to 24 TeV. The absence of strong temporal evolution as well as a significant correlation of the emission strength and the spectral shape in the energy region from 500 GeV to 15 TeV made the determination of a time averaged spectrum astrophysically meaningful. Due to unprecedented [$\gamma$-ray ]{}statistics and the 20% energy resolution of the instrument, it was possible to detect for the first time $\gamma$-rays from an extragalactic source with energies well beyond 10 TeV, and to measure a smooth, curved energy spectrum deeply into the exponential regime. We found, that the spectrum above 0.5 TeV is well described by a power law with an exponential cutoff $dN/dE$ = $\, N_0 \,E^{-1.92}\,\exp{(-E/6.2 \, \rm TeV)}$, the highest recorded photon energies being 16 TeV or more.
The detection of TeV $\gamma$-rays from Mkn 501 leads to the unavoidable conclusion that the observed $\gamma$-radiation is produced in a relativistic jet with a Doppler factor $\delta_{\rm j} \geq 10$. Actually, assuming a pure power-law [production]{} spectrum of $\gamma$-rays we may naturally explain the exponential cutoff in the observed spectrum spectrum by an internal $\gamma$-$\gamma$ absorption in the jet. Because of the strong dependence of the optical depth on the jet’s Doppler factor, this hypothesis gives an accurate determination of the latter, $\delta_{\rm j}=8.5$. Remarkably, the uncertainty of this estimate, which is mainly due to the uncertainty in the value of the Hubble constant, $H_0 \simeq 60^{+40}_{-20} \, \rm
km/c Mpc$ and in the measured energy of the exponential cutoff $E_0=6.2 \, \rm TeV$, does not exceed $20 \%$. However, since there could be other reasons for the steepening of the TeV spectrum, this estimate can only be considered as a robust lower limit on $\delta_{\rm j}$. In particular, the steepening of the $\gamma$-ray spectrum at the highest energies could be attributed to an exponential cutoff in the spectrum of accelerated particles, as well as – in the case of the inverse Compton origin of $\gamma$-rays – to the Klein-Nishina effect.
In addition, a modification of the intrinsic (source) spectrum of TeV $\gamma$-rays takes place during their passage through the intergalactic medium. The recent claims about tentative detections of the diffuse extragalactic background radiation by the DIRBE instrument aboard the COBE at near infrared ($\lambda=3.5 \mu \rm m$) and far infrared ($\lambda=140 \mu \rm m$) wavelengths, both at the $\simeq 10 \, \rm nW/cm^2 sr$ flux level imply a strong effect of intergalactic $\gamma$-$\gamma$ extinction on the observed Mkn 501 spectrum over the entire energy region measured by HEGRA. In particular, the shape of the highest energy part of the observed $\gamma$-ray spectrum combined with the DIRBE flux at $140 \mu \rm m$ requires a very steep ($\nu F_\nu \propto \lambda^{s}$ with $\nu \sim 4$) spectrum of the DEBRA with a characteristic flux at mid infrared wavelengths ($\lambda \sim 10-30 \mu \rm m$) around 1 to 2 $\rm nW/m^2 ster$. Due to the expected flat spectrum of DEBRA at near infrared wavelengths (close to $\nu F_\nu \propto \lambda^{-1}$) the modification of the $\gamma$-ray spectrum is less prominent at energies of a few TeV, although the absolute extinction could be very large. This does not allow us to draw definite conclusions about the absorption effect based on the analysis of the shape of the $\gamma$-ray spectrum. Nevertheless, in the case of Mkn 501 this ambiguity can be significantly reduced by rather general arguments regarding the $\gamma$-ray [*energetics*]{}. Indeed, even relatively modest assumption about the optical depth of about $\tau \sim 3$ which corresponds to a flux of the DEBRA in the K-band ($\lambda=2.2 \mu \rm m$) of about $30 \rm \, nW/m^2 ster$ creates uncomfortable conditions for any realistic models of the high energy radiation from Mrk 501. Thus, this value of the flux of the DEBRA may be considered as a rather strong upper limit comparable with the DIRBE upper limit at this wavelength.
To summarize, our excursion through the nonthermal physics of AGNs shows that [$\gamma$-ray ]{}observations alone do not allow a unique interpretation of a spectrum like the one we have presented for Mkn 501, even though one can make a number of highly interesting inferences. In particular, our current poor knowledge about the intrinsic spectrum of Mkn 501 does not allow us to make definite conclusions about the effect of the intergalactic absorption of TeV $\gamma$-rays. And [*vice versa*]{}, the uncertainty in the density of DEBRA does not allow us to take into account the effect of the intergalactic absorption, and thus to “reconstruct” the intrinsic $\gamma$-ray spectrum. The knowledge of the latter is an obvious condition for the quantitative study of the acceleration and radiation processes in the source. Therefore we believe that decisive progress in this field could be achieved by the analysis of both the spectral and temporal characteristics of X-ray and TeV $\gamma$-ray emissions obtained during the multiwavelength campaigns of several X-ray selected BL Lac objects at different states of activity, and located at different distances within 1 Gpc. Due to the time variability of such sources this requires simultaneous observations together with extensive theoretical modeling. Given these we expect to obtain independent knowledge of the diffuse intergalactic radiation fields which can be compared with direct measurements and models of galaxy formation. We hope that the successful multiwavelength campaigns of Mkn 421 and Mkn 501 in 1998 with participation of several X-ray satellites and the HEGRA IACT system will provide highly interesting results in this area.\
[*Acknowledgments*]{}. The support of the German ministry for Research and technology BMBF and of the Spanish Research Council CYCIT is gratefully acknowledged. We thank the Instituto de Astrophysica de Canarias for the use of the site and for supplying excellent working conditions at La Palma. We gratefully acknowledge the technical support staff of the Heidelberg, Kiel, Munich, and Yerevan Institutes.
Systematic errors on the shape of the spectrum
==============================================
To illustrate the effect of systematic errors for determination of energies and effective areas in more detail, we consider here the following model. For simplicity, we consider a detector consisting of a single Cherenkov telescope, which is located inside a uniformly illuminated Cherenkov light pool of area $A_{\rm pool}$. To good approximation, the light yield $I$ (in photons per m$^2$) is proportional to the shower energy, $$I = \alpha E$$ neglecting logarithmic corrections. The response of the telescope is now characterized by two quantities, namely the total intensity $Q$ detected in the image (usually the [*size*]{}, measured e.g. in units of ADC channels, or, with a conversion factor, in units of ‘photoelectrons’), and the signal $V$ induced in the highest pixel(s), which is fed into the trigger circuitry. Both $Q$ and $V$ should be proportional to $I$, $$Q = \beta I \, ; \, V = \gamma I \, ,$$ but they are influenced by rather different factors. While both $\beta$ and $\gamma$ include the mirror reflectivity, the PMT quantum efficiency, and the PMT gain, $V$ is to a much higher degree sensitive to the point spread function of the mirror, and to the shape of the signal generated by the PMT [^2]. The effective detection area is given by $$A = P(V) A_{\rm pool}$$ where $P$ is the trigger probability for a given value of $V$. The fact that triggering and energy determination are not based on identical quantities is the key origin of systematic errors, in particular in the threshold region.
In the analysis of data, values $\alpha_{\rm MC}$, $\beta_{\rm MC}$ and $\gamma_{\rm MC}$ are assumed for these constants, usually based on Monte Carlo simulations. The values may differ from the true values due to imperfections in the parameterization of the atmosphere ($\alpha$) or of the optics and electronics of the telescope ($\beta$, $\gamma$). The reconstructed energy $\tilde E$ of a shower of true energy $E$ is then, in the absence of fluctuations, $$\tilde E = \alpha_{\rm MC}^{-1} \beta_{\rm MC}^{-1} Q =
\alpha_{\rm MC}^{-1} \beta_{\rm MC}^{-1} \beta \alpha E \equiv f E~.$$ Based on eqs. \[eqa\] and \[eqb\] and using ${\cal R}(E',E)=\delta(E'-f\,E)$ the rate of events with measured energy $\tilde E$ is then $$r(\tilde E) = \phi(\tilde E / f) P(\alpha
\gamma \tilde E /f) A_{\rm pool} / f~.$$ In the analysis, the Monte Carlo simulated rate $r_{MC}$ is used to evaluate the effective area $$r_{MC}(\tilde E) = \phi_{\rm MC}(\tilde E) P(\alpha_{\rm MC}
\gamma_{\rm MC} \tilde E) A_{\rm pool} \, ,$$ resulting in a reconstructed flux $$\phi_{rec}(\tilde E) = \phi_{MC}(\tilde E) { r(\tilde E) \over
r_{MC}(\tilde E)}$$ $$\hspace*{1.4cm} = {1 \over f} \phi(\tilde E / f)
{P(\alpha \gamma \tilde E /f) \over P(\alpha_{MC} \gamma_{MC} \tilde E)}~~~.$$ Incorrect constants used in the simulation may hence result in 1) a factor $f$ modifying the energy scale, but not the shape of the spectra, and 2) a change of the shape of the spectra in particular in the threshold region, where $P(V)$ varies steeply with $V$. The scale factor $f$ cannot be determined from IACT data alone; some external reference is required. The second effect – the distortion of the spectra – disappears provided that $\alpha \gamma /f = \alpha_{MC} \gamma_{MC}$, i.e., $\gamma_{MC} / \beta_{MC} = \gamma / \beta$, assuming that the simulation correctly accounts for the statistical fluctuations determining the shape of $P(V)$. This condition can be checked internally within the data set. In particular, one should compare the distribution in $Q$ for a given value of $V$ (or for a fixed range in $V$, $V > V_0$) in the data and in the simulation (see Sect. 4.3 and Fig. 5). If $\gamma / \beta$ differs between data and simulation, the distribution of $Q$ in the threshold region will be different. This comparison also tests the simulation of the shape of $P(V)$.
If the thresholds in $Q$ are shifted by a factor $1+\epsilon$ in the data relative to the simulation, this implies that $\gamma_{MC} / \beta_{MC}$ is off by the same factor, resulting in the flux error given by Eq. \[eqd\].
[^1]: Replacing the 200 m by a 300 m restriction, the number of events reconstructed between 5 TeV and 10 TeV, and above 10 TeV increases by 20% and 35% after $\gamma$-ray selection cuts, respectively.
[^2]: In HEGRA, like in most other IACTs, the ADCs effectively measure the integral charge delivered by the PMTs, regardless of the exact time-dependence of the signal. In contrast, the trigger circuitry is voltage-sensitive, and a signal of a given charge may or may not trigger, depending on the exact waveform.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'Laboratorium für Festkörperphysik, ETH-Hönggerberg, 8093 Zürich, Switzerland'
- 'National High Magnetic Field Laboratory, Tallahassee, FL 32306-4005, USA '
- 'Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA'
author:
- 'Ch. Wälti and H. R. Ott'
- 'Z. Fisk'
- 'J. L. Smith'
title: 'Spectroscopic evidence for unconventional superconductivity in UBe$_{13}$'
---
In a recent letter [@Letter], we reported the observation of giant zero-bias conductance peaks (ZBCPs) for contacts between Au wires and bulk UBe$_{13}$, which we ascribed to the existence of low-energy Andreev surface bound-states in superconducting UBe$_{13}$. In the comment [@Gloos] preceding this reply, Gloos proposes an alternative interpretation of the ZBCPs observed in Ref. which is based on Wexler’s formula [@Gloos]. Below, we first show that the criticism of our interpretation by Gloos is not relevant for the experimental situation discussed in Ref. and second, we argue that the observed ZBCPs in the UBe$_{13}$–Au contacts are not compatible with the interpretation given in Ref. .
In Ref. , Gloos discusses the properties of normal metal–normal metal (NN) contacts by means of Wexler’s formula, from which he derives a condition claiming that the UBe$_{13}$–Au contacts discussed in Ref. are not in the limit where pure Andreev reflection (AR) [@2498] may occur. Pure AR, or electron-hole reflection, at the interface between a normal conductor and a superconductor (NS) may cause an enhancement of the differential conductivity $G(E)$ for $|E|<\Delta$, where $\Delta$ denotes the amplitude of the superconducting energy gap, but only in the limit of a low potential-barrier at the interface [@2487]. The mean free path $l$ in the contact, which is discussed in Ref. , is large, if the potential-barrier is small and vice versa. However, in our discussion of the observed ZBCPs we do not claim to observe an enhancement of the differential conductivity of the contact due to pure AR. Instead we claim to observe a surface resonance phenomenon, which is caused by the presence of subgap Andreev bound-states. These two AR-type phenomena are simply not the same.
Wexler’s formula, in the form used in Ref. , is, apart from other criteria, based on the assumption that the Fermi-liquid parameters and hence the resistivity $\rho$ of both metals on either side of the contact are at least similar. Since in Ref. one of the metals of the contact is the heavy-electron metal UBe$_{13}$, this assumption is certainly not valid, and Wexler’s formula in the form as employed by Gloos [@Gloos] needs to be revised. In addition, the application of Wexler’s formula for estimating the contact radius $a$ using the contact resistance $R$ and $\rho$ of bulk UBe$_{13}$ is, in our opinion, an invalid step.
The interpretation of our experiments [@Letter] which is proposed in Ref. is based on the assumption that the loss of the electrical resistance of UBe$_{13}$ in its superconducting state is responsible for the huge enhancement of the differential conductivity at zero energy. The discussion given in Ref. only embraces the temperature dependence of the conductivity of NN contacts. Discussing the [*energy*]{}-dependent differential conductivity $G(E)$ of, e.g., UBe$_{13}$–Au contacts, however, would require a model including its [*energy*]{} dependence. This important issue is not contained in the discussion presented in Ref. .
It is well known that superconductivity of UBe$_{13}$ with zero electrical resistance is fully established within a few ten mK below $T_c$. Therefore, in the context of Gloos’ model, one would expect, for $|E|<\Delta$, the differential conductivity of the UBe$_{13}$–Au contact to increase sharply at $T_{c}$, reaching its maximum just below $T_c$. In that case there would be no obvious reason why the differential conductivity at zero energy, $G(0)$, should further increase substantially with decreasing temperature, as it is observed for the data discussed in Ref. .
The shape of the $G(E)$ curve of the UBe$_{13}$–Au contact (see Fig. 3 of Ref. ) does also not match the predictions of the model proposed by Gloos. The electrical resistance of UBe$_{13}$ at $T<T_c$ is zero for $|E|<\Delta$. The model discussed in Ref. does not offer an explanation, why $G(E)$ of the UBe$_{13}$–Au contact is only strongly enhanced in a narrow region around zero bias, but substantially [*reduced*]{} compared to the normal state of the contact at intermediate energies $|E|<\Delta$.
We acknowledge helpful discussions with M. Sigrist.
[10]{}
Ch. Wälti, H. R. Ott, Z. Fisk, and J. L. Smith, Phys. Rev. Lett. [**84**]{}, 5676 (2000).
K. Gloos, preceding Comment, Phys. Rev. Lett. (2000).
H. R. Ott, H. Rudigier, Z. Fisk, and J. L. Smith, Phys. Rev. Lett. [**50**]{}, 1595 (1983).
A. E. Andreev, Zh. Eskp. Teor. Fiz. [**46**]{}, 1823 (1964) \[Sovj. Phys. JETP [**19**]{},1228 (1964)\].
G. E. Blonder, M. Tinkham, and T. M. Klapwijk, Phys. Rev. B [**25**]{}, 4515 (1982).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[Aaron Steven White](http://aswhite.net) [Drew Reisinger](http://pages.jh.edu/~dreisin2/) [Rachel Rudinger](http://www.clsp.jhu.edu/people/graduate-students/rachel-rudinger/)\
[**Kyle Rawlins**](http://sites.krieger.jhu.edu/rawlins/) [**Benjamin Van Durme**](http://www.cs.jhu.edu/~vandurme/)\
Johns Hopkins University
bibliography:
- 'Zotero.bib'
title: Computational linking theory
---
Introduction {#sec:introduction}
============
Related work {#sec:relatedwork}
============
Data {#sec:data}
====
Evaluating linking models {#sec:categoricalfeatural}
=========================
Exploring linking models {#sec:proto}
========================
Conclusion {#sec:conclusion}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The pair- and three-proton interaction potentials for metal-phase hydrogen are calculated. Irreducible three-proton interactions are shown to be essential in the development of the metal hydrogen structure. Possible manifestations of the three-ion interactions in the structure of quickly quenched metallic systems are discussed. Existence of amorphous metals is related to manifestations of the three-ion interactions in non-equilibrium conditions when amorphisation front travels through a liquid metal.'
address: |
$^{1}$[*Polytechnic University, 1 Shevchenko av., Odessa, UA-65044,*]{}\
Ukraine\
$^{2}$[*State University, 2 Dvoryanskaya str., Odessa, UA-65100, Ukraine*]{}
author:
- 'E. V. Vasiliu$^{1}$, S. D. Kaim$^{1}$, N. P. Kovalenko$^{2}$'
title: |
[Calculation of an effective three-ionic interactions potential]{}\
[in metallic hydrogen]{}
---
\[theorem\][Acknowledgement]{} \[theorem\][Algorithm]{} \[theorem\][Axiom]{} \[theorem\][Claim]{} \[theorem\][Conclusion]{} \[theorem\][Condition]{} \[theorem\][Conjecture]{} \[theorem\][Corollary]{} \[theorem\][Criterion]{} \[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Exercise]{} \[theorem\][Lemma]{} \[theorem\][Notation]{} \[theorem\][Problem]{} \[theorem\][Proposition]{} \[theorem\][Remark]{} \[theorem\][Solution]{} \[theorem\][Summary]{}
Introduction
============
The inclusion of many-particle interactions into consideration of equilibrium and kinetics properties is the urgent problem in the theory of condensed matter [@1], [@2]. Calculation of the many-particle interaction potentials is the key to this problem. For simple dielectric and metal liquids and for crystals pair interaction potentials are known at least at a phenomenological level,about the three-particle interaction it cannot be said. This situation is explained by general lack of development of the theory of highly nonuniform electron gas. Thus to calculate the three-particle interactions in various condensed systems one needs to know in an explicit form the three-point nonlinear response functions involving cooperative phenomena in the many-electron system, which is formed from atomic electron shells. Nowadays the problems of calculation of such functions for the majority of systems are not even formulated. It should be stressed that for the most systems even the polarization operator of nonuniform electron gas, which makes up these systems, remains unknown. But there are same exceptions. This refers to the simple metals which can be closely approximated by the uniform gas of free electrons. A discrete nature of the ion subsystem can be taken into account by perturbation theory. The nonlinear response function of uniform gas corresponding to three- and four-pole diagrams have been calculated within the framework of many-particle theory of metals [@3].
Calculations of the structural and thermodynamical equilibrium volume properties of metal-phase hydrogen point up the significance of inclusion of many-particle ion interactions, and suggest the possibility of a metastable metallic state at zero external pressure [@3]. Comparison of energies corresponding to different crystal structures have led to the conclusion of liquidity tendencies in the structure of metal hydrogen. Energy difference of crystal and liquid phases approaches zero with decreasing pressure [@4]. So the liquid state of metal hydrogen is possible to form at zero external pressure.
Calculations of energy and other properties of crystal metals are carried out in the reciprocal space and do not require the explicit form of interaction potentials of ion clusters. In amorphous, liquid and heterogeneous metals the calculations in configurational space are preferable. It provides insight into the nature of local atomic order and gives the assessment of correlations in the positions of atom clusters.
The occurrence of a maximum in the temperature dependence of the third virial coefficient evidences directly the existence of three-particle interactions in simple dielectric liquids [@1]. Taking into account the three-atom interaction in the asymptotic form of the Akselrod-Teller three-dipole interaction potential requires the inclusion of a cutoff parameter (the characteristic length) into the theory and does not explain the wave-vector dependence of the third virial coefficient (on the example of krypton) [@5].
In metals the many-ion interactions are directly proven to exist by the anomalies in phonon spectra which could not be reduced to the Kohn’s ones, and conform with the singularities of the many-pole diagrams of uniform electron gas [@3].
Interaction potential in liquid Na and K have been calculated previously [@6] for three-ion equiangular configurations. The pair-interaction potentials for ions in metal hydrogen have been calculated in [@7], where the third-order contributions with respect to electron-ion interaction were shown to be of fundamental importance for development of attractive part of potential. Pair interactions in a liquid metal hydrogen have been calculated within the context of the density-functional method [@8]. The potentials were formulated in terms of direct correlation functions of electron-ion system.
Here we present the pair- and three-ion interaction potentials in metal hydrogen at the Wigner-Seitz radius $r_{S}=1,65$ calculated on a basis of the many-particle theory of metals [@3] within the third-order perturbation theory. The computed array of values of three-ion interaction potential makes possible the further calculations of the contributions of these interactions into various properties of metal hydrogen. Possible outcomes of the three-ion interactions affecting local atomic order in amorphous metal systems are discussed.
Calculation Results
===================
In the context of many-particle theory of non-transition metals the energy of a metal is calculated by the use of the adiabatic approximation for electron-ion system. The electron gas energy $E_{e}$, provided that ions positions are fixed, can be calculated with the theory of perturbations in the potentials of electron-ion and electron-electron interactions [@3]. The energy $E_{e}$ fulfills the role of an electron term upon the treatment of the properties of ion subsystem in metal. Together with kinetic energy of ions and energy of their direct interaction, the energy $E_{e}$ comprises the effective Hamiltonian of the ion subsystem of metal. $E_{e}$ can be considered as a sum of contributions independent of ion positions, dependent on the locations of separate ions, ion pairs, triplets, etc. [@3]:
$$E_{e}=\varphi _{0}+\sum_{n}\varphi _{1}({\bf R}_{n})\ +\ \frac{1}{2!}%
\sum_{m\neq {n}}\varphi _{2}({\bf R}_{n},{\bf R}_{m})\ +\ \frac{1}{3!}%
\sum_{m\neq {n}\neq {l}}\varphi _{3}({\bf R}_{n},{\bf R}_{m},{\bf R}_{l})\
+\cdots$$
Each term of the series (1) describes interactions of ion groups through the surrounding electron gas and can be represented as a power series in the potential of electron-ion interaction [@3]:
$$\varphi _{2}({\bf R}_{1},{\bf R}_{2})\quad =\quad \sum_{i=2}^{\infty }\Phi
_{2}^{(i)}({\bf R}_{1},{\bf R}_{2}),$$ $$\varphi _{3}({\bf R}_{1},{\bf R}_{2},{\bf R}_{3})\quad =\quad
\sum_{i=3}^{\infty }\Phi _{3}^{(i)}({\bf R}_{1},{\bf R}_{2},{\bf R}_{3}),$$ etc., where $\Phi _{n}^{(k)}({\bf R}_{1},\ldots ,{\bf R}_{n})$ represents the indirect interaction of the $n$ ions through electron gas in the $k$-order perturbation theory in electron-ion interaction.
The indirect interaction of two ions described by the potential $\Phi
_{2}^{(2)}\left( \left| {\bf R}_{1}-{\bf R}_{2}\right| \right) $ is well studied. The potential $\Phi _{2}^{(3)}\left( \left| {\bf R}_{1}-{\bf R}%
_{2}\right| \right) $, corresponding to the energy of indirect two-ion interaction in the third-order in potential of electron-ion interaction, has been calculated previously in [@6] for liquid sodium and potassium, and also for metal hydrogen [@7]. It is not difficult to obtain an expression for the $\Phi _{2}^{(3)}\left( \left| {\bf R}_{1}-{\bf R}%
_{2}\right| \right) $:
$$\Phi _{2}^{(3)}(R)\quad =\quad \frac{3}{4\pi ^{4}}\int\limits_{0}^{\infty
}dq_{1}q_{1}^{2}\int\limits_{0}^{\infty
}dq_{2}q_{2}^{2}\int\limits_{-1}^{1}dxV(q_{1})V(q_{2})V(q_{3})\times$$ $$\times \Gamma ^{(3)}(q_{1},q_{2},q_{3})\frac{\sin (q_{1}R)}{q_{1}R},$$ where $V(q)$ is the form-factor of the electron-ion interaction potential; $%
\Gamma ^{(3)}(q_{1},q_{2},q_{3})$ is the sum of three-pole diagrams.
In the third-order perturbation theory in $V(q)$ the indirect three-ion interaction (the irreducible three-ion interaction) is defined by the expression [@6]:
$$\Phi _{3}^{(3)}(R_{1},R_{2},R_{3})\quad =\quad \frac{3}{2\pi ^{4}}%
\int\limits_{0}^{\infty }dq_{1}q_{1}^{2}\int\limits_{0}^{\infty
}dq_{2}q_{2}^{2}\times$$ $$\times \int\limits_{-1}^{1}dzV(q_{1})V(q_{2})V(q_{3})\Gamma
^{(3)}(q_{1},q_{2},q_{3})\int\limits_{0}^{1}dx\times$$ $$\times \cos \Biggl(x\biggl(q_{1}R_{1}\frac{R_{1}^{2}+R_{2}^{2}-R_{3}^{2}}{%
2R_{1}R_{2}}+q_{2}R_{2}z\biggr)\Biggr)\times$$ $$\times {J_{0}}\Biggl(q_{1}R_{1}\biggl(1-x^{2}\biggr)^{1/2}\biggl(1-\frac{%
(R_{1}^{2}+R_{2}^{2}-R_{3}^{2})^{2}}{4R_{1}^{2}R_{2}^{2}}\biggr)^{1/2}\Biggr)%
\times$$ $$\times {J_{0}\Biggl(q_{2}R_{2}\biggl(1-x^{2}\biggr)^{1/2}\biggl(1-z^{2}%
\biggr)^{1/2}\Biggr)},$$ where $J_{0}(x)$ is the Bessel function of zero order; $z\equiv \cos ({\bf q}%
_{1},{\bf q}_{2})$; $q_{3}=(q_{1}^{2}+q_{2}^{2}+2q_{1}q_{2}x)^{1/2}$; $%
R_{1},\ R_{2},\ R_{3}$- the distances between the vertices of a triangle formed by the protons.
The pair- and three-proton interaction potentials were calculated at the Wigner-Seitz radius $r_{S}=1,65$ which corresponds to zero pressure in the zeroth model of a metal. A permittivity function in the Heldart-Vosko form was employed.
Figure 1 shows computed potentials of the two-proton interactions $%
e^{2}/R+\Phi _{2}^{(2)}(R)$, $\Phi _{2}^{(3)}(R)$, and $\varphi ^{\ast
}(R)=e^{2}/R+\Phi _{2}^{(2)}(R)+\Phi _{2}^{(3)}(R)$. It is obvious that the interaction $\Phi _{2}^{(3)}(R)$ significantly renormalizes potential $%
e^{2}/R+\Phi _{2}^{(2)}(R)$. Noteworthy is a minimum in the repulsive part of the potential $\varphi ^{\ast }(R)$, which arises as a result of the contribution $\Phi _{2}^{(3)}(R)$. At $r_{S}=1,72$ this minimum becomes deeper and as a result the potential $\varphi ^{\ast }(R)$ takes the form typical for simple metals. At $r_{S}<1,65$ the minimum in the repulsive part of $\varphi ^{\ast }(R)$ turns shallow and its position shifts towards smaller $R$.
The potentials of irreducible three-proton interaction $\Phi
_{3}^{(3)}(R_{1},R_{2},R_{3})$ calculated at different lengths of sides of a three-proton triangle are listed in Table. It should be noted that the number of various $(R_{1},R_{2},R_{3})$ sets is restricted by the triangle axiom to which the distances $R_{1},R_{2}$ and $R_{3}$ are subject.
The calculation results can be presented more conspicuously and informatively as plots of versus geometry parameters for selected configurations of protons. Thus, Figure 2 demonstrates the potential curve $%
\Phi _{3}^{(3)}(R,R,R)$, that is, one for proton equiangular triplets. The potential $\Phi _{3}^{(3)}(R,R,R)$ corresponds to strong attraction of the triplet at short distances and oscillates at the long ones.
Figure 3 shows plots of the potential $\Phi _{3}^{(3)}$ for isosceles ion triplets. It is clear that attractability of proton triplet is actually short-range. An equilibrium local atomic order results from the balance of direct proton interactions, the indirect two-proton ones, and of the irreducible three-proton interactions.
Figure 4 presents a proton energy plotted against the distance to the two others (the isosceles proton triplet). Two protons are placed in abscissa axis at the points (1,0,0) and (-1,0,0) (distances in atomic units). The third proton is sited in ordinate axis. In Fig. 4 is shown that inclusion of the irreducible potential $\Phi _{3}^{(3)}$ is of fundamental importance for determination of the equilibrium local order in proton spacing. Attractive nature of the $\Phi _{3}^{(3)}$ potential have to reduce the average interionic distances in equilibrium state of metal.
Discussion and Conclusion
=========================
A comprehensive analysis of the atomic properties of simple metals leads to the conclusion that of basic importance is the consideration of three-ion interactions as in polyvalent metals [@9], and also in metal hydrogen where the electron-ion interaction is free from a non-coulomb part [@3]. Calculations of the two-ion interaction potentials for various simple metals display the similarity of their behavior in all such metals [@10]. Our calculations $\Phi _{3}^{(3)}$ for equilateral proton triplets, along with the results of [@6], demonstrate the similarity of the $\Phi _{3}^{(3)}$ behavior in simple metals.
Results of our calculations suggest certain inferences about the possible manifestations of three-particle interactions through the structure of metal systems. The potential $\Phi _{3}^{(3)}(R_{1},R_{2},R_{3})$ is symmetric about interchange of ions and therefore in equilibrium conditions the most probable three-ion configurations in liquid metals would be the equilateral triplets.
Experiments on the quick quenching of metals and alloys show that amorphous phase is readily obtainable in the polyvalent metals and their alloys [@11] where consideration of many-ion interactions are essential. Under the quick quenching conditions aluminum alloys form quasicrystalline structures with specific 5-fold axes of symmetry forbidden for crystals [@12]. The possible reason of quasicrystal formation is manifestation of the three-ion interactions as the amorphisation front moves through a liquid metal.
When two of the three ions are fixed on a surface or located within amorphous phase and the third ion remains in the liquid, energetically advantageous would be an isosceles triplet rather than equilateral one. Owing to the attractive nature of $\Phi _{3}^{(3)}$ the third ion shifts so that one of the angles in the triplet will be more than $60^{\circ }$. Such isosceles triplets can serve as a basis for generation of amorphons in an amorphous phase. Existence of the 5-fold axes of symmetry in quasicrystals corresponds to the presence of the isosceles triplets with an angle of $%
72^{\circ }$. Thus one of the manifestation of three-ion interactions can be the creation of amorphons while propagating the amorphization front in liquid metal.
The authors are grateful to Prof. Yu. P. Krasny for stimulating discussion.
C. A. Croxton.[* Liquid State Physics. A Statistical Mechanical Introduction*]{} (Cambridge University Press, 1974)
// Ed. by C. A. Croxton (Chichester, New York, Brisbane, Toronto: John Wiley and Sons, 1978)
E. G. Brovman, Yu. M. Kagan. Uspekhi Fiz. Nauk [**112**]{}, 369 (1974) (in Russian)
J. E. Jaffe, N. W. Ashcroft. Phys. Rev. [**B23**]{}, 6176 (1981)
A. Teisma, P. A. Egelstaff. Phys. Rev. [**A21**]{}, 367 (1980)
M. Hasegawa. J. Phys. F: Met. Phys. [**6**]{}, 649 (1976)
N. P. Kovalenko, L. M. Kuz’mina, G. Meier. Pis’ma ZhETF [**28**]{}, 239 (1978) (in Russian)
J. Chihara. J. Phys. [**C19**]{}, 1665 (1986)
V. G. Bar’yahtar, E. V. Zarochentsev, E. P. Troitskaya. Computing methods in a solid state physics. Atomic properties of metals (Naukova dumka, Kiev, 1990) (in Russian)
N. P. Kovalenko, Yu. P. Krasny, S. A. Trigger. The statistical theory of liquid metals (Nauka, Moscow, 1990) (in Russian)
Rapidly Quenched metals // Ed. by S. Steeb, H. Warlimont (North-Holland, Amsterdam, Oxford, New York, Tokyo, 1985)
D. Gratias. Les quasi-cristaux. La Recherche [**178**]{}, 788 (Juin 1986)
[Table]{}
$R_{1}$ $R_{2}$ $R_{3}$ $\Phi _{3}^{(3)}(R_{1},R_{2},R_{3})$
--------- --------- --------- --------------------------------------
0 0 0 -0.95758
1 1 1 -0.23684
1 1 2 -0.07582
1 2 2 -0.01523
1 2 3 -0.00326
1 3 3 -0.00353
1 3 4 -0.00411
1 4 4 -0.00167
1 4 5 0.00130
1 5 5 0.00116
1 5 6 0.00005
1 6 6 -0.00055
2 2 2 0.00117
2 2 3 -0.00032
2 2 4 -0.00283
2 3 3 -0.00184
2 3 4 -0.00161
2 3 5 0.00029
2 4 4 -0.00028
2 4 5 0.00065
2 4 6 0.00010
2 5 5 0.00038
2 5 6 -0.00024
$R_{1}$ $R_{2}$ $R_{3}$ $\Phi _{3}^{(3)}(R_{1},R_{2},R_{3})$
--------- --------- --------- --------------------------------------
2 6 6 -0.00033
3 3 3 -0.00134
3 3 4 -0.00030
3 3 5 0.00048
3 3 6 0.00006
3 4 4 0.00027
3 4 5 0.00025
3 4 6 -0.00019
3 5 5 -0.00007
3 5 6 -0.00021
3 6 6 -0.00007
4 4 4 0.00018
4 4 5 -0.00006
4 4 6 -0.00019
4 5 5 -0.00015
4 5 6 -0.00005
4 6 6 0.00006
5 5 5 -0.00004
5 5 6 0.00007
5 6 6 0.00006
6 6 6 0.00003
(distances in the Table in atomic units, energy in Ry)
[Figure Captions]{}
Fig. 1. Potential of the two-proton interactions: 1 - $e^{2}/R+\Phi
_{2}^{(2)}(R)$; 2 - $\Phi _{2}^{(3)}(R)$; 3 - $%
\varphi ^{\ast }(R)=e^{2}/R+\Phi _{2}^{(2)}(R)+\Phi _{2}^{(3)}(R)$
Fig. 2. Potential $\Phi _{3}^{(3)}(R,R,R)$
Fig. 3. Potential $\Phi _{3}^{(3)}$: 1 - $\Phi _{3}^{(3)}(1\,a.u.,R,R)$; 2 - $\Phi _{3}^{(3)}(2\,a.u.,R,R)$; 3 - $\Phi _{3}^{(3)}(3\,a.u.,R,R)$
Fig 4. Potential $\varphi (R)=2\varphi ^{\ast }(R)+\Phi
_{3}^{(3)}(2\,a.u.,R,R)$: 1 - $2\varphi ^{\ast }(R)$; 2 - $\Phi
_{3}^{(3)}(2\,a.u.,R,R)$; 3 - $\varphi (R)$ (see explanations in the text)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In several smart city applications, multiple resources must be allocated among competing agents that are coupled through such shared resources and are constrained — either through limitations of communication infrastructure or privacy considerations. We propose a distributed algorithm to solve such distributed multi-resource allocation problems with no direct inter-agent communication. We do so by extending a recently introduced additive-increase multiplicative-decrease (AIMD) algorithm, which only uses very little communication between the system and agents. Namely, a control unit broadcasts a one-bit signal to agents whenever one of the allocated resources exceeds capacity. Agents then respond to this signal in a probabilistic manner. In the proposed algorithm, each agent makes decision of its resource demand locally and an agent is unaware of the resource allocation of other agents. In empirical results, we observe that the average allocations converge over time to optimal allocations.'
author:
- 'Syed Eqbal Alam$^\ast$[^1], Robert Shorten$^\dagger$[^2], Fabian Wirth$^\ddagger$[^3], and Jia Yuan Yu$^\ast$'
bibliography:
- 'DistOpt\_bib.bib'
title: '**Communication-efficient Distributed Multi-resource Allocation**'
---
***Keywords—* distributed optimization, optimal control, multi-resource allocation, AIMD algorithm, smart city, Internet of things, multi-camera coordination system**
Introduction
============
Smart cities are built on smart infrastructures like intelligent transportation systems, security systems, smart grids, smart hospitals, smart waste management systems, etc., [@Harrison2010; @Zanella2014]. Internet of things (IoT) are the essential building blocks to develop such smart infrastructures [@Hernandez2011; @Mohanty2016], we call these devices as [*Internet-connected devices (ICDs)*]{}. In several smart city applications, multiple resources must be allocated among competing Internet-connected devices that are coupled through multiple resources. Generally speaking, such problems are more difficult to solve than those with a single resource. This is particularly true when Internet-connected devices are constrained — either through limitations of communication infrastructure, or due to privacy considerations. These distributed optimization problems have numerous applications in smart cities and other application areas. The recent literature is rich with algorithms that are designed for distributed control and optimization applications. While this body of work is too numerous to enumerate, we point the interested readers to the works of Nedic [@Nedic2009],[@Nedic2011]; Cortes [@KIA2015]; Jadbabaie and Morse[@Jadbabaie2003]; Bullo [@Bullo2011]; Pappas [@Pappas2017], Bersetkas [@Bertsekas2011]; Tsitsiklis [@Blondel2005] for recent contributions. A survey of some of the related work is given in [@Wirth2014]. In many instances in smart cities and other areas, network of Internet-connected devices achieve optimal allocation of resources through regular communication with each other and/or with a control unit. Motivated by such scenarios, we propose an algorithm that is tailored for these but does not require inter-device communication due to privacy considerations. The proposed solution is based on the generalization of stochastic [*additive-increase and multiplicative-decrease (AIMD)*]{} algorithm [@Wirth2014]. By way of background, the AIMD algorithm was proposed in the context of congestion avoidance in transmission control protocol (TCP) [@Chiu1989]. The AIMD algorithm is further explored and used in several application domains for example, micro-grids [@Crisostomi2014]; multimedia [@Cai2005]; electric vehicle (EV) charging [@Studli2012]; resource allocation [@Avrachenkov2017], etc. Interested readers can refer the recent book by Corless et al. [@Corless2016] for an overview of some of the applications. The authors of [@Wirth2014] demonstrate that simple algorithms from Internet congestion control can be used to solve certain optimization problems. Roughly speaking, in [@Wirth2014], the iterative distributed optimization algorithm works as follows. Internet-connected devices continuously acquire an increasing share of the shared resource, this phase is called [*additive increase*]{} phase. When the aggregate resource demand of Internet-connected devices exceeds the total capacity of resource, then the control unit broadcasts a one bit [*capacity event*]{} notification to all competing Internet-connected devices and these devices respond in a probabilistic manner to reduce the demand, this phase is called [*multiplicative decrease*]{} phase. By judiciously selecting the probabilistic manner in which Internet-connected devices respond, a portfolio of optimization problems can be solved in a stochastic and distributed manner.
Our contribution here is to demonstrate that the ideas therein [@Wirth2014] extend to a much broader (and more useful) class of optimization problems which can be used in many application domains of smart cities and other areas. Our proposed algorithm builds on the choice of probabilistic response strategies described therein but is different in the sense that we generalize the approach to deal with [*multiple resource constraints*]{} and the cost functions are coupled through multiple resources. We show that the optimal values obtained by proposed algorithm is same as if the optimization problem is solved in a centralized way.
In the proposed solution, for a system with $m$ resources, in the worst case scenario the communication overhead is $m$ bits per time unit, which is very low. We would also like to mention that in the proposed solution, the communication complexity is independent of the number of Internet-connected devices competing for resources in the system. In this paper, we present a use case of a smart city that deploys a multi-camera coordination system, in which several cameras coordinate for the surveillance of the city. Each camera has private cost function which is coupled through allocation of multiple resources. Notice that we use the names agent and Internet-connected device interchangeably in this paper. The paper is organized as follows, Section \[prob\_form\] describes the problem and provides the formulation of the problem, it also describes the conditions for optimality. A brief description of classical AIMD algorithm is presented in Section \[prelim\]. Section \[divisible\_mul\_res\] describes the multi-resource allocation strategies. The numerical results are presented in Section \[results\]. The paper concludes with future directions in Section \[conc\].
Problem formulation {#prob_form}
===================
Suppose that a smart city deploys a [*multi-camera coordination system*]{} described in Figure \[Diag\_camera\], in which several cameras work together for the surveillance of the city, these cameras are deployed at different locations. If a camera observes any unusual activity then it should demand the required amount of resources with higher probability than other cameras, to notify the observed activity immediately. Suppose that there are central servers set up by the city, which store and process the videos sent by all the cameras, these servers also act as a control unit. Each camera requires different amount of network bandwidth, CPU cycles, memory (RAM) and storage to transmit, process and store the videos on the central servers. Assume that a camera decides its demand based on its cost function and its previous allocations.
Suppose that there are $n$ Internet-connected devices that compete for $m$ resources $R^1, R^2, \ldots, R^m$ with capacity $C^1, C^2, \ldots, C^m$, respectively. In this paper we assume that these Internet-connected devices are the cameras that compete for memory (say $R^1$), storage ($R^2$) and network bandwidth ($R^3$). We further assume that each camera receives enough CPU cycles to process its data, for the sake of generality we use $m$ resources here. We denote $\mathcal{N}:=\{1, 2, \ldots, n\}$ and $\mathcal{M}:=\{1, 2, \ldots, m\}$ and use $i \in \mathcal{N}$ as an index for cameras and $j \in \mathcal{M}$ to index the resources. We assume that each camera has a private cost function $f_i: \mathbb{R}^m \to \mathbb R$ which associates a cost to a certain allotment of resources. We assume that $f_i$ is twice continuously differentiable, convex, and increasing in all variables, for all $i$. For all $i$ and $j$, we denote by $x_i^j \in \mathbb{R}_+$ the amount of resource $R^j$ allocated to camera $i$. We are interested in the following optimization problem of [*multi-resource*]{} allocation: $$\begin{aligned}
\label{obj_fn1}
\begin{split}
\min_{{x}^1_1, \ldots, {x}^m_n} \quad &\sum_{i=1}^{n} f_i(x^1_i, x^2_i,
\ldots, x^m_i), \\
\mbox{subject to} \quad
&\sum_{i=1}^{n} x^j_i = C^j, \quad j \in \mathcal{M}, \\
&x^j_i \geq 0, \quad i \in \mathcal{N}, \ j \in \mathcal{M}.
\end{split}\end{aligned}$$ Note that there are $nm$ decision variables $x^j_i$ in this optimization problem, for all $i$ and $j$. We denote the solution to the minimization problem by $x^{*} \in \mathbb{R}_+^{nm}$, where $x^* = (x_1^{*1}, \ldots, x_n^{*m})$. By compactness of the constraint set optimal solutions exist. We also assume strict convexity of the cost function $\sum_{i=1}^{n} f_i$, so that the optimal solution is unique.
Suppose that $\mathbb{N}$ denotes the set of natural numbers and $k \in
\mathbb{N}$ denotes the time steps. To this end, we denote by $x_i^j(k)$ and $\overline{x}^j_i(k)$ (refer ) the amount of resource allocated and average allocation at the (discrete) time step $k$, respectively. The camera can obtain any amount in $[0, C^j]$, for all $j$. We define the average allocation for $i \in \mathcal{N}, \ j \in \mathcal{M}$, and $k \in \mathbb{N}$, as follows: $$\begin{aligned}
\label{average_eqn}
\overline{x}^j_i(k)=\frac{1}{k+1} \sum_{\ell=0}^k x^j_i(\ell).\end{aligned}$$ The goal is to propose a distributed iterative scheme, such that the long-term average allocations converge to the optimal allocations: $$\begin{aligned}
\label{eq:longtermopt}
\lim_{k\to\infty} \overline{x}_i^j(k) \to x_i^{*j}, \quad \text{for $i \in \mathcal{N}$ and $j \in \mathcal{M}$}.\end{aligned}$$ Let $\nabla_j f_i(.)$ be (partial) derivative of $f_i(.)$ with respect to resource $R^j$. Similar to [@Syed2018], we write the Lagrange multipliers of , with careful analysis we obtain that the derivatives of cost functions of all cameras competing for a particular resource should make a consensus at optimal allocations, i.e., the following holds true: $$\begin{aligned}
\label{optimality}
\nabla_j f_i \big( x_i^{*1}, \ldots, x_i^{*m} \big) = \nabla_j f_u \big ( x_u^{*1}, \ldots, x_u^{*m} \big ), \nonumber\\ \mbox{ for all } i,u \in \mathcal{N} \mbox{ and } j \in \mathcal{M},\end{aligned}$$ which satisfies all the Karush-Kuhn-Tucker (KKT) conditions. KKT conditions are necessary and sufficient condition for optimality of convex problem , interested readers may refer Chapter 5.5.3 [@Boyd2004] for a detailed discussion on KKT conditions. Now, to check the efficacy of our results we use the consensus of derivatives of cost function of all cameras with respect to a particular resource and show that the average allocation converge to the optimal allocation.
A primer on AIMD {#prelim}
================
The AIMD algorithm is of interest because it can be tuned to achieve optimal distribution of a single resource among a group of agents. To this end no inter-agent communication is necessary. The agents just receive capacity signals from a control unit and respond to it in a stochastic manner. This response can be tuned so that the long-term average optimality criterion (cf. ) can be achieved. The following is an excerpt from [@Corless2016].
In AIMD algorithm each agent follows two rules of action at each time step: either it increases its share of the resource by adding a fixed amount while total demand is less than the available capacity, or it reduces its share in a multiplicative manner when notified that global capacity has been reached. In the additive increase (AI) phase of the algorithm agents probe the available capacity by continually increasing their share of the resource. The multiplicative decrease (MD) phase occurs when agents are notified that the capacity limit has been reached; they respond by reducing their shares, thereby freeing up the resource for further distribution. This pattern is repeated by every agent as long as the agent is competing for the resource. The only information given to the agents about availability of the resource is a notification when the collective utilization of the resource achieves some capacity constraint. At such times, so called [*capacity events*]{}, some or all agents are instantaneously informed that capacity has been reached. The mathematical description of the basic continuous-time AIMD model is as follows. Let $n$ agents compete for a resource, and suppose that $x_i(t) \in \mathbb{R}_+$ denotes the quantity of the collective resource obtained by agent $i$ at time $t\in \mathbb{R}_+$. Let $C$ denotes the total capacity of the resource available to the entire system (which need not be known by the agents). The capacity constraint requires that $\sum_{i=1}^{n} x_i(t) \le C$ for all $t$. As all agents are continuously increasing their share this capacity constraints will be reached eventually. We denote the times at which this happens by $t_k, k\in \mathbb{N}$. At time $t_k$ the global utilization of the resource reaches capacity, thus $\sum_{i=1}^n x_i(t_k) = C.$ When capacity is achieved, some agents decrease their share of the resource. The instantaneous decrease of the share for agent $i$ is defined by: $$\begin{aligned}
\label{MD} x_i(t_k^{+}) := \lim_{t \rightarrow t_k,
\, t > t_k} x_i(t) = \beta_i x_i(t_k),\end{aligned}$$ where $\beta_i$ is a constant satisfying $ 0\le \beta_i <1.$ In the simplest version of the algorithm, agents are assumed to increase their shares at a constant rate in the AI phase: $$\begin{aligned}
\label{AI} x_i(t ) = \beta_i x_i(t_k) + \alpha_i (t
- t_k), \quad t_k < t \le t_{k+1},\end{aligned}$$ where, $\alpha_i>0$, is a positive constant, which may be different for different agents, $\alpha_i$ is known as the [*growth rate*]{} for agent $i$. By writing $x_i(k)$ for the $i$th agent’s share at the $k$th capacity event as $x_i(k) := x_i(t_k)$ we have: $$\begin{aligned}
x_i(k+1) = \beta_i x_i(k) + \alpha_i T(k),\end{aligned}$$ where $
T(k) := t_{k+1}-t_k,$ is the time between events $k$ and $k+1$. There are situations where not all agents may respond to every capacity event. Indeed, this is precisely the case considered in this paper. In this case agents respond asynchronously to a congestion notification and the AIMD model is easily extended by using our previous formalism by changing the multiplicative factor to $\beta_i = 1$ at the capacity event if agent $i$ does not decrease.
Multi-resource allocation {#divisible_mul_res}
=========================
Let $\delta_j>0$ be a fixed constant, for all $j$ and $\nabla^2 f$ be the matrix of second order partial derivatives of $f$ called Hessian of $f$. Furthermore, let $\mathcal{F}_\delta$ denotes the set of twice continuously differentiable functions defined as follows: $$\label{def_F_delta}
\begin{split}
\mathcal{F}_\delta = \Big\{ f:\mathbb R^m_+ \to \mathbb R
\Big \lvert \Big( x^j > 0 \implies 0 < \delta_j \nabla_j f(x) < x^j \\
\text{ for all $j$}\Big) \text{ and } \nabla^2 f(x) \succeq 0 \text{ for all }x\in \mathbb{R}^m_+ \Big\}.
\end{split}$$ Here, $\nabla^2 f(x) \succeq 0$ represents a positive semi-definite matrix. We observe that $\mathcal{F}_\delta$ is essentially the set of functions that are convex, twice continuously differentiable and increasing in each coordinate. We consider the problem of allocating $m$ resources with capacity $C^j$, for $j\in \mathcal{M}$ among $n$ competing Internet-connected devices, whose cost functions $f_1,\ldots,f_n$ belong to the set $\mathcal F_\delta$. Additionally, each cost function is private and should be kept private. However, we assume that the set $\mathcal F_\delta$ is common knowledge — the control unit needs the knowledge of $\delta_j$ and the Internet-connected devices need to have cost functions from this set. We should make clear that $\mathcal F_\delta$ has a large range of allowed cost functions. By knowing this range, the control unit can not easily guess the actual cost function, thereby giving the Internet-connected device a nontrivial amount of privacy. In this paper, we propose a distributed algorithm that determines instantaneous allocations $\{x_i^j(k)\}$, for all $i, j$ and $k$. Recall that $x^* = (x_1^{*1}, \ldots, x_n^{*m})$ is the solution of . We also show empirically that for every Internet-connected device $i$ and resource $R^j$, the long-term average allocations converge to the optimal allocations i.e., $\overline{x}_i^j(k) \to {x}_i^{*j}$ as $k\to \infty$ (cf. ) to achieve the minimum overall cost to the society called [*social cost*]{}.
Algorithm
---------
In the system, each Internet-connected device runs a distinct distributed AIMD algorithm. We use $\alpha^j>0$ to represent the additive increase factor or growth rate and $0 \leq \beta^j \leq 1$ to represent multiplicative decrease factor, both corresponding to resource $R^j$, for $j \in \mathcal{M}$. We represent $\Gamma^j$ as the [*normalization factor*]{}, chosen based on the knowledge of fixed constant $\delta_j$ to scale probabilities $\lambda_i^j(k)$. Every algorithm is initialized with the same set of parameters $\Gamma^j$, $\alpha^j$, $\beta^j$ received from the control unit of the system. We represent the one-bit *capacity event signals* by $S^j(k) \in \{0,
1\}$ at time step $k$ for resource $R^j$, for all $j$ and $k$. At the start of the system the control unit initializes the capacity event signals $S^j(0)$ with $0$, and updates $S^j(k)=1$ when the total allocation $\sum_{i=1}^n x_i^j(k)$ exceeds the capacity $C^j$ of a resource $R^j$ at a time step $k$. After each update, control unit broadcasts it to Internet-connected devices in the system signaling that the total demand has exceeded the capacity of the resource $R^j$. We describe the algorithm of control unit in Algorithm \[algoCU1\].
Input: $C^{j}$, for $j \in \mathcal{M}$.
Output: $S^{j}(k+1)$, for $j \in \mathcal{M}$, $k \in
\mathbb{N}$.
Initialization: $S^{j}(0) \leftarrow 0$, for $j \in \mathcal{M}$,
broadcast $\Gamma^{j} \in (0,\delta_j]$ according to ;
The algorithm of each Internet-connected device works as follows. At every time step, each algorithm updates its demand for resource $R^j$ in one of the following ways: an [*additive increase (AI)*]{} or a [*multiplicative decrease (MD) phase*]{}. In the additive increase phase, the algorithm increases its demand for resource $R^j$ linearly by the constant $\alpha^j$ until it receives a capacity event signal $S^j(k) =1$ from the control unit of the system at time step $k$ that is: $$\begin{aligned}
x_i^j (k+1) = x_i^j (k) + \alpha^j.\end{aligned}$$ The multiplicative decrease phase occurs when total demand exceeds the capacity of a resource (say $R^j$), and the control unit in response broadcasts a capacity event signal $S^j(k) =1$. In turn, each Internet-connected device $i$ responds with probability $\lambda^j_{i}(k)$ by scaling its demand by $\beta^j$. If $S^j(k) =1$, we thus have: $$\begin{aligned}
x_i^j(k+1)= \left\{
\begin{array}{ll}
\beta^j x_i^j(k) & \mbox{with probability } \lambda^j_{i}(k) , \\
x_i^j(k) & \mbox{with probability } 1-\lambda^j_{i}(k).\\
\end{array}
\right.\end{aligned}$$ The probability $\lambda^j_{i}(k)$ depends on the average allocation and the derivative of cost function with respect to $R^j$ of Internet-connected device $i$, for all $i$ and $j$. It is calculated as follows: $$\begin{aligned}
\label{prob_x} \lambda^j_{i}(k) = \Gamma^j
\frac{{\nabla_j} f_i \big( \overline{x}^1_i(k), \overline{x}^2_i(k), \ldots,
\overline{x}^m_i(k) \big)}{\overline{x}^j_i(k)},\end{aligned}$$ for all $i $, $j$ and $k$. After the reduction of demands, all Internet-connected devices can again start to increase their demands until the next capacity event occurs. This process repeats. It is obviously required that always $ 0 < \lambda^j_{i}(k) < 1$. To this end the normalization factor $\Gamma^j$ is needed which is based on the set $\mathcal F_\delta$. The fixed constant $\delta_j >0$ is chosen such that $\Gamma^j$ satisfies the following: $$\begin{aligned}
\label{gamma_delta}
0 <\Gamma^j \leq \delta_j, \text{ for all } j. \end{aligned}$$ At the beginning of the algorithm the normalization factor $\Gamma^j$ for resource $R^j$ is calculated explicitly as the following and broadcast to all Internet-connected devices in the system: $$\begin{aligned}
\label{gamma}
\Gamma^j =
\inf_{x_1^1,\ldots,x_n^m \in \mathbb{R}_+, f \in
\mathcal{F}_\delta}
\Big(\frac{x^j}{\nabla_j f(x^1,
x^2, \ldots, x^m)} \Big), \text{ for all } j. \end{aligned}$$ To capture the stochastic nature of the response to the capacity signal, we define the following independent Bernoulli random variables: $$\begin{aligned}
\label{bern_var}
b^j_i(k)=
\left\{
\begin{array}{ll}
1 & \mbox{with probability } \ \lambda^j_{i}(k),\\
0 & \mbox{with probability } 1-\lambda^j_{i}(k),
\end{array}
\right. \end{aligned}$$ for all $i$, $j$ and $k$. The following theorem proves that $ 0 < \lambda^j_{i}(k) < 1$.
\[theorem1\] For a given $\delta_j >0$, if $\overline{x}_i^j(k) >0$ and the cost function $f_i$ of Internet-connected device $i$ belongs to $\mathcal{F}_\delta$, then for all $i, j$ and $k$, $\lambda_i^j(k)$ satisfies $0 < \lambda_i^j(k) < 1$.
It is given that $f_i \in \mathcal{F}_\delta$ and $\overline{x}_i^j(k) > 0$ for all $i$, $j$ and $k$ then from , we write as follows: $$\begin{aligned}
\label{eq117}
0 < \delta_j \nabla_j f_i \big( \overline{x}_i^1(k),\overline{x}_i^2(k), \ldots, \overline{x}_i^m(k) \big) < \overline{x}_i^j(k).
\end{aligned}$$ We know that for a fixed constant $\delta_j>0$, the normalization factor $\Gamma^j$ satisfies $0 < \Gamma^j \leq \delta_j$, for all $j$ (cf. ). It is given that $\overline{x}_i^j(k) >0$, dividing by $\overline{x}_i^j(k)$ and substituting $\Gamma^j$ we obtain as follows: $$\begin{aligned}
\label{eq114}
&0 < \frac{\Gamma^j \nabla_j f_i \big( \overline{x}_i^1(k),\overline{x}_i^2(k), \ldots, \overline{x}_i^m(k) \big)}{\overline{x}_i^j(k)} < 1, \\&\text{ for all $i,j$ and $k$} \nonumber .
\end{aligned}$$ Since, for all $i, j$ and $k$, an Internet-connected device $i$ makes a decision to respond the capacity event of a resource $R^j$ with $\lambda_i^j(k)$ (cf. ). Hence, after placing $\lambda_i^j(k)$ in , we obtain $0 < \lambda_i^j(k) < 1$, for all $i,j$ and $k$.
Notice that because of the stochastic nature of the algorithm, an Internet-connected device may reduce its resource demand and fails to complete its current job, but only in cases where other Internet-connected devices derive more benefit than this Internet-connected device. This is done in order to maximize the overall benefit to the society called [*social welfare*]{}.
We present the block diagram of the system in Figure \[Diag\_AIMD\] and the proposed distributed multi-resource allocation algorithm for each Internet-connected device in Algorithm \[algo1\].
Input: $S^{j}(k)$, for $j \in \mathcal{M}, k \in
\mathbb{N}$ and $\Gamma^j$, $\alpha^j, \beta^j$, for $j \in \mathcal{M}$.
Output: $x^j_i(k+1)$, for $j \in \mathcal{M}$, $k \in
\mathbb{N}$.
Initialization: $x^j_i(0) \leftarrow 0$ and $\overline{x}^j_i(0) \leftarrow x^j_i(0)$, for $j \in \mathcal{M}$;
We observe using numerical results in Section \[results\] that the average allocation $\overline{x}_i^j(k)$ converge to the optimal allocation $x_i^{*j}$ of resource $R^j$ over time, for all $i$ and $j$.
Suppose that there are $m$ resources in the system, then communication overhead will be $\sum_{j=1}^{m} S^{j}(k)$ bits at $k^{th}$ time step, for all $k$. In the worst case scenario this will be $m$ bits per time unit, which is quite low. Furthermore, the communication complexity does not depend on the number of Internet-connected devices in the system.
Numerical results {#results}
=================
In this section, we use the multi-camera coordination system described in Section \[prob\_form\]. We illustrate here that the proposed distributed multi-resource allocation algorithm provides optimal allocations to all cameras in long-term average allocations and the city achieves a minimum social cost, these optimal values are same as if the problem is solved in a centralized way.
Now, suppose that there are $60$ cameras in the multi-camera coordination system, each camera has different resolution, frame size and frame generation rate (frames per second), therefore every camera generates different amount of data. For example, a camera with frame size of $30$ KB and frame rate $10$ frames per second, produces $300$ KB video data in one second, hence $1.08$ GB in an hour.
Let us assume that the videos from all the cameras are stored on a server or Cloud. To transmit, process and store the videos on the server or Cloud they require network bandwidth, CPU cycles, memory (RAM), and disk storage. We assume that each camera gets enough CPU cycles to process the data but the server has limited memory (RAM), disk storage and network bandwidth. Let, $R^1$ denotes the memory (RAM), $R^2$ denotes the disk storage and $R^3$ denotes the network bandwidth. We chose capacities of memory, disk storage and network bandwidth as $C^1 = 32$ GB, $C^2 = 200$ GB and $C^3 = 250$ Mbps, respectively. Let $10$ GB is denoted by $\mathrm{GB}^D$ and $10$ Mbps is denoted by $\mathrm{Mbps}^D$, then we write $C^2 = 20$ $\mathrm{GB}^D$ and $C^3 = 25$ $\mathrm{Mbps}^D$, we do so for the sake of uniformity of cost of resources in the cost function. Let $f_i$ be the cost function of camera $i$, each cost function depends on the average allocation of the resources. Our aim is to minimize the total cost incurred in transmitting, storing and processing the video data. For illustrative purpose we use the pricing model of Google compute engine for custom machine types [@Google2018] as shown in Table \[price\_tab\]. We create a dynamic pricing scheme for our simulation, keeping the values of Table \[price\_tab\] into consideration. Notice that in Table \[price\_tab\], for the disk storage we use the price of image storage for $10$ days and the listed prices are for Iowa state. Furthermore, we use the price of bandwidth for North America as listed in [@Prince2018].
------------------------------------------- ---------------------- -- --
Resource type Price per unit (USD)
\[0.5ex\] vCPU 0.132696
RAM (GB) 0.017784
Disk storage ($10$ GB) 0.283333
Network bandwidth (10 Mbps) [@Prince2018] 0.277775
\[1ex\]
------------------------------------------- ---------------------- -- --
: Pricing scheme of Google compute engine custom machines for $4$ hours
\[price\_tab\]
Now, let $a_i, b_i,$ and $c_i$ represent the price for RAM, disk storage and bandwidth, and $d_i$ represents any other costs incurred. For all $i$, let $a_i, b_i,$ $c_i$ and $d_i$ are modeled as uniformly distributed random variables. In the simulation, we use $a_i \in \{10, 11, \ldots, 20\}$, $b_i \in \{25, 26, \ldots, 35\}$, $c_i \in \{22, 23, \ldots, 32\}$ and $d_i \in \{1, 2, \ldots, 5\}$. We use these random variables to generate random costs of each camera at different time steps, as described in . To take vCPU price into consideration, we add a fraction of its price in the price of memory. In the simulation, we chose the following additive increase factors $\alpha^1=25$ MB, $\alpha^2 = 20$ MB and $\alpha^3 = 225$ Kbps. Additionally, we chose the following multiplicative decrease factors $\beta^1=0.70$, $\beta^2 = 0.85$ and $\beta^3 = 0.75$, for the respective resources. Furthermore, we use the normalization factors $\Gamma^1 = \Gamma^2 = \Gamma^3 = 1/90$. Notice that allocation $x_i^1$ is in GB, $x_i^2$ is in $\mathrm{GB}^D$ and $x_i^3$ is in $\mathrm{Mbps}^D$.
$$\begin{aligned}
\label{cost_fn}
f_{i}(x_i^1, x_i^2, x_i^3) =
\begin{cases}
a_i(x_i^1)^2 + c_i(x_i^3)^2 + \frac{1}{2}a_i(x_i^1)^4 + 2b_i(x_i^2)^4 + \frac{1}{2}b_i(x_i^2)^6 + \frac{1}{4} c_i(x_i^3)^4 + \frac{1}{8}d_i(x_i^3)^8 & \text{ w.p. } 1/3 \\
a_i(x_i^1)^2 + b_i(x_i^2)^2 + \frac{1}{2}b_i(x_i^2)^4 + \frac{3}{2} c_i(x_i^3)^4 & \text{ w.p. } 1/3 \\
b_i(x_i^2)^2 + c_i(x_i^3)^2 + \frac{1}{3} a_i(x_i^1)^6 + \frac{1}{6}d_i(x_i^2)^6 + \frac{1}{8}d_i(x_i^3)^4 & \text{ w.p. } 1/3.
\end{cases}
\end{aligned}$$
Here, for the illustrative purpose we use only few cost functions but the proposed algorithm works on a set of cost functions with condition that these are convex, twice differentiable and increasing functions.
\[avg\_aimd\] \[inst\_alloc\_aimd\] \[err\_grad\_aimd\]
\[err\_grad\_BAIMD\_2var\_x1\] \[err\_grad\_BAIMD\_2var\_x2\] \[avg\_BAIMD\_2var\]
The following are some of the results obtained from the simulation. We observe in Figure \[fig1\](a) that the average allocations $\overline{x}_i^j(k)$ converge over time to its respective optimal value $x^{*j}_i$, for all $i$ and $j$. Figure \[fig1\](b) shows the instantaneous allocation $x_i^j(k)$ of all resources over last $100$ time steps, which demonstrates the allocation phases (AI and MD).
\[1a\] \[1b\] \[1c\]
\[sum\_avg\] \[sum\_alloc\] \[instant\_alloc\]
We know that, to achieve optimality, the derivatives of the cost functions of all participating cameras for a particular resource should make a consensus, which satisfies all the Karush-Kuhn-Tucker (KKT) conditions that are necessary and sufficient conditions for optimality of , as described in Section \[prob\_form\]. Figure \[fig1\](c) is the error bar of derivatives $\nabla_jf_i$ of cost functions $f_i$ for single simulation calculated across all cameras, for all $j$. It illustrates that the derivatives of cost functions of all cameras with respect to a particular resource concentrate more and more over time around the same value. Hence, the long-term average allocation of resources for the stated optimization problem is optimal.
For comparison purpose, we solve the optimization problem in a centralized way using the interior-point method and denote the optimal values obtained by $x_i^{*j}$, for all $i$ and $j$. We compare these optimal values with average allocation values at largest time steps in the simulation (long-term average) obtained by our proposed algorithm, we find that the results are approximately equal. Let $K$ be the largest time step used in the simulation, Figure \[fig2\](a) shows the evolution of [*absolute error*]{} which is the absolute difference of average allocation $\overline{x}_i^j(k)$ at time step $k$ and the calculated optimal allocation $x_i^{*j}$, i.e., $|\overline{x}_i^{j}(k) - x_i^{*j}|$. We observe that the absolute error approaches close to zero over time. Additionally, we calculate the [*relative error*]{} which we define as the ratio of sum of absolute errors and the sum of calculated optimal allocations i.e., $\frac{\sum_{i=1}^{n} |x_i^j(k) - x_i^{*j}|}{\sum_{i=1}^{n} x_i^{*j}}$. The evolution of relative error is presented in Figure \[fig2\](b), which decreases with time and is very low, for the described simulation it is below $5\%$. Figure \[fig2\](c) illustrates that the ratio of the sum of cost functions with average allocations and the sum of cost functions with optimal allocations i.e., the ratio of $\sum_{i=1}^{n} f_i(\overline{x}_i^1(K), \overline{x}_i^2(K), \overline{x}_i^3(K))$ and $\sum_{i=1}^{n} f_i(x_i^{*1}, x_i^{*2}, x_i^{*3})$ is close to $1$, which further strengthens our claim. Furthermore, to gather information about absolute errors $|\overline{x}_i^{j}(K) - x_i^{*j}|$ of all cameras at time step $K$, we present their histograms in Figure \[fig3\], we observe that the absolute error of most of the cameras are close to zero. Figure \[fig4\](a) illustrates the sum of average allocations $\sum_{i=1}^{n} \overline{x}_i^j(k)$ over time. We observe that the sum of average allocations at largest time step $K$ is approximately equal to the respective capacity i.e., $\sum_{i=1}^{n} \overline{x}_i^j(K) \approxeq C^j$, for all $j$ (capacities are $C^1 = 32$ GB, $C^2 = 20$ $\mathrm{GB}^D$ and $C^3 = 25$ $\mathrm{Mbps}^D$). Figure \[fig4\](b) shows the sum of instantaneous allocations $\sum_{i=1}^{n} x_i^j(k)$ of resource $R^j$ for last $40$ time steps. We observe that the sum of instantaneous allocations are concentrated around the respective capacities. To reduce the overshoots of total allocations of resource $R^j$, we assume $\gamma^j < 1$ and modify the algorithm of control unit to broadcast the capacity event signal $S^j(k)=1$ when $\sum_{i=1}^{n} x_i^j(k) > \gamma^j C^j$, for all $j$ and $k$. Furthermore, the number of capacity events is the communication overhead of the system to reach the consensus of derivatives of all cameras with respect to a particular resource, which is illustrated in the Figure \[fig4\](c) for several simulations. For example, the number of capacity events broadcast by the control unit in a simulation running for $30000$ time steps are $11427, 11988$ and $8355$, for resources $R^1,R^2$ and $R^3$, respectively, which are the communication overhead of the system in bits for the respective resource. Notice that the communication overhead is very low for each resource. It is also observed that the number of capacity events increases approximately linearly with time steps for different simulations.
Conclusion {#conc}
==========
In this paper a distributed algorithm is proposed. The algorithm solves the multi-variate optimization problems for capacity constraint problems in a distributed manner. It is done by extending a variant of AIMD algorithm. The features of the proposed algorithm are; it involves little communication overhead, there is no inter-device communication needed and each Internet-connected device has its own private cost functions. It is shown in the paper that the long-term average allocation of resources converge to approximately same values as if the optimization problem under consideration is solved in a centralized setting.
It is interesting to solve the following open problems: first is to provide a theoretical basis for the proof of convergence and second is to find the bounds for the rate of convergence, and its relationship with different parameters or the number of occurrence of capacity events. The work can also be extended in several application areas like Cloud computing, smart grids or wireless sensor networks, where sensors have very limited processing power and battery life.
Acknowledgment
==============
The work is supported partly by Natural Sciences and Engineering Research Council of Canada grant RGPIN-2018-05096 and by Science Foundation Ireland grant 16/IA/4610.
[^1]: $^\ast$Concordia Institute for Information Systems Engineering, Concordia University, Montreal, Quebec, Canada, email: sy\[email protected], [email protected]
[^2]: $^\dagger$School of Electrical, Electronic and Communications Engineering, University College Dublin, Dublin, Ireland, email: [email protected]
[^3]: $^\ddagger$Faculty of Computer Science and Mathematics, University of Passau, Passau, Germany, email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the enumerative geometry of algebraic curves on abelian surfaces and threefolds. In the abelian surface case, the theory is parallel to the well-developed study of the reduced Gromov-Witten theory of $K3$ surfaces. We prove complete results in all genera for primitive classes. The generating series are quasimodular forms of pure weight. Conjectures for imprimitive classes are presented. In genus 2, the counts in all classes are proven. Special counts match the Euler characteristic calculations of the moduli spaces of stable pairs on abelian surfaces by Göttsche-Shende. A formula for hyperelliptic curve counting in terms of Jacobi forms is proven (modulo a transversality statement).
For abelian threefolds, complete conjectures in terms of Jacobi forms for the generating series of curve counts in primitive classes are presented. The base cases make connections to classical lattice counts of Debarre, Göttsche, and Lange-Sernesi. Further evidence is provided by Donaldson-Thomas partition function computations for abelian threefolds. A multiple cover structure is presented. The abelian threefold conjectures open a new direction in the subject.
address:
- 'University of British Columbia, Department of Mathematics'
- 'ETH Zürich, Department of Mathematics'
- 'ETH Zürich, Department of Mathematics'
- 'ETH Zürich, Department of Mathematics'
author:
- Jim Bryan
- Georg Oberdieck
- Rahul Pandharipande
- Qizheng Yin
date: December 2016
title: Curve counting on abelian surfaces and threefolds
---
Introduction
============
Vanishings
----------
Let $A$ be a complex abelian variety of dimension $d$. The Gromov-Witten invariants of $A$ in genus $g$ and class $\beta\in H_2(A,{{\mathbb{Z}}})$ are defined by integration against the virtual class of the moduli space of stable maps ${{\overline M}}_{g,n}(A,\beta)$, $$\Big \langle \tau_{a_1}(\gamma_1) \cdots \tau_{a_n}(\gamma_n) \Big \rangle_{g,\beta}^{A}
= \int_{ [{{\overline M}}_{g,n}(A, \beta)]^{\text{vir}} }
\psi_1^{a_1} {\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1) \cdots \psi_n^{a_n} {\mathop{\rm ev}\nolimits}_n^{\ast}(\gamma_n) \label{gw1}\, ,$$ see [@PT3] for an introduction.[[^1]]{} However, for abelian varieties of dimension $d \geq 2$, the Gromov-Witten invariants often vanish for two independent reasons. Fortunately, both can be controlled. The result is a meaningful and non-trivial enumerative geometry of curves in $A$.
The first source of vanishing is the obstruction theory of stable maps. For dimensions $d \geq 2$, the cohomology $$H^{2,0}(A,{{\mathbb{C}}}) = H^{0}(A, \Omega_{A}^2)$$ does not vanish and yields a trivial quotient of the obstruction sheaf. As a consequence, the virtual class vanishes [@KL] for non-zero classes $\beta$.
An alternative view of the first vanishing can be obtained by deformation invariance. A homology class $\beta \in H_2(A,{{\mathbb{Z}}})$ is a [*curve class*]{} if $\beta$ is represented by an algebraic curve on $A$. The Gromov-Witten invariants vanish if $\beta$ is [*not*]{} a curve class since then the moduli space ${{\overline M}}_{g,n}(A,\beta)$ is empty. After generic deformation of $A$, every non-zero curve class $\beta$ acquires a part of $H^{2,0}(A,{{\mathbb{C}}})^{\vee}$ and is no longer the class of an algebraic curve. By deformation invariance, the Gromov-Witten invariants of $A$ then necessarily vanish for all non-zero $\beta$.
Second, an independent source of vanishing arises from the action of the abelian variety $A$ on the moduli space ${{\overline M}}_{g,n}(A,\beta)$ by translation: most stable maps $$f : C {{\ \rightarrow\ }}A$$ appear in $d$-dimensional families. Integrands which are translation invariant almost always lead to vanishing Gromov-Witten invariants. We must therefore impose a $d$-dimensional condition on the moduli space which picks out a single or finite number of curves in each translation class.
Curve classes on $A$ are equivalent to divisor classes on the dual abelian variety $\widehat{A}$. Every curve class $\beta \in H_2(A,{{\mathbb{Z}}})$ has a type[[^2]]{} $$(d_1, \ldots, d_{\text{dim}\, A})\, , \ \ \ \ d_i\geq 0$$ obtained from the standard divisor theory of $\widehat{A}$. A curve class $\beta$ is [*non-degenerate*]{} if $d_i>0$ for all $i$. Otherwise, $\beta$ is a [*degenerate*]{} curve class. The degenerate[[^3]]{} case is studied by reducing the dimension of $A$.
Various techniques have been developed in recent years to address the first vanishing. The result in the non-degenerate case is a [*reduced virtual class*]{} $[ {{\overline M}}_{g,n}(A, \beta) ]^{\text{red}}$ with dimension increased by $h^{2,0}(A)$. Integrals against the reduced class are invariant under deformations of $A$ for which $\beta$ stays algebraic. Up to translation, we expect the family of genus $g$ curves in class $\beta$ to be of dimension $$\begin{aligned}
& {\rm vdim}\, {{\overline M}}_{g}(A, \beta) + h^{2,0}(A) - d \\ ={} & (d-3)(1-g) + \frac{d (d-1)}{2} - d \\ ={} & (d - 3)\left(\frac{d}{2} + 1 - g\right) \,. \label{vdim}
\end{aligned}$$ Hence, for abelian varieties of dimension $1$, $2$, and $3$, we expect families (modulo translation) of dimensions $2g-3$, $g-2$, and $0$ respectively.
For abelian varieties of dimension $d \geq 4$, the reduced virtual dimension is non-negative only if $$g \, \leq \, \frac{d}{2} + 1 \, \leq \, d-1 \,.$$ In the non-degenerate case, a generic abelian variety $A$ admits no proper abelian subvariety and thus admits [*no*]{} map from a curve of genus less than $d$. Hence, all invariants vanish.
The Gromov-Witten theory of elliptic curves has been completely solved by Okounkov and Pandharipande in [@OP1; @OP3]. Some special results are known about abelian surfaces [@BLA; @D; @G; @Ros14]. We put forth here several results and conjectures concerning the complete Gromov-Witten theory of abelian surfaces and threefolds.
Abelian surfaces
----------------
### Basic curve counting {#bcc}
Let $A$ be an abelian surface and let $$\beta \in H_2(A, {{\mathbb{Z}}})$$ be a curve class of type $(d_1, d_2)$ with $d_1, d_2 > 0$. The moduli space $${{\overline M}}_{g,n}(A,\beta)^{\text{FLS}} \subset {{\overline M}}_{g,n}(A,\beta)$$ is the closed substack of ${{\overline M}}_{g,n}(A,\beta)$ parameterizing maps with image in a fixed linear system (FLS) on $A$. Given a curve $C$ in class $\beta$, the FLS condition naturally picks out $(d_1 d_2)^2$ elements in the translation class of $C$. The FLS moduli space carries a reduced virtual fundamental class [@KT; @MP; @STV], $$\left[ {{\overline M}}_{g}(A, \beta)^{\text{FLS}} \right]^{\text{red}}\, ,$$ of virtual dimension $g - 2$. Define $\lambda_k$ to be the Chern class $$\lambda_k = c_k(\mathbb{E})$$ of the Hodge bundle $\mathbb{E} \rightarrow {{\overline M}}_{g,n}(A,\beta)$ with fiber $H^0(C,\omega_C)$ over the moduli point $$[f : C {{\ \rightarrow\ }}A] \in {{\overline M}}_{g,n}(A,\beta)\, .$$
There are no genus 0 or 1 curves on a general abelian surface $A$. The most basic genus $g\geq 2$ Gromov-Witten invariants of $A$ are $$\mathsf{N}^{\text{FLS}}_{g,\beta}
= \int_{[ {{\overline M}}_{g}(A, \beta)^{\text{FLS}} ]^{\text{red}}} (-1)^{g-2} \lambda_{g-2} \,. \label{123}$$ The integrand $(-1)^{g-2} \lambda_{g-2}$ corresponds to the natural deformation theory of curves in $A$ when considered inside a Calabi-Yau threefold. The invariants are therefore precisely the analogs of the genus $g-2$ Gromov-Witten invariants of $K3$ surface which appear in the Katz-Klemm-Vafa formula, see [@MP; @MPT; @PT2].
By deformation invariance, $\mathsf{N}_{g,\beta}^{\text{FLS}}$ depends only on the type $(d_1, d_2)$ of $\beta$. We write $$\mathsf{N}^{\text{FLS}}_{g,\beta} = \mathsf{N}^{\text{FLS}}_{g,(d_1, d_2)} \,.$$ We have the following fully explicit conjecture for these counts.
\[conjA\] For all $g \geq 2$ and $d_1, d_2 > 0$, $$\mathsf{N}^{\textup{FLS}}_{g,(d_1, d_2)}
= (d_1 d_2)^2\, \frac{2 (-1)^{g-2}}{(2g-2)!} \sum_{k | \gcd(d_1,d_2)} \, \sum_{ m | \frac{ d_1 d_2 }{k^2} } k^{2g-1} m^{2g-3} \,.$$
The right hand side incorporates a multiple cover rule which expresses the invariants in imprimitive classes in terms of primitive invariants.[^4] The multiple cover structure is discussed in Section \[mcr\].
\[YZ\_intro\] Conjecture \[conjA\] is true in the following cases:
1. for all $g$ in case $\beta$ is primitive,
2. for all $\beta$ in case $g=2$.
For part (i) concerning primitive $\beta$, our method relies on a degeneration formula for Gromov-Witten invariants of abelian surfaces and calculations in [@MPT]. Via a version of the Gromov-Witten/Pairs correspondence [@PT1], the primitive case also yields an independent proof of the Euler characteristic calculations of relative Hilbert schemes of points by Göttsche and Shende [@GS].
For part (ii) concerning genus $2$, the proof is reduced by a method of Debarre [@D], Göttsche [@G], and Lange-Sernesi [@LS] to a lattice count in abelian groups. Our results reveal a new and surprising multiple cover structure in these counts.
Conjecture \[conjA\] is parallel to the full Katz-Klemm-Vafa conjecture for $K3$ surfaces. Part (i) of Theorem \[YZ\_intro\] is parallel to the primitive KKV conjecture proven in [@MPT]. Part (ii) is parallel to the full Yau-Zaslow conjecture for rational curves on $K3$ surfaces proven in [@KMPS]. While our proof of part (i) involves methods parallel to those appearing in the proof of the primitive KKV conjecture, our proof of part (ii) is completely unrelated to the (much more complicated) geometry used in the proof of the full Yau-Zaslow conjecture.
The full Katz-Klemm-Vafa conjecture is proven in [@PT2]. However, most cases of Conjecture \[conjA\] remain open.
### Point insertions for primitive classes {#ptptpt}
Let ${{\mathsf{p}}}\in H^4(A,{{\mathbb{Z}}})$ be the class of a point. Define the $\lambda$-twisted Gromov-Witten invariants with $k$ point insertions by: $$\mathsf{N}_{g,k,(d_1,d_2)}^{\text{FLS}} =
\int_{[ {{\overline M}}_{g,k}(A, \beta)^{\text{FLS}}]^{\text{red}} } (-1)^{g-2-k} \lambda_{g-2-k} \prod_{i=1}^{k} {\mathop{\rm ev}\nolimits}_i^{\ast}({{\mathsf{p}}})\, ,$$ where $\beta$ is a curve class of type $(d_1,d_2)$. Define the function $${\mathsf{S}}(z,\tau) = - \sum_{d \geq 1} \sum_{m | d} \frac{d}{m} \big( p^m - 2 + p^{-m} \big) q^d\,,$$ considered as a formal power series in the variables $$p = e^{2\pi i z} \quad \text{ and } \quad q= e^{2 \pi i \tau}\, .$$
\[thm\_point\_insertion\] After setting $u=2\pi z$, we have $$\sum_{g \geq 2} \sum_{d \geq 1} \mathsf{N}^{\textup{FLS}}_{g, k, (1, d)} u^{2g-2} q^d
= q \frac{d}{dq}\, \left( \frac{{\mathsf{S}}(z,\tau)^{k+1}}{k+1} \right)\, .$$
For $k=0$, by definition $$\mathsf{N}^{\textup{FLS}}_{g, 0, (1, d)}=
\mathsf{N}^{\textup{FLS}}_{g, (1, d)}\ .$$ Hence, by Theorem \[thm\_point\_insertion\], $$\sum_{g \geq 2} \sum_{d \geq 1} \mathsf{N}^{\textup{FLS}}_{g, (1, d)} u^{2g-2} q^d
= q \frac{d}{dq}\, {\mathsf{S}}(z,\tau) \, ,$$ which is a restatement of the formula of Conjecture \[conjA\] for the classes $(1,d)$. Theorem \[thm\_point\_insertion\] specializes in the $k=0$ case to the primitive part of Theorem \[YZ\_intro\].
### Quasi-modular forms {#qmfo}
Let $\gamma_1, \dots, \gamma_n \in H^{\ast}(A, {{\mathbb{Q}}})$ be cohomology classes. The primitive descendent potential of $A$ with insertions $\tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n)$ is defined[^5] by $$\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )
= \sum_{d \geq 0} \Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,(1,d)}^{A, \text{red}} q^{d} \,,$$ where the coefficients on the right hand side denote the reduced invariants of $A$, $$\Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,\beta}^{A, \text{red}} =
\int_{ [ {{\overline M}}_{g,n}(A,\beta) ]^{\text{red}} } \prod_{i=1}^n {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \psi_i^{a_i} \,.$$
The ring $\text{QMod}$ of holomorphic quasi-modular forms (of level 1) is the free polynomial algebra in the Eisenstein[[^6]]{} series $E_2(\tau)$, $E_4(\tau)$ and $E_6(\tau)$, $$\text{QMod} = {{\mathbb{Q}}}[ E_2, E_4, E_6 ] \,.$$ The ring $\text{QMod}$ carries a grading by weight, $$\text{QMod} = \bigoplus_{k \geq 0} {\text{QMod}}_{2k} \,,$$ where $E_{2k}$ has weight $2k$. Let $$\text{QMod}_{\leq 2k}\subset \text{QMod}$$ be the linear subspace of quasi-modular forms of weight $\leq 2k$.
The series vanish in $g=0$. For $g \geq 1$ and arbitrary insertions, we have the following result.
\[modularity\] The series $\mathsf{F}_{g}^{A}\left( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \right)$ is the Fourier expansion in $q = e^{2\pi i \tau}$ of a quasi-modular form of weight $\leq 2(g-2) + 2n$, $$\mathsf{F}_g^A\left(\tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n)\right) \in
\textup{QMod}_{\leq 2(g-2)+2n}\, .$$
A sharper formulation of Theorem \[modularity\] specifying the weight appears in Theorem \[modularity\_refined\] of Section \[section\_modularity\].
### Hyperelliptic curves {#hyp_intro}
A nonsingular curve $C$ of genus $g\geq 2$ is [*hyperelliptic*]{} if $C$ admits a degree 2 map to ${\mathbb{P}}^1$, $$C \rightarrow {\mathbb{P}}^1\, .$$ A stable curve $C$ is [*hyperelliptic*]{} if $[C]\in \overline{M}_g$ is in the closure of the locus of nonsingular hyperelliptic curves.[[^7]]{} An [*irreducible hyperelliptic curve*]{} of genus $g$ on an abelian surface $A$, $$C \subset A\, ,$$ is the image of a stable map $$f:\widehat{C} \rightarrow C\subset A$$ satisfying the following two conditions:
- $\widehat{C}$ is an irreducible stable hyperelliptic curve of genus $g$,
- $f:\widehat{C} \rightarrow C$ is birational.
By [@Pi], for any abelian surface $A$ and curve class $\beta$, the number of irreducible hyperelliptic curves of genus $g$ in a fixed linear system of class $\beta$ is [*finite*]{}.[^8] We write $\mathsf{h}^{A, \text{FLS}}_{g, \beta}$ for this finite count. Unlike all other invariants considered in the paper, $\mathsf{h}^{A, \text{FLS}}_{g, \beta}$ is [*defined*]{} by classical counting.
Since every genus $2$ curve is hyperelliptic, for generic $A$ and $\beta$ of type $(1, d)$ we have $$\mathsf{h}^{A, \text{FLS}}_{2, \beta} = d^2 \sum_{m|d}m\,$$ by the genus 2 part of Theorem \[YZ\_intro\]. The following result calculates the genus $3$ hyperelliptic counts in generic primitive classes.
\[cor\_hyp3\] For a generic abelian surface $A$ with a curve class $\beta$ of type $(1,d)$, $$\mathsf{h}^{A, \textup{FLS}}_{3, \beta} = d^2 \sum_{m|d} \frac{m ( 3m^2 + 1 - 4d) }{4}
\,.$$
Let $\mathcal{H}_g$ be the stack fundamental class of the closure of nonsingular hyperelliptic curves inside ${{\overline M}}_g$. By [@FPM], $\mathcal{H}_g$ is a tautological[[^9]]{} class of codimension $g-2$. While the restriction of $\mathcal{H}_g$ to $M_g$ is a known multiple of $\lambda_{g-2}$, a closed formula for $\mathcal{H}_g$ on $\overline{M}_g$ in terms of the standard generators of the tautological ring is not known.
For $\beta$ of type $(d_1,d_2)$, we [define]{} a [*virtual*]{} count of hyperelliptic curves in class $\beta$ by $$\label{xxgg} \mathsf{H}^{\text{FLS}}_{g,(d_1,d_2)} = \int_{[ {{\overline M}}_{g}(A, \beta)^{\text{FLS}} ]^{\text{red}} } \pi^{\ast}( \mathcal{H}_g )\, ,$$ where $\pi$ is the forgetful map $$\pi:\overline{M}_g(A,\beta) \rightarrow {{\overline M}}_g\, .$$ Because the integral is deformation invariant, the left side depends only upon $g$ and $(d_1,d_2)$.
For irreducible curve classes $\beta$ of type $(1, d)$ on an abelian surface $A$, consider the following property:
- [*Every irreducible curve in ${\mathbb{P}}^1 \times A$ of class $$(2, \beta) = 2[{\mathbb{P}}^1] + \beta \in H_2({\mathbb{P}}^1 \times A, {{\mathbb{Z}}})$$ is nonsingular.*]{}
We will prove property $(\dag)$ for curves of genus $2$ in case $A$ and $\beta$ are generic. Together with the explicit expression [@HM] for $$\mathcal{H}_3 \in H^2({{\overline M}}_3, {{\mathbb{Q}}})\, ,$$ we can deduce Proposition \[cor\_hyp3\]. The existence of classes $\beta$ of type $(1,d)$ satisfying ($\dag$) is not known for most $d$, but is expected generically for dimension reasons. Define the Jacobi theta function [@C] $$K(z,\tau) = \frac{i\, \vartheta_1(z,\tau)}{\eta(\tau)^3}
= i u \, \exp \bigg( {\sum_{k \geq 1}} \frac{(-1)^k B_{2k}}{2k (2k)!} E_{2k}(\tau) u^{2k} \bigg)\, ,
\label{KKKKKdef}$$ where $u = 2 \pi z$.
\[thm\_hyp\] Let $\beta$ be an irreducible class of type $(1, d)$ on an abelian surface $A$ satisfying $(\dag)$. Then we have:
1. $\displaystyle\sum_{g \geq 2} \mathsf{h}^{A, \textup{FLS}}_{g, \beta} \left( 2 \sin( u/2) \right)^{2g+2}
= \sum_{g \geq 2}\, \mathsf{H}^{\textup{FLS}}_{g,(1,d)} \, u^{2g+2} \,.$
2. After the change of variables $u = 2 \pi z$ and $q = e^{2 \pi i \tau}$, $$\sum_{g \geq 2} \, \mathsf{H}^{\textup{FLS}}_{g,(1,d)} \, u^{2g+2} =
\textup{Coeff}_{q^d} \left[ \bigg( q \frac{d}{dq} \bigg)^2
\frac{ K(z,\tau)^4 }{4}
\right]\,,$$ where $\textup{Coeff}_{q^d}$ denotes the coefficient of $q^d$.
Enumerative results on hyperelliptic curves via Gromov-Witten theory were first obtained for ${\mathbb{P}}^2$ by T. Graber [@Gra] using the Hilbert scheme of points $\text{Hilb}^2({\mathbb{P}}^2)$. The hyperelliptic curve counts on abelian surfaces were first studied by S. Rose [@Ros14] using the orbifold Gromov-Witten theory of $\text{Sym}^2(A)$ and the geometry of the Kummer. Rose derives his results from the crepant resolution conjecture (CRC) [@CRC; @Ruan] and certain geometric genericity assumptions. While our approach is similar, the closed formula (ii) for $\mathsf{H}^{\textup{FLS}}_{g,(1,d)}$ via the theta function $K(z,\tau)$ is new.
Proposition \[cor\_hyp3\] for $\mathsf{h}^{A, \textup{FLS}}_{3, \beta}$ and the formula of Theorem \[thm\_hyp\] for $\mathsf{h}^{A, \textup{FLS}}_{g, \beta}$ in higher genus (obtained by combining parts (i) and (ii)) do [*not*]{} match Rose’s results. The errors in Rose’s genus 3 counts can be repaired to agree with Proposition \[cor\_hyp3\]. We hope the CRC approach will be able to arrive exactly at the formula of Theorem \[thm\_hyp\] for $\mathsf{h}^{A, \textup{FLS}}_{g, \beta}$. The values of $\mathsf{h}^{A, \text{FLS}}_{g, \beta}$ in low genus and degree are presented in Table \[hyptable\] below. The distribution of the non-zero values in Table \[hyptable\] matches precisely the results of Knutsen, Lelli-Chiesa, and Mongardi in [@KLM15 Theorem 1.6]: $\mathsf{h}^{A, \text{FLS}}_{g, \beta}$ is non-zero if and only if $$\label{gmax}
(g - 1) + \bigg\lfloor \frac{g - 1}{4} \bigg\rfloor \bigg((g - 1) - 2 \bigg\lfloor \frac{g - 1}{4} \bigg\rfloor - 2\bigg) \leq d \,.$$ The entry for $g = 4$ and $d=3$ has recently been confirmed in [@BS].
$1$ $2$ $3$ $4$ $5$ $6$ $7$ $8$ $9$ $10$
----- ----- ------ ------ ------- -------- -------- --------- --------- ---------- ----------
$2$ $1$ $12$ $36$ $112$ $150$ $432$ $392$ $960$ $1053$ $1800$
$3$ $0$ $6$ $90$ $456$ $1650$ $4320$ $9996$ $20640$ $36774$ $67500$
$4$ $0$ $0$ $9$ $192$ $1425$ $6732$ $23814$ $68352$ $173907$ $387900$
$5$ $0$ $0$ $0$ $4$ $150$ $1656$ $10486$ $48240$ $174474$ $539200$
$6$ $0$ $0$ $0$ $0$ $0$ $36$ $735$ $6720$ $41310$ $191400$
$7$ $0$ $0$ $0$ $0$ $0$ $0$ $0$ $96$ $1620$ $14700$
$8$ $0$ $0$ $0$ $0$ $0$ $0$ $0$ $0$ $0$ $100$
: The first values for the counts $\mathsf{h}^{A, \text{FLS}}_{g, \beta}$ of hyperelliptic curves of genus $g$ and type $(1,d)$ in a FLS of a generic abelian surface $A$ as predicted by Theorem \[thm\_hyp\].[]{data-label="hyptable"}
Abelian threefolds
------------------
### Donaldson-Thomas theory {#abcal}
Let $X$ be an abelian threefold and let $\beta \in H_2(X,{{\mathbb{Z}}})$ be a curve class. The Hilbert scheme of curves $$\operatorname{Hilb}^{n}(X, \beta) = \{ Z\subset X \ | \ [Z]= \beta, \ \chi({{\mathcal O}}_{Z})=n \}$$ parameterizes 1 dimensional subschemes of class $\beta$ with holomorphic Euler characteristic $n$. The group $X$ acts on $\operatorname{Hilb}^{n}(X, \beta)$ by translation.
If $n \neq 0$, no assumption on $\beta$ is made. If $n = 0$, we assume that $\beta$ is not of type $(d, 0, 0)$ up to permutation. Then, the action of $X$ has finite stabilizers and the stack quotient $$\label{asgdb}
\operatorname{Hilb}^{n}(X, \beta) / X$$ is a Deligne-Mumford stack.
We consider here two numerical invariants of $\operatorname{Hilb}^n(X,\beta)/X$, the topological Euler characteristic $${\widehat{{\operatorname{\mathsf{DT}}}}}_{n, \beta}^X = e \big(\operatorname{Hilb}^{n} (X, \beta)/X\big),$$ and the *reduced Donaldson-Thomas invariant* of $X$ defined as the Behrend function weighted Euler characteristic $$\begin{aligned}
{\operatorname{\mathsf{DT}}}_{n, \beta}^X & = e \big(\operatorname{Hilb}^{n} (X, \beta)/X, \nu \big)\\
&=\sum _{k\in {{\mathbb{Z}}}}k\cdot e\big(\nu ^{-1} (k) \big)\, .\end{aligned}$$ While the Behrend function $$\nu :\operatorname{Hilb}^{n} (X, \beta)/X \to {{\mathbb{Z}}}$$ is integer valued, the topological Euler characteristic $e$ is taken in the orbifold sense and so may be a rational number. Hence, $${\operatorname{\mathsf{DT}}}_{n, \beta}^X \in \mathbb{Q}\,, \quad \quad {\widehat{{\operatorname{\mathsf{DT}}}}}_{n, \beta}^X \in \mathbb{Q} \,.$$
By results of M. Gulbrandsen [@Gul] ${\operatorname{\mathsf{DT}}}_{n,\beta}^X$ is invariant under deformations of the pair $(X,\beta)$ if $n \neq 0$. For $n = 0$ Gulbrandsen’s method breaks down, but deformation invariance is still expected. The numbers ${\widehat{{\operatorname{\mathsf{DT}}}}}^{X}_{n,\beta }$ are not expected to be deformation invariant.
By deformation equivalence, we may compute ${\operatorname{\mathsf{DT}}}^{X}_{n,\beta }$ after specialization to the product geometry $X=A\times E$. We compute ${\widehat{{\operatorname{\mathsf{DT}}}}}_{n, \beta}^{X}$ for the product geometry, and we conjecture a simple relationship there between ${\operatorname{\mathsf{DT}}}^{X}_{n,\beta }$ and ${\widehat{{\operatorname{\mathsf{DT}}}}}^{X}_{n,\beta}$. We then obtain a formula for ${\operatorname{\mathsf{DT}}}^{X}_{n,\beta}$.
Let $A$ be a generic abelian surface with a curve class $\beta_{{{d'}}}$ of type $(1, {{d'}}> 0)$, and let $E$ be a generic elliptic curve. Consider the abelian threefold $$X = A \times E \,.$$ The curve class $$(\beta_{{{d'}}}, d) = \beta_{{{d'}}} + d [E] \in H_2(X, {{\mathbb{Z}}})$$ is of type $(1, {{d'}}, d)$.
The following result determines the invariants ${\widehat{{\operatorname{\mathsf{DT}}}}}_{n,(\beta_{{{d'}}},d)}^X$ in the first two nontrivial cases ${{d'}}=1$ and ${{d'}}=2$.
Let $K$ be the theta function which already appeared in Section \[hyp\_intro\], $$K(p,q)
= (p^{1/2} - p^{-1/2}) \prod_{m \geq 1} \frac{ (1-pq^m) (1-p^{-1}q^m)}{ (1-q^m)^2 } \,.
$$
\[dthatthm\] \[dtthm\] \[eulthm\] For the topological Euler characteristic theory, we have
1. $\displaystyle\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}
{\widehat{{\operatorname{\mathsf{DT}}}}}_{n, (\beta_1, d)}^X\, p^n q^{d} = K(p,q)^2 \,,$
2. $\displaystyle\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}
{\widehat{{\operatorname{\mathsf{DT}}}}}_{n, (\beta_2, d)}^X\, p^n q^{d}
= K(p,q)^4 \cdot \bigg( \frac{1}{2} + \frac{3 p}{(1-p)^2} $ $$\phantom{K(p,q)^4 \cdot \Big\{ \frac{1}{2} + \frac{3 p}{(1-p)^2}} + \sum_{d \geq 1}\sum_{k | d} k \cdot \big( 3 (p^k + p^{-k}) q^d + 12 q^{2d} \big) \bigg) \,.$$
Assuming Conjecture \[conj\_Behrend\] in Section \[subsec: putting in the Behrend function\] below, we obtain the following result for the invariants ${\operatorname{\mathsf{DT}}}_{n,(\beta_{{{d'}}},d)}^X$ in the cases ${{d'}}=1$ and ${{d'}}=2$. Consider the Weierstrass elliptic function $$\wp(p,q) = \frac{1}{12} + \frac{p}{(1-p)^2} + \sum_{d \geq 1}\sum_{m|d} m (p^m - 2 + p^{-m}) q^d$$ expanded in the region $|p|<1$.
\[dtcor\] Assume Conjecture \[conj\_Behrend\] holds. Then we have
1. $\displaystyle\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}
{\operatorname{\mathsf{DT}}}_{n, (\beta_1, d)}^X\, (-p)^n q^{d} = - K(p,q)^2\, ,$
2. $\displaystyle
\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}} {\operatorname{\mathsf{DT}}}_{n, (\beta_2,d)}^X \,(-p)^n q^{d}
= -\frac{3}{2} K(p,q)^4 \wp(p,q) - \frac{3}{8} K(p^{2}, q^{2})^2\, .$
Part (i) of Corollary\* \[dtcor\] verifies an earlier prediction of BPS counts on abelian threefolds by Maldacena, Moore, and Strominger [@MMS].
Theorem \[dthatthm\] concerns the Hilbert schemes of curves on $X$. The Euler characteristics associated to the Hilbert scheme of points of $X$ (via the generalized Kummer construction) have been calculated recently by J. Shen [@Shen] – proving a conjecture of Gulbrandsen [@Gul].
### Gromov-Witten theory
Let $X$ be an abelian threefold, and let $\beta$ be a curve class of type $(d_1, d_2, d_3)$ with $d_1, d_2 > 0$. We consider curves of genus $g \geq 2$.
The translation action of $X$ on ${{\overline M}}_{g}(X, \beta)$ has finite stabilizer. Hence, $${{\overline M}}_{g}(X, \beta) / X$$ is a Deligne-Mumford stack. In Section \[GW3\], we use methods of Kiem and Li [@KL] to construct a reduced virtual class $$\left[ {{\overline M}}_{g}(X, \beta) / X \right]^{\text{red}}$$ on ${{\overline M}}_{g}(X, \beta) / X$ of dimension $0$. We define the *reduced quotient Gromov-Witten invariants of* $X$ by $$\mathsf{N}_{g, \beta} = \int_{ [ {{\overline M}}_{g}(X, \beta) / X ]^{\text{red}} } \mathsf{1} \,.$$ By construction, $\mathsf{N}_{g, \beta}$ is deformation invariant and hence only depends on $g$ and the type $(d_1, d_2, d_3)$ of $\beta$. We write $$\mathsf{N}_{g, \beta} = \mathsf{N}_{g, (d_1, d_2, d_3)} \,.$$ The number $\mathsf{N}_{g, \beta}$ is a virtual count of translation classes of genus $g$ curves in $X$ of class $\beta$ in $X$. In Section \[GW3\], we show that $\mathsf{N}_{g, \beta}$ determines the full reduced descendent Gromov-Witten theory of $X$ in genus $g$ and class $\beta$.
The following conjecture relates the Gromov-Witten invariants $\mathsf{N}_{g, \beta}$ to the Donaldson-Thomas invariants ${\operatorname{\mathsf{DT}}}_{n,\beta}^X$ defined above. Define the generating series $$Z^{\text{GW}}_{\beta}(u) = \sum_{g \geq 2} \mathsf{N}_{g, \beta} u^{2g-2}
\quad \text{ and } \quad
Z^{\text{DT}}_{\beta}(y) = \sum_{n \in {{\mathbb{Z}}}} {\operatorname{\mathsf{DT}}}_{n, \beta}^X y^n \,.$$
\[conjB\] The series $Z^{\textup{DT}}_{\beta}(y)$ is the Laurent expansion of a rational function in $y$ and $$Z^{\textup{DT}}_\beta(y) = Z^{\textup{GW}}_\beta(u) \,.$$ after the variable change $y = -e^{iu}$.
Conjecture \[conjB\] is a Gromov-Witten/Donaldson-Thomas correspondence for reduced theories [@MNOP1; @ObP]. In conjunction with part (i) of Corollary\* \[dtcor\] and the expansion , Conjecture \[conjB\] determines the invariants $\mathsf{N}_{g,(1,1,d)}$ for all $d \geq 0$ by the formula $$\begin{aligned}
\sum_{d \geq 0} \sum_{g \geq 2} \mathsf{N}_{g, (1,1,d)} u^{2g-2} q^{d} &=
-K^2(z,\tau)|_{y=-e^{2\pi i z}, q=e^{2\pi i\tau}}\\
&= (y + 2 + y^{-1}) \prod_{m \geq 1} \frac{(1 + y q^m)^2 (1 + y^{-1} q^m)^2}{(1-q^m)^4} \,.\end{aligned}$$
To capture the invariants $\mathsf{N}_{g,(1,{{d'}},d)}$ for higher ${{d'}}$, we conjecture an additional structure governing the counting. Let $$f_{(d_1,d_2,d_3)}(u) = \sum_{g \geq 2} \mathsf{N}_{g, (d_1, d_2, d_3)} u^{2g-2} \,.$$ The following multiple cover rule expresses the invariants of type $(1,{{d'}},d)$ in terms of those of type $(1,1,d)$.
\[conjC\] For all ${{d'}},d > 0$, $$f_{(1,{{d'}},d)}(u) = \sum_{k | \gcd({{d'}},d)} \frac{1}{k}\, f_{\left(1, 1, \frac{{{d'}}d}{k^2} \right)}(k u) \,.$$
Conjecture \[conjC\] matches the counts of genus $3$ curves by the lattice method of [@D; @G; @LS]. The deepest support for Conjecture \[conjC\] is a highly non-trivial match with part (ii) of Corollary\* \[dtcor\].[^10]
Taken together, Corollary\* \[dtcor\] and Conjectures \[conjB\] and \[conjC\] determine the invariants $\mathsf{N}_{g, \beta}$ for all primitive[^11] classes $\beta$. The discussion is parallel to [@ObP Conjecture A] concerning the virtual enumeration of curves on $K3\times E$ in classes $(\beta,d)$ where $\beta\in H_2(K3,\mathbb{Z})$ is primitive. The latter has been computed for $\langle \beta, \beta \rangle \in \{ -2, 0 \}$ in Donaldson-Thomas theory in [@Bryan-K3xE]. The proof of Theorem \[dtthm\] in Section \[secdt\] closely follows the strategy of [@Bryan-K3xE].
For Theorem \[dtthm\] and Conjecture \[conjB\], the Hilbert scheme of curves can be replaced by the moduli space of stable pairs [@PT1]. For technical aspects of the proof of Theorem \[dtthm\], ideal sheaves are simpler – but there is no fundamental difference in the arguments required here.
A multiple cover formula for BPS counts in classes $(1,d,d')$ was proposed in [@MMS]. However, the formula in [@MMS] is different from ours and does *not* match the genus $3$ counts or Corollary\* \[dtcor\]. An extension of Conjecture \[conjC\] to all curve classes is discussed in Section \[Subsection\_imprimitive\_classes\].
Plan of the paper
-----------------
In Section 1, we recall several classical facts concerning divisors and curves on abelian varieties. Polarized isogenies, which play a central role in the enumeration of low genus curves, are reviewed. Reduced virtual classes are discussed in Section 1.4. Degenerate curves classes are analyzed in Section 1.5.
Part I of the paper (Sections 2-5) concerns the enumeration of curves on abelian surfaces $A$. In Section 2, the genus 2 part of Theorem \[YZ\_intro\] is proven. In Section 3, the proofs of Theorem \[YZ\_intro\] and Theorem \[thm\_point\_insertion\] for primitive classes are completed. A connection with the Euler characteristic calculations of stable pairs moduli spaces on $A$ by Göttsche and Shende [@GS] is discussed in Section 3.7. The quasi-modularity of the primitive descendent potentials of $A$ is studied in Section 4 where a refinement of Theorem \[modularity\] is proven. A parallel refined quasi-modularity result for the reduced Gromov-Witten theory of $K3$ surfaces is presented in Section 4.6. The enumeration of hyperelliptic curves on $A$ and the proof of Theorem \[thm\_hyp\] is given in Section 5.
Part II of the paper (Sections 6-7) concerns the enumeration of curves on abelian threefolds $X$. In Section 6, the topological and Behrend weighted Euler characteristics of the Hilbert scheme of curves in $A\times E$ are studied. For $d'\in \{1,2\}$, the topological Euler characteristic theory is calculated and Theorem \[dtthm\] is proven (except for genus 3 lattice counts which appear in Section 7.4). Conjecture \[conj\_Behrend\] relating the two theories is presented in Section 6.6. In Section 7, the foundations of the quotient Gromov-Witten theory are discussed and the full descendent theory is expressed in terms of the invariants $\mathsf{N}_{g,(d_1,d_2,d_3)}$. The relationship between Theorem \[dtthm\] and Conjectures \[conjB\] and \[conjC\] is studied in Section 7.5. Finally, a multiple cover formula for imprimitive classes is proposed in Section 7.6.
Acknowledgements
----------------
We thank B. Bakker, L. Göttsche, S. Katz, M. Kool, A. Klemm, A. Knutsen, A. Kresch, J. Li, D. Maulik, G. Moore, B. Pioline, M. Raum, S. Rose, J. Schmitt, E. Sernesi, J. Shen, V. Shende, and R. Thomas for discussions about counting invariants of abelian surfaces and threefolds. The results here were first presented at the conference [*Motivic invariants related to $K3$ and abelian geometries*]{} at Humboldt University in February 2015. We thank G. Farkas and the Einstein Stiftung for support in Berlin.
J.B. was partially supported by NSERC Accelerator and Discovery grants. The research here was partially carried out during a visit of J.B. to the Forschungsinstitut für Mathematik at ETH Zürich in November 2014. G.O. was supported by the grant SNF-200021-143274. R.P. was partially supported by SNF-200021-143274, ERC-2012-AdG-320368-MCSK, SwissMap, and the Einstein Stiftung. Q.Y. was supported by the grant ERC-2012-AdG-320368-MCSK.
Abelian varieties
=================
Overview
--------
We review here some basic facts about divisor and curve classes on abelian varieties. A standard reference for complex abelian varieties is [@CAV]. A treatment of polarized isogenies is required for the lattice counting in Sections \[g2c\] and \[secg3lc\]. Using results of Kiem-Li [@KL], we define reduced virtual classes on the moduli spaces of stable maps to abelian varieties. Finally, we show that the (reduced) Gromov-Witten theory of abelian varieties of arbitrary dimensions is determined by the (reduced) theories in dimensions up to $3$.
Curve classes {#cct}
-------------
Let $V={{\mathbb{C}}}^n$. Let $\Lambda \subset V$ be a rank $2n$ lattice for which $$A = V/\Lambda$$ is an $n$-dimensional compact complex torus. Let $L$ be a holomorphic line bundle on $A$. The first Chern class $$c_1(L) \in H^2(A, {{\mathbb{Z}}})$$ induces a Hermitian form $$H : V \times V {{\ \rightarrow\ }}{{\mathbb{C}}}$$ and an alternating form $$E = {\rm Im}\, H : \Lambda \times \Lambda {{\ \rightarrow\ }}{{\mathbb{Z}}}\,.$$ By the elementary divisor theorem, there exists a [*symplectic*]{} basis of $\Lambda$ in which $E$ is given by the matrix $$\begin{pmatrix}
0 & D \\
-D & 0
\end{pmatrix},$$ where $D = \text{Diag}(d_1, \ldots, d_n)$ with integers $d_i \geq 0$ satisfying $$d_1\, |\, d_2\, | \,\cdots\, | \,d_n \,.$$ The tuple $(d_1, \ldots, d_n)$ is uniquely determined by $L$ (in fact by $c_1(L)$) and is called the [*type*]{} of $L$.
A [*polarization*]{} on $A$ is a first Chern class $c_1(L)$ with positive definite Hermitian form $H$ (in particular $d_i > 0$). The polarization is [*principal*]{} if $d_i = 1$ for all $i$. The moduli space of polarized $n$-dimensional abelian varieties of a fixed type is irreducible of dimension $n(n + 1)/2$.
Let $\beta \in H_2(A, {{\mathbb{Z}}})$ be a curve class on $A$. The class corresponds to $$\widehat{\beta} = c_1(\widehat{L}) \in H^2(\widehat{A}, {{\mathbb{Z}}}) \,,$$ where $\widehat{A} = {\mathop{\rm Pic}\nolimits}^0(A)$ is the dual complex torus of $A$ and $\widehat{L}$ is a line bundle on $\widehat{A}$. We define the [*type*]{} $(d_1, \ldots,d_n)$ of $\beta$ to be the type of $\widehat{L}$. The class $\beta$ is primitive if and only if $d_1=1$. For abelian surfaces, we may view $\beta$ as either a curve class or a divisor class: the resulting types are the same.
If $\beta$ is of type $(d_1, \ldots, d_n)$ with $d_i > 0$ for all $i$, then $\widehat{\beta}$ is a polarization on $\widehat{A}$. Hence, all curve classes of a fixed type $(d_1, \ldots, d_n)$ with $d_i>0$ for all $i$ are deformation equivalent.
If $d_i = 0$ for some $i$, then we say that $\beta$ is of [*degenerate*]{} type. Write $k = \max\{ i \, | \, d_i \neq 0 \}$. By [@CAV Theorem 3.3.3], there exists a subtorus $B \subset \widehat{A}$ of dimension $n - k$ with quotient map $$p : \widehat{A} {{\ \rightarrow\ }}\bar{A} = \widehat{A}/B \,,$$ and a polarization $$\bar{\beta} = c_1(\bar{L}) \in H^2(\bar{A}, {{\mathbb{Z}}})$$ of type $(d_1, \ldots, d_k)$, such that $\widehat{\beta} = p^*(\bar{\beta})$. The deformation of $\beta$ is then governed by the deformation of $\bar{\beta}$. As a result, curve classes of a fixed type $(d_1, \ldots, d_k, 0, \ldots, 0)$ are also deformation equivalent.
Let $A$ be the product of $n$ elliptic curves $E_1 \times \cdots \times E_n$. For integers $a_1, \ldots, a_n \geq 0$, consider the curve class $$\beta = a_1[E_1] + \cdots + a_n[E_n]\in H_2(A,{{\mathbb{Z}}}) \,.$$ The type $(d_1, \ldots, d_n)$ of $\beta$ is given by the rank and the invariant factors of the abelian group associated to $(a_1, \ldots, a_n)$: $$\bigoplus_{i=1}^n {{\mathbb{Z}}}/a_i \cong {{\mathbb{Z}}}^m \oplus \bigoplus_{j=1}^k {{\mathbb{Z}}}/d'_j \,.$$ Here, $k, m \leq n$ and $$(d_1, \ldots, d_n)= (\underbrace{1, \ldots, 1}_{n-k-m}, d'_1,\ldots, d'_k, \underbrace{0, \ldots, 0}_m) \,.$$ Later we shall also say that $\beta$ is of type $(a_1, \ldots, a_n)$ without requiring $a_1 | a_2 | \cdots | a_n$.
Two tuples $(a_1, \ldots, a_n)$ and $(b_1, \ldots, b_n)$ are deformation equivalent if and only if $$\bigoplus_{i=1}^n {{\mathbb{Z}}}/a_i \cong \bigoplus_{i=1}^n {{\mathbb{Z}}}/b_i \,.$$ The primitivity of $\beta$ is determined by $\gcd(a_1, \ldots, a_n)$.
Polarized isogenies {#piso}
-------------------
The following discussion is based on [@D; @G; @LS]. Let $$f : C {{\ \rightarrow\ }}A$$ be a map from a nonsingular curve of genus $g$. Suppose the curve class $$\beta = f_*[C]$$ is of type $(d_1, \ldots, d_n)$ with $d_i > 0$ for all $i$. The map $f$ factors as $$C \xrightarrow{\mathsf{aj}} J \xrightarrow{\pi} A \,,$$ where $J$ is the Jacobian of $C$ and $\mathsf{aj}$ is the Abel-Jacobi map, defined up to translation by $J$. By duality, $\pi$ corresponds to $$\widehat{\pi} : \widehat{A} {{\ \rightarrow\ }}J \quad \text{ such that } \quad \widehat{\beta} = \widehat{\pi}^*\theta \,,$$ where $\theta$ is the theta divisor class on $J$ (here we identify $J$ with $\widehat{J}$). When $g = n$, the map $\widehat{\pi}$ is a [*polarized isogeny*]{}.
More generally, consider the isogeny $$\phi_{\widehat{\beta}} : \widehat{A} {{\ \rightarrow\ }}\skew{5}\widehat{\widehat{A}} \cong A \,, \quad\quad x \mapsto t_x^*\widehat{L} \otimes \widehat{L}^{-1} \,,$$ where $\widehat{\beta} = c_1(\widehat{L})$ for some line bundle $\widehat{L}$ and $t_x : \widehat{A} \to \widehat{A}$ is the translation by $x$. The finite kernel of $\phi_{\widehat{\beta}}$ admits a non-degenerate multiplicative alternating form $$\langle\,\,\,, \,\,\rangle : {\mathop{\rm Ker}\nolimits}(\phi_{\widehat{\beta}}) \times {\mathop{\rm Ker}\nolimits}(\phi_{\widehat{\beta}}) {{\ \rightarrow\ }}{{\mathbb{C}}}^* \,,$$ called the [*commutator pairing*]{}. By [@CAV Corollary 6.3.5], there is a bijective correspondence between the following two sets:
- polarized isogenies from $(\widehat{A}, \widehat{\beta})$ to principally polarized abelian varieties $(B, \theta)$,
- maximal totally isotropic subgroups of ${\mathop{\rm Ker}\nolimits}(\phi_{\widehat{\beta}})$.
The cardinality of both sets depends only on the type $(d_1, \ldots, d_n)$ of $\beta$, and is denoted by $$\nu(d_1, \ldots, d_n) \,.$$ In fact, under a suitable basis of $\widehat{\Lambda}$ we have $${\mathop{\rm Ker}\nolimits}(\phi_{\widehat{\beta}}) \cong ({{\mathbb{Z}}}/d_1 \times \cdots \times {{\mathbb{Z}}}/d_n)^2 \,, \label{Kerbeta}$$ and in terms of standard generators $e_1, \ldots, e_n, f_1, \ldots, f_n$ of , $$\langle e_k, f_{\ell} \rangle = e^{\delta_{k\ell}\frac{2\pi i}{d_k}} \,.$$ The number $\nu(d_1, \ldots, d_n)$ can be computed as follows.
We have $$\nu(d_1, \ldots, d_n) = \sum_{K < {{\mathbb{Z}}}/d_1 \times \cdots \times {{\mathbb{Z}}}/d_n} \#{\mathop{\rm Hom}\nolimits}^{\rm sym}(K, \widehat{K}) \,, \label{Lattice}$$ where $\widehat{K} = {\mathop{\rm Hom}\nolimits}(K, {{\mathbb{C}}}^*)$ and ${\mathop{\rm Hom}\nolimits}^{\rm sym}$ stands for symmetric homomorphisms.
A straightforward analysis yields $$\nu(1, \ldots, 1, d) = \sigma(d) = \sum_{k | d} k \,. \label{Primlat}$$ A list of values of $\nu(d_1, \ldots, d_n)$ can be found in [@LS], but some of the entries are incorrect. For example, the number $\nu(2, 4)$ should be $39$ and not $51$.
The counts of polarized isogenies are closely related to the counts of the lowest genus curves on abelian surfaces and threefolds. Moreover, this lattice method is also important in counting higher genus curves in class of type $(1, 2, d)$, see Sections \[g2c\] and \[secg3lc\].
Reduced virtual classes {#cosection}
-----------------------
Let $A$ be an abelian variety of dimension $n \geq 2$, and let $\beta \in H_2(A, {{\mathbb{Z}}})$ be a curve class of type $$(d_1, \ldots, d_k, \underbrace{0, \ldots, 0}_m)$$ with $d_i > 0$ for all $i$. Here, $k > 0$, $m \geq 0$, and $k + m = n$.
By the discussion in Section \[cct\], there exists a subtorus $A' \subset A$ of dimension $k$ and a curve class $\beta' \in H_2(A', {{\mathbb{Z}}})$ of type $(d_1, \ldots, d_k)$ such that $\beta$ is the push-forward of $\beta'$. Write $$\pi : A \to A'' = A/A'$$ for the quotient map.
Consider the moduli space of stable maps ${{\overline M}}_g(A, \beta)$. Using the [*cosection localization*]{} method of Kiem-Li [@KL], we define a (maximally) reduced virtual class $$\left[{{\overline M}}_g(A, \beta)\right]^{\text{red}}$$ on ${{\overline M}}_g(A, \beta)$. The case with marked points is done similarly. The result provides a foundation for the reduced Gromov-Witten theory of abelian varieties.
By [@KL Section 6], every holomorphic 2-form $\theta \in H^0(A, \Omega^2_A)$ induces a map $$\sigma_\theta : \text{Ob}_{{{\overline M}}_g(A, \beta)} {{\ \rightarrow\ }}\mathcal{O}_{{{\overline M}}_g(A, \beta)} \,,$$ where $\text{Ob}_{{{\overline M}}_g(A, \beta)}$ is the obstruction sheaf of ${{\overline M}}_g(A, \beta)$.
The map $\sigma_\theta$ is trivial if $$\theta \in \pi^*H^0(A'', \Omega^2_{A''})$$ and surjective otherwise.
Let $[f : C \to A] \in {{\overline M}}_g(A, \beta)$ be a stable map in class $\beta$. After translation, we may assume $\text{Im}(C) \subset A'$.
By [@KL Proposition 6.4], the map $\sigma_\theta$ is trivial at $[f]$ if and only if the composition $$\label{compo}
T_{C_{\text{reg}}} \xrightarrow{df} f^*T_A |_{C_{\text{reg}}} \xrightarrow{f^*\widehat{\theta}} f^*\Omega_A |_{C_{\text{reg}}}$$ is trivial. Here $C_{\text{reg}}$ is the regular locus of $C$ and $\widehat{\theta} : T_A \to \Omega_A$ is the map induced by $\theta$. Since $\text{Im}(C) \subset A'$, it is clear that is trivial if $\theta \in \pi^*H^0(A'', \Omega^2_{A''})$.
For the surjectivity statement, we identify the tangent space $T_{A, x}$ at $x \in A$ with $T_{A, 0_A}$ by translation. Since $\beta'$ is of type $(d_1, \ldots, d_k)$ with $d_i > 0$, the curve $\text{Im}(C)$ generates $A'$ as a group. By [@Debarre Lemma 8.2], there exists an open dense subset $U \subset \text{Im}(C)_{\text{reg}}$, such that $T_{A', 0_{A'}}$ is spanned by $$T_{\text{Im}(C)_{\text{reg}}, x} \quad \text{ for } \quad x \in U \,.$$ It follows that for any $\theta \in H^0(A, \Omega^2_A) \setminus \pi^*H^0(A'', \Omega^2_{A''})$, there exists a point $x \in U$ with non-trivial at $x$.
Hence, by taking a basis of the quotient $H^0(A, \Omega^2_A) / \pi^* H^0(A'', \Omega^2_{A''})$, we obtain a surjective map $$\sigma : \text{Ob}_{{{\overline M}}_g(A, \beta)} {{\ \rightarrow\ }}\mathcal{O}_{{{\overline M}}_g(A, \beta)}^{\oplus r(k, m)}$$ with $$r(k, m) = \binom{k + m}{2} - \binom{m}{2} = \frac{k(k - 1)}{2} + km \,.$$ Then, by the construction of [@KL], the map $\sigma$ yields a reduced virtual class $[{{\overline M}}_g(A, \beta)]^{\text{red}}$ of dimension $$\text{vdim}\,{{\overline M}}_g(A, \beta) + r(k, m) = (k + m - 3)(1 - g) + \frac{k(k - 1)}{2} + km \,.$$
Gromov-Witten theory in degenerate curve classes {#dcosection}
------------------------------------------------
We now explore the possibilities of obtaining non-trivial reduced Gromov-Witten invariants for $A$ and $\beta$. By deformation invariance, the invariants depend only on the type $$(d_1, \ldots, d_k, \underbrace{0, \ldots, 0}_m)\, .$$ We may then assume $$A = A' \times A''$$ with $A'$ generic among abelian varieties carrying a curve class of type $(d_1, \ldots, d_k)$, and $A''$ a product of $m$ elliptic curves, $$A'' = E_1 \times \cdots \times E_m \,.$$ By the genericity of $A'$, there are no stable maps of genus less than $k = \dim \, A'$ in class $\beta$. Hence, all invariants in genus $< k$ vanish.
We list the following four cases according to the number $k$ of non-zero entries in the type of $\beta$.
**Case** $k = 1$. For $g \geq 1$, stable maps $[f : C \to A] \in {{\overline M}}_g(A, \beta)$ come in $m$-dimensional families via the translation action of $A''$. On the other hand, the translation by the elliptic curve $A'$ fixes $\text{Im}(f)$. The expected dimension modulo the translation by $A''$ is $$\text{vdim}\,{{\overline M}}_g(A, \beta) + r(1, m) - m = (m - 2)(1 - g) \,.$$
Integrals over the reduced class $[{{\overline M}}_g(A, \beta)]^{\text{red}}$ can be evaluated by eliminating the $E$-factors. In each step from $$A' \times E_1 \times \cdots \times E_{i + 1} \quad \text{ to } \quad A' \times E_1 \times \cdots \times E_i \,,$$ we find a surjective map ${\mathbb{E}}^{\vee} \to \mathcal{O}$ where ${\mathbb{E}}$ is the Hodge bundle. We then obtain a copy of the top Chern class of $${\mathop{\rm Ker}\nolimits}({\mathbb{E}}^{\vee} {{\ \rightarrow\ }}\mathcal{O})$$ which is $(-1)^{g - 1}\lambda_{g - 1}$. This follows from a close analysis of the obstruction sheaf and the definition of the reduced class. In the end, we arrive at integrals over $[{{\overline M}}_g(A', \beta')]^{\text{vir}}$ with $$\big((-1)^{g - 1}\lambda_{g - 1}\big)^m$$ in the integrand.
For $m = 1$ ($\dim \, A = 2$), the theory becomes the study of $\lambda_{g - 1}$-integrals on the elliptic curve $A'$. Such Hodge integrals may be expressed [@FP1] in terms of the descendent theory of an elliptic curve [@OP1; @OP3].
For $m \geq 2$ ($\dim \, A \geq 3$), all invariants in genus $g \geq 2$ vanish. By Mumford’s relation for $g\geq 2$, $$\lambda_{g - 1}^2 = 2\lambda_g\lambda_{g - 2}\, ,$$ and $\lambda_g$ annihilates the virtual fundamental class of non-constant maps to the elliptic curve $A'$. In genus $1$, all invariants are multiples of $$\label{gg11} \int_{[{{\overline M}}_1(A', \beta')]^{\text{vir}}} \mathsf{1}
= \frac{\sigma(d_1)}{d_1}$$ for $\beta'$ of type $(d_1)$.
**Case** $k = 2$. For $g \geq 2$, stable maps in ${{\overline M}}_g(A, \beta)$ come in $(2 + m)$-dimensional families via the translation action of $A$. The expected dimension modulo translation is $$\text{vdim}\,{{\overline M}}_g(A, \beta) + r(2, m) - (2 + m) = (m - 1)(2 - g) \,.$$ Similar to the $k = 1$ case, by eliminating each $E$-factor we find a surjective map ${\mathbb{E}}^{\vee} {{\ \rightarrow\ }}\mathcal{O}^{\oplus 2}$, and obtain a copy of the top Chern class of $${\mathop{\rm Ker}\nolimits}({\mathbb{E}}^{\vee} {{\ \rightarrow\ }}\mathcal{O}^{\oplus 2})$$ which is $(-1)^{g - 2}\lambda_{g - 2}$.
The reduced Gromov-Witten theory of the abelian surface $A'$ is the subject of Part I of the paper. For $m = 1$ ($\dim \, A = 3$), we find integrals of the form $$\int_{[{{\overline M}}_g(A', \beta')]^{\text{red}}} (-1)^{g - 2} \lambda_{g - 2}\, \ldots\, ,$$ where the dots stand for further terms in the integrand. Our interest in $\lambda_{g - 2}$-integrals on an abelian surface (see and Theorem \[YZ\_intro\]) is directly motivated by Gromov-Witten theory in degenerate curve classes on abelian threefolds.
For $m \geq 2$ ($\dim \, A \geq 4$), all invariants in genus $g \geq 3$ vanish for dimension reasons. We are then reduced to the genus $2$ invariants of $A'$ and $\beta'$.
**Case** $k = 3$. Similar to the $k = 2$ case, for $g \geq 3$, the expected dimension modulo translation is $$\text{vdim}\,{{\overline M}}_g(A, \beta) + r(3, m) - (3 + m) = m(3 - g) \,.$$ Here by eliminating each $E$-factor we find a surjective map ${\mathbb{E}}^{\vee} {{\ \rightarrow\ }}\mathcal{O}^{\oplus 3}$, and obtain a copy of the top Chern class of $${\mathop{\rm Ker}\nolimits}({\mathbb{E}}^{\vee} {{\ \rightarrow\ }}\mathcal{O}^{\oplus 3})$$ which is $(-1)^{g - 3}\lambda_{g - 3}$.
The reduced Gromov-Witten theory of the abelian threefold $A'$ is studied in Part II of the paper. For $m \geq 1$ ($\dim \, A \geq 4$), all invariants in genus $g \geq 4$ vanish for dimension reasons. We are reduced to the genus $3$ invariants of $A'$ and $\beta'$.
**Case** $k \geq 4$. For $g \geq k$, the expected dimension modulo translation is $$\begin{gathered}
\label{dimcount}
\text{vdim}\,{{\overline M}}_g(A, \beta) + r(k, m) - (k + m) \\
= (k - 3)\bigg(\frac{k}{2} + 1 - g\bigg) + m(k - g) \,.\end{gathered}$$ The right hand side of is always negative for $g \geq k \geq 4$. Hence, all invariants vanish.
In conclusion, the (reduced) Gromov-Witten theory of abelian varieties of arbitrary dimensions is completely determined by the (reduced) Gromov-Witten theories of abelian varieties of dimensions $1\leq d \leq 3$. The analysis here justifies our focus on these low dimensions.
Furthermore, for abelian varieties of dimension at least $4$, only genus $1\leq g \leq 3$ invariants can possibly survive. Exact formulas[[^12]]{} are available for the genus $1\leq g \leq 3$ invariants which arise for abelian varieties of dimension at least 4.
The genus 2 case {#g2c}
================
Quotient Gromov-Witten invariants {#secg2lc}
---------------------------------
Let $A$ be an abelian surface, and let $\beta \in H_2(A, {{\mathbb{Z}}})$ be a curve class of type $(d_1, d_2)$ with $d_1, d_2 > 0$. In Section \[bcc\], we defined invariants $$\mathsf{N}^{\text{FLS}}_{g,\beta} = \mathsf{N}^{\text{FLS}}_{g,(d_1,d_2)}$$ counting genus $g$ curves in a fixed linear system. It is sometimes more natural to count curves up to translation. A reasonable path[^13] to the definition of such invariants is by integrating over the quotient stack $${{\overline M}}_{g}(A, \beta) / A \,. \label{qmbargn}$$
Classically, people have taken a simpler course. Let $$p : {{\overline M}}_{g}(A, \beta) {{\ \rightarrow\ }}{\mathop{\rm Pic}\nolimits}^{\beta}(A) \cong \widehat{A} \,,$$ be the morphism which sends a curve $[f : C \rightarrow A]$ to the divisor class associated with its image curve.[[^14]]{} The map $p$ is equivariant with respect to the actions of $A$ on ${{\overline M}}_{g}(A, \beta)$ by translation and on $\widehat{A}$ by the isogeny $$\phi_{\beta} : A \rightarrow \widehat{A}\, .$$ An element $x \in A$ fixes a linear system of type $(d_1, d_2)$ if and only if $x$ is an element of $${\mathop{\rm Ker}\nolimits}( \phi_{\beta} : A \to \widehat{A} ) \cong ({{\mathbb{Z}}}/d_1 \times {{\mathbb{Z}}}/d_2)^2 \,. \label{KerFLS}$$ The quotient space equals the quotient of ${{\overline M}}_{g}(A, \beta)^{\text{FLS}}$ by the finite group with $(d_1d_2)^2$ elements. Therefore, we [*define*]{} the invariants counting curves up to translation by $$\mathsf{N}^{\text{Q}}_{g,(d_1, d_2)} = \frac{1}{(d_1 d_2)^2} \mathsf{N}^{\text{FLS}}_{g,(d_1, d_2)} \,. \label{defqqq}$$
In genus $2$, the invariants are related to the lattice counts considered in Section \[piso\].
\[G2lattice\] For all $d_1, d_2 > 0$, $$\mathsf{N}^{\textup{Q}}_{2,(d_1, d_2)} = \nu(d_1, d_2) \,.$$
Let $\beta$ be of type $(d_1, d_2)$ and assume ${\mathop{\rm End}\nolimits}_{{{\mathbb{Q}}}}(A) = {{\mathbb{Q}}}$. In particular, $A$ is simple (contains no elliptic curves) and ${\mathop{\rm Aut}\nolimits}(A) = \{\pm 1\}$. It follows that every genus $2$ stable map $$f : C {{\ \rightarrow\ }}A$$ in class $\beta$ has a nonsingular domain $C$. As discussed in Section \[piso\], to such a map $f$, we can associate a polarized isogeny $$(\widehat{A}, \widehat{\beta}) \to (J, \theta)\, ,$$ where $J$ is the Jacobian of $C$.
Conversely, every simple principally polarized abelian surface $(B, \theta)$ is the Jacobian of a unique nonsingular genus $2$ curve $C$. Hence, each polarized isogeny $(\widehat{A}, \widehat{\beta}) \to (B, \theta)$ induces a map $$f : C \xrightarrow{\mathsf{aj}} B \to A \,.$$ The map $f$ is unique up to translation and automorphism of $A$. Moreover, the automorphism $-1$ of $A$ corresponds to the hyperelliptic involution of $C$.
The abelian surface $A$ acts freely[^15] on ${{\overline M}}_2(A, \beta)$ by translation. To prove this, we decompose a genus $2$ map $f : C \to A$ as $$f : C \xrightarrow{\mathsf{aj}} J \xrightarrow{\pi} A \,.$$ First, since $\big[\mathsf{aj}(C)\big]$ is a divisor class of type $(1, 1)$, the only element in $J$ fixing $\mathsf{aj}(C)$ is $0_J$. Second, the preimage $\pi^{-1}\big(f(C)\big)$ is the union $$\bigcup_{x \in {\mathop{\rm Ker}\nolimits}(\pi)} t_x\big(\mathsf{aj}(C)\big) \,, \label{Union}$$ where $t_x : J \to J$ is the translation by $x$. Suppose a point $a \in A$ fixes $f(C)$, and let $b \in \pi^{-1}\{a\}$. By , we have $$t_b\big(\mathsf{aj}(C)\big) = t_x\big(\mathsf{aj}(C)\big)$$ for some $x \in {\mathop{\rm Ker}\nolimits}(\pi)$. In other words, the element $b - x \in J$ fixes $\mathsf{aj}(C)$. Hence, $b - x = 0_J$ and $a = \pi(b) = \pi(b - x) = 0_A$.
It follows that ${{\overline M}}_2(A, \beta) / A$ is precisely a set of $\nu(d_1, d_2)$ isolated reduced points (or equivalently, ${{\overline M}}_2(A, \beta)^{\text{FLS}}$ is a set of $(d_1d_2)^2 \nu(d_1, d_2)$ isolated reduced points).
Genus 2 counts
--------------
For genus $2$, the following result determines the counts in all classes.
\[YZA\] For all $d_1, d_2 > 0$, $$\mathsf{N}^{\textup{Q}}_{2,(d_1, d_2)} = \sum_{k | \gcd(d_1,d_2)} \sum_{ \ell | \frac{ d_1 d_2 }{k^2}} k^{3} \ell \,. \label{YZ}$$
The primitive case (where $\gcd(d_1, d_2) = 1$) was proven in [@D; @G; @LS] via the lattice method discussed in Section \[piso\] and in [@BLA; @Ros14] via geometric arguments. A closer look at the lattice method actually yields a proof of Theorem \[YZA\] in the general case.
To prove for $\nu(d_1, d_2)$, we are immediately reduced to the case $\nu(p^m, p^n)$ where $p$ is a prime number and $m \leq n$. For $m = 0$, we have, by , $$\nu(1, p^n) = \sigma(p^n) = \sum_{k = 0}^n p^k \,.$$ It suffices then to prove the following recursion: $$\nu(p^m, p^n) = \nu(1, p^{m + n}) + p^3 \nu(p^{m - 1}, p^{n - 1}) \label{Rec}$$ for $1 \leq m \leq n$.
The proof uses . Consider the quotient map $$\pi : {{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n {{\ \rightarrow\ }}{{\mathbb{Z}}}/p^{m - 1} \times {{\mathbb{Z}}}/p^{n - 1} \,.$$ For $1\leq r \leq s$, the map $\pi$ induces a bijective correspondence between the following two sets:
- subgroups of ${{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n$ isomorphic to ${{\mathbb{Z}}}/p^r \times {{\mathbb{Z}}}/p^s$,
- subgroups of ${{\mathbb{Z}}}/p^{m - 1} \times {{\mathbb{Z}}}/p^{n - 1}$ isomorphic to ${{\mathbb{Z}}}/p^{r - 1} \times {{\mathbb{Z}}}/p^{s - 1}$.
We also have $$\begin{gathered}
\#{\mathop{\rm Hom}\nolimits}^{\rm sym}({{\mathbb{Z}}}/p^r \times {{\mathbb{Z}}}/p^s, \widehat{{{\mathbb{Z}}}/p^r \times {{\mathbb{Z}}}/p^s}) \\
= p^3 \#{\mathop{\rm Hom}\nolimits}^{\rm sym}({{\mathbb{Z}}}/p^{r - 1} \times {{\mathbb{Z}}}/p^{s - 1}, \widehat{{{\mathbb{Z}}}/p^{r - 1} \times {{\mathbb{Z}}}/p^{s - 1}}) \,. \end{gathered}$$ The remaining subgroups of ${{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n$ are cyclic and isomorphic to ${{\mathbb{Z}}}/p^k$ for some $0 \leq k \leq n$. Moreover, $$\#{\mathop{\rm Hom}\nolimits}^{\rm sym}({{\mathbb{Z}}}/p^k, \widehat{{{\mathbb{Z}}}/p^k}) = \#{\mathop{\rm Hom}\nolimits}({{\mathbb{Z}}}/p^k, \widehat{{{\mathbb{Z}}}/p^k}) = p^k \,.$$ Applying , we find $$\nu(p^m, p^n) = \sum_{k = 0}^n p^k \#\{{{\mathbb{Z}}}/p^k < {{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n\} + p^3 \nu(p^{m - 1}, p^{n - 1}) \,.$$ The numbers of cyclic subgroups can be deduced from classical group theory (see [@SUBG Lemma 1.4.1]): $$\#\{{{\mathbb{Z}}}/p^k < {{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n\} = \sum_{\substack{0 \leq i \leq m, 0 \leq j \leq n \\ \max(i, j) = k}} \varphi(p^{\min(i, j)}) \,,$$ where $\varphi$ is Euler’s phi function.[^16] We have $$\begin{aligned}
& \sum_{k = 0}^n p^k \#\{{{\mathbb{Z}}}/p^k < {{\mathbb{Z}}}/p^m \times {{\mathbb{Z}}}/p^n\} \\
={} & \sum_{k = 0}^n p^k \sum_{\substack{0 \leq i \leq m, 0 \leq j \leq n \\ \max(i, j) = k}} \varphi(p^{\min(i, j)}) \\
={} & \sum_{k = 0}^m p^k \sum_{i = 0}^k \varphi(p^i) + \sum_{k = m + 1}^n p^k \sum_{i = 0}^m \varphi(p^i) + \sum_{k = 1}^m p^k \sum_{j = 0}^{k - 1} \varphi(p^j) \\
={} & \sum_{k = 0}^m p^{2k} + \sum_{k = m + 1}^n p^{m + k} +\sum_{k = 1}^m p^{2k - 1} \\
={} & \sum_{k = 0}^{m + n} p^k = \nu(1, p^{m + n}) \,,\end{aligned}$$ proving the recursion . Theorem \[YZA\] then follows from Lemma \[G2lattice\].
Since $g=2$ is the minimal genus for curve counting on abelian surfaces, Theorem \[YZA\] may be viewed as the analogue of the Yau-Zaslow conjecture [@YZ] for $g=0$ counting on $K3$ surfaces. In the $K3$ case, primitive classes were handled first in [@Beau; @BLA2]. To treat the imprimitive classes, a completely new approach [@KMPS] was required (and came a decade later). For abelian surfaces, the lattice counting for the primitive case is much easier than the complete result of Theorem \[YZA\]. The perfect matching of the lattice counts in all cases with formula of Theorem \[YZA\] appears miraculous.
Multiple cover rule {#mcr}
-------------------
A multiple cover formula in $g=2$ can be extracted from Theorem \[YZA\]. The result follows the structure of the complete multiple cover formula for $K3$ surfaces [@PT2]. We state here the multiple cover conjecture for the invariants $\mathsf{N}^{\text{Q}}_{g,(d_1,d_2)}$ for all $g$.
For $d_1, d_2 >0$, define the generating series of the [*quotient*]{} invariants: $$f_{(d_1, d_2)}(u) = \sum_{g \geq 2} \mathsf{N}^{\text{Q}}_{g,(d_1, d_2)} u^{2g-2} \,.$$ The quotient invariants are defined in terms of the FLS invariants in .
[**Conjecture**]{} ${\mathbf{A'}.}$ [*For all $d_1, d_2 > 0$,*]{} $$f_{(d_1, d_2)}(u) = \sum_{k | \gcd(d_1, d_2)} k f_{\left( 1, \frac{d_1 d_2}{k^2} \right)}(k u) \,.$$
Theorem \[YZA\] implies the $g=2$ case of Conjecture $\mathrm{A'}$. By an elementary check, Conjecture \[conjA\] is [*equivalent*]{} to Conjecture $\mathrm{A'}$ plus the $k=0$ case of Theorem \[thm\_point\_insertion\]. Since Theorem \[thm\_point\_insertion\] is proven in Section \[primc\], Conjectures \[conjA\] and $\mathrm{A'}$ are equivalent.
Primitive classes {#primc}
=================
Overview
--------
Let $A$ be an abelian surface, let $g\geq 2$ be the genus, and let $\beta \in H_2(A,{{\mathbb{Z}}})$ be a curve class of type $(d_1, d_2)$ with $d_1, d_2 > 0$. The class $\beta$ is primitive if $\gcd(d_1,d_2)=1$.
The proof of Theorem \[thm\_point\_insertion\] is presented here. We proceed in two steps. First, we relate the FLS invariants to the reduced Gromov-Witten invariants of $A$ with pure point insertions and $\lambda$ classes in Sections \[prim2\] and \[prim3\]. Next, we degenerate an elliptically fibered abelian surface $A$. Using the degeneration formula in Sections \[prim4\] and \[P1Eevaluation\], we reduce the calculation to an evaluation on ${\mathbb{P}}^1 \times E$. The proof of Theorem \[thm\_point\_insertion\] is completed in Section \[prim6\].
We conclude with an application in Section \[prim7\]: a new proof is presented of a result by Göttsche and Shende [@GS] concerning the Euler characteristics of the moduli spaces of stable pairs on abelian surfaces in irreducible classes.
Notation {#prim2}
--------
Let $\alpha \in H^{\ast}({{\overline M}}_{g.n},{{\mathbb{Q}}})$ be a cohomology class on the moduli space of stable curves ${{\overline M}}_{g,n}$, and let $$\gamma_1, \dots, \gamma_n \in H^{\ast}(A, {{\mathbb{Q}}})$$ be cohomology classes on $A$. The classes $\alpha$ and $\gamma_i$ can be pulled back to the moduli spaces $${{\overline M}}_{g,n}(A,\beta)^{\text{FLS}} \quad \text{ and } \quad {{\overline M}}_{g,n}(A,\beta)$$ via the forgetful map $\pi$ and the evaluation maps ${\mathop{\rm ev}\nolimits}_1, \dots, {\mathop{\rm ev}\nolimits}_n$.
For each marking $i\in \{1,\dots,n\}$, let $L_i$ be the associated cotangent line bundle on ${{\overline M}}_{g,n}(A,\beta)$. Let $$\psi_i = c_1(L_i) \in H^2({{\overline M}}_{g,n}(A,\beta),{{\mathbb{Q}}})$$ be the first Chern class. Since $${{\overline M}}_{g,n}(A,\beta)^{\text{FLS}}\subset {{\overline M}}_{g,n}(A,\beta)\, ,$$ the classes $\psi_i$ restrict to ${{\overline M}}_{g,n}(A,\beta)^{\text{FLS}}$. The reduced Gromov-Witten invariants of $A$ are defined by: $$\Big\langle \alpha \,;\, \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,\beta}^{A, \text{red}} =
\int_{ [ {{\overline M}}_{g,n}(A,\beta) ]^{\text{red}} } \pi^{\ast}(\alpha)\, \prod_{i=1}^n {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i)\, \psi_i^{a_i}\ .
$$ The FLS invariants of $A$ are defined by: $$\Big\langle \alpha \,;\, \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n)
\Big\rangle_{g,\beta}^{A, \text{FLS}} =
\int_{ [ {{\overline M}}_{g,n}(A,\beta)^{\text{FLS}} ]^{\text{red}} } \pi^{\ast}(\alpha) \, \prod_{i=1}^n {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \, \psi_i^{a_i} \,.
$$
The FLS invariants can be expressed in terms of the usual invariants by a result[^17] of Bryan and Leung [@BLA] as follows. Let $$\xi_1,\, \xi_2,\, \xi_3,\, \xi_4 \in H_1(A,{{\mathbb{Z}}})$$ be a basis for which the corresponding dual classes $$\widehat{\xi}_1,\, \widehat{\xi}_2, \, \widehat{\xi}_3,\, \widehat{\xi}_4 \in H^1(\widehat{A},{{\mathbb{Z}}}) \cong H_1(A, {{\mathbb{Z}}})$$ satisfy the normalization $$\int_{\widehat{A}} \widehat{\xi}_1 \cup \widehat{\xi}_2 \cup \widehat{\xi}_3
\cup \widehat{\xi}_4 = 1\, .$$ By first trading the higher descendents $\tau_k( \gamma_i)$ for classes pulled back from ${{\overline M}}_{g,n}$ up to boundary terms, we may reduce to non-gravitational insertions $\tau_0(\gamma_i)$. Then $$\Big\langle \alpha \,;\, \prod_{i=1}^{n} \tau_{0}(\gamma_i) \Big\rangle_{g,\beta}^{A, \text{FLS}}
= \Big\langle \alpha \,;\, \prod_{i=1}^{4} \tau_0(\xi_i) \cdot \prod_{i=1}^{n} \tau_{0}(\gamma_i) \Big\rangle_{g,\beta}^{A, \text{red}} \,,
\label{mmbbbnv}$$ where $\alpha$ on the right side of is viewed to be a cohomology class on ${{\overline M}}_{g,n+4}$ via pull-back along the map which forgets the four new points.
Odd and even classes {#prim3}
--------------------
### Trading FLS for insertions
We prove here the following simple rule which trades the FLS condition for insertions in the reduced Gromov-Witten theory of $A$.
As in Section \[ptptpt\], let ${{\mathsf{p}}}\in H^4(A,{{\mathbb{Z}}})$ be the class of a point.
\[ev\_odd\_prop\] For $g \geq 2$ and $d_1, d_2 > 0$, we have $$\Big \langle \alpha \,;\, \tau_0({{\mathsf{p}}})^k \Big \rangle^{A, \textup{FLS}}_{g, (d_1, d_2)} =
\frac{d_1d_2}{(k+1)(k+2) } \Big \langle \alpha \,;\, \tau_0({{\mathsf{p}}})^{k+2} \Big \rangle^{A, \textup{red}}_{g,(d_1, d_2)}$$ for all $\alpha \in H^*(\overline{M}_{g},{{\mathbb{Q}}})$ and $k\geq 0$.
The proof uses the action of $A$ on the moduli space ${{\overline M}}_{g,n}(A,\beta)$ to produce relation among various Gromov-Witten invariants. The argument is a modification of an elliptic vanishing argument introduced in [@OP3].
### Abelian vanishing {#ababab}
Let $\beta \in H_2(A, {{\mathbb{Z}}})$ be any curve class. For $n \geq 1$, let $${\mathop{\rm ev}\nolimits}_1 : {{\overline M}}_{g,n}(A, \beta) {{\ \rightarrow\ }}A$$ be the first evaluation map. Denote the fiber of the identity element $0_A \in A$ by $${{\overline M}}_{g,n}^{0}(A, \beta) = {\mathop{\rm ev}\nolimits}_1^{-1}(0_A) \,.$$ We have the product decomposition $$\label{333}
{{\overline M}}_{g,n}(A, \beta) = {{\overline M}}_{g,n}^{0}(A, \beta) \times A \,.$$ The reduced virtual class on ${{\overline M}}_{g,n}(A, \beta)$ is pulled back from a class on ${{\overline M}}_{g,n}^0(A, \beta)$. Consider the commutative diagram $$\xymatrix{
{{\overline M}}_{g,n}(A, \beta) \ar[d]^{\text{pr}} \ar[r]^-{{\mathop{\rm ev}\nolimits}} & A^n \ar[d]^{p} \\
{{\overline M}}^{0}_{g,n}(A, \beta) \ar[r] & A^{n-1}
}$$ where $\text{pr}$ is projection onto the first factor of and $$p( x_1, \dots, x_n ) = (x_2 - x_1, \dots, x_n - x_1) \,.$$
\[abelvan\] Let $\alpha \in H^{\ast}({{\overline M}}_{g,n},{{\mathbb{Q}}})$ and $\gamma\in H^{\ast}(A^{n-1},{{\mathbb{Q}}})$ be arbitrary classes. For any $\gamma_1 \in H^{\ast}(A, {{\mathbb{Q}}})$ of degree $\deg(\gamma_1) \leq 3$, $$\int_{ [ {{\overline M}}_{g,n}(A, \beta) ]^{\textup{red}} } \pi^{\ast}(\alpha)
\cup {\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1) \cup {\mathop{\rm ev}\nolimits}^{\ast} p^{\ast}(\gamma) = 0 \,.$$
The class $$\Big( \pi^{\ast}(\alpha) \cup {\mathop{\rm ev}\nolimits}^{\ast}p^{\ast}(\gamma) \Big) \cap [ {{\overline M}}_{g,n}(A, \beta) ]^{\text{red}}$$ is the pull-back via $\text{pr}$ of a class $\theta$ on ${{\overline M}}_{g,n}^{0}(A, \beta)$. Using the push-pull formula, we have $$\text{pr}_{\ast}\Big( \text{ev}_1^{\ast}(\gamma_1) \cap \text{pr}^{\ast}(\theta) \Big) = \text{pr}_{\ast} {\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1) \cap \theta = 0 \,.$$ The last equality holds for dimension reasons since $\gamma_1\in H^{\leq 3}(A,{{\mathbb{Q}}})$ and the fibers of $\text{pr}$ are $A$.
### Proof of Proposition \[ev\_odd\_prop\]
We study the split abelian surface $$A = E_1 \times E_2,$$ where $E_1$ and $E_2$ are two generic elliptic curves. Consider the curve class $$(d_1, d_2) = d_1 [E_1] + d_2 [E_2] \in H_2(A, {{\mathbb{Z}}}) \,.$$
For $i \in \{1,2\}$, let ${{\omega}}_i \in H^2(E_i, {{\mathbb{Z}}})$ be the class of a point on $E_i$, and let $${\mathsf{a}}_i, {\mathsf{b}}_i \in H^{1}(E_i, {{\mathbb{Z}}})$$ be a symplectic basis. We use freely the identification induced by the Künneth decomposition $$H^{\ast}(E_1 \times E_2, {{\mathbb{Z}}}) = H^{\ast}(E_1, {{\mathbb{Z}}}) \otimes H^{\ast}(E_2, {{\mathbb{Z}}}) \,.$$
The proof of Proposition \[ev\_odd\_prop\] follows directly from and the following two Lemmas.
\[Lemma\_even\_odd\_1\] For $\alpha\in H^*(\overline{M}_g,{{\mathbb{Q}}})$, we have $$\begin{gathered}
d_2 \,\Big\langle \alpha \,;\, \tau_0({{\mathsf{p}}})^{k+2} \Big\rangle^{A, \textup{red}}_{g, (d_1, d_2)} \\ =
(k+2) \cdot \Big\langle \alpha \,;\, \tau_0( {\mathsf{a}}_1 \omega_2) \tau_0( {\mathsf{b}}_1 \omega_2 )
\tau_0( {{\mathsf{p}}})^{k+1} \Big\rangle^{A, \textup{red}}_{g, (d_1, d_2)} \,.\end{gathered}$$
Consider the class $$\gamma = {\mathsf{b}}_1 \omega_2 \otimes {{\mathsf{p}}}^{\otimes k+1 } \in H^{\ast}( A^{k+2},{{\mathbb{Q}}}) \, .$$ In the notation of Section \[ababab\] for $n=k+3$, we have $$\label{jjjj}
p^{\ast} \big( \mathsf{1}^{\otimes i-1} \otimes v \otimes \mathsf{1}^{\otimes k+2-i} \big) = - v \otimes \mathsf{1}^{\otimes k+2} + \mathsf{1}^{\otimes i} \otimes v \otimes \mathsf{1}^{\otimes k+2-i}$$ for all $i=1,\dots, k+2$ and all $v \in \{ {\mathsf{a}}_1, {\mathsf{b}}_1, {\mathsf{a}}_2, {\mathsf{b}}_2\}$.
Denote the projection onto the first factor of $A^{k+3}$ by $$\pi_1: A^{k+3} \rightarrow A\, .$$ Let $u={\mathsf{a}}_1 \omega_2$. Via several applications of , we find $$\begin{aligned}
\pi_1^{\ast}(u) \cup p^{\ast}(\gamma) & = {\mathsf{a}}_1 \omega_2 \otimes {\mathsf{b}}_1 \omega_2 \otimes {{\mathsf{p}}}^{\otimes k+1} - {{\mathsf{p}}}\otimes \omega_2 \otimes {{\mathsf{p}}}^{\otimes k+1} \\
& \qquad - \sum_{i=1}^{k+1} {{\mathsf{p}}}\otimes {\mathsf{b}}_1 \omega_2 \otimes {{\mathsf{p}}}^{\otimes i-1} \otimes {\mathsf{a}}_1 \omega_2 \otimes {{\mathsf{p}}}^{\otimes k+1-i} \,.\end{aligned}$$ After applying the abelian vanishing of Lemma \[abelvan\], we obtain $$\begin{gathered}
\Big\langle \alpha \,;\, \tau_0(\omega_2) \tau_0({{\mathsf{p}}})^{k+2} \Big\rangle^{A, \text{red}}_{g, (d_1, d_2)}
\\ = (k+2) \, \Big\langle \alpha \,;\, \tau_0( {\mathsf{a}}_1 \omega_2) \tau_0( {\mathsf{b}}_1 \omega_2 ) \tau_0( {{\mathsf{p}}})^{k+1} \Big\rangle^{A, \text{red}}_{g, (d_1, d_2)} \,.\end{gathered}$$ The Lemma then follows from the divisor equation.
\[Lemma\_even\_odd\_2\] For $\alpha\in H^*(\overline{M}_g,{{\mathbb{Q}}})$, we have $$\begin{gathered}
d_1 \, \Big\langle \alpha \,;\, \tau_0( {\mathsf{a}}_1 \omega_2) \tau_0( {\mathsf{b}}_1 \omega_2 ) \tau_0( {{\mathsf{p}}})^{k+1} \Big\rangle^{A, \textup{red}}_{g, (d_1, d_2)} \\
= (k+1) \Big\langle \alpha \,;\,
\tau_0( {\mathsf{a}}_1 \omega_2) \tau_0( {\mathsf{b}}_1 \omega_2 ) \tau_0( \omega_1 {\mathsf{a}}_2 ) \tau_0( \omega_1 {\mathsf{b}}_2 )
\tau_0( {{\mathsf{p}}})^{k} \Big\rangle^{A, \textup{red}}_{g, (d_1, d_2)} \,.\end{gathered}$$
Consider the class $$\gamma = \omega_1 {\mathsf{b}}_2 \otimes {\mathsf{a}}_1 \omega_2 \otimes {\mathsf{b}}_1 \omega_2 \otimes {{\mathsf{p}}}^{\otimes k} \,$$ and let $u = \omega_1 {\mathsf{a}}_2$. We apply the abelian vanishing just as in the proof of Lemma \[Lemma\_even\_odd\_1\]. Every term with an insertion of the form $\tau_0( v w )$ for $v \in \{ {\mathsf{a}}_1, {\mathsf{b}}_1 \}$ and $w \in \{ {\mathsf{a}}_2, {\mathsf{b}}_2 \}$ contributes zero by the divisor equation. We obtain $$\begin{gathered}
\Big\langle \alpha \,;\, \tau_0(\omega_1) \tau_0({\mathsf{a}}_1 \omega_2 ) \tau_0({\mathsf{b}}_1 \omega_2) \tau_0({{\mathsf{p}}})^{k+1} \Big\rangle^{A, \text{red}}_{g, (d_1, d_2)} \\
= (k+1)
\Big\langle \alpha \,;\, \tau_0( {\mathsf{a}}_1 \omega_2) \tau_0( {\mathsf{b}}_1 \omega_2 ) \tau_0( \omega_1 {\mathsf{a}}_2 ) \tau_0( \omega_1 {\mathsf{b}}_2 ) \tau_0({{\mathsf{p}}})^k \Big\rangle^{A, \text{red}}_{g, (d_1, d_2)} \,,\end{gathered}$$ which yields the result by an application of the divisor equation.
Degeneration formula {#prim4}
--------------------
Let the abelian surface $A$ $$A = E_1 \times E_2$$ be the product of two generic elliptic curves $E_1$ and $E_2$, and let $$E = 0_{E_1}\times E_2$$ be a fixed fiber of the projection to $E_1$. The degeneration of $A$ to the normal cone of $E\subset A$ is the family $$\epsilon : X = {\rm Bl}_{E \times 0} ( A\times {\mathbb{P}}^1 ) {{\ \rightarrow\ }}{\mathbb{P}}^1 \,. \label{normal_cone}$$ For $\xi \in {\mathbb{P}}^1\setminus \{ 0\}$, the fiber $X_\xi = \epsilon^{-1}(\xi)$ is isomorphic to $A$. For $\xi=0$, we have $$X_{0} = A \, \cup_{E} \, ({\mathbb{P}}^1 \times E) \, .$$ We will apply the degeneration formula of Gromov-Witten theory [@Junli1; @Junli2] to the family .
For our use, the degeneration formula must be modified for the reduced virtual class. More precisely, the degeneration formula expresses the reduced Gromov-Witten theory of $A$ in terms of the reduced relative Gromov-Witten theory of $A/E$ and standard relative Gromov-Witten theory of $({\mathbb{P}}^1 \times E)/E$. The technical point in the modification of the degeneration formula is to define a reduced virtual class on the moduli space $${{\overline M}}_{g,n}(\epsilon, \beta)$$ of stable maps to the fibers of $\epsilon$. Note that every fiber $X_\xi$ and every expanded degeneration $\widehat{X}_\xi$ maps to the abelian surface $A$. The pull-back of the symplectic form of $A$ to $\widehat{X}_\xi$ then yields a 2-form on $\widehat{X}_\xi$ which vanishes on all components except $A$. With the usual arguments [@MP; @MPT], we obtain a quotient of the obstruction sheaf which only changes the obstruction sheaf on the $A$ side. The outcome is the desired degeneration formula. A parallel argument can be found in [@MPT Section 6].
The surface ${\mathbb{P}}^1 \times E$ {#P1Eevaluation}
-------------------------------------
For the trivial elliptic fibrations $$p : A {{\ \rightarrow\ }}E_1 \quad \text{ and } \quad \widehat{p} : {\mathbb{P}}^1\times E \rightarrow {\mathbb{P}}^1\, ,$$ we denote the section class by $B$ and the fiber class by $E$. We also write $$(d_1, d_2) = d_1 B + d_2 E$$ for the corresponding classes in $H_2(A,{{\mathbb{Z}}})$ and $H_2({\mathbb{P}}^1 \times E,{{\mathbb{Z}}})$.
We will use the standard bracket notation $$\begin{gathered}
\Big\langle \ \mu\ \Big| \ \alpha\, \prod_{i} \tau_{a_i}(\gamma_i) \ \Big| \ \nu \ \Big\rangle^{{\mathbb{P}}^1 \times E}_{g, (1,d)} \\
= \int_{ [\overline{M}_{g,n}( ({\mathbb{P}}^1\times E)/\{0,\infty\}, (1,d))_{\mu, \nu}]^{\text{vir}} } \alpha\, \cup\, \prod_{i} \psi_i^{a_i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i)\end{gathered}$$ for the Gromov-Witten invariants of ${\mathbb{P}}^1 \times E$ relative to the fibers over $0,\infty\in {\mathbb{P}}^1$. The integral is over the moduli space of stable maps $$\overline{M}_{g,n}( ({\mathbb{P}}^1 \times E)/\{0,\infty\},\, (1,d))$$ relative to the fibers over $0,\infty \in {\mathbb{P}}^1$ in class $(1,d)$. Here, $$\mu \in H^{\ast}(0\times E) \quad \text{ and } \quad \nu \in H^{\ast}
(\infty \times E)$$ are cohomology classes on the relative divisors. The integrand contains $\alpha \in H^*(\overline{M}_{g,n},{{\mathbb{Q}}})$ and the descendents.
We form the generating series of relative invariants $$\begin{gathered}
\Big\langle \ \mu \ \Big| \ \alpha\, \prod_{i} \tau_{a_i}(\gamma_i) \ \Big| \ \nu \ \Big\rangle^{{\mathbb{P}}^1 \times E} \\
= \sum_{g \geq 0} \sum_{d \geq 0} u^{2g-2}q^d
\Big\langle \ \mu \ \Big| \ \alpha\, \prod_{i} \tau_{a_i}(\gamma_i) \ \Big| \ \nu \ \Big\rangle^{{\mathbb{P}}^1 \times E}_{g, (1,d)} \,.\end{gathered}$$ Similar definitions apply also to the case of a single relative divisor and/or the case of the abelian surface $A$ (with respect to the reduced virtual class).
We will require several exact evaluations. Let $${\mathbb{E}}^\vee(1) = c({{\mathbb{E}}}^{\vee})$$ denote the total Chern class of the dual of the Hodge bundle, and let $\omega$ be the class of a point on the relative divisors of $A$ and $E\times {\mathbb{P}}^1$.
\[P1Eeval344\] We have $$\begin{aligned}
&\Big\langle \ \mathsf{1}\ \Big|\ {\mathbb{E}}^{\vee}(1)\, \tau_0(\mathsf{p})\ \Big| \ \mathsf{1}\ \Big \rangle^{ {\mathbb{P}}^1 \times E} = \frac{1}{u^2}\, , \\
&\Big\langle \ \mathsf{1}\ \Big|\ {\mathbb{E}}^{\vee}(1)\, \tau_0(\mathsf{p})\ \Big| \ \omega\ \Big \rangle^{ {\mathbb{P}}^1 \times E} =
\frac{1}{u^2} \sum_{d \geq 1} \sum_{m | d} \frac{d}{m} \Big( 2 \sin(mu/2) \Big)^2 q^d\, , \\
&\Big\langle \ \omega \ \Big| \ {\mathbb{E}}^\vee(1)\, \tau_0({{\mathsf{p}}}) \ \Big| \ \omega \ \Big
\rangle^{{\mathbb{P}}^1 \times E} = 0\, .\end{aligned}$$
The first equation is obtained by exactly following the proof of [@MPT Lemma 24]. The second equation follows from [@MPT Lemmas 25 and 26]. For the third equality, the point conditions on the relative divisors can be chosen to be different. Then, since the degree over ${\mathbb{P}}^1$ is 1 (and there are no nonconstant maps from ${\mathbb{P}}^1$ to $E$), the moduli space with the relative conditions imposed is empty.
\[Aeval344\] For $g \geq 0$ and $d \geq 0$, $$\Big\langle\ {\mathbb{E}}^{\vee}(1)\ \Big|\ {{\omega}}\ \Big\rangle^{A, \textup{red}}_{g,(1,d)} = \delta_{g,1} \delta_{d,0} \, .$$
By dimension reasons only the term $$(-1)^{g-1} \lambda_{g-1} = c_{g-1}({{\mathbb{E}}}^{\vee})$$ contributes in the evaluation of Lemma \[Aeval344\].
*Case $d > 0$.* We will prove the vanishing of $\lambda_{g-1}$ on ${{\overline M}}_{g}(A / E, \beta)$ by giving two linearly independent sections of ${{\mathbb{E}}}$.
Let $\gamma, \gamma' \in H^0(A, \Omega_A)$ be the pull-backs to $A$ of non-zero global differential forms on $E_1$ and $E_2$ respectively. Let $$\pi : {{\mathcal C}}{{\ \rightarrow\ }}{{\overline M}}_{g}(A / E, \beta)$$ be the universal curve and $f : {{\mathcal C}}{{\ \rightarrow\ }}A$ be the universal map. We have the induced sequence $${{\mathcal O}}_{{\mathcal C}}^{2} \xrightarrow{(\gamma, \gamma')} f^{\ast} \Omega_A {{\ \rightarrow\ }}\Omega_\pi {{\ \rightarrow\ }}\omega_\pi \,.$$ By pushforward via $\pi$ we obtain the sequence $$s: {{\mathcal O}}_{{{\overline M}}_{g}(A / E, \beta)}^2 {{\ \rightarrow\ }}\pi_{\ast} \omega_{\pi} = {{\mathbb{E}}}\,.$$
If $s$ does not define a 2-dimensional subbundle of ${{\mathbb{E}}}$, there exists a map $f_0 : C {{\ \rightarrow\ }}A$ in class $(1,d)$ and nonzero elements $c, c' \in {{\mathbb{C}}}$ such that $$f_0^{\ast}( c \gamma + c' \gamma' ) = 0 \,.$$ Thus $f_0$ must map to a (translate of a) 1-dimensional abelian subvariety $V$ inside $E_1 \times E_2$. Because $f$ has degree $(1,d)$, the subvariety $V$ induces an isogeny between $E_1$ and $E_2$. But $E_1$ and $E_2$ were choosen generic, which is a contradiction. Hence $s$ is injective and $\lambda_{g-1} = 0$.
*Case $d = 0$.* Then, we have the factor $R^1 \pi_{\ast} (f^{\ast} T_{E_2})$ inside the obstruction sheaf, which yields an additional class $(-1)^{g - 1}\lambda_{g-1}$ after reduction. Therefore, $$\Big\langle\ {\mathbb{E}}^{\vee}(1)\ \Big|\ {{\omega}}\ \Big\rangle^{A, \text{red}}_{g,(1,0)} = \int_{[ {{\overline M}}_g(E_1/0,1) ]^{\text{vir}}} \lambda_{g-1}^2 \,.$$ For $g \geq 2$, we have $\lambda_{g-1}^2 = 2 \lambda_g \lambda_{g-2}$ by Mumford’s relation. By pulling-back the global non-zero $1$-form from $E_1$, we obtain a 1-dimensional subbundle of ${{\mathbb{E}}}$. Therefore, $\lambda_g = 0$ and the integral vanishes.
Finally, for $g = 1$ and $d = 0$ the moduli space ${{\overline M}}_1(E_1/0,1)$ is a single point and the invariant is $1$.
Proof of Theorem \[thm\_point\_insertion\] {#prim6}
------------------------------------------
We are now able to evaluate the invariants $\mathsf{N}^{\text{FLS}}_{g,k,(1,d)}$ of Section \[ptptpt\] and prove Theorem \[thm\_point\_insertion\]. By definition, $$\label{vv45}
\sum_{g\geq 2} \sum_{d\geq 1} \mathsf{N}^{\text{FLS}}_{g,k,(1,d)} u^{2g-2} q^d=
\sum_{g \geq 2} \sum_{d \geq 1} u^{2g-2} q^d \Big\langle {{\mathbb{E}}}^{\vee}(1) \, \tau_0({{\mathsf{p}}})^k \Big\rangle_{g, (1,d)}^{A,\text{FLS}} \, .$$ By Proposition \[ev\_odd\_prop\], the right side of equals $$q\frac{d}{dq}\ \frac{1}{(k+1)(k+2)} \sum_{g \geq 0} \sum_{d \geq 0} u^{2g-2} q^d \Big \langle {{\mathbb{E}}}^{\vee}(1) \, \tau_0({{\mathsf{p}}})^{k+2} \Big\rangle_{g, (1,d)}^{A, \textup{red}} \, .$$ Next, we apply the degeneration formula. Only one term satisfies the dimension constraints: $$q\frac{d}{dq}\ \frac{u^2}{(k+1)(k+2)} \cdot \Big \langle\ {{\mathbb{E}}}^{\vee}(1)\ \Big|\ \omega\ \Big\rangle^{A, \text{red}}
\cdot \Big\langle\ \mathsf{1}\ \Big|\ {{\mathbb{E}}}^{\vee}(1) \, \tau_0({{\mathsf{p}}})^{k+2} \ \Big\rangle^{{\mathbb{P}}^1
\times E} \, .$$ An application of Lemma \[Aeval344\] then yields $$q\frac{d}{dq}\ \frac{u^2}{(k+1) (k+2)} \cdot \Big\langle\ \mathsf{1}\ \Big|\ {{\mathbb{E}}}^{\vee}(1) \, \tau_0({{\mathsf{p}}})^{k+2} \ \Big\rangle^{{\mathbb{P}}^1\times E}\, .$$ We degenerate the base ${\mathbb{P}}^1$ to obtain a chain of $k+3$ surfaces isomorphic to ${\mathbb{P}}^1 \times E$. The first $k+2$ of these each receive a single insertion $\tau_0({{\mathsf{p}}})$. Using the vanishing of Lemma \[P1Eeval344\] and the evaluation of the last ${\mathbb{P}}^1\times E$ by [@MPT Lemma 24], we obtain $$\begin{gathered}
q\frac{d}{dq}\ \frac{u^2 (k+2)}{(k+1) (k+2)} \cdot
\Big\langle \ \mathsf{1} \ \Big|\ {\mathbb{E}}^{\vee}(1)\, \tau_0(\mathsf{p})\ \Big| \ \mathsf{1}\ \Big\rangle^{ {\mathbb{P}}^1 \times E} \\
\cdot \bigg( u^2 \Big\langle \ \mathsf{1}\ \Big|\ {\mathbb{E}}^{\vee}(1)\, \tau_0(\mathsf{p})\ \Big| \ \omega\ \Big \rangle^{ {\mathbb{P}}^1 \times E} \bigg)^{k+1}\, . \end{gathered}$$ A further application of Lemma \[P1Eeval344\] yields $$q\frac{d}{dq}\ \frac{1}{k+1} \Big( \sum_{d \geq 1} \sum_{m | d} \frac{d}{m} \Big( 2 \sin(mu/2) \Big)^2 q^d \Big)^{k+1} \, .$$ Rewriting the result in the variables $$p=e^{2\pi i z}\, , \ \ \ q= e^{2\pi i \tau}$$ and $u=2\pi z$, we obtain $$q\frac{d}{dq}\, \frac{1}{k+1} \Big( {- \sum_{d \geq 1}} \sum_{m | d} \frac{d}{m} (p^m - 2 + p^{-m}) q^d \Big)^{k+1}
= q\frac{d}{dq}\, \frac{1}{k+1} \mathsf{S}(z,\tau)^{k+1} \, .$$ The proof of Theorem \[thm\_point\_insertion\] is thus complete.
Relation to stable pairs invariants {#prim7}
-----------------------------------
Let $A$ be an abelian surface and let $\beta$ be an *irreducible* curve class of type $(1,d)$. Let $$P_n(A, \beta)$$ be the moduli space of stable pairs $(F,s)$ on $A$ in class $\beta$ and with Euler characteristic $\chi(F) = n$, see [@PT1]. The moduli spaces $P_n(A, \beta)$ are isomorphic to the relative Hilbert scheme over the universal family of curves in class $\beta$. It is nonsingular of dimension $2d + n +1$.
Consider the Hilbert-Chow map $$p : P_n(A, \beta) {{\ \rightarrow\ }}{\mathop{\rm Pic}\nolimits}^{\beta}(A) \cong \widehat{A} \,.$$ with sends a stable pair $(F,s)$ to the divisor class associated to the support of $F$. The map is equivariant with respect to the action of $A$ and is an isotrivial étale fibration. The fiber of $p$ over $0_{\widehat{A}}$ is denoted by $$P_n(A, \beta)^{\text{FLS}} = p^{-1}(0_{\widehat{A}}),$$ We define the FLS stable pairs invariants in class $\beta$ by the signed Euler characteristic $$\mathsf{P}_{n,\beta}^{\text{FLS}}
= (-1)^{2d + n - 1} e \big( P_n(A, \beta)^{\text{FLS}} \big) \,. \label{stable_pair_xxkt}$$ The definition agrees with the definition of residue stable pairs invariants of the threefold $A \times {{\mathbb{C}}}$ using torus localization, see [@MPT].
By deformation invariance of the Euler characteristic under deformations with smooth fibers, $\mathsf{P}_{n,\beta}^{\text{FLS}}$ only depends on the type $(1,d)$ of the irreducible class $\beta$. We write $$\mathsf{P}_{n,\beta}^{\text{FLS}} = \mathsf{P}_{n,(1,d)}^{\text{FLS}} \,.$$ The Euler characteristics (in fact the $\chi_y$-genus) have been computed by Göttsche and Shende in [@GS]:
\[GS\_thm\] We have $$\sum_{d \geq 1} \sum_{n \in {{\mathbb{Z}}}} \mathsf{P}_{n,(1,d)}^{\textup{FLS}} (-p)^n q^d
=
- \sum_{d \geq 1} \sum_{m | d} \frac{d^2}{m} ( p^{m} - 2 + p^{-m} ) q^d \,.
\label{GS_gen}$$
The $k=0$ case of Theorem \[thm\_point\_insertion\] yields a second proof of this result: using an analog of the abelian vanishing relation (Lemma \[abelvan\]) for stable pairs, we may express the FLS condition by point insertions on the full moduli space $P_n(A,\beta)$ as in Proposition \[ev\_odd\_prop\]. After degenerating the abelian surface $A$ to ${\mathbb{P}}^1 \times E$, we can apply the Gromov-Witten/Pairs correspondence [@PaPix2], which yields the result.
In [@MPT], a parallel study of the reduced invariants of $K3$ surfaces was undertaken. The Gromov-Witten/Pairs correspondence and the Euler characteristic calculations of Kawai-Yoshioka [@KY] were together used to evaluate the Gromov-Witten side to prove the primitive Katz-Klemm-Vafa conjecture. The analogue of the Kawai-Yoshioka calculation for abelian surfaces is Theorem \[GS\_thm\]. However, for abelian surfaces we are able to evaluate the Gromov-Witten side directly without using input from the stable pairs side.
Quasi-modular forms {#section_modularity}
===================
Descendent series {#secds}
-----------------
Let $A = E_1 \times E_2$ be the product of two generic elliptic curves $E_1$ and $E_2$, and let $$(d_1, d_2) = d_1 [E_1] + d_2 [E_2] \in H_2(A, {{\mathbb{Z}}})\, .$$ For $i \in\{ 1,2\}$, let ${{\omega}}_i \in H^2(E_i, {{\mathbb{Z}}})$ be the class of a point on $E_i$ and let $${\mathsf{a}}_i, {\mathsf{b}}_i \in H^{1}(E_i, {{\mathbb{Z}}})$$ be a symplectic basis. As before, we use freely the identification induced by the Künneth decomposition $$H^{\ast}(E_1 \times E_2, {{\mathbb{Z}}}) = H^{\ast}(E_1, {{\mathbb{Z}}}) \otimes H^{\ast}(E_2, {{\mathbb{Z}}}) \,.$$ A class $\gamma \in H^{\ast}(A, {{\mathbb{Q}}})$ is [*monomial*]{} if $\gamma$ is a product $$\gamma = {\mathsf{a}}_1^{i} {\mathsf{b}}_1^{j} {\mathsf{a}}_2^{k} {\mathsf{b}}_2^{l}\, ,\ \ \ \ i,j,k, l \in \{ 0,1 \}\, .$$ A basis of the cohomology of $A$ is given by monomial classes. For a monomial class $\gamma$, we denote by $v_s(\gamma)$ the exponent of $s \in \{ {\mathsf{a}}_1, {\mathsf{b}}_1, {\mathsf{a}}_2, {\mathsf{b}}_2 \}$ in $\gamma$. Hence, $$\gamma = {\mathsf{a}}_1^{v_{{\mathsf{a}}_1}(\gamma)} {\mathsf{b}}_1^{v_{{\mathsf{b}}_1}(\gamma)} {\mathsf{a}}_2^{v_{{\mathsf{a}}_2}(\gamma)} {\mathsf{b}}_2^{v_{{\mathsf{b}}_2}(\gamma)} \,.$$
Let $\gamma_1, \dots, \gamma_n \in H^{\ast}(A, {{\mathbb{Q}}})$ be monomial classes, and let $$a_1, \dots, a_n \in {{\mathbb{Z}}}^{\geq 0}$$ be non-negative integers. The primitive descendent potential of $A$ with insertions $\tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n)$ is defined by $$\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )
= \sum_{d \geq 0} \Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,(1,d)}^{A, \text{red}} q^{d} \,.$$ Theorem \[modularity\] states a modularity[[^18]]{} property for $\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )$. The following refined result will be proven here.
\[modularity\_refined\] The primitive descendent potential satisfies the following properties for all $g\geq 1$:
1. $\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )$ vanishes unless $$\sum_{k=1}^{n} v_{{\mathsf{a}}_i}(\gamma_k) = \sum_{k=1}^{n} v_{{\mathsf{b}}_i}(\gamma_k) \,, \ \ \ \
i\in \{1,2\}\, .$$
2. $\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )
\in \text{\em QMod}_{2(g-2) + 2 \ell}$ for $\ell = \sum_{k=1}^{n} v_{{\mathsf{a}}_2}(\gamma_k)$.
Part (i) of Theorem \[modularity\_refined\] is a basic balancing condition. Part (ii) is a homogeneity property which refines the statement of Theorem \[modularity\].
Having the precise weight is useful in applications. For example, by part (ii) of Theorem \[modularity\_refined\], the series $\mathsf{F}_2^A( \tau_1( {{\mathsf{p}}}) )$ is a quasi-modular form of weight $2$, and hence a multiple of $E_2(\tau)$. The constant coefficient is given by $$\int_{[ {{\overline M}}_{2,1}(A, (1,0)) ]^{\text{red}} } {\mathop{\rm ev}\nolimits}_1^{\ast}({{\mathsf{p}}}) \psi_1
= \int_{ [ {{\overline M}}_{2,1}(E_1, 1) ]^{\text{vir}} } (- \lambda_1) {\mathop{\rm ev}\nolimits}_1^{\ast}(\omega_1) \psi_1
= - \frac{1}{12} \,.$$ We conclude $$\mathsf{F}_2^A( \tau_1( {{\mathsf{p}}}) ) = - \frac{1}{12} E_2(\tau) = - \frac{1}{12} + 2 \sum_{d \geq 0} \sigma(d) q^d \,.$$
For genus $g=1$, both parts of Theorem \[modularity\_refined\] are easy to see. The contributions of curve classes $(1,d>0)$ vanish for $g=1$ since the moduli space of maps is [*empty*]{}: an element of $$\overline{M}_{1,n}(E_1 \times E_2, (1,d>0))$$ would yield an isogeny between $E_1$ and $E_2$ contradicting the genericity of $E_1$ and $E_2$. The series $\mathsf{F}_{1}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) ) $ therefore has only a constant term in $q$. For curve classes of degree $(1,0)$, the moduli space factors as $${{\overline M}}_{1,n}(E_1\times E_2, (1,0)) \cong {{\overline M}}_{1,n}(E_1,1) \times E_2\ .$$ The balancing condition of part (i) then follows by the separate balancing on the two factors. For nonvanishing invariants $$\Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n)
\Big\rangle_{1,(1,0)}^{A,\text{red}}\, ,$$ we must have $$\ell = \sum_{k=1}^{n} v_{{\mathsf{a}}_2}(\gamma_k)=1.$$ Hence, part (ii) correctly predicts a quasi-modular form of weight $$2(g-2) + 2 \ell = 2 (1 - 2) + 2 \cdot 1 = 0 \,.$$
Tautological classes
--------------------
The first step in the proof of Theorem \[modularity\_refined\] is to recast the result in terms of tautological classes on the moduli space of curves. For $2g-2+n > 0$, let $$R^*({{\overline M}}_{g,n}) \subset H^{\ast}({{\overline M}}_{g,n},\mathbb{Q})$$ be the tautological ring. Let $$\pi : {{\overline M}}_{g,n}(A, \beta) {{\ \rightarrow\ }}{{\overline M}}_{g,n}$$ be the forgetful map. For $\alpha \in R^*({{\overline M}}_{g,n})$, we define $\alpha$-twisted reduced invariants of $A$ by $$\Big\langle \alpha \, ; \, \gamma_1, \dots, \gamma_n \Big\rangle_{g,\beta}^{A,\text{red}} =
\int_{[ {{\overline M}}_{g,n}(A,\beta) ]^{\text{red}} } \pi^{\ast}(\alpha) \cup \prod_{i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i)\, .$$ Here, $\beta$ is a curve class on $A$ and $$\gamma_1, \dots, \gamma_n \in H^{\ast}(A,\mathbb{Q})\,$$ are monomial classes. The associated primitive potential is defined by $$\mathsf{F}^{A}_{g}( \alpha ; \gamma_1, \dots, \gamma_n )
=
\sum_{d \geq 0} \Big\langle \alpha \, ; \, \gamma_1, \dots, \gamma_n
\Big\rangle_{g,(1,d)}^{A,\text{red}} q^d \,.$$
\[modularity\_proposition\] The primitive $\alpha$-twisted potential satisfies the following properties for all $g\geq 1$:
1. $\mathsf{F}^{A}_{g}( \alpha ; \gamma_1, \dots, \gamma_n )$ vanishes unless $$\sum_{k=1}^{n} v_{{\mathsf{a}}_i}(\gamma_k) = \sum_{k=1}^{n} v_{{\mathsf{b}}_i}(\gamma_k) \,, \ \ \ \
i\in \{1,2\}\, .$$
2. $\mathsf{F}^{A}_{g}( \alpha ; \gamma_1, \dots, \gamma_n ) \in \text{\em QMod}_{2(g-2) + 2 \ell}$ for $\ell = \sum_{k=1}^{n} v_{{\mathsf{a}}_2}(\gamma_k)$.
The cotangent line classes $$\psi_1, \dots, \psi_n\in H^2({{\overline M}}_{g,n}(A,\beta),\mathbb{Q})$$ can be expressed as pull-backs of the corresponding cotangent line classes from ${{\overline M}}_{g,n}$ up to boundary corrections.[[^19]]{} Integration over the boundary corrections is governed by the splitting formula for reduced invariants. The boundary corrections yield integrals of lower genus or fewer marked points. Arguing inductively, the descendent series $\mathsf{F}_{g}^{A}( \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) )$ can therefore be expressed in terms of the series $$\big\{ \mathsf{F}^{A}_{g'}( \alpha' ; \gamma'_1, \dots, \gamma'_{n'} )\big\}$$ for various $\alpha', g', \gamma'_1, \dots, \gamma'_{n'}$. By a simple verification, the splitting formula preserves the vanishing and modularity statements of Theorem \[modularity\_refined\]. Hence, Proposition \[modularity\_proposition\] implies Theorem \[modularity\_refined\].
The balancing condition of part (i) of Proposition \[modularity\_proposition\] follows easily from a Hodge theoretic argument. Alternatively, the balancing condition can be obtained inductively via the proof of part (ii) of Proposition \[modularity\_proposition\].
Proof of Proposition \[modularity\_proposition\]
-------------------------------------------------
By [@Tat], or by a direct Mumford-Tate group calculation in case $E_1$ and $E_2$ are generic, the subring of Hodge classes on $A^n = (E_1 \times E_2)^n$, $$\text{Hdg}^*(A^n) \subset H^*(A^n, {{\mathbb{Q}}})\, ,$$ is generated by divisor classes.[^20]
Hence, under the Künneth decomposition, the ring $\text{Hdg}^*(A^n)$ is generated by pull-backs of divisors classes in $$H^2(E_i, {{\mathbb{Q}}}) \,, \quad H^1(E_i, {{\mathbb{Q}}}) \otimes H^1(E_i, {{\mathbb{Q}}}) \,, \quad H^1(E_1, {{\mathbb{Q}}}) \otimes H^1(E_2, {{\mathbb{Q}}}) \,,$$ for $i = 1, 2$. We have $H^2(E_i, {{\mathbb{Q}}}) = \langle \omega_i \rangle = \langle {\mathsf{a}}_i{\mathsf{b}}_i \rangle$. By the genericity of $E_1$ and $E_2$, all divisor classes in $H^1(E_i, {{\mathbb{Q}}}) \otimes H^1(E_i, {{\mathbb{Q}}})$ are multiples of $${\mathsf{a}}_i \otimes {\mathsf{b}}_i - {\mathsf{b}}_i \otimes {\mathsf{a}}_i \,,$$ and there are no non-zero divisor classes in $H^1(E_1, {{\mathbb{Q}}}) \otimes H^1(E_2, {{\mathbb{Q}}})$.
It follows that all classes in $\text{Hdg}^*(A^n)$ are linear combinations of $$\delta_1 \otimes \cdots \otimes \delta_n$$ where $\delta_k \in H^*(A, {{\mathbb{Q}}})$ are monomial classes as defined in Section \[secds\], such that $$\sum_{k=1}^{n} v_{{\mathsf{a}}_i}(\delta_k) = \sum_{k=1}^{n} v_{{\mathsf{b}}_i}(\delta_k) \,, \ \ \ \
i\in \{1,2\}\, .$$
Consider the evaluation map $${\mathop{\rm ev}\nolimits}: {{\overline M}}_{g, n}(A, \beta) {{\ \rightarrow\ }}A^n \,.$$ By the push-pull formula, we have $$\Big\langle \alpha \, ; \, \gamma_1, \dots, \gamma_n
\Big\rangle_{g,(1,d)}^{A,\text{red}} = \int_{{\mathop{\rm ev}\nolimits}_*(\pi^*(\alpha) \cap [{{\overline M}}_{g, n}(A, \beta)]^{\text{red}})} \gamma_1 \otimes \cdots \otimes \gamma_n \,. \label{doubleb}$$ By the algebraicity of $\alpha$ and $[{{\overline M}}_{g, n}(A, \beta)]^{\text{red}}$, the integral is zero unless $$\pushQED{\qed} \sum_{k=1}^{n} v_{{\mathsf{a}}_i}(\gamma_k) = \sum_{k=1}^{n} v_{{\mathsf{b}}_i}(\gamma_k) \,, \ \ \ \
i\in \{1,2\}\, . \qedhere\popQED$$
Proof of Proposition \[modularity\_proposition\] (ii): Base cases
-----------------------------------------------------------------
We argue by induction on the pair $(g,n)$, where $g \geq 1$ is the genus of the source curve and $n \geq 0$ is the number of marked points. We order the pairs $(g,n)$ lexicographically: we define $(g', n') < (g,n)$ if and only if
- $g' < g$ or
- $g' = g$, $n' < n$.
[**[Base cases:]{}**]{} $g = 1$ and $n\geq 0$.
We have already observed that Theorem \[modularity\_refined\] holds in all $g=1$ cases. Proposition \[modularity\_proposition\] holds in $g=1$ by the same argument. We discuss the $n=0$ and $n=1$ cases as examples.
In case $(g,n) = (1,0)$, the only series is $$\mathsf{F}^{A}_{1}(\,\, ;\,) =0\, ,$$ since the reduced virtual dimension is 1 and there are no insertions.
In case $(g,n) = (1,1)$, the moduli space ${{\overline M}}_{1,1}(A, (1,d))$ has reduced virtual dimension 2. We must have either $\alpha \in R^1({{\overline M}}_{1,1})$ or $\tau({{\mathsf{p}}})$ as integrands. Such an $\alpha$ is a multiple of the class of a point on ${{\overline M}}_{1,1}$. Because a generic elliptic curve does not admit a non-vanishing map to $E_1 \times E_2$, the integral vanishes. In the second case, we evaluate $$\label{paap}
\mathsf{F}^A_1( \mathsf{1} ; {{\mathsf{p}}}) = \big\langle \tau_0({{\mathsf{p}}}) \big\rangle_{1, (1,0)}^{A, \text{red}} = 1 \,,$$ which is a quasi-modular form of weight 0.
Proof of Proposition \[modularity\_proposition\] (ii): Induction
----------------------------------------------------------------
Consider a pair $(g,n)$ satisfying $g\geq 2$ and assume Proposition \[modularity\_proposition\] is proven in [*all*]{} cases $(g',n')$ where $$(g',n') < (g,n)\, .$$
Let $\gamma_1, \dots, \gamma_n \in H^{\ast}(A,\mathbb{Q})$ be monomial classes, and let $\alpha \in R^*({{\overline M}}_{g,n})$ be a tautological class of pure degree. We must prove Proposition \[modularity\_proposition\] holds for the series $$\label{vppq}
\mathsf{F}_g^A( \alpha ; \gamma_1, \dots, \gamma_n )\, .$$
We may assume the dimension constraint $$2g + 2n = \deg(\alpha) + \sum_{i=1}^{n} \deg(\gamma_i) \label{dim_constraint}$$ is satisfied, since the series vanishes otherwise. In the dimension constraint , $\deg$ denotes the *real* cohomological degree of a class (both on $\overline{M}_{g,n}$ and $A$).
**Case (i):** $\sum_{i=1}^{n} \deg(\gamma_{i}) \leq 2 n$.
From dimension constraint , we find $$\deg(\alpha) = 2g + 2n - \sum_i \deg(\gamma_i) \geq 2g \,,$$ or equivalently, $$\alpha \in R^{\geq g}({{\overline M}}_{g,n}) \,.$$ Using the strong form of Getzler-Ionel vanishing proven in [@FPM Proposition 2], we can find a class $$\widetilde{\alpha} \in R^*( \partial {{\overline M}}_{g,n})$$ such that $\iota_{\ast} \widetilde{\alpha} = \alpha$, where $\iota : \partial {{\overline M}}_{g,n} {{\ \rightarrow\ }}{{\overline M}}_{g,n}$ is the inclusion of the boundary. By the splitting formula, the series $$\mathsf{F}^A_g( \alpha ; \gamma_1, \dots, \gamma_n )
=
\mathsf{F}^A_g( \iota_{\ast}\widetilde{\alpha} ; \gamma_1, \dots, \gamma_n )$$ is expressed in terms of a linear combination of series $$\mathsf{F}^A_{g'}( \widetilde{\alpha} ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'} ) \ \text{ for which }\ (g',n') < (g,n)\, .$$ The induction hypothesis for the latter implies Proposition \[modularity\_proposition\] holds for $\mathsf{F}^A_g( \alpha ; \gamma_1, \dots, \gamma_n )$.
**Case (ii):** $\sum_{i=1}^{n} \deg(\gamma_{i}) > 2 n$.
Consider the moduli space $$M_{g,n} \subset {{\overline M}}_{g,n}$$ of nonsingular genus $g\geq 2$ curves with $n$ marked points. The tautological ring $R^*(M_{g,n})$ is generated by classes pulled-back from $M_{g}$ via the forgetful map $$p : M_{g,n} \to M_g$$ and the cotangent line classes $$\psi_1, \dots, \psi_n\in H^2(M_{g,n},\mathbb{Q}) \,.$$ A class $\alpha \in R^*({{\overline M}}_{g,n})$ can therefore be written as a sum of classes of the form:
1. $\iota_{\ast}(\widetilde{\alpha})$ for $\widetilde{\alpha} \in R^*(\partial {{\overline M}}_{g,n})$,
2. $\psi_1^{k_1} \cup \ldots \cup \psi_n^{k_n} \cup p^{\ast}(\zeta)$ for $k_1, \dots, k_n \geq 0$ and $\zeta \in R^*({{\overline M}}_{g})$.
Here, we let $\psi_1, \dots, \psi_n$ denote also the cotangent line classes on ${{\overline M}}_{g,n}$.
A summand of the form $\iota_{\ast}(\widetilde{\alpha})$ is expressed by the splitting formula in lower order terms, see Case (i). Hence, we may assume $$\alpha = \psi_1^{k_1} \cup \ldots \cup \psi_n^{k_n} \cup p^{\ast}(\zeta) \,. \label{gghh}$$ **Case (ii-a):** There exists an $i$ for which $k_i > 0$.
We assume $k_1>0$. If $\deg(\gamma_1) \leq 3$, we first apply the vanishing of Lemma \[abelvan\] for $\gamma_1$ and $$\gamma = \gamma_2 \otimes \dots \otimes \gamma_n \,.$$ Using the abelian vanishing relation (see also ), we find, that $\mathsf{F}^A_g( \alpha ; \gamma_1, \gamma_2, \dots, \gamma_n )$ can be expressed as a sum of series $$\pm \mathsf{F}^A_g( \alpha ; \gamma_1 \cup \delta, \gamma'_2, \dots, \gamma'_n )$$ for various monomial classes $\delta, \gamma'_2, \dots, \gamma'_n \in
H^{\ast}(A,\mathbb{Q})$ with $\deg(\delta) \geq 1$.
The above relation [*increases*]{} the degree of $\gamma_1$. By induction on $\deg(\gamma_1)$, we may assume $\deg(\gamma_1) = 4$, or equivalently, $\gamma_1 = {{\mathsf{p}}}$.
We use next the degeneration of $$A = E_1 \times E_2$$ to the normal cone of an elliptic fiber $E = E_2$ of the projection to $E_1$, $$A \rightsquigarrow A \, \cup_{E} \, ({\mathbb{P}}^1 \times E) \,.$$ We choose the point class $\gamma_1$ to lie (after degeneration) on the component ${\mathbb{P}}^1 \times E$. The distribution of the remaining classes $\gamma_2, \dots, \gamma_n$ can be chosen freely.
For classes $\widetilde{\alpha} \in R^*({{\overline M}}_{g,r})$, $\xi \in H^{\ast}(E,\mathbb{Q})$, and $\widetilde{\gamma}_i$ on $A$ and ${\mathbb{P}}^1\times E$ respectively, we define $$\begin{aligned}
\mathsf{F}^{A/E}_{g} (\widetilde{\alpha} ; \widetilde{\gamma}_{{1}},\dots,
\widetilde{\gamma}_{{r}} ; \xi )
&=
\sum_{d \geq 0}
\Big\langle \ \widetilde{\alpha} \, ; \, \prod_{i=1}^{r} \tau_0( \widetilde{\gamma}_{{i}}) \ \Big|\ \xi\ \Big\rangle^{A/E,\text{red}}_{g, (1,d)} q^d\, ,\\
\mathsf{F}^{({\mathbb{P}}^1 \times E)/E}_{g}( \widetilde{\alpha} ; \widetilde{\gamma}_{{1}},\dots,\widetilde{\gamma}_{{r}} ; \xi )
&=
\sum_{d \geq 0}
\Big\langle \ \widetilde{\alpha} \, ; \, \prod_{i=1}^{r} \tau_0(\widetilde{\gamma}_{{i}}) \ \Big|\ \xi\ \Big\rangle^{({\mathbb{P}}^1 \times E)/E}_{g,(1,d)} q^d \,.\end{aligned}$$ where we use the bracket notation defined in Section \[P1Eevaluation\].
For $\gamma_1 = {{\mathsf{p}}}$, the degeneration formula then yields: $$\begin{gathered}
\label{de9mxSA}
\mathsf{F}^A_g( \alpha ; \gamma_1, \dots, \gamma_n ) = \\
\sum_{(g',n') \leq (g,n-1)}
\mathsf{F}_{g',n'}^{A/E}( \alpha' ; (\, \dots \, ) ; \xi ) \cdot
\mathsf{F}_{g'',n''}^{({\mathbb{P}}^1 \times E)/E}( \alpha'' ; \gamma_1, ( \, \dots \, ) ; \xi^{\vee} )\, .\end{gathered}$$ The summation here is over splittings $$g= g'+g''\, , \quad n= n'+n''\, ,$$ and distributions $(\, \dots \, )$ of the insertions $\gamma_2, \dots, \gamma_{n}$. The marking numbers $n'$ and $n''$ are placed in the subscripts of the generating series inside the sum for clarity.[[^21]]{} The class $\alpha$ determines $\alpha'$ and $\alpha''$ by restriction. Finally, there is also a sum over all relative conditions $$\xi \in \{ \mathsf{1}, {\mathsf{a}}_2, {\mathsf{b}}_2, \omega_2 \}\,$$ where $\xi^{\vee}$ denotes the class dual to $\xi$.
\[2uVMV\] The primitive potential for $({\mathbb{P}}^1 \times E)/E$ satisfies the following properties for all $g''\geq 0$:
1. $\mathsf{F}_{g''}^{({\mathbb{P}}^1 \times E)/E}( \alpha'' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n''} ;
\xi^\vee )$ vanishes unless $$v_{{\mathsf{a}}_2}( \xi^\vee) + \sum_{k=1}^{n''} v_{{\mathsf{a}}_2}({{\widetilde{\gamma}}}_k) = v_{{\mathsf{b}}_2}( \xi^\vee) + \sum_{k=1}^{n''} v_{{\mathsf{b}}_2}({{\widetilde{\gamma}}}_k)\,.$$
2. $\mathsf{F}_{g''}^{({\mathbb{P}}^1 \times E)/E}( \alpha'' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n''} ; \xi^\vee ) \in \textup{QMod}_{2g'' + 2 \ell'' - 2}$ for $$\ell'' = v_{{\mathsf{a}}_2}( \xi^\vee) + \sum_{k=1}^{n''} v_{{\mathsf{a}}_2}({{\widetilde{\gamma}}}_k)\,.$$
Because there is only algebraic cohomology on ${\mathbb{P}}^1$, the vanishing statement (i) follows from the fact that the virtual class is algebraic.
In [@MP2], the relative invariants of $({\mathbb{P}}^1\times E)/E$ were effectively expressed in terms of the absolute descendent invariants of ${\mathbb{P}}^1 \times E$ through relations obtained by the following operations:
- degeneration of ${\mathbb{P}}^1 \times E$ to ${\mathbb{P}}^1 \times E \cup_{E} {\mathbb{P}}^1 \times E$,
- localization on the ${\mathbb{P}}^1$ factor,
- rubber calculus.
A simple verification shows the resulting relations respect modularity of the weight specified by (ii) for each of these operations. Hence, we are reduced to the case of absolute descendent invariants of ${\mathbb{P}}^1 \times E$.
Then, virtual localization on the ${\mathbb{P}}^1$-factor expresses the descendents invariants of ${\mathbb{P}}^1\times E$ in terms of the descendent invariants of $E$. Finally, [@MPT Proposition 28] yields the required modularity property (ii).
For the last step, instead of localization, the product formula [@Beh] may be used to reduce ${\mathbb{P}}^1 \times E$ to the case of the descendent invariants of the elliptic curve $E$ since the the Gromov-Witten classes of ${\mathbb{P}}^1$ are known to be tautological [@GP].
Lemma \[2uVMV\] controls the factor on the right inside the sum of . Part (ii) of Lemma \[2uVMV\] is a refinement of [@MPT Lemma 30]. However, the proof is the same as given in [@MPT].
The factor on the left inside the sum of is more difficult to control. We will consider the terms of the sum corresponding to $$(g',n') < (g, n-1)\ {\text{ and }}\ (g',n') = (g,n-1)$$ separately. Lemma \[quasi\_mod\_relative\_lemma\] below shows how to apply the inductive hypothesis to the terms in the sum with $(g',n') < (g,n-1)$. The case $(g',n') = (g,n-1)$ will be considered afterwards.
\[quasi\_mod\_relative\_lemma\] Let $(g', n') < (g,n-1)$. The primitive potential for $A/E$ satisfies the following properties:
1. $\mathsf{F}_{g'}^{A/E}( \alpha' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'} ; \xi )$ vanishes unless $$v_{{\mathsf{a}}_i}( \xi) + \sum_{k=1}^{n'} v_{{\mathsf{a}}_i}( {{\widetilde{\gamma}}}_k )
=
v_{{\mathsf{b}}_i}( \xi) + \sum_{k=1}^{n'} v_{{\mathsf{b}}_i}( {{\widetilde{\gamma}}}_k ) \,, \ \ \ \
i\in \{1,2\}\, .$$
2. $\mathsf{F}_{g'}^{A/E}( \alpha' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'} ; \xi ) \in \textup{QMod}_{2 (g' - 2) + 2 \ell'}$ for $$\ell' = v_{{\mathsf{a}}_2}( \xi) + \sum_{k=1}^{n'} v_{{\mathsf{a}}_2}( {{\widetilde{\gamma}}}_k) \,.$$
We apply the degeneration formula to $$\label{cwwc}
\mathsf{F}_{g'}^A( \alpha' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'}, \omega_1 \xi )$$ with the last point degenerating to ${\mathbb{P}}^1 \times E$ and the specialization by pull-back for the other insertions. Since $(g',n')<(g,n-1)$, we have $$(g',n'+1) < (g,n)$$ so the induction hypothesis applies to the series .
The degeneration formula yields a relation involving the relative geometries $A/E$ and $({\mathbb{P}}^1 \times E)/E$, $$\begin{gathered}
\label{iiop}
\mathsf{F}_{g'}^A( \alpha'; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'}, \omega_1 \xi ) = \\
\sum_{(g'_\circ, n'_\circ) \leq (g',n')}
\mathsf{F}_{g'_\circ,n'_{\circ}}^{A/E}( \alpha'_\circ ; (\, \dots \, ) ; \widetilde{\xi} ) \cdot
\mathsf{F}_{g'_\bullet,n'_\bullet}^{({\mathbb{P}}^1 \times E)/E}( \alpha'_\bullet ; ( \, \dots \, ), \omega_1 \xi ; \widetilde{\xi}^{\vee} ) \,.\end{gathered}$$ The summation here is over splittings $$g'= g'_\circ+g'_\bullet\, , \quad n'+1 = n'_\circ+n'_\bullet\, ,$$ and distributions $(\, \dots \, )$ of the insertions ${{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'}$. The class $\alpha'$ determines $\alpha'_\circ$ and $\alpha'_\bullet$ by restriction. There is also a sum over all relative conditions $\widetilde{\xi} \in \{ \mathsf{1}, {\mathsf{a}}_2, {\mathsf{b}}_2, \omega_2 \}.$
We now analyse the $(g'_\circ,n'_\circ)=(g',n')$ term of the sum in , $$\label{fbbf}
\mathsf{F}_{g'}^{A/E}( \alpha' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'} ; \widetilde{\xi} )
\cdot \mathsf{F}_{0,1}^{({\mathbb{P}}^1 \times E)/E}( \omega_1 \xi ; \widetilde{\xi}^{\vee} )
\, .$$ Since genus 0 stable maps do not represent classes of type $(1,d>0)$ on ${\mathbb{P}}^1\times E$, $$\mathsf{F}_{0,1}^{({\mathbb{P}}^1 \times E)/E}( \omega_1 \xi ; \widetilde{\xi}^{\vee} )
=
\langle\, \widetilde{\xi}^{\vee}\, |\, \tau_0( \omega_1 \xi )\, \rangle_{0, (1,0)}^{({\mathbb{P}}^1 \times E) / E} = \delta_{\xi,\widetilde{\xi}} \, .$$ The second equality is obtained by the identification of the moduli space of maps by the location of the relative point, $$\label{prrp}
\overline{M}_{0,0}(({\mathbb{P}}^1\times E)/E,(1,0)) \cong E\, .$$ Taken together, we can rewrite as simply $$\mathsf{F}_{g'}^{A/E}( \alpha' ; {{\widetilde{\gamma}}}_1, \dots, {{\widetilde{\gamma}}}_{n'} ; {\xi} ) \, .$$
We find the series $\mathsf{F}^A$ for the absolute geometry can be expressed in terms of the series $\mathsf{F}^{A/E}$ for the relative geometry by a transformation matrix $\mathsf{M}$ which is upper triangular with $1$’s on the diagonal. By Lemma \[2uVMV\], the off-diagonal terms of $\mathsf{M}$ are given by quasi-modular forms. By inverting the unipotent matrix $\mathsf{M}$ and applying the induction hypothesis to $\mathsf{F}^A$, we find that the relative invariants $\mathsf{F}^{A/E}$ are quasi-modular forms.
The weight and vanishing statement can now be deduced from a careful consideration of the entries of $\mathsf{M}^{-1}$. Alternatively, we may argue via a (second) induction on $(g',n')$. In case $(g',n') = (1,1)$, there are no lower order terms in , and we are done. If the statement is true for all $(g'_\circ,n'_\circ) < (g',n')$, then the statement follows directly from and the induction hypothesis.
We now turn to the $(g',n')=(g,n-1)$ term in the sum : $$\mathsf{F}_{g,n-1}^{A/E}( \alpha ; \gamma_2, \dots, \gamma_n ; \omega ) \cdot \mathsf{F}_{0,1}^{({\mathbb{P}}^1 \times E)/E}( \gamma_1;\mathsf{1} )$$ As we have seen above, only curves in class $(1,0)$ contribute to the series $\mathsf{F}_{0,1}^{({\mathbb{P}}^1 \times E)/E}(\gamma_1;\mathsf{1} )$. By the identification of the moduli space , $$\mathsf{F}_{0,1}^{({\mathbb{P}}^1 \times E)/E}( \gamma_1;\mathsf{1}) = \langle\, \mathsf{1}\, |\, \tau_0({{\mathsf{p}}})\, \rangle_{0, (1,0)}^{({\mathbb{P}}^1 \times E) / E} = 1\,.$$ Hence, the $(g',n')=(g,n-1)$ term is $$\mathsf{F}_{g,n-1}^{A/E}( \alpha ; \gamma_2, \dots, \gamma_n ; \omega ) \, , \label{szszsz}$$ where the class $\alpha\in R^*(\overline{M}_{g,n})$ is pulled-back to ${{\overline M}}_{g,n-1}(A / E, (1,d))$ via the map $$\pi: {{\overline M}}_{g,n-1}(A / E, (1,d)) \rightarrow \overline{M}_{g,n}$$ which takes the [*relative*]{} point on the left to the marking 1 on the right. We must prove the induction hypothesis implies
1. $\mathsf{F}_{g,n-1}^{A/E}( \alpha ; \gamma_2, \dots, \gamma_{n} ; \omega )$ vanishes unless $$v_{{\mathsf{a}}_i}( \xi) + \sum_{k=2}^{n} v_{{\mathsf{a}}_i}( \gamma_k )
=
v_{{\mathsf{b}}_i}( \xi) + \sum_{k=2}^{n} v_{{\mathsf{b}}_i}( \gamma_k ) \,, \ \ \ \
i\in \{1,2\}\, .$$
2. $\mathsf{F}_{g,n-1}^{A/E}( \alpha ; \gamma_2, \dots, \gamma_{n} ; \omega ) \in \textup{QMod}_{2 (g - 2) + 2 \ell}$ for $$\ell = v_{{\mathsf{a}}_2}( \xi) + \sum_{k=2}^{n} v_{{\mathsf{a}}_2}( \gamma_k) \,.$$
We proceed by studying the cotangent lines. Let $$L_{\text{rel}} \, \rightarrow\, {{\overline M}}_{g,n-1}(A / E, (1,d))\, , \ \ \ \
L_{1} \, \rightarrow\, {{\overline M}}_{g,n}$$ denote the respective cotangent lines at the relative point and the first marking.
After pull-back via $\pi$, we have an isomorphism $$L_{\text{rel}} \cong \pi^*L_1$$ on ${{\overline M}}_{g,n-1}(A / E, (1,d))$. \[dssxq\]
Let $C$ be the $n$-pointed domain of a map $$f: C \rightarrow \widetilde{A}$$ parameterized by the moduli space ${{\overline M}}_{g,n-1}(A / E, (1,d))$. The $n$ points consist of the relative point together with the $n-1$ standard markings. The target $\widetilde{A}$ is a possible accordian destabilization of $A$ along $E$. The Lemma is a consequence of the following claim: [*the $n$-pointed curve $C$ is Deligne-Mumford stable*]{}.
Since $g\geq 1$ and $n\geq 1$, to prove the stability of the $n$-pointed curve $C$, we need only consider the nonsingular rational components $P\subset C$:
1. If $f(P) \subset A$, then $f$ is constant on $P$, $$f(P) \in A \setminus E\, ,$$ and $P$ must carry at least $3$ special points by the definition of map stability.
2. If $f(P)$ is contained in a rubber bubble over $E$, $P$ is mapped to a point of $E$ and therefore must map to a fiber of the bubble. Stability of the bubble then requires the existence of at least $3$ special points of $P$.
Since the $n$-pointed curve $C$ is stable, there is no contraction of components associated to the map $\pi$. Hence, the cotangent lines are isomorphic.
The relative divisor $E \subset A$ is the fiber over a point $0_{E_1}\in E_1$. Let $$f: C \rightarrow \widetilde{A}$$ be a stable map parameterized by the moduli space ${{\overline M}}_{g,n-1}(A / E, (1,d))$, and let $p_{\text{rel}}\in C$ be the relative point. Composition of the canonical projections $$\epsilon: \widetilde{A} \rightarrow A \rightarrow E_1$$ yields a map $$\epsilon f: C \rightarrow E_1\ \ \ \ \text{with} \ \ \ \ \epsilon f (p_{\text{rel}}) = 0_{E_1}\, .$$ The cotangent line $L_{\text{rel}}$ carries a canonical section via the differential of $\epsilon f$, $$s:\mathbb{C} =T_{0_{E_1}}^*(E_1) \rightarrow L_{\text{rel}}\, .$$ The vanishing locus[[^22]]{} of the section $s$ is the boundary of moduli space ${{\overline M}}_{g,n-1}(A / E, (1,d))$ corresponding to the first bubble over $E$.
Since $\alpha$ is of the form with $k_1 > 0$, a factor $\psi_1= c_1(L_1)$ can be extracted from $\alpha$, $$\alpha = \psi_1 \cdot \widetilde{\alpha}\, .$$ After pull-back via $\pi$, we have $$\pi^*(\psi_1) = c_1(L_{\text{rel}})$$ by Lemma \[dssxq\]. Via the vanishing locus of the section $s$, we obtain the following equation for the series : $$\begin{gathered}
\label{iiww}
\mathsf{F}_{g,n-1}^{A/E}( \alpha; \gamma_2, \dots, \gamma_{n}; \omega ) = \\
\sum_{(g', n') < (g,n-1)}
\mathsf{F}_{g',n'}^{A/E}( \widetilde{\alpha}' ;\, (\, \dots \, )\, ; {\xi} ) \cdot
\mathsf{F}_{g'',n''}^{\text{Rub}({\mathbb{P}}^1 \times E)}( \widetilde{\alpha}'';\,
( \, \dots \, )\, ; {\xi}^{\vee},\omega ) \,.\end{gathered}$$ The summation here is over splittings $$g= g'+g''\, , \quad n-1= n'+n''\, ,$$ and distributions $(\, \dots \, )$ of the insertions $\gamma_2, \dots, \gamma_{n}$. Only insertions $\gamma_i$ satisfying $$v_{{\mathsf{a}}_1}(\gamma_i)=v_{{\mathsf{b}}_1}(\gamma_i)=0$$ can be distributed to the rubber series $$\begin{gathered}
\mathsf{F}^{\text{Rub}({\mathbb{P}}^1 \times E)}_{g'',n''}( \widetilde{\alpha}'' ;\,
(\, \dots \, )\, ; \xi^{\vee},\omega ) =
\sum_{d \geq 0}
\Big\langle \xi^\vee \ \Big| \ \ \widetilde{\alpha}'' \, ; (\, \dots \, )
\ \Big|\ \omega\ \Big\rangle^{\text{Rub}({\mathbb{P}}^1 \times E)}_{g'',(1,d)} q^d\, .\end{gathered}$$ By stability of the rubber, either $g''>0$ or $n''>0$. The class $\widetilde{\alpha}$ determines $\widetilde{\alpha}'$ and $\widetilde{\alpha}''$ by restriction. Finally, there is also a sum on the right side of over relative conditions $\xi$.
In the sum on the right side of , the balancing and modularity of the first factor $$\mathsf{F}_{g',n'}^{A/E}( \widetilde{\alpha}' ;\, (\, \dots \, )\, ; {\xi} )$$ is obtained from Lemma \[quasi\_mod\_relative\_lemma\]. The balancing and modularity of the rubber factor $$\mathsf{F}_{g'',n''}^{\text{Rub}({\mathbb{P}}^1 \times E)}( \widetilde{\alpha}'';\,
( \, \dots \, )\, ; {\xi}^{\vee},\omega )$$ follows from the rubber calculus and an argument parallel to the proof of Lemma \[2uVMV\].[[^23]]{} The results together imply the required balancing and modularity for the series .
We now control the balancing and modularity of all terms in the sum on the right of . As a consequence, Proposition \[modularity\_proposition\] holds for $$\mathsf{F}^A_g( \alpha ; \gamma_1, \dots, \gamma_n )\, .$$ The proof of the induction step for Case (ii-a) is complete.
**Case (ii-b):** $\alpha = p^{\ast}(\zeta)$ for some $\zeta \in R^*({{\overline M}}_{g})$.
We may assume $\gamma_1$ is of minimal degree: $$\deg(\gamma_1) \leq \deg(\gamma_i)\ \ \ \ {\text{for all}} \ \ \ \ i \in \{ 2, \dots, n \} \, .$$ Below, we will distinguish several subcases depending upon $\deg(\gamma_1)$.
Consider the map ${{\overline M}}_{g,n}(A,(1,d)) \rightarrow {{\overline M}}_{g,n-1}(A,(1,d))$ forgetting the first marking. The coefficients of the series $$\mathsf{F}^{A}_{g}( p^{\ast}(\zeta) ; \gamma_1, \dots, \gamma_n ) \label{hhjj}$$ are integrals where all classes in the integrand, except for ${\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1)$, are pull-backs via the map forgetting the first marking.
**Case** $\deg(\gamma_1) \leq 1$. The series vanishes by the push-pull formula since the fiber of the forgetful map has (complex) dimension $1$.
**Case** $\deg(\gamma_1) = 2$. We use the divisor equation for $\gamma_1$ and find $$\begin{aligned}
\mathsf{F}^{A}_{g}( p^{\ast}(\zeta) ; \gamma_1, \dots, \gamma_n ) =
\begin{cases}
\mathsf{F}^{A}_{g}( p^{\ast}(\zeta) ; \gamma_2, \dots, \gamma_n ) & \text{if } \gamma_1 = {\mathsf{a}}_1 {\mathsf{b}}_1 \\
q \frac{d}{dq} \mathsf{F}^{A}_{g}( p^{\ast}(\zeta) ; \gamma_2, \dots, \gamma_n ) & \text{if } \gamma_1 = {\mathsf{a}}_2 {\mathsf{b}}_2 \\
0 & \text{otherwise}\,.
\end{cases}\end{aligned}$$ Because the differential operator $$q \frac{d}{d q} = \frac{1}{2 \pi i} \frac{\partial}{\partial \tau}$$ preserves $\text{QMod}_{\ast}$ and is homogeneous of degree $2$, Proposition \[modularity\_proposition\] holds for by the induction hypothesis.
**Case** $\deg(\gamma_1) = 3$. Since $\sum_i \deg(\gamma_i)$ is even, we must have $n \geq 2$. We order the classes $$\gamma_2,\, \dots\, , \gamma_n$$ so that $\gamma_2, \dots, \gamma_k$ are point classes and $\gamma_{k+1}, \dots, \gamma_n$ are classes of degree $3$, for some $1 \leq k < n$.
We will use the abelian vanishing of Lemma \[abelvan\] for $\gamma_1$ and $$\gamma = \gamma_2 \otimes \dots \otimes \gamma_n \,.$$ Let $s \in \{{\mathsf{a}}_1,{\mathsf{b}}_1,{\mathsf{a}}_2,{\mathsf{b}}_2 \}$ be the factor with $v_s(\gamma_1) = 0$. Using the abelian vanishing relation, we find $$\begin{gathered}
\sum_{i=1}^{k} \mathsf{F}^A_g( p^{\ast}(\zeta) ; \underbrace{{{\mathsf{p}}},\ldots,{{\mathsf{p}}}}_{i-1}, \gamma_1, \underbrace{{{\mathsf{p}}},\ldots,{{\mathsf{p}}}}_{k-i}, \gamma_{k+1}, \dots, \gamma_n ) \\
= \sum_{\substack{ i = k+1 \\ v_s(\gamma_i) = 1}}^{n}
\pm \, \mathsf{F}^A_g( p^{\ast}(\zeta) ; {{\mathsf{p}}}, \gamma_2, \dots, \widetilde{\gamma}_i, \dots, \gamma_n ) \,,\end{gathered}$$ where $\widetilde{\gamma}_i = \gamma_i/s$ denotes the class $\gamma_i$ with the factor $s$ removed. The plus signs in the terms on the left hand side require a careful accounting of the signs. Since the class $p^{\ast}(\zeta)$ is symmetric with respect to interchanging markings, the above equation simplifies to $$k \cdot \mathsf{F}^A_g( p^{\ast}(\zeta) ; \gamma_1, \dots, \gamma_n )
= \sum_{\substack{ i = k+1 \\ v_s(\gamma_i) = 1}}^{n}
\pm \, \mathsf{F}^A_g( p^{\ast}(\zeta) ; {{\mathsf{p}}}, \gamma_2, \dots, \widetilde{\gamma}_i, \dots, \gamma_n ) \,,$$ Since $\widetilde{\gamma}_i$ is of degree $2$, we may apply the divisor equation to each summand on the right side. As result, the right side is reduce to terms of lower order, see Case $\deg(\gamma_1)=2$ above. By the induction hypothesis, Proposition \[modularity\_proposition\] holds for .
**Case** $\deg(\gamma_1) = 4$. All the insertions $\gamma_1, \dots, \gamma_n$ must be point classes. If $n = 1$, the dimension constraint implies $\deg(\alpha) = 2g-2$ and hence $$\zeta \in R^{g-1}({{\overline M}}_{g})$$ Using the strong form of Looijenga’s vanishing $$R^{\geq g-1}(M_g) = 0$$ proven in [@FPM Proposition 2], there exists a class $$\widetilde{\zeta} \in R^*( \partial {{\overline M}}_g)$$ such that $\iota_{\ast} \widetilde{\zeta} = \zeta$. After pulling back via $$p : {{\overline M}}_{g,n} \to {{\overline M}}_g\, ,$$ $p^{\ast}(\zeta)$ can be written as the push forward of a tautological class on the boundary $\partial{{\overline M}}_{g,n}$. Proposition \[modularity\_proposition\] holds for by the splitting formula and the induction hypothesis, see Case (i).
If $n \geq 2$, we use the degeneration $$A \rightsquigarrow A \, \cup_{E} \, ({\mathbb{P}}^1 \times E) \,,$$ which already appeared in Case (ii-a) above. We choose the point classes $\gamma_1$ and $\gamma_2$ to lie after degeneration on the component ${\mathbb{P}}^1 \times E$.
The degeneration formula then yields: $$\begin{gathered}
\label{number_two}
\mathsf{F}^A_g( \alpha ; \gamma_1, \dots, \gamma_n ) = \\
\sum_{(g',n') \leq (g,n)}
\mathsf{F}_{g',n'}^{A/E}( \alpha' ; (\, \dots \, ) ; \xi ) \cdot
\mathsf{F}_{g'',n''}^{({\mathbb{P}}^1 \times E)/E}( \alpha'' ; \gamma_1, \gamma_2, ( \, \dots \, ) ; \xi^{\vee} )\, ,\end{gathered}$$ where the sum is as in .
If $g'= g$ in the sum , then the second factor is $$\label{xxttyz}
\mathsf{F}_{0,n''}^{({\mathbb{P}}^1 \times E)/E}( \alpha'' ; \gamma_1, \gamma_2, ( \, \dots \, ) ; \xi^{\vee} )\, .$$ Genus 0 stable maps do not represent classes of type $(1,d>0)$ on ${\mathbb{P}}^1\times E$, hence only the curve class $(1,0)$ need be considered. Since there are no curves of type $(1,0)$ through two general points of ${\mathbb{P}}^1\times E$, vanishes. As a result, only $g'<g$ terms appear in the sum . Proposition \[modularity\_proposition\] holds for by Lemmas \[2uVMV\] and \[quasi\_mod\_relative\_lemma\].
The proof of the induction step has now been established in all cases. The proof of Proposition \[modularity\_proposition\] is complete.
$K3$ surfaces
-------------
Theorem \[modularity\] of Section \[qmfo\] for abelian surfaces is exactly parallel to the modularity results [@MPT Theorem 4 and Proposition 29] for the primitive descendent potential for $K3$ surfaces. Though the argument for abelian surfaces is more difficult because of the presence of odd cohomology, several aspects are similar.
The refined modularity of Theorem \[modularity\_refined\] for abelian surfaces is strictly stronger than the statements of [@MPT] for $K3$ surfaces. In fact, the proof of [@MPT] also yields the parallel refined statement for $K3$ surfaces. The crucial point is to use the refined modularity of Lemma \[2uVMV\] part (ii) instead of the weaker modularity of [@MPT Lemma 30]. We state the refined modularity for $K3$ surfaces below following the notation of [@MPT].
Let $S$ be a nonsingular, projective, elliptically fibered $K3$ surface, $$S \rightarrow {\mathbb{P}}^1,$$ with a section. Let $\mathbf{s}, \mathbf{f} \in H_2(S,\mathbb{Z})$ be the section and fiber class. The primitive descendent potential for the reduced Gromov-Witten theory of $S$ is defined by $$\mathsf{F}^{S}_{g}\big(\tau_{a_1}(\gamma_{1}) \cdots
\tau_{a_n}(\gamma_{n})\big)=
\sum_{d \geq 0}
\Big\langle \tau_{a_1}(\gamma_{1}) \cdots
\tau_{a_n}(\gamma_{n})\Big\rangle^{S,\text{red}}_{g,\mathbf{s}+ d \mathbf{f} }
\ q^{d-1}$$ for $g\geq 0$.
We define a new degree function $\underline{\text{deg}}(\gamma)$ for classes $\gamma\in H^*(S,\mathbb{Q})$ by the following rules:
1. $\gamma \in H^0(S,\mathbb{Q})$ $\mapsto$ $\underline{\text{deg}}(\gamma)=0$,
2. $\gamma \in H^4(S,\mathbb{Q})$ $\mapsto$ $\underline{\text{deg}}(\gamma)=2$.
For classes $\gamma\in H^2(S,\mathbb{Q})$, the degree is more subtle.
Viewing the section and fiber classes also as elements of cohomology, we define $$V = \mathbb{Q} \mathsf{s} \oplus \mathbb{Q} \mathsf{f}\subset H^2(S,\mathbb{Q})\, .$$ We have a direct sum decomposition $$\label{jj11} \mathbb{Q} \mathsf{f} \, \oplus \, V^\perp \, \oplus \, \mathbb{Q}(\mathsf{s}+\mathsf{f}) \cong
H^2(S,\mathbb{Q})$$ where $V^\perp$ is defined with respect to the intersection form. We consider only classes $\gamma \in H^2(S,\mathbb{Q})$ which are [pure]{} with respect to the decomposition . Then,
1. $\gamma \in \mathbb{Q} \mathsf{f} $ $\mapsto$ $\underline{\text{deg}}(\gamma)=0$,
2. $\gamma \in V^\perp$ $\mapsto$ $\underline{\text{deg}}(\gamma)=1$,
3. $\gamma \in \mathbb{Q} (\mathsf{s}+\mathsf{f}) $ $\mapsto$ $\underline{\text{deg}}(\gamma)=2$.
The modularity of [@MPT Theorem 4 and Proposition 29] is refined by the following result.
\[qqq\] For $\underline{\text{\em deg}}$-homogeneous classes $\gamma_i \in H^*(S,\mathbb{Q})$, $$\mathsf{F}_{g}^S\big(\tau_{a_1}(\gamma_{1})
\cdots
\tau_{a_n}(\gamma_{n})\big) \in \frac{1}{\Delta(q)} \text{\em QMod}_{\ell}$$ for $\ell= 2g+ \sum_{i=1}^n \underline{\text{\em deg}}(\gamma_i)$.
The discriminant modular form entering in Theorem \[qqq\] is $$\Delta(q)= q \prod_{n=1}^\infty (1-q^n)^{24}\,.$$
Hyperelliptic curves {#hyperelliptic-curves}
====================
Overview
--------
The correspondence between hyperelliptic curves on a surface $S$ and rational curves in $\operatorname{Hilb}^2(S)$ has been used often to enumerative hyperelliptic curves on $S$, see [@Gra] for $S = {\mathbb{P}}^2$ and [@Pon07] for $S = {\mathbb{P}}^1 \times {\mathbb{P}}^1$. The main difficulty in applying the correspondence is the need of a non-degeneracy result concerning curves in ${\mathbb{P}}^1 \times S$. For abelian surfaces, the required non-degeneracy, stated as ($\dag$) in Section \[hyp\_intro\], is expected to hold generically, but is not known in most cases. The above correspondence then yields only conditional or virtual results on the number of hyperelliptic curves on an abelian surface, as pursued, for example, by S. Rose in [@Ros14].
We proceed with our study of hyperelliptic curves in three steps. In Section \[hyp\_2\], we provide several equivalent descriptions of ($\dag$) and a proof in genus $2$ for a generic abelian surface. In Section \[hyp\_3\], we prove an unconditional formula for the first non-trivial case of genus $3$ hyperelliptic curves via explicit Gromov-Witten integrals, a boundary analysis, and the genus $2$ result proven in Section \[hyp\_2\].
In Section \[hyp\_4\], we assume the existence of abelian surfaces $A$ and irreducible curve classes $\beta$ satisfying property ($\dag$) in all genera. Employing the correspondence above, we find a closed formula for the $\mathsf{h}_{g, \beta}^{A, \textup{FLS}}$. While a similar strategy has been used in [@Ros14] assuming the crepant resolution conjecture, our closed formula is new. Together with the strong modularity result of Theorem \[modularity\_refined\], we obtain a formula for the Gromov-Witten numbers $\mathsf{H}_{g,(1,d)}^{\textup{FLS}}$ which agrees with the genus $3$ counts.
Non-degeneracy for abelian surfaces {#hyp_2}
-----------------------------------
We briefly recall the correspondences between hyperelliptic curves in $S$, curves in ${\mathbb{P}}^1 \times S$, and rational curves in $\operatorname{Hilb}^2(S)$. For simplicity, we restrict to the case of abelian surfaces $S = A$, see [@FKP; @Gra; @GO] for the general case. Let $A$ be an abelian surface, and let $$f : C {{\ \rightarrow\ }}A$$ be a map from a nonsingular hyperelliptic curve. Let $$p : C \to {\mathbb{P}}^1$$ be the double cover. Since $A$ contains no rational curves, $f$ does not factor through $p$. Consider the map $$(p, f) : C {{\ \rightarrow\ }}{\mathbb{P}}^1 \times A \,.$$ The image $\bar{C} = \text{Im}(C)\subset {\mathbb{P}}^1 \times A$ is an irreducible curve, (flat) of degree $2$ over ${\mathbb{P}}^1$, and has normalization $C \to \bar{C}$.
Let $\operatorname{Hilb}^2(A)$ be the Hilbert scheme of $2$ points of $A$, and let $$\Delta \subset \operatorname{Hilb}^2(A)$$ denote the subvariety parameterizing non-reduced length $2$ subschemes of $A$. By the universal property of the Hilbert scheme, the curve $\bar{C}$ induces a map $$\phi : {\mathbb{P}}^1 {{\ \rightarrow\ }}\operatorname{Hilb}^2(A)$$ such that the image is not contained in $\Delta$.
Conversely, let $\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$ be a map whose image is not contained in $\Delta$. Since $A$ contains no rational curves, by pulling back the universal family $$Z \subset \operatorname{Hilb}^2(A) \times A\, ,$$ we obtain an irreducible curve of degree $2$ over ${\mathbb{P}}^1$. The normalization $C \to \bar{C}$ is hyperelliptic and induces a map $f : C \to A$.
Hence, there are bijective correspondences between
- maps $f : C \to A$ from nonsingular hyperelliptic curves,
- irreducible curves $\bar{C} \subset {\mathbb{P}}^1 \times A$ of degree $2$ over ${\mathbb{P}}^1$,
- maps $\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$ with image not contained in $\Delta$.
The correspondences allow us to reformulate the non-degeneracy property ($\dag$).
\[lem\_trans\] Under the correspondences above, the followings are equivalent.
1. The differential of $f$ is injective at the Weierstrass points of $C$, and no conjugate non-Weierstrass points are mapped to the same point on $A$.
2. The curve $\bar{C}$ is nonsingular.
3. The map $\phi$ meets $\Delta$ transversally.
Using a recent result of Poonen and Stoll [@PS14], we verify that property ($\dag$) holds generically in genus $2$:
\[nondeg2\] Let $A$ be a generic abelian surface with a curve class $\beta$ of type $(1,d)$, and let $f : C \to A$ be a map from a nonsingular genus $2$ curve in class $\beta$. Then $f:C \to A$ satisfies condition of Lemma \[lem\_trans\].
The condition on the differential of $f$ is automatically satisfied by [@LS0 Proposition 2.2] for genus $2$ curves on abelian surfaces.
Now suppose there exists a nonsingular genus $2$ curve $C$, two conjugate non-Weierstrass points $x, y \in C$, and a map $f : C \to A$ in class $\beta$ such that $f(x) = f(y)$. Up to translation we may assume that $f$ maps a Weierstrass point $q \in C$ to $0_A \in A$. Then $f$ factors as $$C \xrightarrow{\mathsf{aj}} J \xrightarrow{\pi} A \,,$$ where $J$ is the Jacobian of $C$ and $\mathsf{aj}$ is the Abel-Jacobi map with respect to $q$. The hyperelliptic involution of $C$ corresponds to the automorphisms $-1$ of $J$ and $A$. For $x, y$ conjugate, this implies that $f(x) = f(y)$ is a $2$-torsion point on $A$.
Since $C$ is of genus $2$ and $\beta$ is of type $(1, d)$, the map $\pi$ is an isogeny of degree $d$. It follows that both $\mathsf{aj}(x)$ and $\mathsf{aj}(y)$ are $2d$-torsions on $J$.
In genus $2$, the assumption that $A$ is generic implies $C$ is generic. However, by [@PS14 Theorem 7.1], a generic (Weierstrass-pointed) hyperelliptic curve $C$ meets the torsions of $J$ only at the Weierstrass points. Hence the points $x, y$ do not exist.
The proof of Lemma \[nondeg2\] works for any type $(d_1, d_2)$ with $d_1, d_2 > 0$. However, since multiple covered curves may arise, statement $(\dag)$ is false for imprimitive classes in higher genus. The most basic counterexample is constructed by taking an étale double cover $$C_3 \rightarrow C_2$$ of a nonsingular genus $2$ curve $C_2\subset A$. Then, $C_3$ is nonsingular and hyperelliptic of genus $3$, but $C_2$ contains a Weierstrass point whose preimage in $C_3$ is a pair of conjugate non-Weierstrass points.
Furthermore, by the proof of [@KLM15 Theorem 1.6], for generic $A$ and $\beta$ of type $(1, d)$, property $(\dag)$ also holds in the maximal geometric genus $g_d$. The value of $g_d$ is determined by the inequality . It is also shown that for every $g \in \{2, \ldots, g_d\}$, there exists at least one nonsingular genus $g$ curve in ${\mathbb{P}}^1 \times A$ of class $(2, \beta)$.
Genus 3 hyperelliptic counts {#hyp_3}
----------------------------
We prove here Proposition \[cor\_hyp3\]. We proceed in two steps. First, we evaluate $\mathsf{H}^{\text{FLS}}_{3, (1,d)}$. Then we identify the contributions from the boundary of the moduli space.
For all $d \geq 1$, $$\mathsf{H}^{\textup{FLS}}_{3, (1,d)} = d^2 \sum_{m | d} \frac{ 3m^2 - 4 dm }{4} \,.$$
On $\overline{M}_3$, let $\lambda_1$ be the first Chern class of the Hodge bundle, $\delta_0$ the class of the curves with a nonseparating node, and $\delta_1$ the class of curves with a separating node. By [@HM], $$\mathcal{H}_3 = 9 \lambda_1 - \delta_0 - 3 \delta_1 \,.$$ The Lemma will be proven by the following three evaluations $$\Big\langle \lambda_1 \Big\rangle_{3,(1,d)}^{A, \textup{FLS}} = \frac{d^2}{12} \sum_{m | d} m^3 \,, \quad
\Big\langle \delta_0 \Big\rangle_{3,(1,d)}^{A, \textup{FLS}} = d^3 \sum_{m|d} m \,, \quad
\Big\langle \delta_1 \Big\rangle_{3,(1,d)}^{A, \textup{FLS}} = 0 \,.$$ The first equation follows directly from Theorem \[YZ\_intro\]. For the second, we have $$\Big\langle \delta_0 \Big\rangle_{3,(1,d)}^{A, \text{FLS}}
= \frac{1}{2} \Big\langle \tau_0( \Delta ) \Big\rangle_{2, (1,d)}^{A, \text{FLS}}
= \frac{1}{2} \cdot 2d \cdot \Big\langle \mathsf{1} \Big\rangle_{2, (1,d)}^{A, \text{FLS}}
= d \cdot d^2 \sum_{m|d} m \,,$$ where $\Delta$ denotes the class of the diagonal in $A \times A$. The divisor $\delta_1$ is associated to the locus of curves which split into genus $1$ and genus $2$ components. Since generically $A$ contains no genus $1$ curves, the class on the genus $1$ component must be 0. Since $[{{\overline M}}_{1,1}(A, 0)]^{\text{vir}} = 0$, we obtain the third evaluation.
Let $A$ and $\beta$ be generic. By Lemma \[nondeg2\] the only contribution to $\mathsf{H}^{\text{FLS}}_{g, (1,d)}$ from maps $f : C \to A$ with $C$ nodal arises from the locus in ${{\overline M}}_3$ with a separating node. The maps are of the form $$f : B \cup C' {{\ \rightarrow\ }}A$$ where $B$ is a genus $2$ curve and $C'$ is an elliptic tail glued to $B$ along one of the $6$ Weierstrass points of $B$. Under $f$, the curve $B$ maps to a genus $2$ curve in $A$, while $C'$ gets contracted. By a direct calculation (or examining the case $d=1$), we find that each genus $2$ curve in the FLS contributes $$6 \cdot \frac{1}{2} \int_{{{\overline M}}_{1,1}} c_1({\rm Ob}) = - \frac{1}{4} \,,$$ where ${\rm Ob}$ denotes the obstruction sheaf. Therefore, $$\mathsf{h}^{A, \text{FLS}}_{3, \beta} = \mathsf{H}^{\text{FLS}}_{3, (1,d)} + d^2 \sigma(d) \cdot \frac{1}{4} =
d^2 \sum_{m|d} \frac{m ( 3m^2 + 1 - 4d) }{4} \,. \qedhere$$
A formula for all genera {#hyp_4}
------------------------
Consider the composition $$\operatorname{Hilb}^2(A) \to \operatorname{Sym}^2(A) \to A \,, \label{addition_map}$$ of the Hilbert-Chow morphism and the addition map. The fiber of $0_A \in A$ is the Kummer $K3$ surface of $A$, denoted ${\operatorname{Km}}(A)$. Alternatively, ${\operatorname{Km}}(A)$ can be defined as the blowup of $A / \pm 1$ at the 16 singular points.
In the notation of Section \[hyp\_2\], a map $$\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$$ not contained in $\Delta$ maps to ${\operatorname{Km}}(A)$ if and only if the corresponding hyperelliptic curve $f : C \to A$ maps a Weierstrass point of $C$ to a $2$-torsion point of $A$.
By Nakajima’s theorem on the cohomology of Hilbert schemes, we have a natural decomposition $$H_2( \operatorname{Hilb}^2(A) ; {{\mathbb{Z}}}) = H_2(A, {{\mathbb{Z}}}) \oplus \wedge^2 H_1(A, {{\mathbb{Z}}}) \oplus {{\mathbb{Z}}}\cdot X \,,$$ where $X$ is the class of an exceptional curve. A hyperelliptic curve $f : C \to A$ in class $\beta$ corresponds to a map $\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$ not contained in $\Delta$, which has class $$\beta + \gamma + k X \in H_2( \operatorname{Hilb}^2(A) , {{\mathbb{Z}}})$$ for some $\gamma \in \wedge^2 H_1(A, {{\mathbb{Z}}})$ and with $$k = \chi( {{\mathcal O}}_{\bar{C}} ) - 2 = -1 - g_a( \bar{C} ) \,,$$ where $g_a( \bar{C} )$ is the arithmetic genus of $\bar{C}$, see [@GO Section 1.3].
\[hyp\_prop\] Let $\beta$ be an irreducible curve class of type $(1,d)$ on an abelian surface $A$ satisfying $(\dag)$. Then, after the change of variables $y= -e^{2 \pi i z}$ and $q = e^{2 \pi i \tau}$, $$\sum_{g \geq 2} \mathsf{h}^{A, \textup{FLS}}_{g, \beta} ( y^{1/2} + y^{-1/2} )^{2g+2}
= \frac{d^2}{16} \, \textup{Coeff}_{q^d}\big[ 4 \, K(z,\tau)^4 \big] \,,$$ where $\textup{Coeff}_{q^d}$ denotes the coefficient of $q^d$.
For every hyperelliptic curve $f : C \to A$ in class $\beta$, the map $$(p,f) : C \to \bar{C}$$ is an isomorphism by ($\dag$). In particular, the arithmetic genus of $\bar{C}$ is equal to the genus of $C$.
Hence, there is a bijective correspondence between
- maps $f : C \to A$ from nonsingular hyperelliptic curves of genus $g$ and class $\beta$ for which a Weierstrass point of $C$ is mapped to a $2$-torsion point of $A$,
- maps $\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$ with image not contained in $\Delta$ of class $$\beta + \gamma - (g + 1) X$$ for some $\gamma \in \wedge^2 H_1(A, {{\mathbb{Z}}})$ and with image in ${\operatorname{Km}}(A)$.
Let $\mathsf{h}_{g, \beta}^{A, \operatorname{Hilb}}$ be the finite number of such curves.
In every translation class of a hyperelliptic curve $f : C \to A$ in class $\beta$, there are $d^2$ members (up to automorphisms) in a given fixed linear system, and $16$ members (up to automorphisms) with a Weierstrass point of $C$ mapping to a $2$-torsion point. Hence $$\mathsf{h}_{g, \beta}^{A, \textup{FLS}} = \frac{d^2}{16} \mathsf{h}_{g, \beta}^{A, \operatorname{Hilb}} \,.$$
By assumption ($\dag)$ and Lemma \[lem\_trans\], every map $\phi : {\mathbb{P}}^1 \to \operatorname{Hilb}^2(A)$ as in (ii) meets $\Delta$ transversely and is isolated. In this situation, Graber in [@Gra Sections 2 and 3] has explicitly determined the relationship between the genus $0$ Gromov-Witten invariants of $\operatorname{Hilb}^2(A)$ and the number of these rational curves.
Let $ p : {{\overline M}}_0(\operatorname{Hilb}^2(A)) \to A $ be the map induced by . Then, $$\begin{gathered}
\sum_{g \geq 0} \mathsf{h}^{A, \operatorname{Hilb}}_{g, \beta} (y^{1/2} + y^{-1/2})^{2g+2} = \\
\sum_{k \in {{\mathbb{Z}}}} \,\sum_{\gamma \in \wedge^2 H_1(A, {{\mathbb{Z}}})} y^k \int_{[ {{\overline M}}_0( \operatorname{Hilb}^2(A) , \beta + \gamma + k A) ]^{\text{red}} } p^{\ast}( 0_A ) \,.
\end{gathered}$$
The integral on the right hand side reduces to the genus $0$ invariants of the Kummer $K3$ surfaces and is determined by the Yau-Zaslow formula. Direct calculations and theta function identities, see [@GO] for details, then provide the closed evaluation $$\sum_{\substack{k \in {{\mathbb{Z}}}\\ \gamma \in \wedge^2 H_1(A, {{\mathbb{Z}}})}} y^k \int_{[ {{\overline M}}_0( \operatorname{Hilb}^2(A) , \beta + \gamma + k A) ]^{\text{red}} } p^{\ast}( 0_A )
= \textup{Coeff}_{q^d} \big[ 4 K(z,\tau)^4 \big] \,. \qedhere$$
We are now ready to prove Theorem \[thm\_hyp\].
Let $\beta$ be an irreducible class of type $(1,d)$ on an abelian surface $A$ satisfying ($\dag$). The only contribution to $\mathsf{H}_{g,(1,d)}^{\textup{FLS}}$ from maps $f : C \to A$ with $C$ nodal are of the form $$f : B \cup C_1 \cup \dots \cup C_{2h+2} {{\ \rightarrow\ }}A \label{5556}$$ where:
- $f : B \to A$ is a map in class $\beta$ from a nonsingular hyperelliptic curve $B$ of some genus $h < g$,
- the $C_i$ are pairwise disjoint curves, that are glued to the $i$-th Weierstrass point $x_i$ of $B$, and are contracted under $f$,
- the genera $g_i$ of $C_i$ satisfy $h + g_1 + \dots + g_{2h+2} = g$,
- if $g_i \geq 2$, then $C_i$ is a hyperelliptic curve and $x_i$ is a Weierstrass point of $C_i$.
By stability, the case $g_i = 0$ does not appear.
For $g \geq 2$, let $$\mathcal{H}_{g,1} \in A^{g-1}( {{\overline M}}_{g,1} )$$ be the stack fundamental class of the closure of nonsingular hyperelliptic curves with marked point at a Weierstrass point. By convention, we set $$\mathcal{H}_{1,1} = \frac{1}{2} [ {{\overline M}}_{1,1} ]\,.$$ Then, the contribution of a nonsingular hyperelliptic curve $f : B \to A$ of genus $h$ in class $\beta$ to $\mathsf{H}_{g,(1,d)}^{\textup{FLS}}$ is $$\sum_{\substack{ g_1, \dots, g_{2h+2} \geq 0 \\ g_1 + \dots + g_{2h+2} = g-h }} \,
\prod_{\substack{i = 1 \\ g_i > 0}}^{2h+2} \int_{{{\overline M}}_{g_i,1}} \mathcal{H}_{g_i,1} \cup c( {\rm Ob}) \label{5555}$$ where ${\rm Ob}$ denotes the obstruction sheaf. Analyzing the tangent obstruction sequence, we obtain $$c( {\rm Ob}) = \frac{ c( {{\mathbb{E}}}^{\vee} )^2 }{1 - \psi_1} \,.$$
Define the generating series $$F(u)
= u + \sum_{g \geq 1} u^{2g+1} \int_{{{\overline M}}_{g,1}}
\frac{ \mathcal{H}_{g,1} \cup c( {{\mathbb{E}}}^{\vee} )^2}{ 1 - \psi_1 } \,,$$ Then from relation and the definition of $\mathsf{h}_{g, \beta}^{A, \text{FLS}}$, we obtain $$\sum_{h \geq 2} \mathsf{h}^{A, \textup{FLS}}_{h, \beta} F(u)^{2h+2} =
\sum_{g \geq 2}\, \mathsf{H}^{\textup{FLS}}_{g,(1,d)} \, u^{2g+2} \,.$$ The series $F(u)$ has been computed by J. Wise using orbifold Gromov-Witten theory [@Wise]. The result is $$F(u) = 2 \sin(u/2) = u - \frac{1}{24} u^3 + \frac{1}{1920} u^5 \pm \dots \, .$$ Together with Proposition \[hyp\_prop\], the claim follows.
The calculation of the invariants $\mathsf{H}_{g,(1,d)}^{\text{FLS}}$ is similar to the calculations of the orbifold genus $0$ Gromov-Witten theory of the second symmetric product of a nonsingular surface as pursued in [@Ros14; @WiseP2]. We expect a connection can be made to their work.
The main step in the proof of Theorem \[thm\_hyp\] is the evaluation of the generating series $F(u)$. Below we will give a second proof of Theorem \[thm\_hyp\] under slightly stronger assumptions. The main new input here is the refined modularity statement of Theorem \[modularity\_refined\]. Using the modularity property, the evaluation of $F(u)$ will follow automatically from the theory of modular forms.
For the second proof, we will assume the following holds:
- [*For every $d\geq 1$, there exists an abelian surface $A$ and an irreducible curve class $\beta$ of type $(1,d)$ satisfying property $(\dag)$.*]{}
For all $d \geq 1$, let $\beta_d$ be an irreducible class of type $(1,d)$ on an abelian surface $A_d$ satisfying ($\dag$).
*Step 1.* Define the generating series $$\varphi_{g}(q) = \sum_{d \geq 1} \mathsf{h}^{A_d, \textup{FLS}}_{g, \beta_d} q^d \,.$$ By Proposition \[hyp\_prop\], we have, after the change of variables $u = 2\pi z$ and $y = - e^{iu}$, $$\sum_{g \geq 2} (y^{1/2} + y^{-1/2})^{2g+2} \varphi_g(q) =
\left( q \frac{d}{dq} \right)^2 \frac{K(z,\tau)^4}{4} = \sum_{m \geq 2} u^{2m} f_m(q) \,,
\label{sosos}$$ where $f_m(q)$ are quasi-modular forms of weight $2m$, $$f_m(q) \in \text{QMod}_{2m} \,.$$ Let $r = - (y^{1/2} + y^{-1/2}) = 2 \sin(u/2)$, and let $$u = 2 \arcsin( r/2 ) = r + \frac{1}{24} r^3 + \frac{3}{640} r^5 + \dots$$ be the inverse transform. After inserting into , we obtain $$\sum_{g \geq 2} \varphi_{g}(q) r^{2g+2}
= \sum_{m \geq 2} \bigg(r + \frac{1}{24} r^3 + \dots\ \bigg)^{2m} f_m(q) \,.$$ Hence, $\varphi_g(q)$ is a quasi-modular form with highest weight term $f_{g+1}(q)$: $$\varphi_{g}(q) = f_{g+1}(q) + R(q) \label{remainder_eqn}$$ for $R(q) \in \text{QMod}_{\leq 2g}$.
*Step 2.* By trading of the FLS for insertions as in , the vanishing of the $d=0$ term and deformation invariance, $$\label{xxxxxxx123}
\mathsf{F}_{g}^{E_1 \times E_2}( {{\mathcal H}}_g ; {\mathsf{a}}_1 \omega_2, {\mathsf{b}}_1 \omega_2, \omega_1 {\mathsf{a}}_2, \omega_1 {\mathsf{b}}_2 )
=
\sum_{d \geq 1} \mathsf{H}^{\textup{FLS}}_{g,(1,d)} q^d,$$ where we use the notation of Section \[section\_modularity\]. Applying Theorem \[modularity\_refined\], the series is hence a quasi-modular form of pure weight $2g+2$.
*Step 3.* By assumption ($\dag$) and the discussion after , the Gromov-Witten invariant $\mathsf{H}_{g,(1,d)}^{\textup{FLS}}$ equals the sum $$\mathsf{H}^{\textup{FLS}}_{g,(1,d)} = \sum_{2 \leq g' \leq g} c_{g',g} \mathsf{h}^{A_d, \textup{FLS}}_{g', \beta_d} \label{hHrelation}$$ for coefficients $c_{g',g} \in {{\mathbb{Q}}}$. Summing up over all $d$, we obtain $$\sum_{d \geq 1} \mathsf{H}^{\textup{FLS}}_{g,(1,d)} q^d
= \sum_{2 \leq g' \leq g} c_{g',g} \varphi_{g'}(q) \,.$$ The left hand side is homogeneous of weight $2g+2$, hence must equal the weight $2g+2$ part of the right hand side. Therefore, by , $$\sum_{d \geq 1} \mathsf{H}^{\textup{FLS}}_{g,(1,d)} q^d = f_{g+1}(q) \,.$$ By the definition of the $f_{g+1}(q)$ this shows part (ii) of the Theorem: $$\sum_{g \geq 2} u^{2g+2} \sum_{d \geq 1} \mathsf{H}^{\textup{FLS}}_{g,(1,d)} q^d =
\sum_{g \geq 2} u^{2g} f_g(q) = \left( q \frac{d}{dq} \right)^2 \frac{1}{4} K(z,\tau)^4 \,.
\label{lkjh}$$ Comparing with , also part (i) follows.
Donaldson-Thomas theory {#secdt}
=======================
Overview
--------
Let $A$ be a generic abelian surface with a curve class $\beta_{{{d'}}}$ of type $(1, {{d'}}> 0)$, and let $E$ be a generic elliptic curve. Throughout Section \[secdt\], we will work with the abelian threefold $$X = A \times E \,.$$
Here we compute the topological Euler characteristic of the stack $\operatorname{Hilb}^n(X, (\beta_{{{d'}}}, d)) / X$ in the cases ${{d'}}\in \{ 1, 2 \}$ proving Theorem \[dtthm\]. Next, we present a conjectural relationship between the Behrend function weighted Euler characteristic and the topological Euler characteristic via a simple sign change, and show how it implies Corollary\* \[dtcor\]. We discuss the motivation and plausibility for the conjecture. Our computation here is parallel to the computation of the reduced Donaldson-Thomas invariants for $K3\times E$ in [@Bryan-K3xE]. We will frequently refer to results of [@Bryan-K3xE]. The technique used was developed by Bryan and Kool in [@Bryan-Kool].
Notation {#notation}
--------
Since the translation action of $X$ on $\operatorname{Hilb}^n(X, (\beta_{{{d'}}}, d))$ has finite stabilizer, the reduced Donaldson-Thomas invariants $${\operatorname{\mathsf{DT}}}_{n, (\beta_{{{d'}}}, d)}
=
e \big(\operatorname{Hilb}^{n} (X, (\beta_{{{d'}}}, d))/X,\nu \big)
=
\sum _{k\in {{\mathbb Z}}} k\cdot e \big(\nu ^{-1} (k) \big)$$ and the topological (unweighted) Euler characteristics $${\widehat{{\operatorname{\mathsf{DT}}}}}_{n, (\beta_{{{d'}}}, d)} = e \big(\operatorname{Hilb}^{n}(X, (\beta_{{{d'}}}, d) ) /X\big) \,.$$ are well-defined. We have dropped the superscript $X$ in the notation for the Donaldson-Thomas invariants of Section 0.3.1.
We also use the short hand notation $$\operatorname{Hilb}^{n, {{d'}},d}(X) = \operatorname{Hilb}^n(X, (\beta_{{{d'}}}, d))$$ and the following bullet convention:
[**Convention.**]{} [*When an index in a space is replaced by a bullet $(\bullet)$, we sum over the index, multiplying by the appropriate variable. The result is a formal series with coefficients in the Grothendieck ring.*]{}
For example, we let $$\operatorname{Hilb}^{\bullet ,{{d'}},\bullet } (X)/X = \sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}\, [\operatorname{Hilb}^{n,{{d'}},d} (X)/X] \, p^{n}q^{d} \,,$$ which we regard as an element in $K_{0} ({\rm DM}_{{{\mathbb{C}}}}) ((p))[[q]]$, the ring of formal power series in $q$, Laurent in $p$, with coefficients in the Grothendieck ring of Deligne-Mumford stacks over ${{\mathbb{C}}}$.
Define the Donaldson-Thomas partition functions of $X$, $$\begin{aligned}
{\operatorname{\mathsf{DT}}}_{{{d'}}} &= \sum_{d \geq 0} \sum_{n \in {{\mathbb{Z}}}} {\operatorname{\mathsf{DT}}}_{n, (\beta_{{{d'}}} ,d)} \, (-p)^n q^{d} \,, \\
{\widehat{{\operatorname{\mathsf{DT}}}}}_{ {{d'}}} &= \sum_{d \geq 0} \sum_{n \in {{\mathbb{Z}}}} {\widehat{{\operatorname{\mathsf{DT}}}}}_{n, (\beta_{{{d'}}} ,d)} \, p^n q^{d} \,.\end{aligned}$$ By the bullet convention, $${\widehat{{\operatorname{\mathsf{DT}}}}}_{{{d'}}} = e\big( \operatorname{Hilb}^{\bullet, {{d'}}, \bullet}(X) / X \big) \,,$$ where we extend the Euler characteristic $$e : K_0({\rm DM}_{{{\mathbb{C}}}}) \to {{\mathbb{Q}}}$$ termwise to the ring of formal power series in $p$ and $q$ over $K_0({\rm DM}_{{{\mathbb{C}}}})$.
Vertical and diagonal loci
--------------------------
Let $p_A$ and $p_E$ be the projections of $X = A \times E$ onto the factors $A$ and $E$ respectively. We say an irreducible curve $C \subset X$ is
- *vertical*, if $p_E : C \to E$ has degree $0$,
- *horizontal*, if $p_A : C \to A$ has degree $0$,
- *diagonal*, if $p_A, p_E$ have both non-zero degree.
The various definitions are illustrated in Figure \[fig: diag, vert, and horiz curves\].
\[canvas is yz plane at x=0\] (0,0) rectangle (3,5);
\[canvas is xz plane at y=0\] (0,0) rectangle (4,5);
(0,0) rectangle (4,3);
(3,1,0)–(3,1,5) (1.0,2.5,0)– (1.0,2.5,5); (2 ,0 ,5) to \[out=90,in=-90\] (2 ,0.6 ,5) to \[out=90,in=-90\] (1.5,1.5 ,5) to \[out=90,in=-90\] (2 ,2.4 ,5) to \[out=90,in=-90\] (2 ,3 ,5);
at (0,1.5,5) [$A$]{}; at (4.2,0,3) [$E$]{}; at (4.2,0,5) [$z_{0}$]{};
\[canvas is yz plane at x=4\] (0,0) rectangle (3,5); plot (,[5\*pow(sin(28.8\*pi\*),2)]{});
\[canvas is xz plane at y=3\] (0,0) rectangle (4,5);
(0,0,5) rectangle (4,3,5); (0,0,5) rectangle (4,3,5);
Consider a subscheme $C \subset X$ which defines a point in $\operatorname{Hilb}^{n, {{d'}},d}(X)$. Since the class $p_{A \ast}[C] = \beta_{{{d'}}}$ is irreducible, there is a unique irreducible component of $C$ of dimension $1$, which is either vertical or diagonal. All other irreducible components of $C$ of dimension $1$ are horizontal.
Consider the sublocus $$\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X) \subset \operatorname{Hilb}^{n, {{d'}}, d}(X) \,, \label{6_2_vertdef}$$ parametrizing subschemes $C \subset X$ with $$C_0 \times \{ z_{0} \} \subset C$$ for some $z_{0} \in E$ and for some curve $C_0 \subset A$ of class $\beta_{d'}$. We endow $\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)$ with the natural scheme structure. It is a closed subscheme of $\operatorname{Hilb}^{n, {{d'}}, d}(X)$.
Let $\operatorname{Hilb}^{n,{{d'}}, d }_{{\mathsf{diag}}} (X)$ be the complement of the inclusion , $$\operatorname{Hilb}^{n,{{d'}}, d }_{{\mathsf{diag}}} (X) = \operatorname{Hilb}^{n, {{d'}}, d}(X) \setminus \operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X) \,.$$ Hence, every point in $\operatorname{Hilb}^{n,{{d'}}, d }_{{\mathsf{diag}}} (X)$ corresponds to a subscheme $C \subset X$, which contains a diagonal component.
Since the condition defining the subscheme is invariant under the translation action of $X$, we have an induced action of $X$ on $\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)$ and its complement. We exhibit the stack $\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)/X$ as a global quotient stack of a scheme by a finite group of order ${{d'}}^{2}$ as follows.
Let $L \to A$ be a fixed line bundle on $A$ with $c_1(L) = \beta_{{{d'}}}$, and let $z_0 \in E$ be a fixed point. Consider the subscheme $$\operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,{{d'}}, d } (X) \subset \operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)$$ parametrizing subschemes $C\subset X$ with $C_0 \times \{ z_0 \} \subset C$ for some $$C_0 \in |L| \,. \label{dt_fix_cdn}$$ The stabilizer of $\operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,{{d'}}, d } (X)$ under the translation action of $X$ is the subgroup $$\label{6_2_subgrp}
{\mathop{\rm Ker}\nolimits}(\phi :A\to \widehat{A} )\subset A \,,$$ where $\phi :a\mapsto L\otimes t^{*}_{a}L^{-1}$ and $t_{a}:A\to A$ denotes the translation by $a\in A$. By , the subgroup is isomorphic to ${{\mathbb Z}}_{{{d'}}} \times {{\mathbb Z}}_{{{d'}}}$. Hence, we have the stack equivalence $$\label{6_2_abc}
\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)/X \cong \operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,{{d'}}, d }(X)/({{\mathbb Z}}_{{{d'}}} \times {{\mathbb Z}}_{{{d'}}}) \,.$$
Proof of Theorem \[eulthm\] (i)
-------------------------------
Let $L$ be a line bundle on $A$ with $c_{1}(L)= \beta_{1}$, and let $$C_0 \in |L|$$ be the unique nonsingular genus $2$ curve in $|L|$. Since $L$ has type $(1,1)$, the class $c_1(L)$ is a principal polarization of $A$. In particular, $A$ is isomorphic to the Jacobian $J$ of $C_0$.
*Step 1.* Every irreducible diagonal curve $C \subset X$ in class $(\beta_1, d)$ maps isomorphically to $C_0$ and, therefore, induces a non-constant map $$C_0 \to E \,.$$ Dualizing, we obtain a non-constant map $E \to J(C_0) \cong A$, whose image is an abelian subvariety of $A$ of dimension $1$. Hence, by the genericity of $A$, [*no*]{} diagonal curve exists and $\operatorname{Hilb}^{n,{{d'}}, d }_{{\mathsf{diag}}} (X)$ is empty.
Since there are no diagonal curves, we write $$\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,{{d'}}, d }(X) = \operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,{{d'}}, d }(X) \,.$$ Then, by the equivalence with ${{d'}}= 1$, $$e( \operatorname{Hilb}^{n,{{d'}}, d } (X) / X )
= e(\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,{{d'}}, d }(X) ) \,.$$ Using the bullet convention, we find $${\widehat{{\operatorname{\mathsf{DT}}}}}_{1} =e\big(\operatorname{Hilb}^{\bullet,1,\bullet }_{{\mathsf{fixed}}} (X)
\big) \,.$$
*Step 2.* Let ${\widehat{X}}_{C_{0}\times E}$ be the formal completion of $X$ along the closed subvariety $C_{0}\times E$, and let $$U=X \setminus C_{0}\times E$$ be the open complement. The subschemes $\{{\widehat{X}}_{C_{0}\times E},U \}$ forms a fpqc cover of $X$. By fpqc descent, subschemes in $X$ are determined by their restrictions to ${\widehat{X}}_{C_{0}\times E}$ and $U$. Since subschemes parameterized by $\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} (X)$ are disjoint unions of components contained entirely in ${\widehat{X}}_{C_{0}\times E}$ or $U$, see Figure \[figure\_6\_2\_11curve\], there is no overlap condition for descent.
\[canvas is yz plane at x=0\] (0,0) rectangle (3,5);
\[canvas is xz plane at y=0\] (0,0) rectangle (4,5);
(0,0) rectangle (4,3); in [2]{} in [2.5]{} [ (,,0)– (,,5); in [0,0.1,...,5.1]{} (,,) – ([+0.15\*cos(24\*pi\*)]{},[+0.15\*sin(24\*pi\*)]{},) ; ]{} in [3.2]{} in [0.7]{} [ (,,0)– (,,5); in [0,0.1,...,5.1]{} (,,) – ([+0.15\*cos(36\*pi\*)]{},[+0.15\*sin(36\*pi\*)]{},) ; ]{}
(2,0.5,0)–(2,0.5,5) (1.0,2.5,0)– (1.0,2.5,5); (2 ,0 ,5) to \[out=90,in=-90\] (2 ,0.6 ,5) to \[out=90,in=-90\] (1.5,1.5 ,5) to \[out=90,in=-90\] (2 ,2.4 ,5) to \[out=90,in=-90\] (2 ,3 ,5); in [(1.6,1.2,5),(2.5,2,0), (3,1,0),(1,2.5,4.5),(2,2.5,5),(2,2.5,3)]{} circle (0.1); at (0,1.5,5) [$A$]{}; at (4.2,0,3) [$E$]{}; at (4.2,0,5) [$z_{0}$]{}; at (2,0,5)[$C_{0}$]{}; (0,0,5) rectangle (4,3,5);
\[canvas is yz plane at x=4\] (0,0) rectangle (3,5);
\[canvas is xz plane at y=3\] (0,0) rectangle (4,5);
Consequently, we can stratify $\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} (X)$ by locally closed subsets isomorphic to the product of Hilbert schemes of ${\widehat{X}}_{C_{0}\times E}$ and $U$ respectively. The result is succinctly expressed as an equality in the Grothendieck ring $K_{0} ({\rm Var}_{{{\mathbb{C}}}}) ((p))[[q]]$: $$\label{6_3_aaaa}
\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet } (X) = \operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet } ({\widehat{X}}_{C_{0}\times E}) \cdot \operatorname{Hilb}^{\bullet ,0,\bullet } (U)\, ,$$ where we regard $\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} ({\widehat{X}}_{C_{0}\times E})$ and $\operatorname{Hilb}^{n,0,d} (U)$ as subschemes of $\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} (X)$ and $\operatorname{Hilb}^{n,0,d} (X)$ respectively. Taking Euler characteristics in , we obtain $$\label{6_3_dtdt}
{\widehat{{\operatorname{\mathsf{DT}}}}}_{1}
= e\big(\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet } ({\widehat{X}}_{C_{0}\times E}) \big)
\cdot e\big(\operatorname{Hilb}^{\bullet ,0,\bullet } (U) \big) \,.$$
*Step 3.* We calculate the second factor $e\big(\operatorname{Hilb}^{\bullet ,0,\bullet } (U) \big)$.
The $E$ action on $U$ induces an action of $E$ on $\operatorname{Hilb}^{n ,0,d } (U)$. This new $E$ action exists because the fixed condition only concerns the $\operatorname{Hilb}^{n ,1,d}_{{\mathsf{fixed}}} ({\widehat{X}}_{C_{0}\times E})$ factors and is independent of $U$ and $\operatorname{Hilb}^{n ,0,d } (U)$.
Since a scheme with a free $E$ action has trivial Euler characteristic, we have $$e \big( \operatorname{Hilb}^{n,0,d} (U)\big) =e \big( \operatorname{Hilb}^{n,0,d}(U)^{E}\big) \,,$$ where $\operatorname{Hilb}^{n,0,d} (U)^{E}$ is the fixed locus of the $E$-action on $\operatorname{Hilb}^{n,0,d} (U)$. Every element of $\operatorname{Hilb}^{n,0,d} (U)^{E}$ corresponds to an $E$-invariant subscheme, or equivalently, is of the form $Z\times E$ for a zero-dimensional subscheme $Z\subset A \setminus C_{0}$ of length $d$. Since $\chi ({\mathcal{O}}_{Z\times E})=0$ for every such $Z$, we find $$\label{UUU_result_123}
\begin{aligned}
e\big(\operatorname{Hilb}^{\bullet ,0,\bullet } (U) \big)
&= e\Big(\sum _{d \geq 0} \operatorname{Hilb}^{d} (A \setminus C_{0})\, q^{d} \Big) \\
&= \Big(\prod _{m \geq 1} (1-q^{m})^{-1} \Big)^{e (A \setminus C_{0})} \\
&=\prod _{m \geq 1} (1-q^{m})^{-2} \,.
\end{aligned}$$ We have used Göttsche’s formula for the Euler characteristic of the Hilbert scheme of points of a surface [@Gottsche].
*Step 4.* We calculate the first factor $e\big(\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet } ({\widehat{X}}_{C_{0}\times E}) \big)$.
Consider the constructible morphism[^24] $$\label{hhhrhomapdef}
\rho _{d} : \operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} ({\widehat{X}}_{C_{0}\times E}) \to \operatorname{Sym}^{d} (C_{0}) \,,$$ defined as follows. Let $[C] \in \operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} ({\widehat{X}}_{C_{0}\times E})$ be a scheme with curve support $C_{0} \times z_0
\cup _{i} (x_i \times E)$ and multiplicity $a_{i}$ along $x_i \times
E$. Then $$\rho_d([C]) = \sum_{i}a_{i}x_{i}\in \operatorname{Sym}^{d} (C_{0}) \,.$$ Hence, $\rho_d([C])$ records the intersection (with multiplicities) of $C_0$ with all horizontal components of $C$, see Figure \[fig\_6\_2\_lkl\].
(0,0,-0.3) rectangle (5,4,-0.3); (0,0,-.3)– (0,0,.3) (5,0,-.3)– (5,0,.3) (0,4,-.3)– (0,4,.3) (5,4,-.3)– (5,4,.3); (2.5,-0.2)– (2.5,-1.3); (0,-1.5)– (5,-1.5); (5.2,2)node \[above\][${\widehat{X}}_{C_{0}\times E}$]{}; (5.2,-1.5)node\[above\][$C_{0}$]{}; (3,0,0)– (3,4,0); (3,4.0,0) node\[left\][$x_{i}\times E$]{}; (3,-1.5) circle\[radius=.1\] node\[right\][$a_{i}x_{i}$]{}; in [0.4,1.5,4.2,4.5]{} [(,0,0)– (,4,0); (,-1.5) circle\[radius=.1\] ;]{}
(0,1)node\[below,black\][$ C_{0}\times \{z_{0} \}$]{}– (5,1);
in [3]{} [ (,0)– (,4); in [0,0.1,...,4.1]{} (,,0 ) – ([+0.25\*cos(64\*pi\*)]{},,[0.25\*sin(64\*pi\*)]{} ) ; ]{}
in [(1,1),(0.4,2.2),(2,2), (3,3),(3.8,.5)]{} circle (0.1);
(0,0,.3) rectangle (5,4,.3);
We determine the Euler characteristic of $\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet,1,\bullet } ({\widehat{X}}_{C_{0}\times E})$ by computing the Euler characteristic of $\operatorname{Sym}^{d} (C_{0})$, weighted by the constructible function given by the Euler characteristic of the fibers of $\rho_{d}$. Hence, we write $$\begin{aligned}
e\big(\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} ({\widehat{X}}_{C_{0}\times E}) \big) &= \int _{\operatorname{Hilb}_{{\mathsf{fixed}}}^{n,1,d} ({\widehat{X}}_{C_{0}\times E}) } 1\,\, de\\
&=\int _{\operatorname{Sym}^{d}C_{0}} (\rho _{d})_{*} (1) \,\,de \,,\end{aligned}$$ where $de$ is the measure on constructible subsets induced by the Euler characteristic and $\rho_{d \ast}(1)$ denotes integration along the fiber. By writing $$\operatorname{Sym}^{\bullet } C_{0} = \sum _{d \geq 0}\operatorname{Sym}^{d}C_{0}\,q^{d}$$ and extending the integration to the $\bullet$ notation termwise, we obtain $$\label{6_3_jfhg}
e\big(\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet} ({\widehat{X}}_{C_{0}\times E}) \big) =\int _{\operatorname{Sym}^{\bullet } C_{0}} \rho_{*} (1) \,\,de$$ where the measureable function $\rho _{*} (1)$ is given by $$\rho _{*} (1) \big( \sum _{i}a_{i}x_{i} \big) = e\Big(\rho ^{-1} \big(\sum_{i}a_{i}x_{i}\big) \Big) \in {{\mathbb Z}}((p)) \,.$$
The following result shows that $\rho _{*} (1)$ only depends on the underlying partition of the point in the symmetric product.
\[DT1\_prop1\] We have $$\rho _{*} (1) \big(\sum _{i}a_{i}x_{i}\big) = \big(p^{1/2} (1-p)^{-1}
\big)^{e (C_{0})}\prod _{i}F (a_{i})
$$ where $$\sum _{a \geq 0} F (a)\,q^{a} = \prod _{m \geq 1} \frac{(1-q^{m})}{(1-pq^{m}) (1-p^{-1}q^{m})} \,.$$
The proof of Proposition \[DT1\_prop1\] is identical to the proof of [@Bryan-K3xE Proposition 4.1 and Lemma 4.3] with $e(C_0) = -2$ here (instead of Euler characteristic $2$ in [@Bryan-K3xE]).
We apply the following result regarding weighted Euler characteristics of symmetric products.
\[DT1\_lem1\] Let $S$ be a scheme, and let $\operatorname{Sym}^{\bullet } (S)=\sum _{d \geq 0} \operatorname{Sym}^{d} (S)\, q^{d}$. Let $G$ be a constructible function on $\operatorname{Sym}^{\bullet}(S)$ such that $$G \big( \sum _{i}a_{i}x_{i} \big)=\prod _{i}g(a_{i}) \,,$$ for a function $g$ with $g(0) = 1$. Then $$\int _{\operatorname{Sym}^{\bullet }S} G\,\, de = \Big(\sum _{a \geq 0} g
(a)\, q^{a} \Big)^{e (S)} \,.$$
An elementary proof of Lemma \[DT1\_lem1\] is given in [@Bryan-Kool], but see also [@Bryan-K3xE Lemma 4.2].
After applying Proposition \[DT1\_prop1\] and Lemma \[DT1\_lem1\] to , we obtain $$\begin{aligned}
e\big(\operatorname{Hilb}_{{\mathsf{fixed}}}^{\bullet ,1,\bullet} ({\widehat{X}}_{C_{0}\times E}) \big)
& = p^{-1} (1-p)^{2}\Big(\sum _{a \geq 0} F(a)\,q^{a} \Big)^{-2} \\
& = p^{-1} (1-p)^{2} \prod _{m \geq 1}\frac{(1-pq^{m})^2 (1-p^{-1}q^{m})^2}{(1-q^{m})^2} \,.\end{aligned}$$ Using , , and the definition of $K(p,q)$, we obtain the evaluation of part (i) of Theorem \[eulthm\].
Proof of Theorem \[eulthm\] (ii)
--------------------------------
Let $A$ be a generic abelian surface with curve class $\beta_2$ of type $(1,2)$, and let $L \to A$ be a fixed line bundle with $c_1(L) = \beta_2$. The linear system $$|L| = {\mathbb{P}}^1$$ is a pencil of irreducible genus $3$ curves. The generic curve in the pencil is nonsingular, but there are exactly 12 singular curves (each of which has a single nodal), see [@BLA].
By the disjoint union $$\operatorname{Hilb}^{n,{{d'}}, d } (X) = \operatorname{Hilb}^{n,{{d'}}, d }_{{\mathsf{diag}}} (X)
\, \sqcup \,\operatorname{Hilb}_{{\mathsf{vert}}}^{n,{{d'}}, d } (X)$$ and the isomorphism , we have $$\operatorname{Hilb}^{n,2,d} (X)/X
= \operatorname{Hilb}^{n,2,d}_{{\mathsf{vert}}, {\mathsf{fixed}}} (X)/({{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2) \sqcup \operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}}} (X)/X \,.$$ Using the bullet convention, it follows $${\widehat{{\operatorname{\mathsf{DT}}}}}_{2} =
\frac{1}{4} e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big)
+
e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{diag}}} (X) / X \big) \,.
\label{6_2_dt2sum_start}$$
*Step 1.* We begin to evaluate $e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big)$. Consider the map $$\tau :\operatorname{Hilb}^{\bullet ,2,\bullet }_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \to |L| = {{\mathbb{P}}}^{1} \,,$$ which maps a subscheme $C$ to the divisor in $|L|$ associated to $p_A(C)$. The fiber of $\tau$ over a point $C \in |L|$, denoted $$\operatorname{Hilb}_{C}^{n,2,d} (X)\subset \operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,2,d} (X) \,,$$ is the sublocus of $\operatorname{Hilb}_{{\mathsf{vert}},{\mathsf{fixed}}}^{n,2,d} (X)$ which parametrizes curves which contain the curve $C \times \{ z_0 \}$.
As we have done in , we may write $$e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big) = \int_{|L|} \tau_{\ast}(1) de$$ where $\tau_{\ast}(1)$ denotes the constructible function obtained by integration along the fiber: $$\tau_{\ast}(1)( [C] ) = e\big( \operatorname{Hilb}_{C}^{\bullet,2,\bullet} (X) \big) \,.$$
*Step 2.* Let $C \subset A$ be a curve in $|L|$. Following a strategy similar to the proof of part (i), we will compute explicit expressions for $\tau_{\ast}(1)( [C] )$ depending only upon whether $C$ is nodal or not.
Following Step 2 of the proof of part (i), we have $$\operatorname{Hilb}^{\bullet ,2,\bullet }_{C} (X) =\operatorname{Hilb}^{\bullet ,2,\bullet }_{C} ({\widehat{X}}_{C\times E} ) \cdot \operatorname{Hilb}^{\bullet ,2,\bullet }_{C} ( X \setminus C\times E ) \,.$$ Using the extra $E$ action on the second factors, we obtain $$e\big(\operatorname{Hilb}_{C}^{\bullet ,2,\bullet } (X) \big) =e\big(\operatorname{Hilb}_{C}^{\bullet ,2,\bullet } ({\widehat{X}}_{C\times E}) \big)\cdot \prod _{m \geq 1} (1-q^{m})^{- e(A \setminus C)} \,.
\label{snd_factor}$$
For the first factor, we use the map $$\rho : \operatorname{Hilb}^{\bullet ,2,\bullet }_{C}({\widehat{X}}_{C\times E} ) \to \operatorname{Sym}^{\bullet } (C)$$ which records the location and multiplicity of the horizontal components (and has already appeared in ).
*Step 3.* If $C$ is nonsingular, we apply Proposition \[DT1\_prop1\] with $C$ in place of $C_0$ for the integration along the fiber of $\rho$. By Lemma \[DT1\_lem1\], we have $$\begin{aligned}
e\big(\operatorname{Hilb}_{C}^{\bullet ,2,\bullet }({\widehat{X}}_{C\times E}) \big)
&=\int _{\operatorname{Sym}^{\bullet }C} \rho _{*} (1) \,de\\
&= \big(p^{1/2} (1-p)^{-1} \big)^{e (C)}\Big(\sum_{a\geq0} F (a)q^{a} \Big)^{e (C)}\\
&=p^{-2} (1-p)^{4} \prod_{m\geq1} \frac{(1-pq^{m})^{4} (1-p^{-1}q^{m})^{4}}{(1-q^{m})^{4}} \,.\end{aligned}$$ Using with $e(A \setminus C) = 4$, we find $$\tau_{\ast}(1)([C]) = e\big(\operatorname{Hilb}_{C}^{\bullet ,2,\bullet } (X) \big) = K (p,q)^{4} \,. \label{tau_CC}$$
\[canvas is yz plane at x=0\] (0,0) rectangle (3,5);
\[canvas is xz plane at y=0\] (0,0) rectangle (4,5);
(0,0) rectangle (4,3); in [2.8]{} in [1.5]{} [ (,,0)– (,,5); in [0,0.1,...,5.1]{} (,,) – ([+0.15\*cos(24\*pi\*)]{},[+0.15\*sin(24\*pi\*)]{},) ; ]{} (3,.5,0)–(3,.5,5) (1.0,2.5,0)– (1.0,2.5,5) (1.5,1,0)– (1.5,1,5); (3 ,0 ,5) to \[out=90,in=0\] (2.3 ,1.8 ,5) to \[out=180,in=90\] (2 ,1.5 ,5) to \[out=270,in=180\] (2.3 ,1.2 ,5) to \[out=0,in=270\] (3 ,3 ,5);
in [ (2.8,1.5,5), (3,.5,.5), (3,.5,2.5), (2.95,1,5), (2.8,1.5,2) ]{} circle (0.1); at (3,3,5) [$N$]{};
at (0,1.5,5) [$A$]{}; at (4.2,0,3) [$E$]{}; at (4.2,0,5) [$z_{0}$]{};
\[canvas is yz plane at x=4\] (0,0) rectangle (3,5);
\[canvas is xz plane at y=3\] (0,0) rectangle (4,5);
(0,0,5) rectangle (4,3,5); (0,0,5) rectangle (4,3,5);
*Step 4.* Let $C = N \in |L|$ be a curve with a nodal point $z \in C$. The corresponding moduli space $\operatorname{Hilb}_{N}^{n,2,d} (X)$ is depicted in Figure \[fig: subschemes parameterized by Hilb\_N\]. We have the following result.
\[DT2\_prop2\] Let $x_{1},\dotsc ,x_{l}\in N \setminus \{z \}$, then $$\label{dt2_prop2_eqn}
\rho _{*} (1)\Big(bz + \sum _{i=1}^{l}a_{i}x_{i}\Big) =p^{-2} (1-p)^{4}N (b)\prod _{i=1}^{l}F (a_{i})$$ where $$\sum _{b\geq0} N (b) q^{b} =\prod _{m\geq1} (1-q^{m})^{-1}
\cdot \bigg(1+\frac{p}{(1-p)^{2}} + \sum _{d\geq1}\sum _{k|d} k (p^{k}+p^{-k})q^{d}\bigg) \,.$$
The proof is identical to the proof of the corresponding statement for contributions of nodal curves in the $K3\times E$ geometry of [@Bryan-K3xE Section 5]. The only difference is that in our case $e (N \setminus \{z \})=-4$, whereas in the $K3$ case $e(N \setminus \{z \})=0$. The different Euler characteristic results in the different prefactor $p^{-2} (1-p)^{4}$ in . The prefactor in general is $$\big(p^{1/2} (1-p)^{-1} \big)^{e (N \setminus \{z\})} \,.$$
The geometry of the term $N(b)$ arises as the contribution $$N (b) = e\big(\operatorname{Hilb}^{\bullet ,2,b}({\widehat{X}}_{\{z \}\times E} ) \big) \,.$$ In [@Bryan-K3xE], the right hand side is expressed in terms of the topological vertex. By results of [@Bloch-Okounkov], we obtain the closed form of Proposition \[DT2\_prop2\].
By Proposition \[DT2\_prop2\] and Lemma \[DT1\_lem1\], we obtain $$\begin{aligned}
& e\big(\operatorname{Hilb}_{N}^{\bullet ,2,\bullet } ({\widehat{X}}_{N\times E}) \big)
= \int _{\operatorname{Sym}^{\bullet }N} \rho _{*} (1) \, de \\
={} & p^{-2 } (1-p)^{4} \int _{\operatorname{Sym}^{\bullet } (N \setminus \{z \})} \prod _{i} F (a_{i})\, de \int _{\operatorname{Sym}^{\bullet } (\{z \})} N (b)\, de \\
={} & p^{-2} (1-p)^{4}\Big(\sum_{a \geq 0} F (a) q^{a} \Big)^{e(N \setminus \{z \})}\cdot \Big(\sum _{b \geq 0} N (b)q^{b} \Big)\\
={} &p^{-2} (1-p)^{4} \bigg( \prod _{m \geq 1}\frac{(1-pq^{m})^{4} (1-p^{-1}q^{m})^{4}}{(1-q^{m})^{5}} \bigg) \\
& \quad \quad \quad \quad \quad \quad \quad \quad\cdot
\bigg(1+\frac{p}{(1-p)^{2}} + \sum _{d \geq 1}\sum _{k|d} k (p^{k}+p^{-k})q^{d}\bigg) \,.
$$ By with $e(A \setminus N) = 3$, we find $$\label{tau_NN}
\begin{aligned}
& \tau_{\ast}(1)([N]) = e\big(\operatorname{Hilb}_{N}^{\bullet ,2,\bullet } (X) \big) \\
={} & K(p,q)^4 \cdot \bigg(1+\frac{p}{(1-p)^{2}} + \sum _{d \geq 1}\sum _{k|d} k (p^{k}+p^{-k})q^{d}\bigg) \,.
\end{aligned}$$
*Step 5.* We complete the calculation of $e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big)$.
By and , the function $\tau_{\ast}(1)([C])$ only depends upon whether $C \in |L|$ is nodal or not. Therefore, $$\begin{aligned}
& e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big)
= \int_{|L|} \tau_{\ast}(1) de \\
={} & e( {\mathbb{P}}^1 \setminus 12 \text{ points} ) \cdot K(p,q)^4 + e(12 \text{ points}) \cdot \tau_{\ast}(1)(N) \\
={} & {-10} K(p,q)^4 + 12 K(p,q)^4 \cdot \bigg(1+\frac{p}{(1-p)^{2}} + \sum _{d\geq1}\sum _{k|d} k (p^{k}+p^{-k})q^{d}\bigg) \\
={} & K(p,q)^4 \cdot \bigg(2+12 \frac{p}{(1-p)^{2}} + 12 \sum _{d\geq 1}\sum _{k|d} k (p^{k}+p^{-k})q^{d}\bigg) \,.\end{aligned}$$
*Step 6.* We compute the contribution $e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{diag}}} (X)/X \big)$ arising from the locus of curves with a diagonal component.
By Lemma \[graph\_curves\] of Section \[secg3lc\], there are $$12\, \sigma \left( \frac{d}{2} \right)\, \delta_{d, \text{even}}$$ isolated translation classes of diagonal curves of class $(\beta_{2},d)$. Moreover, the translation action of $X$ on each translation class is free.
Choose one representative from each $X$-orbit of the diagonal classes. Let $$\label{gg55}
\operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}},{\mathsf{fixed}}} (X) \subset \operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}}} (X)$$ be the subscheme parameterizing curves, who contain one of the chosen representatives. The moduli space defines a slice for the action of $X$ on $\operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}}} (X)$, $$\operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}}} (X)/X\cong \operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}},{\mathsf{fixed}}} (X) \,.$$
The contribution of such subschemes to the Euler characteristic is computed precisely as the contribution with a genus $3$ vertical component in Step 3 above. Taking into account the number of diagonal curves and their degree in the horizontal direction, we find $$\begin{aligned}
e\big(\operatorname{Hilb}^{\bullet ,2,\bullet }_{{\mathsf{diag}},{\mathsf{fixed}}} (X) \big)
&=e\big(\operatorname{Hilb}_{C}^{\bullet ,2,\bullet } (X) \big) \cdot \Big( 12 \sum_{d \geq 1}\sum_{k|d} k q^{2d} \Big) \\
&= K(p,q)^4 \cdot \Big( 12 \sum_{d \geq 1}\sum_{k|d} k q^{2d} \Big) \,.\end{aligned}$$
*Step 7.* We have calculated all terms in the sum in Steps 5 and 6. After summing, the proof of part (ii) of Theorem \[eulthm\] is complete.
The Behrend function. {#subsec: putting in the Behrend function}
---------------------
In the cases ${{d'}}\in \{1, 2 \}$, we conjecture that the Behrend function weighted Euler characteristic of the Hilbert schemes differs from the ordinary Euler characteristic by a factor of $\pm (-1)^{n}$. Here, $n$ is the holomorphic Euler characteristic, and the overall sign depends upon whether the component of the Hilbert scheme corresponds to subschemes with diagonal curves or vertical curves.
The Behrend function on the quotient $$\nu :\operatorname{Hilb}^{n,{{d'}},d} (X)/X \to {{\mathbb{Z}}}$$ induces, by our identification of the various components with different slices of the $X$-action, constructible functions on $$\operatorname{Hilb}^{n,1,d}_{{\mathsf{fixed}}} (X) \,, \quad \operatorname{Hilb}^{n,2,d}_{{\mathsf{vert}},{\mathsf{fixed}}} (X)\,, \quad \text{and } \quad \operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}},{\mathsf{fixed}}} (X) \,.$$ We will denote these functions by $\nu$ as well and write $e( \cdot, \nu )$ for the topological Euler characteristic weighted by $\nu$.
\[conj\_Behrend\] We have $$\begin{aligned}
e \big(\operatorname{Hilb}^{n,1,d}_{{\mathsf{fixed}}} (X)\big)& = - (-1)^{n} e \big(\operatorname{Hilb}^{n,1,d}_{{\mathsf{fixed}}} (X),\nu \big) \,,\\
e \big(\operatorname{Hilb}^{n,2,d}_{{\mathsf{vert}},{\mathsf{fixed}}} (X)\big)& = - (-1)^{n} e \big(\operatorname{Hilb}^{n,2,d}_{{\mathsf{vert}},{\mathsf{fixed}}} (X),\nu \big) \,,\\
e \big(\operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}},{\mathsf{fixed}}} (X)\big)& = + (-1)^{n} e \big(\operatorname{Hilb}^{n,2,d}_{{\mathsf{diag}},{\mathsf{fixed}}} (X),\nu \big) \,.\end{aligned}$$
Assuming Conjecture \[conj\_Behrend\], we prove Corollary\* \[dtcor\].
In case ${{d'}}= 1$, we have, by Conjecture \[conj\_Behrend\], $${\operatorname{\mathsf{DT}}}_1= - {\widehat{{\operatorname{\mathsf{DT}}}}}_1 \,.$$ Part (i) of Corollary\* \[dtcor\] hence follows from part (i) of Theorem \[eulthm\]. In case ${{d'}}= 2$, we have, following , $$\begin{gathered}
{\operatorname{\mathsf{DT}}}_2 =\\
\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}
\bigg(
\frac{1}{4} e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) , \nu \big)
+
e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{diag}},{\mathsf{fixed}}} (X) , \nu \big)
\bigg)
(-p)^n q^d \,.\end{gathered}$$ By Conjecture \[conj\_Behrend\], the right side equals $$\sum_{d \geq 0}\sum_{n \in {{\mathbb{Z}}}}
\bigg(
{-\frac{1}{4}} e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{vert}},{\mathsf{fixed}}} (X) \big)
+
e\big(\operatorname{Hilb}^{\bullet ,2,\bullet}_{{\mathsf{diag}},{\mathsf{fixed}}} (X) \big)
\bigg)\,
p^n q^d \,.$$ These terms have been calculated in Steps 5 and 6 of the proof of Theorem \[eulthm\] (ii). Summing up, we obtain $$K(p,q)^4 \cdot \bigg( {-3} \wp(p,q) - \frac{1}{4} + 6 \sum_{d \geq 1} \sum_{k | d} k (2 q^{2d} - q^d) \bigg) \,,$$ where $$\wp(p,q) = \frac{1}{12} + \frac{p}{(1-p)^2} + \sum_{d \geq 1}\sum_{k|d} k (p^k - 2 + p^{-k}) q^d$$ is the Weierstrass elliptic function expanded in $p$ and $q$. Rewriting $$\begin{gathered}
{- \frac{1}{4}} + 6 \sum_{d \geq 1}\sum_{ k | d} k (2 q^{2d} - q^d) \\
= -\frac{1}{4}
\Big( 1 + 24 \sum_{d \geq 1}\sum_{ k | d} k q^d - 24 \sum_{d \geq 1} \sum_{\substack{k|d \\ k \text{ even}}} k q^{d} \Big) =
-\frac{1}{4} \vartheta_{D_4}(q)\end{gathered}$$ where $$\vartheta_{D_4}(q) = 1 + 24 \sum_{d \geq 1} \sum_{\substack{k|d \\ k \textup{ odd}}} k q^d$$ is the theta function of the $D_4$ lattice, we find $${\operatorname{\mathsf{DT}}}_2 = - K(p,q)^4 \cdot \bigg( 3 \wp(p,q) + \frac{1}{4} \vartheta_{D_4}(q) \bigg) \,.$$ Hence, part (ii) of Corollary\* \[dtcor\] follows from Lemma \[modular\_identity\] below.
\[modular\_identity\] We have $$K(p,q)^4 \cdot \bigg( 3 \wp(p,q) + \frac{1}{4} \vartheta_{D_4}(q) \bigg)
=
\frac{3}{2}K (p,q)^{4}\wp(p,q)
+\frac{3}{8}K (p^{2},q^{2})^{2} \,.$$
The Lemma is stated as an equality of formal power series. Since both sides converge for the variables $$p = e^{2 \pi i z} \quad \text{ and } \quad q = e^{2 \pi i \tau}$$ with $z \in {{\mathbb{C}}}$ and $\tau \in {{\mathbb{H}}}$, we may work with the actual functions $K(z,\tau), \wp(z,\tau)$, and $\vartheta_{D_4}(\tau)$.
The statement of the Lemma is then equivalent to $$\varphi(z,\tau) = \frac{K(2z, 2\tau)^2 }{K(z,\tau)^4} - 4 \wp(z,\tau) = \frac{2}{3} \vartheta_{D_4}(\tau) \,. \label{xuenq}$$ From the definition of $K(z,\tau)$, we obtain $$K(z+ \lambda \tau + \mu, \tau) = (-1)^{\lambda + \mu} q^{-\lambda/2} p^{-\lambda} K(z,\tau)$$ for all $\lambda, \mu \in {{\mathbb{Z}}}$. Combined with the double-periodicity of the Weierstrass $\wp$-function, this implies $$\varphi(z + \lambda \tau + \mu, \tau) = \varphi(z, \tau)$$ for all $\lambda ,\mu \in {{\mathbb{Z}}}$. Since $$K(z,\tau) = 2 \pi i z + O(z^3) \quad \text{ and } \quad \wp(z,\tau) = \frac{1}{(2 \pi i z)^2} + O(1) \,,$$ the function $\varphi(z,\tau)$ has no pole at $z = 0$. Because the only zero of $K(z,\tau)$ and $\vartheta_1(z,\tau)$ and the only pole of $\wp(z,\tau)$ in the fundamental region are at $z = 0$, the function $\varphi(z,\tau)$ is entire. By double-periodicity, $\varphi(z,\tau)$ is hence a constant only depending on $\tau$.
We evaluate $\varphi(z,\tau)$ at $z = 1/2$. We have $$\wp\bigg( \frac{1}{2}, \tau \bigg) = - \frac{1}{6} \vartheta_{D_4}(\tau) \,.$$ Since $K(1/2, \tau) \neq 0$, but $K(1, \tau) = 0$, this shows $$\varphi(z,\tau) = \varphi\bigg( \frac{1}{2}, \tau\bigg) = -4 \cdot \bigg( {- \frac{1}{6}} \vartheta_{D_4}(\tau) \bigg) = \frac{2}{3} \vartheta_{D_4}(\tau) \,.\qedhere$$
Discussion of Conjecture \[conj\_Behrend\]
------------------------------------------
The phenomenon proposed by Conjecture \[conj\_Behrend\] is parallel to the phenomenon exhibited by the Donaldson-Thomas invariants of toric Calabi-Yau threefolds. In the case of toric Calabi-Yau threefolds, the only subschemes which contribute to the DT invariants are the torus fixed subschemes, namely those which are locally given by monomial ideals. The value of the Behrend function at such a subscheme $Z$ is given by $\pm (-1)^{n}$ where $n=\chi ({\mathcal{O}}_{Z})$ and the overall sign depends only on the 1-dimensional component of $Z$ (and not on the embedded points) [@MNOP1].
One route to prove Conjecture \[conj\_Behrend\] would be to show the following two properties.
1. The motivic methods of the previous section are compatible with the Behrend function, specifically that the group actions defined on the various substrata of $\operatorname{Hilb}(X)/X$ respect the Behrend function.
2. The value of the Behrend function at a subscheme $Z$ which is formally locally given by monomial ideals is given by $\pm (-1)^{n}$ where the overall sign is positive if $Z$ contains a diagonal curve and negative if $Z$ contains a vertical curve.
Gromov-Witten theory {#GW3}
====================
Overview
--------
Let $X$ be an abelian threefold, let $g\geq 2$ be the genus, and let $\beta \in H_2(X , {{\mathbb{Z}}})$ be a curve class of type $(d_1, d_2, d_3)$ with $d_1, d_2 > 0$.
In Section \[GW3\_sec1\], we define a virtual fundamental class on the quotient stack $${{\overline M}}_{g}(X, \beta) / X \, .$$ The degree of the virtual class is the *quotient Gromov-Witten invariant* of $X$.
The reduced Gromov-Witten invariants of $X$ are defined by integration against the 3-reduced virtual class (defined in Section \[cosection\]) on the moduli space ${{\overline M}}_{g,n}(X, \beta)$. In Section \[GW3\_sec2\], we prove that these invariants are fully determined by the quotient Gromov-Witten invariants and classical intersections.
In Section \[GW3\_sec3\], we relate the quotient invariants in genus $3$ to the lattice counts of Section \[piso\]. We also prove the crucial Lemma \[graph\_curves\] needed in Section \[secdt\]. In Section \[GW3\_sec4\], we use Jacobi form techniques to show that Conjectures \[conjB\] and \[conjC\] are consistent with Theorem \[dtthm\].
Finally, we extend Conjecture \[conjC\] to all curve classes in Section \[Subsection\_imprimitive\_classes\].
Quotient invariants {#GW3_sec1}
-------------------
Since $g\geq 2$, $X$ acts on ${{\overline M}}_{g}(X, \beta)$ with finite stabilizers. Let $$q : {{\overline M}}_{g}(X, \beta) \to {{\overline M}}_g(X,\beta)/X \,. \label{GW3_101}$$ be the quotient map.
Let $0_X \in X$ be the identity element, let $${\mathop{\rm ev}\nolimits}: {{\overline M}}_{g,1}(X, \beta) \to X$$ be the evaluation map, let $\psi_1$ be the first Chern class of the cotangent line $L_1 \to {{\overline M}}_{g,1}(X, \beta)$ and let $$\pi : {{\overline M}}_{g,1}(X, \beta) \to {{\overline M}}_{g}(X, \beta)$$ be the forgetful map.
We define the reduced virtual class on ${{\overline M}}_g(X,\beta)/X$ by $$[ {{\overline M}}_g(X,\beta)/X ]^{\text{red}} = \frac{1}{2g-2} \, (q \circ \pi)_{\ast}\Big( \bigl( {\mathop{\rm ev}\nolimits}^{-1}(0_X) \cup \psi_1 \bigr) \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} \Big) \,.$$ The definition is justified by the following Lemma.
\[GW3\_pullback\_lemma\] Let $p : {{\overline M}}_{g,n}(X, \beta) \to {{\overline M}}_g(X, \beta)/X$ be the composition of the forgetful with the quotient map. Then, $$p^{\ast} [ {{\overline M}}_{g}(X, \beta)/X ]^{\textup{red}} = [ {{\overline M}}_{g,n}(X, \beta) ]^{\textup{red}} \,.$$
The map $p$ factors as $$p : {{\overline M}}_{g,n}(X, \beta) \overset{p'}{\longrightarrow} {{\overline M}}_g(X, \beta) \overset{q}{\longrightarrow} {{\overline M}}_g(X,\beta)/X$$ where $p'$ is the forgetful and $q$ is the quotient map . Since we have $$p'^{\ast} [ {{\overline M}}_g(X, \beta) ]^{\text{red}} = [ {{\overline M}}_{g,n}(X, \beta) ]^{\text{red}} \,,$$ it is enough to prove $$q^{\ast} [ {{\overline M}}_{g}(X, \beta)/X ]^{\text{red}} = [ {{\overline M}}_{g}(X, \beta) ]^{\text{red}} \,.$$
Consider the product decomposition $${{\overline M}}_{g,1}(X, \beta) = {{\overline M}}_{g,1}^{0}(X, \beta) \times X \label{GW3_102} \,,$$ where $${{\overline M}}_{g,1}^{0}(X, \beta) = {\mathop{\rm ev}\nolimits}^{-1}(0_X) \,.$$ Under the decomposition , write $$q':{{\overline M}}_{g,1}(X,\beta) \rightarrow {{\overline M}}_{g,1}^0(X,\beta)$$ for the projection to the first factor. Since the obstruction theory of ${{\overline M}}_{g,1}(X,\beta)$ is $X$-equivariant, we have $$\psi_1 \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} = q^{\prime \ast} \alpha$$ for some class $\alpha$ on ${{\overline M}}_{g,1}^{0}(X, \beta)$.
Consider the inclusion $$\iota : {{\overline M}}_{g,1}^{0}(X, \beta) \to {{\overline M}}_{g,1}(X,\beta)$$ defined by ${{\overline M}}_{g,1}^{0}(X, \beta) \times 0_X$ under and the fiber diagram $$\xymatrix{
{{\overline M}}_{g,1}(X, \beta) \ar@<+.5ex>[r]^{q'} \ar[d]^{\pi} & {{\overline M}}_{g,1}^{0}(X, \beta) \ar@<+.5ex>[l]^{\iota} \ar[d]^{\pi'} \\
{{\overline M}}_{g}(X, \beta) \ar[r]^q & {{\overline M}}_{g}(X, \beta) / X \,,
}$$ where $\pi'$ is the map induced by $\pi$. Then $$\begin{aligned}
& (2g-2) q^{\ast} [ {{\overline M}}_{g}(X, \beta)/X ]^{\text{red}} \\
={} & q^{\ast} \pi'_{\ast} q'_{\ast} \Big( \bigl({\mathop{\rm ev}\nolimits}^{-1}(0_X) \cup \psi_1 \bigr) \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} \Big) \\
={} & \pi_{\ast} q^{\prime \ast} q'_{\ast} \Big( \bigl( {\mathop{\rm ev}\nolimits}^{-1}(0_X) \cup \psi_1\bigr) \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} \Big) \\
={} & \pi_{\ast} q^{\prime \ast} \iota^{\ast} \bigl( \psi_1 \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} \bigr) \\
={} & \pi_{\ast} q^{\prime \ast} \alpha \\
={} & \pi_{\ast} \bigl( \psi_1 \cap [ {{\overline M}}_{g,1}(X, \beta) ]^{\text{red}} \bigr) \,.\end{aligned}$$ The Lemma now follows directly from the dilaton equation.
We define the [*quotient Gromov-Witten invariants*]{} of $X$ by $$\label{ccc3}
\mathsf{N}_{g, \beta} = \int_{ [ {{\overline M}}_{g}(X, \beta) / X ]^{\text{red}} } \mathsf{1} \,.$$
Reduced Gromov-Witten invariants {#GW3_sec2}
--------------------------------
Let $g \geq 2$ and let $\beta$ be a curve class of type $(d_1, d_2, d_3)$ with $d_1, d_2 > 0$. Let $$[ {{\overline M}}_{g,n}(X, \beta) ]^{\text{red}}$$ be the 3-reduced virtual class on the moduli space ${{\overline M}}_{g,n}(X,\beta)$ constructed in Section \[cosection\]. The *reduced Gromov-Witten invariants of* $X$ are defined by $$\Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,\beta}^{X, \text{red}} =
\int_{ [ {{\overline M}}_{g,n}(X,\beta) ]^{\text{red}} } \prod_{i=1}^n {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \cup \psi_i^{a_i} \,,
\label{X_GW_def}$$ for $\gamma_1, \dots, \gamma_n \in H^{\ast}(X, {{\mathbb{Q}}})$ and $a_1, \dots, a_n \geq 0$.
By the definition of the virtual class $[ {{\overline M}}_{g}(X, \beta) / X ]^{\text{red}}$ and the quotient Gromov-Witten invariants , we have $$\mathsf{N}_{g, \beta} = \frac{1}{2g-2} \cdot \big\langle \tau_1( {{\mathsf{p}}}) \big\rangle_{g,\beta}^{X, \text{red}} \,,
$$ where ${{\mathsf{p}}}\in H^6(X, {{\mathbb{Z}}})$ is the class of a point. The invariants $\mathsf{N}_{g, \beta}$ will be shown to determine all reduced Gromov-Witten invariants of $X$.
We first determine all primary Gromov-Witten invariants in terms of $\mathsf{N}_{g,\beta}$. Consider the translation action $$t : X^{n+1} \to X, \ \ \ \ (a, x_1, \dots, x_n) \mapsto (x_1 - a, \dots, x_n - a) \,. \label{GW3_translation_map}$$
\[GW3\_primary\] For $\gamma_1, \dots, \gamma_n \in H^{\ast}(X, {{\mathbb{Q}}})$, $$\Big\langle \tau_{0}(\gamma_1) \dots \tau_{0}(\gamma_n) \Big\rangle_{g,\beta}^{X, \textup{red}}
=
\mathsf{N}_{g, \beta} \cdot
\int_{t_{\ast}([X] \otimes \beta^{\otimes n})} \gamma_1 \otimes \ldots \otimes \gamma_n \,.
\label{GW3_xxgjg}$$
Let $\gamma_1, \dots, \gamma_n \in H^{\ast}(X, {{\mathbb{Q}}})$ be homogeneous classes. We may assume the dimension constraint $$\sum_{i=1}^{n} \deg(\gamma_i) = 2( 3 + n) \,$$ holds[[^25]]{} – otherwise both sides of vanish.
For every $k$, let $$\pi_k : {{\overline M}}_{g,k}(X,\beta) \to {{\overline M}}_g(X) / X$$ be the composition of the map that forgets all markings with the quotient map. By Lemma \[GW3\_pullback\_lemma\], $$[ {{\overline M}}_{g,n}(X,\beta) ]^{\text{red}} = \pi_n^{\ast} [ {{\overline M}}_{g}(X, \beta) / X ]^{\text{red}} \,,$$ hence by the push-pull formula $$\Big\langle \tau_{0}(\gamma_1) \dots \tau_{0}(\gamma_n) \Big\rangle_{g,\beta}^{X, \text{red}}
=
\int_{ [ {{\overline M}}_{g}(X, \beta) / X ]^{\text{red}} } \pi_{n \ast}\Big( \prod_{i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \Big) \,.$$ Since the map $\pi_n$ is of relative dimension $3+n$, the cohomology class $$\pi_{n \ast}\Big( \prod_{i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \Big)$$ has degree $0$. To proceed, we evaluate $\prod_{i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i)$ on the fibers of $\pi_n$.
Let $f : C \to X$ be a stable map of genus $g$ and class $\beta$, let $$[ f] \in {{\overline M}}_{g}(X, \beta) / X$$ be the associated point, and let $F$ be the (stack) fiber of $\pi_n$ over $[f]$.
By the definition of ${{\overline M}}_{g}(X,\beta)/X$ as a quotient stack [@R05], we may identify $$X = \pi_0^{-1}( [f ] ), \label{GW3_iden}$$ where the induced map $X \to {{\overline M}}_{g}(X,\beta)$ is $x \mapsto (f-x)$. Under , let $$b_0 : F \to X$$ be the map which forgets all markings. For $i \in \{ 1, \dots, n \}$, let $$b_i : F \to C$$ be the map which forgets the map and all except the $i$-th marking. The induced map $$b = (b_0, \dots, b_n) : F \to X \times C^n$$ is birational on components.
The evaluation map ${\mathop{\rm ev}\nolimits}: F \to X^n$ factors as $$F \xrightarrow{b} X \times C^n \xrightarrow{(\operatorname{id}, f, \dots, f)} X^{n+1} \xrightarrow{t} X^{n} \,,$$ where $t$ is the translation map . We find $$\begin{aligned}
\int_{F} {\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1) \cdots {\mathop{\rm ev}\nolimits}_n^{\ast}(\gamma_n)
& = \int_{ {\mathop{\rm ev}\nolimits}_{\ast} [F] } \gamma_1 \otimes \ldots \otimes \gamma_n \\
& = \int_{ t_{\ast} (\operatorname{id}, f^n)_{\ast} ([X] \otimes [C]^{\otimes n}) } \gamma_1 \otimes \ldots \otimes \gamma_n \\
& = \int_{t_{\ast}( [X] \otimes \beta^{\otimes n} ) } \gamma_1 \otimes \ldots \otimes \gamma_n \,.\end{aligned}$$ Since this only depends on $\beta$ and the $\gamma_i$, we conclude $$\pi_{n \ast}\Big( \prod_{i} {\mathop{\rm ev}\nolimits}_i^{\ast}(\gamma_i) \Big) = \Big( \int_{t_{\ast}( [X] \otimes \beta^n ) } \gamma_1 \otimes \ldots \otimes \gamma_n \Big) \cdot \mathsf{1} \,.$$ The claim of the Lemma follows.
We state the abelian vanishing relation for abelian threefolds. Let $$p : X^n \to X^{n-1} \,, \quad (x_1, \dots, x_n) \mapsto (x_2 - x_1, \dots, x_n - x_1) \,.$$
\[GW3\_abelvan\] Let $\gamma\in H^{\ast}(X^{n-1},{{\mathbb{Q}}})$ and let $a_1, \dots, a_n \geq 0$. For any $\gamma_1 \in H^{\ast}(X, {{\mathbb{Q}}})$ of degree $\deg(\gamma_1) \leq 5$, $$\int_{ [ {{\overline M}}_{g,n}(X, \beta) ]^{\textup{red}} }
{\mathop{\rm ev}\nolimits}_1^{\ast}(\gamma_1) \cup {\mathop{\rm ev}\nolimits}^{\ast} p^{\ast}(\gamma) \cup \prod_i \psi_i^{a_i} = 0 \,.$$
The proof is identical to the proof of Lemma \[abelvan\].
The full reduced descendent Gromov-Witten theory of $X$ in genus $g$ and class $\beta$ is determined from $\mathsf{N}_{g,\beta}$ by the following operations:
- the string, dilaton, and divisor equations,
- the abelian vanishing relation of Lemma \[GW3\_abelvan\],
- the evaluation by Lemma \[GW3\_primary\] of primary invariants,
- the evaluation $\big\langle \tau_1( {{\mathsf{p}}}) \big\rangle_{g,\beta}^{X, \textup{red}}=
(2g-2) \cdot \mathsf{N}_{g,\beta}$.
Let $\gamma_1, \dots, \gamma_n \in H^{\ast}(X, {{\mathbb{Q}}})$ be homogeneous classes. We must determine the Gromov-Witten invariant $$\Big\langle \tau_{a_1}(\gamma_1) \dots \tau_{a_n}(\gamma_n) \Big\rangle_{g,\beta}^{X, \text{red}} \,\label{GW3_gwdesc}$$ for $a_1, \dots, a_n \geq 0$. We may assume the dimension constraint $$\sum_{i=1}^{n} \deg(\gamma_i) + 2 a_i = 2(3+n) \,, \label{GW3_dim_constraint}$$ where $\deg( \cdot )$ is the real degree of a class in $X$. In particular, $n \geq 1$.
We proceed by induction on $n$. In case $n=1$, the insertion must be $$\tau_1({{\mathsf{p}}})\, , \ \tau_2( \gamma)\, , \ \tau_3( \gamma') \, , \ \text{or }\ \tau_4( \gamma'' )$$ for classes $\gamma, \gamma', \gamma''$ of degrees $4,2,0$ respectively. The case $\tau_1({{\mathsf{p}}})$ follows from (iv). The cases $\tau_2(\gamma)$, $\tau_3(\gamma')$, and $\tau_4(\gamma'')$ all vanish by the abelian vanishing relation (ii).
Suppose $n > 1$ and assume the Proposition is true for all $n' < n$. If $a_i = 0$ for all $i$, the statement follows from the evaluation (iii). Hence we may assume $a_1 > 0$. If $\deg(\gamma_1) < 6$, we first apply the vanishing of Lemma \[GW3\_abelvan\] for $\gamma_1$ and $$\gamma = \gamma_2 \otimes \dots \otimes \gamma_n \,.$$ We find, that can can be expressed as a sum of series $$\pm \Big\langle \tau_{a_1}(\gamma_1 \cup \delta) \tau_{a_2}(\gamma'_2) \cdots \tau_{a_n}(\gamma'_n) \Big\rangle_{g,\beta}^{X, \text{red}}$$ for homogeneous classes $\delta, \gamma'_2, \dots, \gamma'_n \in
H^{\ast}(X,\mathbb{Q})$ with $\deg(\delta) \geq 1$. The above relation [*increases*]{} the degree of $\gamma_1$. By induction on $\deg(\gamma_1)$, we may assume $\deg(\gamma_1) = 6$. By the dimension constraint , we have $$\sum_{i=2}^{n} \deg(\gamma_i) + 2 a_i = 2(n-a_1) \,,$$ hence there exists a $k \in \{ 2, \dots, n \}$, such that $\deg(\gamma_k) + 2 a_k \leq 2$.
If $a_k = 1$, then $\deg(\gamma_k) = 0$ and we use the dilaton equation. If $a_k = 0$ and $\deg(\gamma_i) \in \{ 0, 1 \}$, we use the string equation. If $a_k = 0$ and $\deg(\gamma_i) = 2$, we use the divisor equation. In each case, we reduce to Gromov-Witten invariants with less then $n$ marked points. The proof of the Proposition now follows from the induction hypothesis.
In , we defined quotient invariants $\mathsf{N}^{\text{Q}}_{g,(d_1,d_2)}$ counting genus $g$ curves on an abelian surface $A$ in class of type $(d_1, d_2)$, with $g \geq 2$ and $d_1, d_2 > 0$. By trading the FLS condition for insertions, moving the calculation to the threefold $A \times E$ via the $k=2$ case of Section \[dcosection\], and by the evaluation of Lemma \[GW3\_primary\], one obtains $$\mathsf{N}^{\text{Q}}_{g,(d_1,d_2)} = \mathsf{N}_{g, (d_1,d_2,0)} \,.$$ Hence, the quotient invariants of abelian surfaces agree with the degenerate case of the quotient invariants of abelian threefolds.
Genus 3 counts {#secg3lc}
--------------
\[GW3\_sec3\] We determine the genus $3$ invariants of $X$ using the lattice method of Section \[piso\]. The strategy is similar to the proof of Lemma \[G2lattice\].
\[G3lcc\] For all $d_1, d_2, d_3 > 0$, $$\mathsf{N}_{3, (d_1, d_2, d_3)} = 2 \nu(d_1, d_2, d_3)\,.
$$
Let $\beta$ be a curve class of type $(d_1, d_2, d_3)$ on a generic abelian threefold $X$. Since $X$ is simple, every genus $3$ stable map $$f : C \to X$$ in class $\beta$ has a nonsingular domain $C$, and induces a polarized isogeny $(\widehat{X}, \widehat{\beta}) \to (J, \theta)$.
Conversely, every simple principally polarized abelian threefold $(B, \theta)$ is the Jacobian of a unique nonsingular genus $3$ curve $C$. Hence, each polarized isogeny $(\widehat{X}, \widehat{\beta}) \to (B, \theta)$ induces a map $$f : C \xrightarrow{\mathsf{aj}} B \to X \,.$$ However, for a generic abelian threefold $X$ we have ${\mathop{\rm Aut}\nolimits}(C) = \{1\}$ and ${\mathop{\rm Aut}\nolimits}(X) = \{\pm 1\}$. The composition $$f^- = (-1) \circ f : C \to X$$ is [*not*]{} translation equivalent to $f$ and the given polarized isogeny corresponds to *two* genus $3$ stable maps up to translation.[^26]
The argument in the proof of Lemma \[G2lattice\] also shows that $X$ acts freely on ${{\overline M}}_3(X, \beta)$. The only point to verify is that given a nonsingular genus $3$ curve $C$ and the Abel-Jacobi map $\mathsf{aj} : C \to J$, the only element in $J$ fixing $\mathsf{aj}(C)$ is $0_J$. For this we consider the map $$\operatorname{Sym}^2(\mathsf{aj}) : \operatorname{Sym}^2(C) {{\ \rightarrow\ }}J \,.$$ The image of $\operatorname{Sym}^2(C)$ is a theta divisor, and is only fixed by $0_J \in J$. Then, if a point $a \in J$ fixes $\mathsf{aj}(C)$, it must also fix the image of $\operatorname{Sym}^2(C)$ under $\operatorname{Sym}^2(\mathsf{aj})$. Hence $a = 0_J$.
It follows that ${{\overline M}}_3(X, \beta) / X$ is a set of $2\nu(d_1, d_2, d_3)$ isolated reduced points.
By Lemma \[G2lattice\], Theorem \[YZA\], and Lemma \[G3lcc\], $$\mathsf{N}_{3,(1, d, d')} = 2 \mathsf{N}^{\text Q}_{2, (d, d')} = 2\sum_{k | \gcd(d,d')} \sum_{ m | \frac{ d d' }{k^2}} k^{3} m \,. \label{YZ'}$$ The right hand side of matches precisely the genus $3$ predictions of Conjectures \[conjB\] and \[conjC\].
Further, the lattice method can be adjusted to count diagonal curves in the $X = A \times E$ setting. Let $A$, $E$, and $(\beta_{{{d'}}}, d)$ be as in Section \[secdt\]. Recall that an irreducible curve $C \subset X$ is diagonal if both projections $p_A : C \to A$ and $p_E : C \to E$ are of non-zero degree.
\[graph\_curves\] For even $d$, there are $$12 \sigma\left( \frac{d}{2} \right) = 12 \sum_{k | \frac{d}{2}} k$$ isolated diagonal curves in class $(\beta_2, d)$ up to translation. All diagonal curves are nonsingular of genus $3$. The translation action of $A \times E$ on the diagonal curves is free.
Let $C$ be a diagonal curve in class $(\beta_2, d)$. Since $\beta_2$ is irreducible, the projection $p_A : C \to A$ is generically injective. The image $$C_0 = p_A(C) \subset A$$ is either a nonsingular genus $3$ curve or a nodal genus $2$ curve. We claim that the latter does not happen.
Suppose it does, and let $q : \widetilde{C} \to C$ be the normalization map. Then $p_A \circ q : \widetilde{C} \to A$ factors through an isogeny $J(\widetilde{C}) \to A$, where $J(\widetilde{C})$ is the Jacobian of the genus $2$ curve $\widetilde{C}$. We also know that $p_E \circ q : \widetilde{C} \to E$ factors through $J(\widetilde{C}) \to E$, which is surjective since $d > 0$. This contradicts the assumption that $A$ is simple.
Hence, $C_0$ is nonsingular of genus $3$ and so is $C$. As before, every such $C$ induces a polarized isogeny $$\big(\widehat{A \times E}, \widehat{(\beta_2, d)}\big) {{\ \rightarrow\ }}(J, \theta) \,.$$ Conversely, every principally polarized abelian threefold $(B, \theta)$ is either
- the Jacobian of a unique nonsingular genus $3$ curve, or
- the product of a principally polarized abelian surface and an elliptic curve, with the product polarization.
Given $(\beta_2, d)$ of type $(1, 2, d)$, we know exactly which maximal totally isotropic subgroups of ${\mathop{\rm Ker}\nolimits}(\phi_{\widehat{(\beta_2, d)}}) \cong ({{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/d)^2$ correspond to polarized isogenies $$\big(\widehat{A \times E}, \widehat{(\beta_2, d)}\big) {{\ \rightarrow\ }}(B, \theta) \label{PolisoSE}$$ to Jacobians $(B, \theta)$. They are precisely the subgroups [*not*]{} of the form $$G_1 \times G_2 \quad \text{ with } \quad G_1 < ({{\mathbb{Z}}}/2)^2 \,, \ G_2 < ({{\mathbb{Z}}}/d)^2 \,.$$ In particular, $d$ must be even for these subgroups to exist. In terms of , there are the following two possibilities:
1. $K = {{\mathbb{Z}}}/2k$ for some $k | \frac{d}{2}$, generated by $$\bigg(1, \frac{d}{2k}\bigg) \in {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/d \,,$$ together with an arbitrary element in ${\mathop{\rm Hom}\nolimits}^{\rm sym}(K, \widehat{K})$,
2. $K = {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2k$ for some $k | \frac{d}{2}$, generated by $$(1, 0),\, \bigg(0, \frac{d}{2k}\bigg) \in {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/d \,,$$ together with a non-diagonal element in ${\mathop{\rm Hom}\nolimits}^{\rm sym}(K, \widehat{K})$.
Summing up (i) and (ii), we find $$\sum_{k | \frac{d}{2}} 2k + \sum_{k | \frac{d}{2}} 2 \cdot 2k = 6 \sum_{k | \frac{d}{2}} k$$ polarized isogenies to Jacobians. We claim that each of the isogenies corresponds to [*two*]{} diagonal curves up to translation.
We have seen that a diagonal curve $C \subset A \times E$ is isomorphic to its image $C_0 \subset A$. By [@CAV Section 10.8 (1)], every nonsingular genus $3$ curve $C' \subset A$ in class $\beta_2$ admits a double cover to an elliptic curve $E'$. In particular, the Jacobian $J(C')$ is isogenous to $A \times E'$. Hence, $${{\mathbb{Z}}}/2 \subset {\mathop{\rm Aut}\nolimits}(C_0) = {\mathop{\rm Aut}\nolimits}(C) \,.$$ On the other hand, we have generically ${\mathop{\rm Aut}\nolimits}(A \times E) = {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2$. Since the Jacobian $J$ of $C$ is isogenous to $A \times E$, we also have $${\mathop{\rm Aut}\nolimits}(J) \subset {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2 \,.$$
A strong form of the Torelli theorem (see [@CAV Section 11.12 (19)]) says $${\mathop{\rm Aut}\nolimits}(C) = \begin{cases} {\mathop{\rm Aut}\nolimits}(J, \theta) & \text{ if $C$ is hyperelliptic} \\ {\mathop{\rm Aut}\nolimits}(J, \theta) / \{\pm 1\} & \text{ if $C$ is not hyperelliptic} \,. \end{cases}$$ In our case this means $${\mathop{\rm Aut}\nolimits}(C) = \begin{cases} {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2 & \text{ if $C$ is hyperelliptic} \\ {{\mathbb{Z}}}/2 & \text{ if $C$ is not hyperelliptic} \,. \end{cases}$$ To see that $C$ is generically not hyperelliptic, recall from Section \[hyp\_3\] that up to translation there are three[^27] hyperelliptic genus $3$ curves $$C'_1, \, C'_2, \, C'_3 \subset A$$ in class $\beta_2$ with ${{\mathbb{Z}}}/2$-stabilizers. For $i = 1, 2, 3$, the Jacobian $J(C'_i)$ is isogenous to $A \times E'_i$ for some $E'_i$. Hence, by taking $E$ non-isogenous to $E'_1, E'_2, E'_3$, we find that the diagonal curve $C \subset A \times E$ is not isomorphic to $C'_1, C'_2, C'_3$.
To conclude, we have ${\mathop{\rm Aut}\nolimits}(C) = {{\mathbb{Z}}}/2$ and ${\mathop{\rm Aut}\nolimits}(J) = {\mathop{\rm Aut}\nolimits}(A \times E) = {{\mathbb{Z}}}/2 \times {{\mathbb{Z}}}/2$. Therefore, each polarized isogeny gives two diagonal curves up to translation. We find in total $$2 \cdot 6 \sum_{k | \frac{d}{2}} k = 12 \sum_{k | \frac{d}{2}} k$$ diagonal curves up to translation. The proof that $A \times E$ acts freely is identical to the one given in the proof of Lemma \[G3lcc\].
Consistency check of Conjectures \[conjB\] and \[conjC\] {#GW3_sec4}
--------------------------------------------------------
Conjecture \[conjC\] expresses the invariants $\mathsf{N}_{g, (1,{{d'}},d)}$ in terms of the invariants $\mathsf{N}_{g, (1,1,d)}$. By Conjecture \[conjB\], we obtain a prediction for the Donaldson-Thomas invariants of type $(1,2,d)$ in terms of those of $(1,1,d)$. We show here that these predictions match the calculations of Theorem \[dtthm\].
For $d \geq 0$, let $f_d(p), g_d(p) \in {{\mathbb{Q}}}((p))$ be the unique Laurent series with $$\label{GW3_Kfns}
\begin{aligned}
\sum_{d \geq 0} f_d(p) q^d & = K(z,\tau)^2 \\
\sum_{d \geq 0} g_d(p) q^d & = \frac{3}{2} K(z,\tau)^4 \wp(z,\tau) + \frac{3}{8} K(2 z, 2 \tau)^2
\end{aligned}$$ under the variable change $$p = e^{2 \pi i z} \quad \text{and} \quad q = e^{2 \pi i \tau} \,.$$ The functions on the right hand side of are exactly the negative of the functions appearing in Corollary\* \[dtcor\]. The following Lemma shows that Corollary\* \[dtcor\] is consistent with Conjectures \[conjB\] and \[conjC\].
We have $$g_d(p) =
\begin{cases}
f_{2d}(p) & \text{ if } d \text{ is odd } \\
f_{2d}(p) + \frac{1}{2} f_{d/2}(p^2) & \text{ if } d \text{ is even} \,.
\end{cases}$$
We use basic results from the theory of Jacobi forms [@EZ]. We will work with the actual variables $p = e^{2 \pi i z}$ and $q = e^{2 \pi i \tau}$, where $z \in {{\mathbb{C}}}$ and $\tau \in {{\mathbb{H}}}$.
Let $\varphi_{-2,1}(z,\tau)$ be the weight $-2$, index $1$ generator of the ring of weak Jacobi forms defined in [@EZ Section 9]. We have the basic identity $$\varphi_{-2,1}(z,\tau) = K(z,\tau)^2 \,,$$ see, for example, [@DMZ Equation 4.29]. Applying the Hecke operator $\big{|}_{-2,1} V_{2}$ defined in [@EZ Section 4], we obtain the weak weight $-2$, index $2$ Jacobi form $$(\varphi_{-2,1} |_{-2,1} V_{2} )(z,\tau) = \sum_{d \geq 0} \left( f_{2d}(p) + \frac{f_{d/2}(p^2)}{2^3} \right) q^d \,,$$ where $f_a(p) = 0$ whenever $a$ is fractional. Using [@EZ Theorem 9.3] and comparing the first coefficients, we find $$\sum_{d \geq 0} \left(f_{2d}(p) + \frac{f_{d/2}(p^2)}{2^3}\right) q^d
= \frac{3}{2} K(z,\tau)^4 \wp(z,\tau) \,.$$ We conclude $$\begin{aligned}
\sum_{d \geq 0} \left( f_{2d}(p) + \frac{f_{d/2}(p^2)}{2} \right) q^d
& = \frac{3}{2} K(z,\tau)^4 \wp(z,\tau) + \sum_{d \geq 0} \frac{3}{8} f_{d/2}(p^2) q^d \\
& = \frac{3}{2} K(z,\tau)^4 \wp(z,\tau) + \frac{3}{8} K(2z, 2 \tau)^2 \\
& = \sum_{d \geq 0} g_d(p) q^d \,. \qedhere\end{aligned}$$
A formula for imprimitive classes {#Subsection_imprimitive_classes}
---------------------------------
We conjecture a multiple cover formula in all classes for the quotient invariants $\mathsf{N}_{g, (d_1, d_2, d_3)}$. The shape of the formula already appeared in the physics approach of [@MMS]. However, [@MMS] does not match the invariants $\mathsf{N}_{g, \beta}$ and our formula below is different.
Define the function $$\mathsf{n}(d_1,d_2,d_3,k) = \sum_{\delta} \delta^2$$ where $\delta$ runs over all divisors of $$\gcd\left( k, d_1, d_2, d_3, \frac{d_1 d_2}{k}, \frac{d_1 d_3}{k}, \frac{d_2 d_3}{k}, \frac{d_1 d_2 d_3}{k^2} \right) \,.$$
\[ConjMC\] For all $g \geq 2$, $d_1, d_2 > 0$, and $d_3 \geq 0$, $$\mathsf{N}_{g,(d_1,d_2,d_3)}
= \sum_{k} \mathsf{n}(d_1,d_2,d_3,k) k^{2g - 3}
\mathsf{N}_{g,\left( 1,1,\frac{d_1d_2d_3}{k^2} \right)}$$ where $k$ runs over all divisors of $\gcd( d_1 d_2, d_1 d_3, d_2 d_3 )$ such that $k^2$ divides $d_1d_2d_3$.
Recall the quotient Donaldson-Thomas invariants ${\operatorname{\mathsf{DT}}}_{n, \beta}$. Assuming deformation invariance, we write $${\operatorname{\mathsf{DT}}}_{n, \beta} = {\operatorname{\mathsf{DT}}}_{n, (d_1,d_2,d_3)}$$ if $\beta$ is of type $(d_1, d_2, d_3)$. The invariants ${\operatorname{\mathsf{DT}}}_{n, (d_1,d_2,d_3)}$ are defined whenever $n \neq 0$ or if at least two of the $d_i$ are positive.
Translating the multiple cover rule of Conjecture \[ConjMC\] via the conjectural GW/DT correspondence yields the following.
[**Conjecture**]{} ${\mathbf{E'}.}$ [*Assume $n > 0$ or at least two of the integers $d_1, d_2, d_3$ are positive. Then*]{} $${\operatorname{\mathsf{DT}}}_{n,(d_1,d_2,d_3)}
= \sum_{k} \frac{1}{k}\,
\mathsf{n}(d_1,d_2,d_3,k) (-1)^{n-\frac{n}{k}}
{\operatorname{\mathsf{DT}}}_{\frac{n}{k}, \left(1,1,\frac{d_1d_2d_3}{k^2} \right)}$$ [*where $k$ runs over all divisors of $\gcd( n, d_1 d_2, d_1 d_3, d_2 d_3 )$ such that $k^2$ divides $d_1d_2d_3$.*]{}
While Conjecture \[ConjMC\] only applies for $d_1, d_2 > 0$, we have stated Conjecture $\mathrm{E'}$ also for the degenerate cases $(0,0,d)$. Unraveling the definition yields $${\operatorname{\mathsf{DT}}}_{n,(0,0,d)}
= \frac{(-1)^{n-1}}{n} \sum_{k|\mathrm{gcd}(n,d)} k^2 \,,$$ which for $d=0$ is in perfect agreement with [@Shen], and for $d > 0$ is proven in [@OSred].
[99]{}
A. Beauville, [*Counting rational curves on $K3$ surfaces*]{}, Duke Math. J. [**97**]{} (1999), no. 1, 99–108.
K. Behrend, [*The product formula for Gromov-Witten invariants*]{}, J. Algebraic Geom. [**8**]{} (1999), no. 3, 529–541.
C. Birkenhake and H. Lange, [*Complex abelian varieties*]{}, Second edition, Grundlehren Math. Wiss., [**302**]{}, Springer-Verlag, Berlin, 2004, xii+635 pp.
S. Bloch and A. Okounkov, [*The character of the infinite wedge representation*]{}, Adv. Math. [**149**]{} (2000), no. 1, 1–60.
P. Borówka and G. K. Sankaran, [*Hyperelliptic genus $4$ curves on abelian surfaces*]{}, [arXiv:1606.09301](http://arxiv.org/abs/1606.09301).
J. Bryan, [*The Donaldson-Thomas theory of $K3 \times E$ via the topological vertex*]{}, [arXiv:1504.02920](http://arxiv.org/abs/1504.02920).
J. Bryan and T. Graber, [*The crepant resolution conjecture*]{}, in [*Algebraic geometry—Seattle 2005*]{}, Part 1, 23–42, Proc. Sympos. Pure Math., [**80**]{}, Part 1, Amer. Math. Soc., Providence, RI, 2009.
J. Bryan and M. Kool, [*Donaldson-Thomas invariants of local elliptic surfaces via the topological vertex*]{}, [arXiv:1608.07369](http://arxiv.org/abs/1608.07369).
J. Bryan and N. C. Leung, [*Generating functions for the number of curves on abelian surfaces*]{}, Duke Math. J. [**99**]{} (1999), no. 2, 311–328.
J. Bryan and N. C. Leung, [*The enumerative geometry of $K3$ surfaces and modular forms*]{}, J. Amer. Math. Soc. [**13**]{} (2000), no. 2, 371–410.
L. M. Butler, [*Subgroup lattices and symmetric functions*]{}, Mem. Amer. Math. Soc. [**112**]{} (1994), no. 539, vi+160 pp.
K. Chandrasekharan, [*Elliptic functions*]{}, Grundlehren Math. Wiss., [**281**]{}. Springer-Verlag, Berlin, 1985, xi+189 pp.
A. Dabholkar, S. Murthy, and D. Zagier, [*Quantum black holes, wall crossing, and mock modular forms*]{}, [arXiv:1208.4074](http://arxiv.org/abs/1208.4074).
O. Debarre, [*On the Euler characteristic of generalized Kummer varieties*]{}, Amer. J. Math. [**121**]{} (1999), no. 3, 577–586.
O. Debarre, [*Complex tori and abelian varieties*]{}, SMF/AMS Texts and Monographs, [**11**]{}, Amer. Math. Soc., Providence, RI; Soc. Math. France, Paris, 2005, x+109 pp.
M. Eichler and D. Zagier, , Progress in Mathematics, [**55**]{}, Birkhäuser Boston, Inc., Boston, MA, 1985, v+148 pp.
C. Faber and R. Pandharipande, [*Hodge integrals and Gromov-Witten theory*]{}, Invent. Math. [**139**]{} (2000), no. 1, 173–199.
C. Faber and R. Pandharipande, [*Relative maps and tautological classes*]{}, J. Eur. Math. Soc. (JEMS) [**7**]{} (2005), no. 1, 13–49.
C. Faber and R. Pandharipande, , in [*Handbook of moduli*]{}, Vol. I, 293–330, Adv. Lect. Math. (ALM), [**24**]{}, Int. Press, Somerville, MA, 2013.
F. Flamini, A. L. Knutsen, G. Pacienza, [*On families of rational curves in the Hilbert square of a surface*]{}, with an appendix by E. Sernesi, Michigan Math. J. [**58**]{} (2009), no. 3, 639–682.
L. Göttsche, [*The Betti numbers of the Hilbert scheme of points on a smooth projective surface*]{}, Math. Ann. [**286**]{} (1990), no. 1-3, 193–207.
L. Göttsche, [*A conjectural generating function for numbers of curves on surfaces*]{}, Comm. Math. Phys. [**196**]{} (1998), no. 3, 523–533.
L. Göttsche and V. Shende, [*The chi-y genera of relative Hilbert schemes for linear systems on Abelian and $K3$ surfaces*]{}, Algebr. Geom. [**2**]{} (2015), no. 4, 405–421.
T. Graber, [*Enumerative geometry of hyperelliptic plane curves*]{}, J. Algebraic Geom. [**10**]{} (2001), no. 4, 725–755.
T. Graber and R. Pandharipande, [*Localization of virtual classes*]{}, Invent. Math. [**135**]{} (1999), no. 2, 487–518.
M. G. Gulbrandsen, [*Donaldson-[T]{}homas invariants for complexes on abelian threefolds*]{}, Math. Z. [**273**]{} (2013), no. 1-2, 219–236.
J. Harris and D. Mumford, [*On the Kodaira dimension of the moduli space of curves*]{}, with an appendix by W. Fulton, Invent. Math. [**67**]{} (1982), no. 1, 23–88.
T. Kawai and K. Yoshioka, [*String partition functions and infinite products*]{}, Adv. Theor. Math. Phys. [**4**]{} (2000), no. 2, 397–485.
Y.-H. Kiem and J. Li, [*Localizing virtual cycles by cosections*]{}, J. Amer. Math. Soc. [**26**]{} (2013), no. 4, 1025–1050.
A. Klemm, D. Maulik, R. Pandharipande, and E. Scheidegger, [*Noether-Lefschetz theory and the Yau-Zaslow conjecture*]{}, J. Amer. Math. Soc. [**23**]{} (2010), no. 4, 1013–1040.
A. L. Knutsen, M. Lelli-Chiesa, and G. Mongardi, [*Severi varieties and Brill-Noether theory of curves on abelian surfaces*]{}, J. Reine Angew. Math. (to appear), [arXiv:1503.04465](http://arxiv.org/abs/1503.04465).
M. Kool and R. P. Thomas, [*Reduced classes and curve counting on surfaces I: theory*]{}, Algebr. Geom. [**1**]{} (2014), no. 3, 334–383.
H. Lange and E. Sernesi, [*Severi varieties and branch curves of abelian surfaces of type $(1,3)$*]{}, Internat. J. Math. [**13**]{} (2002), no. 3, 227–244.
H. Lange and E. Sernesi, [*Curves of genus $g$ on an abelian variety of dimension $g$*]{}, Indag. Math. (N.S.) [**13**]{} (2002), no. 4, 523–535.
J. Li, [*Stable morphisms to singular schemes and relative stable morphisms*]{}, J. Differential Geom. [**57**]{} (2001), no. 3, 509–578.
J. Li, [*A degeneration formula for Gromov-Witten invariants*]{}, J. Differential Geom. [**60**]{} (2002), no. 2, 199–293.
E. Looijenga, [*On the tautological ring of $M_g$*]{}, Invent. Math. [**121**]{} (1995), no. 2, 411–419.
J. Maldacena, G. Moore, and A. Strominger, [*Counting BPS blackholes in toroidal type II string theory*]{}, [arXiv:hep-th/9903163](http://arxiv.org/abs/hep-th/9903163).
D. Maulik, N. Nekrasov, A. Okounkov, and R. Pandharipande, , [Compos. Math.]{}, [**142**]{} (2006), no. 5, 1263–1285.
D. Maulik and R. Pandharipande, [*A topological view of Gromov-Witten theory*]{}, Topology [**45**]{} (2006), no. 5, 887–918.
D. Maulik and R. Pandharipande, [*Gromov-Witten theory and Noether-Lefschetz theory*]{}, in [*A celebration of algebraic geometry*]{}, 469–507, Clay Math. Proc., [**18**]{}, Amer. Math. Soc., Providence, RI, 2013.
D. Maulik, R. Pandharipande, and R. P. Thomas, [*Curves on $K3$ surfaces and modular forms*]{}, with an appendix by A. Pixton, J. Topol. [**3**]{} (2010), no. 4, 937–996.
G. Oberdieck, [*Gromov-Witten invariants of the Hilbert scheme of points of a $K3$ surface*]{}, [arXiv:1406.1139](http://arxiv.org/abs/1406.1139).
G. Oberdieck and R. Pandharipande, [*Curve counting on $K3 \times E$, the Igusa cusp form $\chi_{10}$, and descendent integration*]{}, in [*$K3$ surfaces and their moduli*]{}, 245–278, Progr. Math., [**315**]{}, Birkhäuser/Springer, 2016.
G. Oberdieck and J. Shen, [*Curve counting on elliptic Calabi-Yau threefolds via derived categories*]{}, [arXiv:1608.07073](http://arxiv.org/abs/1608.07073).
G. Oberdieck and J. Shen, [*Reduced Donaldson-Thomas invariants and the ring of dual numbers*]{}, [arXiv:1612.03102](http://arxiv.org/abs/1612.03102).
A. Okounkov and R. Pandharipande, [*Gromov-[W]{}itten theory, [H]{}urwitz theory, and completed cycles*]{}, Ann. of Math. (2) [**163**]{} (2006), no. 2, 517–560.
A. Okounkov and R. Pandharipande, [*Virasoro constraints for target curves*]{}, Invent. Math. [**163**]{} (2006), no. 1, 47–108.
R. Pandharipande and A. Pixton, [*Gromov-Witten/Pairs correspondence for the quintic 3-fold*]{}, J. Amer. Math. Soc. (to appear), [arXiv:1206.5490](http://arxiv.org/abs/1206.5490).
R. Pandharipande and R. P. Thomas, , Invent. Math. [**178**]{} (2009), no. 2, 407–447.
R. Pandharipande and R. Thomas, [*The Katz-Klemm-Vafa conjecture for $K3$ surfaces*]{}, Forum Math. Pi [**4**]{} (2016), e4, 111 pp.
R. Pandharipande and R. P. Thomas, , in [*Moduli spaces*]{}, 282–333, London Math. Soc. Lecture Note Ser., [**411**]{}, Cambridge Univ. Press, Cambridge, 2014.
G. P. Pirola, [*Curves on generic Kummer varieties*]{}, Duke Math. J. [**59**]{} (1989), no. 3, 701–708.
D. Pontoni, [*Quantum cohomology of $\operatorname{Hilb}^2({\mathbb{P}}^1 \times {\mathbb{P}}^1)$ and enumerative applications*]{}, Trans. Amer. Math. Soc. [**359**]{} (2007), no. 11, 5419–5448.
B. Poonen and M. Stoll, [*Most odd degree hyperelliptic curves have only one rational point*]{}, Ann. of Math. (2) [**180**]{} (2014), no. 3, 1137–1166.
M. Romagny, [*Group actions on stacks and applications*]{}, Michigan Math. J. [**53**]{} (2005), no. 1, 209–236.
S. C. F. Rose, [*Counting hyperelliptic curves on an [A]{}belian surface with quasi-modular forms*]{}, Commun. Number Theory Phys. [**8**]{} (2014), no. 2, 243–293.
Y. Ruan, [*The cohomology ring of crepant resolutions of orbifolds*]{}, in [*Gromov-Witten theory of spin curves and orbifolds*]{}, 117–126, Contemp. Math., [**403**]{}, Amer. Math. Soc., Providence, RI, 2006.
T. Schürg, B. Toën, and G. Vezzosi, [*Derived algebraic geometry, determinants of perfect complexes, and applications to obstruction theories for maps and complexes*]{}, J. Reine Angew. Math. [**702**]{} (2015), 1–40.
J. Shen, [*The Euler characteristics of generalized Kummer schemes*]{}, Math. Z. [**281**]{} (2015), no. 3-4, 1183–1189.
J. Tate, [*Algebraic cycles and poles of zeta functions*]{}, in [*Arithmetical algebraic geometry*]{}, (Proc. Conf. Purdue Univ., 1963), 93–110, Harper & Row, New York, 1965.
J. Wise, [*A hyperelliptic Hodge integral*]{}, Port. Math. [**73**]{} (2016), no. 3, 207–218.
J. Wise, [*The genus zero Gromov-Witten invariants of the symmetric square of the plane*]{}, Comm. Anal. Geom. [**19**]{} (2011), no. 5, 923–974.
S.-T. Yau and E. Zaslow, [*BPS states, string duality, and nodal curves on $K3$*]{}, Nuclear Phys. B [**471**]{} (1996), no. 3, 503–512.
[^1]: The domain of a [*stable map*]{} is always taken here to be connected.
[^2]: If $A = E_1 \times \dots \times E_d$, the product of elliptic curves $E_i$, the class $\beta = \sum_{i} d_i [E_i]$ has type $(d_1, \ldots,d_{\text{dim}\,A})$. See Section \[cct\] for a full discussion.
[^3]: A detailed discussion of the degenerate case is given in Section \[dcosection\]. We focus in the Introduction on the non-degenerate case.
[^4]: The class is primitive if and only if $\gcd(d_1, d_2) = 1$ (or, equivalently, the class can be deformed to type $(1, d)$), see Section \[cct\].
[^5]: Unlike in the FLS setting, the degenerate type $(1, 0)$ is included here.
[^6]: The Eisenstein series are defined by $$E_{2k}(\tau)
= 1 - \frac{4k}{B_{2k}} \sum_{m | d} m^{2k-1} q^d$$ where $B_{2k}$ is the Bernoulli number.
[^7]: The closure can be described precisely via the theory of admissible covers [@HM].
[^8]: On the other hand, generic abelian varieties of dimension $\geq 3$ contain [*no*]{} hyperelliptic curve at all, see [@Pi].
[^9]: See [@FP13] for an introduction to tautological classes on the moduli spaces of curves.
[^10]: Recently, further evidence has been obtained in [@OS].
[^11]: The class is primitive if and only if it can be deformed to type $(1, {{d'}}, d)$, see Section \[cct\].
[^12]: For genus 1, the formula is . For genus 2, the formula is given by Theorem \[YZ\_intro\]. For genus 3, the formula appears in Lemma \[G3lcc\].
[^13]: See Section \[GW3\_sec1\] for such a treatment for abelian threefolds.
[^14]: The construction of $p$ relies upon the Hilbert-Chow morphism.
[^15]: The action is in general [*not*]{} free in genus $> 2$. For example, maps from nonsingular genus $3$ hyperelliptic curves in class of type $(1, 2)$ have ${{\mathbb{Z}}}/2$-stabilizers, see Sections \[hyp\_3\] and \[secg3lc\].
[^16]: $\varphi(1) = 1$ and $\varphi(p^k) = p^k - p^{k - 1}$ for $k \geq 1$.
[^17]: See [@KT Section 4] for an algebraic proof.
[^18]: The vector space of quasi-modular forms was defined in Section \[qmfo\].
[^19]: The unstable $g=0$ cases do not play a role. For abelian varieties, genus 0 invariants (standard or reduced) are non-vanishing only in the constant map case where stability requires 3 special points.
[^20]: In particular, all Hodge classes on $A^n$ are algebraic.
[^21]: The relative points are [*not*]{} included in the counts $n'$ and $n''$.
[^22]: The geometry is pulled-back from the Artin stack of degenerations of $A/E$.
[^23]: We leave the details here to the reader.
[^24]: A constructible morphism is a map which is regular on each piece of a decomposition of its domain into locally closed subsets. Because we work with Euler characteristics and the Grothendieck group, we only need to work with constructible morphisms.
[^25]: Here, $\deg( \cdot )$ denotes the real degree of a class in $X$.
[^26]: This fact was overlooked in [@LS].
[^27]: Each corresponds to a degree $2$ polarized isogeny $(A, \beta_2) \to (B, \theta)$, with $B$ the Jacobian of a genus $2$ curve. The hyperelliptic curve is obtained by taking the preimage of the genus $2$ curve.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'There are three main aims of this paper. 1- I explain reasons why I await life to lie significantly deeper in chaos than Kauffman approach does, however still in boundary area near ‘the edge of chaos and order’. The role of negative feedbacks in stability of living objects is main of those reasons. In Kauffman’s approach regulation using negative feedbacks is not considered sufficiently, e.g. in gene regulatory model based on Boolean networks, which indicates therefore not proper source of stability. Large damage avalanche is available only in chaotic phase. It models death in all living objects necessary for Darwinian elimination. It is the first step of my approach leading to structural tendencies which are effects of adaptive evolution of dynamic complex (maturely chaotic) networks. 2- Introduction of $s \geq 2$ equally probable variants of signal (state of node in Kauffman network) as interpretively based new statistical mechanism (RSN) instead of the bias $p$ - probability of one of signal variants used in RBN family and RNS. It is also different than RWN model. For this mechanism which can be treated as very frequent, ordered phase occurs only in exceptional cases but for this approach the chaotic phase is investigated. Annealed approximation expectations and simulations of damage spreading for different network types (similar to CRBN, FSRBN and EFRBN but with $s \geq 2$) are described. Degree of order in chaotic phase in dependency of network parameters and type is discussed. By using such order life evolve. 3- A simplified algorithm called ‘reversed-annealed’ for statistical simulation of damage spreading is described. It is used for simulations presented in this and next papers describing my approach.'
address: |
Centre for Ecological Research Polish Academy of Sciences,\
M. Konopnickiej 1, Dziekanow Lesny, 05-092 Lomianki, Poland
author:
- Andrzej Gecow
title: ' More than two equally probable variants of signal in Kauffman networks as an important overlooked case, negative feedbacks allow life in chaos'
---
Boolean network ,damage spreading ,chaos ,living system ,algorithm.
Introduction - Differences with Kauffman’s approach {#ch.1}
=====================================================
The Kauffman’s approach [@Kauf69; @Derrida86; @Kauf90; @ooKauf; @Luque05; @Iguchi07] to description of living objects (or a system designed by human) using dynamical directed networks is now the most attractive and promising idea, currently developed by many authors. Reading his ‘The Origins of Order’ [@ooKauf] I was highly excited and full of admiration. However, my understanding of few basic assumptions and conclusions is different and in those areas I develop a modified approach.
Contents of the paper in brief {#ch.1.1}
--------------------------------
A. I estimate that living objects or a system designed by human are purely chaotic, however still in ‘liquid region’ - that is area near ‘the edge of chaos and order’ as suggested by Kauffman. My estimation is based mainly on:\
1- simple observation of damage spreading in reality when initial change is not of ‘known, prevented type’ (typically - damage avalanche possible only in chaotic phase and interpreted as death) (ch.1.3.2);\
2- lack of negative feedbacks in adequate level in Kauffman’s investigation of stability especially of gene regulatory model (ch.1.2.2) which makes Kauffman’s approach expectation of ‘life on the edge of order’ based on stability less convincing;\
3- finding that $s$ - number of equally probable signal variants typically should be greater than two which leads to chaos (ch.1.5);\
4- remarking that the ordered case ($s=2,K=2$) is an extreme exception in the ‘sensible’ (without fixed $K=1$) network parameter space (ch.1.5.4, ch.2.1);\
B. The main aim of this paper is to introduce [@paris; @hof; @Luque97; @Luque04; @Luque05] more than two equally probable signal variants ($s\geq 2$) to the Kauffman networks instead of only two variants in Boolean networks with bias $p$ differing their probability (thus expanding Random Boolean Network RBN to Random Signal Network RSN). The reasons are as follows:\
1- It is an important overlooked case which has a good interpretative base - it is probably typical case (second variant contains all remaining real but not interesting variants) ch.1.5.2. Equal probability of variants is a natural and useful first approximation which allows for wide investigation, e.g. to expand the expectation made for $s=2$ in annealed approximation by @Derrida86 also for $s>2$ (ch.2.2, fig.1).\
The statistical mechanisms for such case are different than the ones described using bias $p$ in the classic Boolean case and lead to different results ch.1.5.3.\
2- It is also different than sister networks using more signal variants than two: RWN (Random Walk Network) construction [@Luque04; @Luque05] which contains memory in its nodes’ states or RNS [@Luque97; @Sole00] (Random Network with multiple States) where bias $p$ is used (for one of signal variants, making it different from the other ones) for reaching edge of chaos.\
3- Boolean networks are not generally adequate. They can describe all cases but the requirement of exactly two signal variants introduces not real states or other discrepancies to described reality ch.1.5.1.\
In this paper I am going to show that such a solution can be better and lead to more adequate description and predictions. Interpretation is the base of the proposed assumption $s>2$, but I also show that using it in the Derrida ‘annealed approximation’ and the ‘semi-quenched’ model (named here ‘reversed-annealed’ ch.3.2) significantly different results are obtained (ch.2 fig.1 and ch.3.3 fig.6).
C. The ‘reversed-annealed’ algorithm (described in ch.3.2) used for simulation is a special simplified algorithm dedicated for statistical investigation of damage spreading in Kauffman and similar networks in synchronous mode. It is not an optimisation algorithm but it can be used to study an adaptation process. The main properties of this algorithm are:\
1- Only damaged nodes are calculated here.\
2- The process stops when damage fades out or achieves the equilibrium level.\
3- Therefore it is much more optimal algorithm. In the classic method two full systems are calculated - the disturbed one and the undisturbed one and at each time step they are compared to measure the damage.\
4- This simplification has small meaning for investigations presented in this paper, however, it is a critical key to the area of ‘structural tendencies’ in adaptive evolution (shortly described in [@dgec]) which are the aim of the path starting with this paper. See also [@bgec].\
5- The main feature of this algorithm is that it gives a single particular output state, however only statistically correct, instead of long cyclic attractors. In my approach avoiding attractors is useful in opposition to Kauffman’s approach where properties of attractors are very interesting.\
In ch.2 theoretical consequences of expansion to $s\geq 2$ are discussed. A simple and intuitive ‘coefficient of damage propagation’ $w$ is introduced and the Derrida’s annealed approximation is expanded.
In ch.3 influence of parameter $s$ on networks features is studied. Simulation and its algorithm for different network types are described.
Spontaneous order, negative feedbacks {#ch.1.2}
-------------------------------------
### Is the ‘spontaneous order’ a real choice of order? {#ch.1.2.1}
One of two basic theses of Kauffman’s approach is that ‘spontaneous order’ is a significant part of observed order in living objects. I can agree but what is ‘spontaneous order’? (See Appendix B.) When we are going to understand living objects, it makes sense to deal with inanimate objects or systems first. We can define our currently expected ‘abiotic’ set of possibilities - i.e. set of possible objects and their approximated distribution of probability within inanimate objects. Next, we compare it to the observed reality (also inanimate objects only) and we encounter certain differences, e.g. observed objects cover only a small part of our set. This discrepancy is the observed ‘spontaneous order’. It is here a subjective phenomenon, not a real one. It depends on our expectation which turns out to be inadequate.
Let’s consider an example[^1] from Ref[@ooKauf]. For Boolean network which Kauffman investigated and $K=2$ inputs per node (ordered case) the expected median state cycle length and the number of state cycle attractors is $N^{1/2}$. It means that a system of size $N=$100 000 (comparable to the human genome) would have about only 317 alternative asymptotic attractors which use similar number of states, but number of available states is $2^N$. It is spontaneous order that is shown there. Why we expect so many possibilities ($2^N$) if we know that there can only be so few (317)? It is property of any random Boolean network with $K=2$, not only a living one. For other low $K>2$ this ‘spontaneous order’ is also very significant, however, not so extreme.
I do not suggest that a snow star, the beautiful fractal and symmetrical structure, is not really ordered spontaneously, but what is the real set of possible shapes from which this spontaneous order makes its choice? Does it contain a square? If it contains only possible shapes of snow stars, then there is no choice between ordered or not ordered shapes. What is a difference if we find a metallic spur gear with 6 teeth? Each time when we meet a visible pattern, then we are looking for a goal. For the snow star there is no goal, there is a mechanism of its creation only, but for spur gear a goal is essential - it is the base of creation mechanism.
Now we observe living (adapted) objects with certain low $K$, we also encounter the same ‘spontaneous order’ but we find a more specific distribution. The remaining part of the observed order is a real one, i.e. it is independent of our expectations. It is the discrepancy between the abiotic state and the observed one. This part should be treated as ‘purposeful information’[@R1; @R1rgec] or ‘biological information’[@Kuppers] and Darwinian selection is its only source as stated by Kauffman. The problem is: which observed order (pattern) is the purposeful information, i.e. source of stability, because stability is a goal of living object. An overestimation of spontaneous order and connected with it homeostatic and structural stability (small initial change gives small effected change, see Appendix B) leads to type of creationism without God, i.e. certain purposeful information is taken as created (spontaneously) without natural mechanism which can create such information. (Purposeful information is a choice of cause which leads to assumed effect, i.e. goal. Only for the goal: ‘continuation of existence’ such purposeful information can be spontaneously collected using Darwinian mechanism [@R1; @R1rgec].)
Note, in such an approach there is by definition no place for real ‘spontaneous order’. However, reality of ‘spontaneous order’ shouldn’t be accepted if we can find such reality (as above) to compare (without living objects) and we cannot continue treating our large set as a real abiotic one. For complex dynamical networks it can be a problem to find such reality and someone can state that abiotic set doesn’t contain such complex networks. If so, then the same, earlier subjective phenomena stay similar to real ‘biological information’ as is in the Kauffman’s view. Note, however, that order which is connected with ‘biological information’ is connected also with ‘purposeful information’ [@R1; @R1rgec] but in the case of ‘spontaneous order’ such connection doesn’t exist which has basic interpretative importance. This change of point of view is only allowed as long as we don’t find real objects described by complex random Boolean network which are not living ones. Snow stars exist, however, they probably cannot be described so simply. I incline to the view that complex networks created without adaptive condition do not have to describe living objects only. (In the life process I include whole human activity and human products.)
### Is the ‘ultrastability’ considered in Kauffman approach? {#ch.1.2.2}
To keep a specific state, i.e. not a typical spontaneous state, there are necessary some ‘stability’ mechanisms which block spontaneous transitions to a more probable spontaneous state. The biological information is such a specific state, it can be lose and therefore it needs such a stability mechanisms. These mechanisms typically are some regulators based on negative feedbacks. Creation of these mechanisms is an effect of Darwinian selection. It is commonly observed as extremely high concentration of negative feedbacks in living objects which is even used for life definition [@Benio01; @Benio05] as its specific property.
In Kauffman approach such ‘stability’ mechanisms are named ‘ultrastability’ and are taken as one of base. They are, however, considered only on random normal average level of appearance without any preferences, which is a very large simplification leading to non-realistic modelling of stability of living objects.
Following @Ashby, Kauffman uses ‘essential variable’ to describe ultrastability: “In the context of Boolean networks, keeping the essential variables in bounds corresponds most simply to holding them fixed” (page 211 in [@ooKauf]). Later Kauffman looks for systems with ‘frozen’ areas and finds them near the phase transition. In the ordered phase frozen areas percolate leaving inside small isolated lakes of activity which leads to ‘homeostatic stability’. In the chaotic phase there are isolated islands of frozen areas.
I have checked how the above assumption works using simple known example of thermostat in fridge. The details can be found in Appendix A. In conclusion I state that assumption of fixed states for essential variables is a too strong simplification for description of negative feedbacks. In effect which occurs opposite to intention, negative feedbacks are removed out of such a network. This way ‘homeostatic stability’ based on ‘spontaneous order’, i.e. typically ‘structural stability’, is taken as an explanation of stability of gene regulatory network [@Kauf71; @Wagner01; @Serra04; @Shmul05; @Ramo06; @Serrajtb07; @Serra10] instead of the neglected homeostasis based on negative feedbacks. This is the above mentioned (ch.1.2.1) overestimation of spontaneous order. As can be found in [@RekAlb03] as conclusion of real gene network investigation, “steady states are determined by the topology of the network and the type of regulatory interactions between components”, however, the term ‘feedbacks’ is not yet used there.
Kauffman’s gene regulatory model[@Kauf71] or even[@Kauf69] is still alive and developed[@Serra10; @Serrajtb07; @Sole00; @Luque05] mainly because of its high level of abstraction. Its results gave new perspective which bear fruit of other gene regulatory models, e.g. more connected with genetic reality (less abstractive) GRN’s which are developed basing on @Banzhaf03 work. Evolution of such networks is based on Darwinian mechanisms e.g. [@Knabe06] where biological internal clock emerges. Such simulations together with real network mapping gives base to measure of fraction of negative feedbacks which is crucial to define proper source of stability.
Estimating living objects as chaotic {#ch.1.3}
------------------------------------
### ‘Not fully random’ - ‘known’ changes {#ch.1.3.1}
Homeostasis built by Darwinian selection and based typically on negative feedbacks leads to stability for certain set of initial changes. Such changes may randomly occur but they are ‘known’ to system which means that the system reacts to them in a non-random way. The remaining changes can be treated as ‘fully random’ and when they occur, the system can behave in a chaotic or ordered way. We should focus our investigation on this set of ‘fully random’ changes but considering stability of system we cannot neglect ‘known’ changes.
Current living objects ‘know’ most of typical changes (in their typical environment) and therefore exhibit high stability even though for fully random (unknown) changes they are chaotic. Similar idea of splitting network body by removing certain special part (frozen) which covers the feature of normal rest was applied in [@Bilke01] in ‘decimation algorithm’ for frozen nodes in ordered phase. @Bornholdt00 tends in opposite direction (preferring such nodes) but less openly. It cannot be implemented to prefer the later discussed part of small changes which are selected only from peak of ordered behaviour because these peaks are not parts of the network (they are formed by change events, not network nodes).
Which changes are ‘fully random’ (not predicted in the system structure)? For systems designed by human a designer knows the answer but for living objects it is a problem. We can expect that surgical operation on the brain or heart, or strange environment influences are not predicted in our structure as living objects. Small defects in DNA copying are predicted and certain set of repair or other safety mechanisms are prepared - it is predicted that a defect will occur even though any specific defect is unpredicted. Conditional specification characteristic of all vertebrates can neutralize lot of different events which seem to be random but therefore they are not (fully) random. What range of effected change should we typically expect when particular initial change is random and unpredicted in the above sense? Can we expect homeostatic stability based on spontaneous order or structural stability (see Appendix B) as was the case for ordered systems? I believe that there can be only one answer. If the answer ‘yes’ is considered, then regeneration of body plan in lower animals is probably invoked. Note in such a case that evolution to later ‘higher’ animals lost this feature, but following Kauffman’s idea it should rather have neared to edge of chaos and order (see ch.1.4) and intensified such feature.
### Large damage avalanche modelling death typical for ‘fully random’ changes is available only in chaotic system {#ch.1.3.2}
High stability of ordered systems does not allow for damage to evolve into a large avalanche. (See Appendix B.) The typical behaviour of damage spreading for fully random (not predicted in system designing process) initial change is the main criterion of my estimation that the typical living object or a system designed by human are chaotic. I estimate that such initiation causes with high probability a large avalanche of damage which may be observed only in chaotic systems. This estimation, however, is only intuitive, I will not prove it in this paper but an investigation can be made to verify it.
What can be the interpretation of large damage avalanche? - It can be only death and elimination. Any living object can die. It is common important feature of living objects necessary for Darwinian elimination which creates purposeful (biological) information. It must be an element of the model.
Life is a continuous maintenance of equilibrium at a high level, which is a semi-unstable equilibrium - it will collapse into a large damage avalanche after a single false move. This view correctly describes chaotic state but not an ordered one with its ‘homeostatic stability’ implied from ‘spontaneous order’. Therefore observed stability cannot be spontaneous - means ‘ordered’ but carefully collected which only Darwinian selection can find and not in form suggested by Kauffman (place near ordered phase) but in form active regulation typically - negative feedbacks.
In the ordered phase there is no such radically different possibility which can model death. (See fig.8 in Appendix B.) Large damage avalanche cannot be a transition to a different acceptable attractor. It is such a large change of function, that most of the mechanisms of homeostasis based on negative feedbacks (biological information) lose their parameters needed to function and cannot work. Such ‘biological information’ disappears. Natural criterion of identity [@R1; @R1rgec] based on small change defines such changed system as not ‘the same’. Note, we discuss initial and effective change in system ‘structure and function’ space, not in the fitness space, however, change in the ‘structure and function’ typically causes a change in fitness.
Living objects are self-maintained systems. After damage avalanche in simulated network we still have the same system (structure doesn’t change) but in other basin of attraction. We, however, should remember that large change of living object in function cause large change in maintenance and in effect large change in system structure. It may be interpreted that part of node states describe structure of evolving network. After death structure of earlier living object drastically change and its function is also drastically different. It becomes absolutely other system.
After the above discussion, let us come back to the @Hughes00 experiment analysed in [@Wagner01; @Serra04; @Ramo06; @Serrajtb07; @Serra10] using Kauffman gene regulatory model. This analysis is considered as the main argument behind the ‘life on the edge of chaos’ hypothesis. In this experiment only those cases with knocked out gene are analysed which grow after this disturbance (that is, which are still alive). There are no cases of death in this set. It means, that all measured effects of disturbances are within the range of homeostatic answer. In this range living systems indeed behave like ordered ones with frozen essential variable, which is not suprising.
Living objects there are on opposite shore of liquid region near edge of chaos {#ch.1.4}
--------------------------------------------------------------------------------
The area between chaotic and ordered phase exhibits highest ‘structural stability’ (small initial change gives small effected change) which is useful for adaptive evolution. Maximum of spontaneous order also occurs there. Kauffman isolates this area as third region named ‘liquid’. Evolution using random walk in the space of network parameters should tend to this area because there effects of evolution are larger. This expectation of Kauffman’s approach is known as ‘life on the edge of chaos’. Kauffman expects of living systems rather on ordered shore. (See Appendix B.)
I can agree that evolution tends to this area because of structural stability, however, I am afraid that nearing to the edge of chaos is hard to reach and life is still far from this ‘optimal’ (with respect to the single property of structural stability in this simple model) area. I estimate above that living objects are chaotic. However, in chaotic even large networks a part of initial changes doesn’t effect in chaotic ‘damage avalanche’ but does effect in small change available for evolution. Fraction of such small changes is higher if parameters of network are nearer phase transition between order and chaos. This view can be included into Kauffman’s idea of ‘liquid region near phase transition’, however, living systems exist there on opposite shore. This theme is developed in this paper in ch.3.3 where in simulation results in fig.6 parameter $r$ is discussed.
It is interesting what specific properties these small effective changes in chaotic regime have. I pick up this question in my research which leads to ‘structural tendencies’ [@dgec]. They are the main goal of my approach which allows to explain old classic regularities in ontogeny evolution which were still waiting for explanation. Structural tendencies bring us back to the base of Kauffman’s ideas of structural stability and life on the edge of chaos. This base is the obvious condition of small changes constituting adaptive evolution. I call it ‘small change tendency’ [@paris; @hof; @krab; @dgec]. In ch.3 and in conclusion of this paper we find that the size of effected change initialised in chaotic systems has a distribution with two peaks - for large changes (equilibrium level of damage avalanche in circular attractors) and for very small changes which are an effect of real fade-out of damage in first few steps. The real fadeout is connected to ordered behaviour, damage avalanche (pseudo-fadeout in my algorithm) - to chaotic behaviour. In the next papers on my approach it will be found (see also [@dgec]) that small change tendency selects for adaptive evolution only small changes from the peak of ordered behaviour (as Kauffman correctly expected), but chaotic systems do evolve.
Definitions of particular structural tendencies needs a certain frame of reference in network body. The network external outputs are good such frame, especially when fitness is defined on them which is in my opinion more adequate than the suggestion made in Boolean NK model [@ooKauf] where fitness is a fraction of nodes with ‘proper’ state. The path to results of simulations of structural tendencies isn’t short and it exceeds the frame of this paper (which is only the first step on this path). In this paper we will not yet consider fitness nor external outputs but an extended base (B. i.e. $s>2$ and C. ‘reversed-annealed algorithm’) for investigation of chaotic (A.) network will be defined.
More than two equally probable signal variants {#ch.1.5}
----------------------------------------------
### Boolean networks are not generally adequate {#ch.1.5.1}
As it implies from example in Appendix A Boolean network appears not adequate to precise description of a simple typical thermostat. In reality the temperature is split into three sections $a,b,c$. In one time point a temperature can be only inside one of them. Using three variants signal (state) is natural and leads to simple, correct network (fig.7.2), but requirement to use only two variants (alternatives) leads to losing an important aspect in description (fitg.7.1) or to creation of dummy states (fig.7.3).
### Two variants are often subjective {#ch.1.5.2}
Typically while describing adaptive systems we encounter two alternatives but they usually have very different probabilities. (Typically parameter $p$ called ‘bias’ was used as the probability of one of these alternatives.) It can be, however, defined in a very subjective way, I have a certain explanation how it can happen: One of the typical ways leading to two alternatives is our concentration on one particular event and collecting all the remaining events as the second alternative. This second alternative is NOT the first one, and we obtain two alternatives. There are lots of alternatives in such a case in the reality and this is important for the statistical mechanism but we are only interested in one of these alternatives. Typically it is ‘the proper event’ when we are concerned with systems which adapt. Note, for system which adapt the notions: ‘proper’ and ‘correct’ are defined using fitness but it has nothing to do with the statistical mechanism and such simplification remains subjective. This is the main, however, simple and important cause of introducing more than two alternatives. Note, I estimate that this cause of observation of two alternatives is typical.
If we are going to describe e.g. the long process leading from gene mutations to certain properties assessed directly using fitness, then more than two alternatives for the description of mechanism of such a case seems much more adequate. We should remark that there are 4 nucleotides, 20 amino acids and other unclear spectra of similarly probable alternatives. In this set of spectra of alternatives, the case of as few as two alternatives seem to be an exception, however, for gene regulatory network it seems to be adequate in the first approximation (active or inactive gene). Investigators of real gene networks suggest: “While the segment polarity gene network was successfully modelled by a simple synchronous binary Boolean model, other networks might require more detailed models incorporating asynchronous updating and/or multi-level variables (especially relevant for systems incorporating long-range diffusion).”[@RekAlb03] In second approximations which are RNS [@Luque97; @Sole00] and RWN [@Luque04; @Luque05] models with more than two variants are used but in a different way than here (RSN).
### Equal probability of $s\geq 2$ variants of signal as alternative to bias $p$ and RWN {#ch.1.5.3}
Using equal probability of alternatives is probably the only way to define from within the model a probability necessary for prediction and calculation. In such a way we obtain $s$ (which can be more than two) equally probable signal variants ($s \geq 2$) which I am going to introduce in this paper into the Kauffman networks used for general description of real adaptive systems.
Parameter $s$ describes an alternative (XOR) new mechanism other than the bias $p$ - probability of one of two alternatives (or first one of more than two in RNS [@Luque97; @Sole00]). For $s>2$ (and $K \geq 2$) damage should always statistically grow (whenever it has room to grow) which my ‘coefficient of damage propagation’ shows, and we should always obtain chaos. For extreme bias $p$ and small $K>2$, however, we obtain order [@Derrida86; @Aldana03; @Fronczak08].
Similar to bias $p$ is $P$ - the ‘internal homogeneity’ in Boolean functions [@ooKauf]. Note that parameters $s$ and $P$ lie into opposite direction when they differ to their typical, smallest value. Higher $s$ causes chaos but higher $P$ allows to avoid chaos, however both of them are connected to similar problem of equal probability of two variants of signal. They describe different aspects of this idealisation. An other deviation of random function drawing can be found in the literature. It may have some connection to considered theme of signal diversity. A Boolean function is said to be canalising if at least one value of one of it’s inputs uniquely determines it’s output, no matter what the other input values are. Real gene networks show a proportion of canalising functions much higher than the one corresponding to a uniform choice so it is used the set of canalising Boolean functions only (i.e. for K=2 without XOR and NOT XOR) [@Harris02; @Serra04].
In RNS these signal variants are not equally probable. There bias $p$ plays an important role allowing investigation of phase transition to chaos like in the whole Kauffman’s approach. It is not a mechanism substituting bias $p$ although using $p=1/s$ RNS formally contains my RSN. In my earlier publication [@paris; @hof], however, more than two signal variants also can be found and they are equally probable like here and like in a small paragraph in [@3Aldana]. Typically the case of more than two variants which is taken as interpretatively better [@3Aldana], is rejected [@3Aldana] or not developed as contradictory with the expectation of ‘life at the edge of chaos’ which we reject here. Construction of RWN also is an effect of this condition. Its functions perform a shift of node state and this way create memory. This shifting reaches upper or lower boundaries and the result is random but in a complicated way. (All networks discussed in this paper are deterministic in their function, randomness is used to define the construction and the beginning state. The annealed approximation model concerns probabilistic automata [@Derrida86W], it is not deterministic. ) Mechanisms described using RWN are different than ours, they have to be described using differential equations. It is a step towards regulators and negative feedbacks, however, in this model homeostatic stability based on negative feedbacks has no pressure to emerge.
@Sousa05 considers the scale-free network and more than two different opinions and he obtains a vote distribution which is in better agreement with reality. Similarly [@Stauffer2004; @Jaco2005; @Stauffer2006] consider $Q$ opinion states.
### ‘Coefficient of damage propagation’ simply shows that case $s=2$ is extreme {#ch.1.5.4}
As typically in literature, let $K$ be the number of node inputs and $k$ - the number of node outputs (outgoing links). I assume constant $K$ i.e. equal for all nodes of the particular network as in early @Kauf69 [@ooKauf] works. I neglect the extreme strange case of $K=1$ in my investigation of network features (however in example in Appendix A it appears - fig.7.1 and we can find the case $K=1$ in the literature [@ooKauf; @Wagner01] and $1 < K <2$ in newer works [@Iguchi07] where $K$ is flexible for particular network) therefore $K$ has to be greater or equal $2$. We consider an autonomous network therefore $\langle k\rangle=K$ and $\langle k\rangle\geq 2$ but for a particular node $k=1$ and $k=0$ can happen. For completeness I repeat: $s$ is the number of equally probable variants of signal. Note, that using $s$ we know that they are equally probable. Bias $p$ is replaced by this new parameter $s$.
Up till now the term ‘Kauffman network’ was synonymous with ‘Boolean network’, but not anymore. The term ‘Boolean’ must be limited to two variants of signal but ‘Kauffman network’ can and should contain more signal variants as is used here (RSN) or in RWN or in RNS models.
I introduce [@paris; @hof; @3Aldana] (ch.2) a simple intuitive indicator of the ability of damage to explode. This is $w=\langle k\rangle(s-1)/s$ ‘coefficient of damage propagation’ (or ‘damage multiplication’ on one element of system if only one input signal is changed) which indicates how many output signals of a node will be changed on the average if one input signal is changed (for the random function used by nodes to calculate outputs from the inputs). (I assume minimal $P$ - internal homogeneity [@ooKauf] for the whole of this paper and approach.) In the [@3Aldana] similar equation (6.2): $K_c = s /(s- 1)$ is given for condition $w=1$ i.e. to keep the system within the ordered phase. I use it for an opposite purpose. $K_c$ is here the critical connectivity which for the minimal $s=2$ takes its maximal value of 2.
Coefficient $w$ is interesting for the whole network or for its part, not for a single particular node. However, it is easier to discuss it on a single, averaged node. Therefore I have started my approach using aggregate of automata [@paris; @hof; @krab] (ch.2.3) where $K=k$ and each outgoing link of node has its own signal. It is different than in case of Kauffman network where all outgoing links transmit the same signal. Neglecting $K=1$ for the whole network the value $w\leq 1$ occurs only in case of $s=2$ and $K=2$. For all other cases there is $w>1$ and a small initial damage should statistically explode onto a large part of the system, which means that the system is chaotic. For a part of network the average node degree $k$ can be less than two and locally there can be $w<1$. For definition of such network part, distance to external output of the network can be used (this theme, however, will be discussed in another papers, see also [@dgec; @bgec]).
The case $s=2$ is extreme - there is no smaller sensible (natural) value for $s$ - but it is even more extreme than we have seen. If we simultaneously use $\langle k\rangle=2$ in the same way, then we obtain an especially extreme case - order instead of chaos. It is the case $K=2$ for autonomous Boolean networks which is known [@Derrida86; @ooKauf] as the critical one. Such extreme values of parameters assumed for the model should have special known causes, other than useful simplicity of the model, especially if we need chaos as we estimate above. Such a special cause could be satisfying the assumption of ‘life on the edge of chaos’ (which we reject here). Generally a safer approach is to use not so extreme values for an unknown parameter.
### The damage equilibrium levels for $s>2$ are significantly higher {#ch.1.5.5}
Number $s$ of equally probable variants of signals is the next main parameter of system, like bias $p$, Kauffman’s $K$ - number of inputs per element and $P$ - the internal homogeneity in Boolean functions. These parameters define a system as chaotic or ordered.
Next we examine the differences of effects obtained using $s \geq 2$ and the $s = 2$ only. I find two important differences.
I find the first of them by expanding the Derrida [@Derrida86; @ooKauf] method (annealed approximation) of calculation of equilibrium level of damage (ch.2.2) to cases of $s>2$. The levels for $s>2$ are very different from the case of $s=2$ (see fig.1). The parameter $s$ has a much stronger influence on these levels (moving up their upper limit two times) than the $K$ parameter used up till now for exploration of chaotic regime, therefore $s>2$ cannot be neglected and substituted by $K>2$ for this exploration.
### Importance of parameter $s$ from simulation {#ch.1.5.6}
The result of simulation shows the second important difference (ch.3.3). The parameter $s$, especially for lower values, has a significant influence on the behaviour of different network types (see fig.6), especially scale-free networks, in the first crucial period of damage growth.
Both these differences in effects of the assumptions $s=2$ and $s>2$ (even neglecting the possibility of phase transition from chaos to order) enter the range of qualitative differences. This confirms the importance of this choice.
The assumptions of two variants and their equal probability are also used in a wide range of similar models like e.g. cellular automata, Ising model or spin glasses [@Jan94]. They are typically applied as safe, useful simplifications which should be used for preliminary recognition. But just like in the case of Boolean networks these assumptions may not be so safe and should be checked carefully. In the original application of Ising model and spin glasses to physical spin they are obviously correct, but these models are nowadays applied to a wide range of problems, from social (e.g. opinion formation [@Kos06]) to biological ones, where such assumptions are typically simplifications.
Expectations for $s>2$ - Coefficient $w$ of Damage Propagation and Annealed Approximation {#ch.2}
=========================================================================================
Definition and meaning {#ch.2.1}
----------------------
Let us define a coefficient $w$ of damage propagation:\
$w=\langle k\rangle(s-1)/s$ where $k$ is the number of node outputs. Kauffman in the first step considers constant [@Kauf69; @ooKauf] (for a particular network) number $K$ of node inputs, and considers an autonomous network which has no external inputs or outputs. For autonomous networks $\langle k\rangle=K$: the average $k$ in the network is equal to the fixed $K$. The coefficient $w$ shows how many output signals of a node are changed on the average if at least one of its input signals is changed. In the case of ‘one changed input signal’ this coefficient can be named ‘coefficient of damage multiplication on one node’. If it is greater than one ($w>1$) then damage should statistically grow and create an avalanche which spreads onto a large part of a system. It is similar (especially in the first, crucial period of a process) to the coefficient of neutron multiplication in a nuclear chain reaction - if it is less than one then we have a nuclear power station, if it is greater than one then an atomic bomb explodes.
The coefficient of damage multiplication depends on the functions draw - if functions are not properly random then the coefficient $w$ may be greater or less than the above. The coefficient of damage multiplication is a simple and intuitive indicator of the possibility of damage avalanche and therefore of the system’s place on the chaos-order axis but it is only the first approximation as we will show later. Note that case $w\leq 1$ appears only for case $\langle k\rangle\leq 2$ and $s=2$ and both these parameters are here in their smallest values. For all other cases where $s>2$ or $\langle k\rangle >2$ we have $w>1$ and in such a case damage statistically should explode onto a large part of the system which therefore is chaotic.
The number $s$ of equally probable variants of signals is the next among the main parameters of a system which define the system as chaotic or ordered. It is similar to Kauffman’s $K$ - number of node inputs and $P$ - the internal homogeneity in Boolean functions or to bias $p$ (- probability of first of alternatives) which it substitutes. In the first theoretical approximation the coefficient $w$ can substitute two of them ($s$ and $K$) in this role but other important features of a system depend on the parameters $s$ and $K$ individually and differ although the coefficient $w$ is the same. One of such features is the level of damage equilibrium for chaotic networks which differs much stronger in dependency on the parameter $s$ than on the parameter $K$ (next, ch.2.2). The second such features occurs when we investigate various types of networks, differing mainly in the distribution of node degree $k$: we obtain different results in ‘real fade out’ especially for low $s$ and scale-free network than for higher $s$. Such conclusion is an effect of simulation described at the end of this paper (ch.3.3).
$w^t$ describes first critical period of damage spreading, annealed approximation expanded for $s>2$ {#ch.2.2}
-----------------------------------------------------------------------------------------------------
![Theoretical damage spreading calculated using annealed approximation. (1) Damage change at one time step in synchronous calculation of system known as the ‘Derrida plot’, extended for the case $s>2$ and for $aa$ network type. The crossing of curves $d_{t+1}(d_t)$ with line $d_{t+1}=d_t$ shows equilibrium levels $dmx$ up to which damage can grow. These levels are reached in (2) on the left which shows damage size in time dependency. A simplified expectation $d(t)=d_0w^t$ using coefficient $w$ is shown (Three short curves to the left of the longer reaching equilibrium). This approximation is good for the first critical period when $d$ is still small. (4) shows examples of experimental curves in comparison to their theoretical expectations (see ch.3.3. and table 1). In (3) the increase of damage in consecutive time steps is shown. Experimental curves are similar but wider, the small difference at the point of maximum is shown in table 1.[]{data-label="fig:1"}](a1.eps){width="8.5cm"}
When the avalanche is still small and the range of interaction is a whole and big system (large number $N$ of elements of system) then the probability of more than one changed input signal is also small and damage is well described by $w$ as $d(t)=d_0w^t$ which is shown in fig.1.2. This is a critical period of time $t$, when damage is still so small that probability of its fade out is not to be neglected. Later it practically cannot fade out but the cases of more than one changed input signal occur more and more often and the real multiplication of damage becomes smaller and smaller up to the moment of achieving a stable level of damage (fig.1.1 and fig.1.2). These figures are calculated in a theoretical way based on annealed approximation [@Derrida86] described in @ooKauf book (p.199 and fig.5.8 known as ‘Derrida plot’ for $s=2$), expanded to case $s>2$: If $a$ denotes a part of system $B$ with the same states of nodes as an undisturbed system $A$, then $a^K$ is the probability that the node has all its $K$ inputs with the same signals in both systems. Such nodes will have the same state in the next time point $t+1$. The remaining $1-a^K$ part of nodes will have a random state, which will be the same as in the second system $A$ with probability $1/s$. The part of system which does not differ in $t+1$ is therefore $a^K+(1-a^K)/s$. It is the same as for RNS [@Sole00]. The damage $d=1-a$. For $K=2$ we obtain $d_2= w d_1 - w d_1^2/2 $ where for small $d_1$ we can neglect the second element.
Aggregate of automata - the simplest case of networks for coefficient ‘$w$’ {#ch.2.3}
---------------------------------------------------------------------------
![The basic elements of Kauffman network (1) and aggregate of automata - $aa$ network type (2). Nodes - ovals, signals - rectangles, links - arrows. Each node transforms incoming (input) signals into output signals using a function, these signals are transmitted through links to the next nodes as their input signals. $K$ - number of input signals (or links) of a particular node. $k$ - number of output links of a particular node (node degree). For a particular node of Kauffman network (case 1 on the left) there is one output signal (state of node) which is sent by $k$ output links. $s$ - number of equally probable variants of signal values (in Boolean network $s=2$, i.e. true and false). In the case (2) of the $aa$ network $k$ is fixed and each output link has its own signal, possibly different from others.[]{data-label="fig:2"}](a2.eps){width="7cm"}
In the Kauffman networks all $k$ outputs of a node transmit the same signal - it is the state of node, the value of its function (fig.2.1). To understand the coefficient $w$ of damage multiplication we must average by lots of nodes. It is much simpler and more intuitive (which is important for introducing such a method into biology) if each output link of a node has its own signal to transmit, which need not be the same as on other outputs in the same node. In such a case the function’s argument and value are a $K$- and $k$-dimensional vectors (fig.2.2). Due to function uniformity it is useful to fix $K=k$. I have introduced such a network in [@paris; @hof; @krab] where I have named it ‘aggregate of automata’ ($aa$). For this network if $K=2$ then $d_2=d_1*w-d_1^2*(s-1)^2/(s+1)/s$ which is obtained similarly as above. Note, that for small $d_1$ we can neglect element with $d_1^2$. Theoretical curves for aggregate of automata for case $s=4$ and $K=k=2$ are also included in fig.1.1-3.
These figures show that the level of damage equilibrium for aggregate of automata is much higher than for Kauffman networks. To expect $a_{aa,t+1}$ - the part of nodes in $aa$ network which does not differ in $t+1$ we can use expectation for Kauffman networks shown in previous chapter. Such $a_{Kauff,t+1}$ describes signals on links of $aa$, not states of nodes of $aa$ network which contain $K$ signals:\
$a_{aa,t+1}=a_{ Kauff,t}^K+(1-a_{ Kauff,t}^K)/s^K$
Network Features Dependency on $s$ from Simulation Using Reversed-Annealed Algorithm {#ch.3}
====================================================================================
Project formulation and networks type definition {#ch.3.1}
------------------------------------------------
### Currently investigated network types in RBN range {#ch.3.1.1}
The Kauffman formula gives useful ability to differentiate $k$ within the network and to investigate different types of networks which differ in the distribution of node degree $P(k)$ like Erdős-Rényi random networks (RBN or better CRBN [@Serra04A]), on which Kauffman had worked, or nowadays famous Barabási-Albert scale-free networks. This is because the definition of function does not change if $k$ changes. Due to this reason $K$ was fixed and for a directed network only the $k$ parameter was used [@Kauf69; @ooKauf] as the degree of a node. Currently both the parameters of the investigated Kauffman networks - $k$ and $K$ are typically (not in e.g. [@Serra04A]) flexible [@Aldana03; @Kauf04; @Iguchi07]. Scale-free networks typically turn out to be more adequate to describe reality [@sf99; @sf03; @Kos04; @Kos06; @Crucitti2004; @Serra04A; @Fortunato2005] than old Erdős-Rényi ‘Random Networks’. Most of contemporary authors stay with Boolean networks due to Kauffman’s ‘life on the edge of chaos’ approach and consider SFRBN (Scale-Free Random Boolean Network) [@Serra04A; @Iguchi07] and EFRBN (Exponential-Fluctuation Random Boolean Network) [@Iguchi07] together with old classic RBN (CRBN [@Serra04A]).
If we change our approach to prefer chaotic networks and use simply omitted $s \geq 2$, two interesting questions appear: 1- Are there any significant differences in damage spreading in the chaotic area (which was not interesting up till now) in different network types? 2- What is the effect of higher $s$ on damage spreading, i.e. is this parameter important?
### Network types used in simulation of RSN {#ch.3.1.2}
I have investigated these questions in a simulation using my simplified algorithm named ‘reversed-annealed’, dedicated for statistical analysis of damage spreading. The consistency of its result with annealed model expectation makes this simplification trustworthy. This algorithm will be used later (in the other papers) for a simple and intuitive definition of complexity threshold (connected with maturation of chaotic features of network [@bgec]), useful at the end of this path to investigation (also using this algorithm) of structural tendencies in adaptive evolution of a complex system. These tendencies describe - among other phenomena - some known and interesting regularities in ontogeny development or human activity which however have not been explained until now. In this paper we compare damage spreading in five types of autonomous networks: ‘$er$’ - random [@er], ‘$sf$’ - scale-free (“Barabási-Albert” [@sf99; @sf03]), ‘$ss$’ - single-scale [@ss] (for $s=2$ known as respectively: RBN / CRBN, SFRBN and EFRBN), ‘$aa$’ - aggregate of automata and ‘$ak$’ - a network similar to $aa$ with fixed $K=k$, but using Kauffman formula where one state of a node is transmitted by all its outputs.
Why don’t I use consequently SFRSN, EFRSN, CRSN and so on? There are two reasons for a different choice: The first one is that such names are too long, there are figures where there is only room for one letter to indicate network type, I use there only second letter of my names which are different. First letters of my names indicate a group. The second reason is that the number of considered network types is large (in the next article there will be ten of them) and named differences become another character, not so stable to fix all names for longer time. Therefore level of RSN, RWN, RNS and RBN inside article should be separated. To close problem of naming I include to RSN the family of aggregate of automata together with the Kauffman network family.
### Rules of networks growth {#ch.3.1.3}
Construction of the network simulation has two stages: construction of the network and damage investigation in the constant network. Construction of the network depends on the chosen network type. Except for the type ‘$er$’ - Random networks, all networks have a rule of growth. Aggregate of automata ‘$aa$’ and ‘$ak$’ needs to draw $K$ links in order to add a new node. These links are broken and their beginning parts become inputs to the new node and their ending parts become its outputs (fig.3.1.).
![Additions patterns for aggregate of automata $aa$ (1) and Kauffman networks $ss$ and $sf$ (2). ($K=2$) Links g and h (and function) of node are drawn. Node j is drawn directly instead of link h for $ss$. For $K>2$ additional inputs are constructed like the right ones (h or j). The $ak$ network is maintained as $aa$ but there is only one output signal c (d=c).[]{data-label="fig:3"}](a3.eps){width="7.5cm"}
For ‘$ss$’ - single-scale network the new node is connected to the node present in the network with equal probability for each existing node. For ‘$sf$’ - scale-free network the new node is connected with another node with probability proportional to its node degree, i.e. to the $k$ of this existing node. For both types we draw one link first and we break it like for $aa$ and $ak$ to define one output and its destination node and the first input. For $sf$ type at least one such output is necessary to participate in further network growth. Later we draw the remaining inputs according to the rules described above - for $ss$ by drawing the node directly, for $sf$ by drawing a link and using its source node (fig.3.2).
Damage can take various forms, e.g. in complex computational networks [@nowostawski] but damage spreading in scale free networks describes typically: epidemic spreading [@Kos04], opinion formation [@Fortunato2005; @Kos06; @Sousa05] and attack or error effects [@Crucitti2004; @Gallos04]. However, these networks typically are not directed networks and their important aspect is the spatial description which uses a particular lattice shape. They also are constructed in a different way, not only using preferential attachment [@Kos04; @Kos06]. A partially directed scale-free network was used in [@Stauffer2006] preceded by [@Stauffer2004; @Jaco2005]. These networks describe opinion agreement process. In this approach the direction of links is used for construction of a social network and consequently - initiative to contact; however, during opinion exchange an information flows in both directions and each of the talking nodes randomly takes the opinion of its partner. This second aspect is more similar to signals flow in Kauffman network, but here it is undirected and therefore this approach is not similar to Kauffman networks. The dynamics of Boolean networks with scale free topology were studied by @Aldana03 and @Kauf04, now @Iguchi07. They look for the difference between the dynamics of $er$ (there called: RBN) and the scale-free random Boolean network (SFRBN). @Iguchi07 used also ‘directed exponential-fluctuation networks EFRBN’ similar to our $ss$. All they use only $s$=2, flexible $K$ and $k$, therefore their networks differ from ours RSN.
Simplified algorithm ‘reversed-annealed’ of damage spreading {#ch.3.2}
------------------------------------------------------------
### The main assumption {#ch.3.2.1}
The classic method of damage observation uses two exact processes (quenched) which are compared: $A$ for an unchanged system and $B$ for system with damage initiation [@Jan94]. I observe one process - only damage spreading, but this process is only statistically correct in a particular range of situations. The main assumption is: We consider chaotic system with random function of nodes without memory where damage can fade out only when it is still small, but when it is large, then it grows up to an equilibrium level, where in our algorithm it also stops using the same mechanism which keeps the damage at the equilibrium level (pseudo fade out). In reality (quenched model) it fluctuates around this level infinitely. Our algorithm gives one particular damage size which also fluctuates around this level but for this it needs numerous particular cases of processes and their stops. I have found these levels above (ch.2.2, fig.1) using annealed model. They are the validation of presented algorithm. Annealed approximation, however, simplifies the aspect of structure and doesn’t consider real fade-out which is important for adaptive evolution.
### Limitation {#ch.3.2.2}
The two cases of fade-out described above are mixed, however they have different interpretations. Processes which stop (fade-out) between them with a middle damage $d$ have no interpretation and can be permissible only in negligible frequency. Such cases occur only for $s=2$ (especially for $sf$ 2,3 ($type$ $s,K$) network (fig.5.1) in high level and for $sf$ 2,4 and $ss$ 2,3 in small but visible level) which confirms that $s=2$ is extreme. The case $s,K = 2,2$ is for every network type out of range of permissible levels of middle damage and we cannot use our algorithm for its investigation - it consists mostly of real fade out cases (fig.1.1) or very low damage equilibrium levels but its long tail for higher $d$ is strongly incorrect (too short) in the simulation. Too small network size also leads to incorrect results which will be studied in Ref. [@bgec]. The real fadeout is connected to ordered behaviour and pseudo-fadeout - to chaotic behaviour.
### Simplifications {#ch.3.2.3}
In this algorithm, we only calculate nodes with one or more changed input signals. To detect changed input signals, signals are first calculated and memorized during network growth. If we detect a changed input signal as the effect of damage initiation, we do not care what remaining input signals are. They can be changed before or after the calculation of this particular node (in damage calculation), e.g. as effect of feedbacks loop. If a node is affected by damage, which means that at least one input signal was changed, then the node function is calculated (if necessary but not in situations described in this paper) using ‘old’ remaining input signals, but only once (i.e. a node is never calculated twice).
The damaged part will become a tree. In this paper we also will not use concrete functions for nodes. If the input state is changed, then the output state is random. This calculation gives an answer, whether output signals of this node have undergone any changes. If its input signals change later, then it will not be calculated next time - for statistically correct damaged area it is not needed. Any initiation of a particular node in a particular network should lead to statistically the same damaged area but in each particular case it may be different. We, however, are not interested in a particular case but in a statistical result. Such an algorithm works fast and gives correct statistical effects.
### Intuition behind {#ch.3.2.4}
An intuition behind this algorithm can be found when we consider a network without feedbacks, where each signal on the node output is equal to the value of the function of current signals on the node inputs. It is not a typical system state - in the next time step (e.g. in the synchronous mode) nothing will be changed. Now we introduce a disturbance changing one node function (permanently) and we calculate this node. To obtain a new stable state of the system we must calculate only nodes whose at least one input signal had changed (as a result of damage). For this calculation we can use the old signals on the remaining inputs if for a given node they do not depend on the remaining nodes waiting for calculation. Such a node will always exist because a node does not depend on itself. After a finite number of node calculations all the node states will be equal to the function value of current node inputs as was the case at the beginning, however lots of node states will be different than at the beginning, i.e. - damaged. As was mentioned above we do not use concrete functions, therefore we need not to check a dependency and we can calculate any waiting node.
In the case with feedbacks sometimes an already calculated node gets a damaged input signal for a second time. For measuring the statistical effect only it is not necessary to examine its initiation for the second time, however, if such second initiation will be processed, then the process may never stop.
### Experiment structure {#ch.3.2.5}
In this paper we investigate the damage in a system of a particular size. When a network achieves the assumed number $N$ of nodes we stop the growth and we start to initiate damage: we change the output state of each node using all remaining possibilities as damage initiation. It is the smallest initiation and in the first few steps the damage can fade out. It is a real fade out of damage. In this short way damage can meet an already damaged node which is not calculated for the second time (which helps damage to fade out), however, such an event has a very small probability. We assume that if damage fades out when it is small, then it is not due to meeting an already calculated node. This is a simplification of our algorithm. In this case the number of damaged nodes is interpreted as the number of damaged nodes during the whole process from initiation to real fade out. If coefficient $w>1$ then on average the damage grows. If damage is great, i.e. the number of nodes with changed output state is large (the number of calculated nodes due to their input state was changed is also large), then it practically cannot fade out (probability of such events is very low, we neglect them), but during this damage growth there are less and less nodes which are not reached by damage yet. Therefore the avalanche of damage must slow down and stop (the growth). It looks like a fadeout, but it is equivalent to the achievement of the stable level by the damage which appears at the end of curves in fig.1.2 or on cross of curves with line ‘$d_{t+1}=d_t$’ in fig.1.1. This level is an equilibrium state, as fig.1.1 shows. In our simplification the process stops at this level due to the ‘pseudo-fade out’ on already damaged nodes. Now the number of damaged nodes is interpreted as the equilibrium level and it describes the statistical state of the system at any one time step after the equilibrium is reached.
The number of nodes with changed output state (i.e. number of damaged nodes) divided by $N$ is equivalent to the damage size $d$, despite the fact that they are damaged during the whole process (using our algorithm), not only in the last time step. Note, if damage fades out, regardless of the way it happens (pseudo or real fade-out), then in the last time step damage size looks like zero. However, such a view is incorrect in the case of pseudo fade out, it does not take into consideration the fact that we do not recalculate a damaged node when its input signal changes for a second time. Such a false suggestion appears due to the simplification of our algorithm.
### Reversed-annealed approximation {#ch.3.2.6}
We calculate the damage using a fi-fo queue (first in - first out) for nodes with changed input signals waiting for calculation. The queue length in time $t$ (as in the synchronous mode) dependency is very similar to the one shown in fig.1.3. Small differences at the point of maximum are shown in table 1. The time step number $t$ is defined observing this queue but for control of the process it is not necessary. Simplification used for our algorithm is a step from the ‘quenched model’ [@ooKauf] in the direction similar to ‘annealed model’ [@Derrida86; @Derrida86W], however it is a small deviation only and in ‘reversed’ direction. These simplifications are: 1 - only nodes with damaged input signal are calculated;\
2 - second initiation of already calculated node is neglected; 3 - when function is not used, then new node state is random, else for the remaining, not damaged input signals old values are used. A ‘reversed annealed simulation’ was performed: We randomly define new state for all nodes keeping old network structure and we obtain very similar results. Because the input sites remain constant over time they are quenched [@Derrida86W], however, the functions are simplified therefore the algorithm is not quenched but (reversed-) annealed. The dynamics of such a net is thus approximated in a non-deterministic way. Note, the task which we define for this algorithm is not an adaptation or optimisation of network function but the investigation of damage spreading (in this paper) and (in next papers [@bgec]) - of statistical properties of changes leading to smaller or larger damage which concerns changes constituting adaptive evolution. It will look there like a greedy algorithm, however, ‘greedy’ property is not well adequate for its description. Our algorithm is much simpler than the quenched one, therefore it is computed much faster. By yielding one particular result on external outputs (instead of a circular attractor) it allows to define fitness in a simple way and so this algorithm opens new areas of investigation (e.g. structural tendencies).
Simulation effects and comparison to theoretical expectations {#ch.3.3}
-------------------------------------------------------------
### Simulation parameters {#ch.3.3.1}
The computer program realizing the above algorithm is prepared for $s=$ 2, 3, 4, 8, 16, 32, 64 and $K=$ 2, 3, 4. I have investigated the whole of this area but the most interesting part is the area near the phase transition from chaos to order where differences are larger. For comparison of dependency in main parameters $s$ and $K$ I choose five cases described as $s,K$: 2,3; 2,4; 3,2; 4,2; 4,3 for five network types described above: $er, ss, sf, ak, aa$. In this set there are: $K=3$ and $4$ for $s=2$, next: $s=3$ and $4$ for $K=2$. Similarly for $K=3$ and $s=4$ the second parameter has two variants. Cases 2,3 and 4,2 have the same $w=1.5$. The coefficient $w$ is the smallest for case 3,2 (1.33) and the largest in the shown set for 4,3 (2.25). The simulation results are shown in fig.4, fig.5 and fig.6, also in table 1 but only in the main fig.6 the whole set of the above enumerated cases is shown. For the most interesting networks $sf$ and $er$ the full set of combination $s$ and $K$ in range of values 2, 3, 4 is show in fig.6.3. Each simulation consists of 600 000 damage initiations - e.g. for $s=4$ (excluding $aa$) 100 different networks grow randomly up to $N=2000$ (or $N=3000$) nodes and later each node has its output state changed 3 times (2 times). For $aa$ and $s=4$ I use 20 networks and the output state is changed 15 times.
### Comparison results to expectations {#ch.3.3.2}
[rrrrrrrrrrrrrrrr]{} net &2,3 d1 & d2 & t1 & t2 & t3&4,2 d1 & d2 & t1 & t2 & t3&4,3 d1 & d2 & t1 & t2 & t3\
$aa$ & 1338 & 1319 & 16 & 35 & 28 & 1668 & 1672 & 17 & 32 & 27& 1938 &1908 & 10 & 16 & 15\
$ak$ & 772 & 690 & 14 & 34 & 25 & 1335 &1307 & 16 & 31 &26& 1473 & 1515 & 10 & 15 & 14\
$er$ & 770 & 880 & 14 & 34 & 25 & 1335 & 1336 & 15 & 34 & 24& 1473 & 1441 & 10 & 17 & 14\
$ss$ & 774 & 654 & 13 & 40 & 24 & 1335 & 1202 & 16 & 36 & 25 & 1476 & 1495 & 10 & 19 & 14\
$sf$ & 818 & 877 & 12 & 45 & 21 & 1344 & 1217 & 16 & 43 & 26 & 1476 & 1485 & 10 & 26 & 15\
th. & 764 & & 17 & 26 & &$aa$1667 & 1333 & 18 & 28& &1472 & & 10 & 14\
As the first effect of simulation I show a comparison to the expectation presented in fig.1.3 and fig.1.2. The obtained distributions of increase of damage in consecutive time steps in synchronous mode of calculation are very similar to the ones shown in fig.1.3. (A small difference at the point of maximum is shown in table 1). However this similarity conceals some statistical diversity of speed of damage spreading process. This diversity appears when we try to compare obtained results especially for networks $sf$ and $ss$ (e.g. fig.1.4) to fig.1.2. In fig.1.4 a case for $ss$ 2,3 is presented and a similar case $sf$ 4,2; they are both less rough than $sf$ 2,3 and $sf$ 3,2. We can identify a few independent processes with different speeds of damage growth. This diagram is plotted for each particular case of damage initiation replacing the old one. When damage reaches the level of equilibrium it stops (in our simplified algorithm) and there is no data for later time steps. In this later area, therefore, we can observe other processes which are earlier and slower. For network types scale-free and single-scale this speed is strongly connected with the time of reaching the hubs by damage. For network types $er$ and obviously $ak$ and $aa$ there are no hubs and the obtained curves are more uniform and similar to the theoretical ones shown in fig.1.2.
### Distribution $P(t)$ of time of damage fades out {#ch.3.3.3}
{width="16cm"}
Fig.4 shows the distribution of time of damage fadeout in both real and ‘pseudo’ cases. There are two peaks in this distribution: one for early fadeout in the first steps (we interpret them as real fadeout) and the second one later, for ‘pseudo-fadeout’ when damage reaches the equilibrium level which is high (at the last time steps in our algorithm). For the network cases with wide range of node degrees like $sf$ and $ss$ with a large fraction of $k=1$ nodes the probability of early fadeout is much greater especially for small $s=2$. If $K=3$ then 60% nodes for $sf$ and 33% for $ss$ have $k=1$ but there are 11% and 20% nodes of $k>4$ which have 55% and 46% outgoing links respectively. If $K=2$ then respectively for $sf$ and $ss$ networks, 67% nodes and 50% nodes have $k=1$, there are 7% and 6% nodes of $k>4$ which have 34% ($sf$) and 19% ($ss$)outgoing links . For $s=2$ nodes with $k=1$ have $w=1/2$ but for $s=4$ there is $w=3/4$. To reach the threshold value 1 for $s=2$ there remains two times more (1/2) than for $s=4$ (1/4). Only nodes with higher $k$ can increase $w$ and reach this threshold but they are rare as shown above. Therefore locally it seems to be $w<1$ (like for the ordered regime) more often for $s=2$ and early fadeout is more probable then than for $s=4$. Hubs are present in this case. The biggest hub ($k=955$) appears in $sf$ when $K=4$, for $K=3$ it reaches $k=520$. This single hub takes 12% (the second 9%) of all the outgoing links. Hubs decrease the average $k$ and in effect also the average $w$ for remaining nodes, which makes the remaining area ‘ordered’ and helps damage to fade out before the first hub is achieved. For $er$ network even $k=0$ occurred but nodes with $k<2$ constitute less than 1/4 of all the nodes. If $s$ is small, e.g. $s=2$, then the coefficient $w$ is locally especially low. Note that we have used local coefficient $w$ for the explanation. On the opposite end (only of Kauffman mode, i.e. excluding $aa$ network) the case of $ak$ 4,3 lies where $k<2$ and hubs are absent. In this case the coefficient $w$ of damage propagation is high and equal for all the nodes, therefore the early fadeout is very small and most of the damage grows until the equilibrium level is reached.
Fig.4 consist of four different distributions. They are all plotted for networks of $N=2000$ nodes, but this sequence looks like a sequence of distributions of stage of e.g. $ak$ 4,3 growth. (We investigate this process in Ref.[@bgec] in much more detail. It can be a base for the definition of complexity threshold.) This means that $sf$ 2,3 in left (fig.4.1) looks like small, not yet mature networks $ak$ or $er$. In an $sf$ network 2,3 in the most cases (80%) the damage fades out (real fade out) without reaching hubs. Such damage behaviour is more typical of ‘ordered’ networks rather than ‘chaotic’ ones. This is ‘liquid region’ available for adaptive evolution. For $ss$ 2,3 and $sf$ 4,2 both the peaks have a similar area (fig.4.2) but the left peak is much higher. Later, (fig.4.3) for $sf$ 4,3 and (fig.4.4) $ak$ 4,3 the left peak has a smaller area and the right peak which describes chaotic behaviour contains most of the cases. This looks like advanced stages of transition from order to chaos. In the percolation theory maximum of right peak describe size of large cluster. It appears much after percolation point in which only left peak exists and it just reached power-law shape.
### Distribution $P(d)$ of fade out in ‘damage size’ {#ch.3.3.4}
{width="16cm"}
The phenomenon of different speed of damage spreading described above is the reason behind the large width of these peaks and lack of sharp boundary between them here (fig.4.1-3). This remark suggests that the variable $t$ - time of damage fadeout is not the best choice. However, the variable $t$ is interesting in practice and therefore often used [@Jan94]. A similar distribution of damage fadeout in the variable: damage size $d$, shown in fig.5.1-3 appears much better suited for the description and understanding of the underlying mechanisms. @Serra04 have introduced term ‘avalanche’ which is consequently used later [@Shmul05; @Ramo06; @Serra10], for damage size measured in damaged node number. Such a variable is used in fig.5, but above used parameter $d$ differ from ‘avalanche’ only in normalisation ($d=avalanche/N$) and we will stay with $d$ only. Using such a variable we also obtain the same two peaks but this time they are very narrow and a big segment of exact zero frequency lies between them typically. The only exception from this rule is the extreme case of $sf$ 2,3 (fig.5.1) but we have discussed the causes of this exception above. The case of $sf$ 3,2 and (fig.5.2.) $ss$ 2,3 follows the rule but the second peak is not very narrow and we can find some single cases between the peaks. (These cases have no clear interpretation in our algorithm.) All the remaining cases are similar to the last $sf$ 4,3 shown in fig.5.3, small differences concern proportion of both peaks and the peaks’ width.
Position of maximum of the second peak is exactly equal to the theoretical point of equilibrium of damage size ($dmx$), it is obviously much more exact than the values of maximum which can be read from the distribution like in fig.1.4 which are for one particular process. The comparison of this 3 values for different cases is shown in table 1.
### $w$ doesn’t substitute consideration of $s$ and $K$ separately {#ch.3.3.5}
{width="16cm"}
As it was discussed above, the new network types, especially the scale-free networks, due to concentration of a great part of links in a few hubs, exhibit significant differences in behaviour of damage spreading. These differences appear especially near the boundary of chaos and order and are more intensive for $s=2$. To summarize these differences I show fig.6.1 where we can compare average (for all initiations) damage size $d$ for each simulated case of network type and $s,K$.
Depicted data have 3 decimal digits of precision, therefore the shown differences are not statistical fluctuations. As it can be seen, using higher $s=4$ for $K=2$ causes different behaviour of damage spreading than for $s=2$ and $K=3$, especially for $er$ network type, despite the same value of coefficient $w=1.5$, therefore these both parameters cannot substitute each other, i.e. we cannot limit ourselves to one of them or to the coefficient $w$.
### ‘Degree of order’ of network {#ch.3.3.6}
Fig.6.1 contains two different causes which differ results. One of them was already described at the end of ch.2.2. It is the different level of damage equilibrium visible in the theoretically obtained fig.1. If we remove this aspect from fig.6.1, then what remains is the second cause which is connected to the early (real) fadeout. This cause depends mainly on the network type and the $s$ parameter which is depicted in fig.6.2. This figure shows how big is the part of initiations which end in real fadeout. They are separated using threshold on $d=250/N$. Compare the point for $sf$ 2,3 to the fig.4.1. This aspect contains the mechanism of $er$ distinctness which results from the events $k=0$.
Fig.6.2 has an important interpretation: the depicted parameter $r$ is a ‘degree of order’ of a network in a certain aspect. It describes the probability of real fadeout of damage. However, for application to living object description it should be noted, that this fadeout only occurs in a random way which does not consider negative feedbacks collected by adaptive evolution.
These investigations using simulations of different network types were designed and included in this paper to show that the parameter $s$ is important and we cannot limit ourselves to the parameter $K$ only when we study chaotic behaviour. The dependency on $s$ is about as strong as the dependency on $K$ but it also differs from dependency on $K$ for different network types. In the aggregate of automata the state of a node has $s^K$ variants and this network type has obviously stronger (and different) dependency on these parameters than Kauffman networks. The $ss$ and $ak$ networks exhibit symmetrical dependency in $s$ and $K$ but for the most interesting $sf$ and $er$ network types there is no symmetry, which is depicted in the fig.6.3. For $sf$ the dependency on $s$ is stronger but for $er$ - weaker than the dependency on $K$. For each network type two sets of two ‘lines’ are shown. Each ‘line’ consists of three cases for the same value of parameter $s$ or $K$ and all three values (2, 3, 4) of the second parameter. E.g. we compare the line consisting of (s,K) cases 3,2 3,3 3,4 and the line of 2,3 3,3 4,3 cases. These two lines have one common case 3,3 indicated in the figure. We are interested in the slopes of both compared lines. The relative slopes appear to be approximately constant for both parts of each lines. For both network types shown the slopes for parameters $s$ and $K$ are significantly different which already for two steps leads to significantly different values. These differences are not big but may be important. The significantly lower damage size for $sf$ network which can be seen in fig. 5 is known [@Gallos04; @Crucitti2004] as the higher tolerance of a scale-free network of attack. Also @Iguchi07 state: “It is important to note that the SFRBN is more ordered than the RBN compared with the cases with $K = \langle k\rangle$” which in our fig.6.2 is clearly visible, however in the area shifted into chaotic direction.
Conclusion {#ch.4}
==========
There are three different types of conclusion of this paper.
The first is mainly interpretative - systems describing living objects should be chaotic, not ordered. More exactly - they should be much more chaotic, even purely (maturely) chaotic (with possibility of damage avalanche which models death) than Kauffman expects in his “bold hypothesis: Living systems exist in solid regime near the edge of chaos”. There, living objects are on opposite shore of Kauffman’s ‘liquid region near edge of chaos’. In this region, a part of small changes, available for adaptive evolution, exists, which I show as ‘degree of order’ in simulation results. Death is possible in all living objects and is necessary for Darwinian elimination. It should be an element of the model. Damage avalanche which can only model death is available only in chaotic phase. The stability of gene regulatory network is the main experimental argument behind Kauffman’s localization of living systems on order-chaos axis but in this model negative feedbacks are not considered sufficiently.
The second, main conclusion of this paper is that the parameter $s$ (number of equally probable variants of signal i.e. states of nodes in Kauffman network) is an important one. It approximates other statistical mechanisms than using bias $p$ for two variants only. This new mechanism is named here Random Signal Network (RSN). (Full name: ‘Random equally probable Signal variants Network’ is too long.) Together with RNS and RWN it expands the notion of ‘Kauffman network’ which was the synonym of ‘Boolean network’. For the investigation of damage spreading behaviour in chaotic phase, especially for scale-free networks, parameter $s$ cannot be neglected or substituted by the parameter $K$ (number of node inputs) and others like $P$ (internal homogeneity) and bias $p$ (the probability of one of two signals). It ($s>2$) leads to much higher levels of damage equilibrium which can be found theoretically using extended annealed approximation. It also leads to different ‘degree of order’ for different network types which is seen in simulation. Next arguments of various types are collect that especially for adaptive systems describing real living or human-designed systems, the $s$ parameter should generally be greater than two which placed these systems (with constant $K>1$) in the chaotic area. I believe that the typically observed case of two alternatives with different probability has typically bases whose description using bias $p$ is incorrect and $s$ should be used instead.
Along the way of investigation of properties of the parameter $s$ using simulation, I have found that the contemporary first BA scale-free network has significantly different behaviour in damage spreading than the Erdős-Rényi random network which was used in Kauffman’s path and these differences are largest for $s=2$. Generally, the networks with higher frequency of $k<2$ nodes ($k$ is node degree - the number of node outputs, for autonomous systems $K=\langle k\rangle$ ) have higher chances of damage fade out in the critical beginning period. If hubs are present then this chance also increases because they decrease the average $k$ for remaining nodes which helps the damage to fade out before the first hub is reached.
For this simulation I design and describe here a special simplified algorithm which I name ‘reversed annealed’. I also use it for the next investigation, of ‘complexity threshold’ and ‘structural tendencies’. This algorithm uses only calculation of damage spreading up to reaching the equilibrium level instead of calculation and comparison of two systems - damaged and undisturbed one as in the classic method. Therefore it is much more optimal algorithm. And now the third conclusion of this article - the technical one. This simplification has small meaning for investigations presented in this paper, however, it is a critical key to the area of ‘structural tendencies’ in adaptive evolution which are the aim of the path starting with this paper (see also [@bgec]). The main feature of this algorithm is that it gives a single particular output state, however only statistically correct, instead of long cyclic attractors. In my approach avoiding attractors is useful in opposition to Kauffman’s approach where properties of attractors are very interesting.
Another technical useful result of this paper is the coefficient of damage propagation, which I introduce here. It connects two main parameters $K$ and $s$ and describes the first, critical period of damage spreading (in the first theoretical approximation). It is simple and intuitive and it easily shows when damage should explode: damage should explode always if $k>2$ or $s>2$. The ability to produce an explosion of damage is one of the definitions of a chaos which I used following Kauffman for chaotic systems. The above mentioned influence of the hubs suggests that such a coefficient should be considered more locally, e.g. in the area of damage spreading.
Acknowledgments {#acknowledgments .unnumbered}
===============
Research financed by Polish government as grant 3 T11F 035 30.
Appendix A - Thermostat {#ch.Appendix A .unnumbered}
=======================
![Thermostat of fridge described using Kauffman networks as example of regulation based on negative feedbacks. Case (2) describes it just as it is in reality - temperature $T2$ is split into three sections $a$ - too cold, $b$ - accurate, $c$ - too hot but therefore this case is not Boolean. To hold signal in Boolean range we can neglect temperature state $b$ - case (1) or split node $T2$ into two nodes with separate states - case (3) which together describe all temperature states but this way dummy variant ($a,c$) of temperature state is introduced. Node $V$ decides power for aggregate: v - on, 0 - off. Common tables of functions for nodes (left one) and for consecutive state (right one) are attached. In left table for case (1) and (2) certain lines of $V$ states should be omitted. Similarly in right table for case (1) there are only four lines.[]{data-label="fig:7"}](a7.eps){width="8cm"}
We try to check how Kauffman’s assumptions for regulation based on negative feedbacks work. Let us analyze a simple known example of thermostat in fridge. If fridge leaves proper temperature of range $b$ as result of environment influence and enters too high temperature of range $c$, then power supply for aggregate is turned on and temperature inside fridge goes down. It passes range $b$ and reaches too cold range $a$, then power is made off and temperature slowly grows through $b$ section. Case (2) in fig.7 properly describes this regulation mechanism in Kauffman network terms. However, there are three states of temperature $a,b,c$ which are described by three variants of state of node $T2$ and therefore this case of Kauffman network is not a Boolean network.
To hold signal describing temperature in Boolean range (two variants only) we can neglect temperature state $b$. This is case (1) but here, the most important proper temperature state, where fridge stays most of time, is missing. Almost any time we check the state of a real fridge this state is not present in such description. Reading such description we find that wrong temperature $a$ - too cold occurs directly after wrong temperature $c$ - too hot and similarly in opposite direction. Is this fully correct information and description?
Splitting node $T2$ into two nodes $Ta$ and $Tc$ with separate states - case (3) is the second method to hold Boolean signals. Two separate Boolean signals create together four variants but temperature takes only three. New dummy state emerges: $a$ - too cold and $c$ -too hot simultaneously. It has no sense and never appears in reality but function should be defined for such a state. In table shown in fig.7 the functions values for nodes of case (3) for such dummy input state are marked by ‘?’. Anything can be put there, however, for statistical investigation it is taken as real proper state.
Do cases (1) and (3) describe reality adequately? I’m afraid that the answer is ‘NO’. It is because Boolean networks are not generally adequate. We can describe all what we need using Boolean networks but in lot of cases we will introduce dummy states or we will simplify something which we don’t like to simplify. In both cases statistical investigation will be not correctly connected to reality. The only way is to use real number of signal variants and not to limit ourselves to only two Boolean alternatives.
Even in simplest case (1), but also in remaining cases, temperature as ‘essential variable’ varies and this is necessary for regulation. For feedback mechanism it cannot be fixed. See right table in fig.7 which shows consecutive states for nodes in all cases. If we fix temperature constant e.g. in range $b$ which is a good approximation measuring it frequency in time, then it can be treated as effect of regulation mechanism work but it deprives of job existing in network feedbacks. Such mechanisms keeping it constant are there outside such a network. In effect of fixing ‘essential variables’ values Kauffman removed stability mechanisms of these variables from such networks and cannot look for cause of this stability inside investigated networks. Found ‘spontaneous order’ and ‘homeostatic stability’ of ordered phase is not this searched cause.
Appendix B - main Kauffman’s terms {#ch.Appendix B .unnumbered}
===================================
This paper discuss certain details in Kauffman’s approach and can be comprehensible only for readers familiar with this area, however, for completeness I introduce some short remaining of main terms which may be understood a little bit differently.
I use the term ‘chaos’ as @ooKauf does. It differs from the more common definition [@Schuster84] used for continuous arguments of a function. In networks there is a finite number $N$ of nodes, each one with a small natural number $s$ of possible states. Such chaotic systems differ from the ordered ones in damage spreading.
Damage (or ‘perturbations’ [@Serra04; @Ramo06]) is the difference between two identical functioning (dynamical) systems which appears as an effect of some disturbance in one of these systems [@Jan94]. Typically a very small change is investigated which initiates damage, e.g. a change of state of one element of the system. For chaotic systems a small initiation of damage typically causes a large avalanche of damage which spreads onto a big part of the system, however, it ends at an equilibrium level. The existence of this equilibrium level as the limit of damage growth is the main difference between this ‘chaos’ and the more commonly used definition. Levels of damage equilibrium are discussed in ch.2.2 basing on annealed approximation [@Derrida86; @ooKauf]. High stability of ordered systems does not allow for damage to evolve into a large avalanche. There could be very small avalanche in ‘small unfrozen lakes of activity’ which I don’t name ‘avalanche’ (however @Serra04 do), especially - ‘large avalanche’.
There are four similar terms used in the Kauffman approach which should be clearly separated: ‘structural stability’, ‘homeostatic stability’, ‘ultrastability’ and ‘spontaneous order’.
As it is described in Ref.[@ooKauf], the ‘structurally stable systems’ evolve in correlated landscape (e.g. of fitness) which typically allows the small initial change to give a small change as its effect. This landscape is considered on a space of system parameters where the nearest neighbours vary by the smallest change of connections, function or states. It means that such neighbours are similar and they function similarly. Chaotic systems, which are not ‘structurally stable’, evolve in uncorrelated landscape where small change typically causes crossing many walls of bifurcation, radically changing the system’s properties (e.g. adding a new basin of attraction). Function of neighbours typically are not similar. Adaptive evolution needs small changes. Large change typically lost collected aptness. Kauffman approach looks therefore for such ‘structurally stable’ area and found that it occurs between chaotic and ordered phase. It is the cause why in the Kauffman’s approach the phase transition between chaos and order is one of the most important themes of investigation. Kauffman isolates this area basing on implication for evolvability and structure of fitness landscape as third region named ‘liquid’ between ‘solid’ (because frozen - ch.1.2.2) ordered region and ‘gas’ - chaotic one. “Landscape is very rugged in the chaotic regime. This ruggedness is a direct consequence of the fact that damage spreads widely in networks in the chaotic regime. Almost any single mutation will dramatically alter landscape structure” [@ooKauf]. Note - not all mutation. Liquid region lies on formally ordered and chaotic regions and its boundaries are smooth and not well defined. [@Holland98] named this region ‘the edge-of-chaos membrane’.
When initiation of damage occurs then effected damage (change) can be large or small. Minimizing of initiation’s effect is the homeostatic feature. Typical homeostatic mechanisms based on negative feedbacks are named ‘ultrastability’ in Kauffman approach and are considered separately (see ch.1.2.2). Term ‘homeostatic stability’ Kauffman uses for general resistance of system to disturbance. This term is not formally introduced in [@ooKauf], it is absent in index there, however, it is used mainly for stability which emerges spontaneously together with spontaneous order which are main topic of Kauffman researches. Due to practical rejection of ultrastability from this researches (ch.1.2.2 and Appendix A) homeostatic stability contain there only aspect of spontaneous resistance to disturbance.
The ordered area, where maximum of spontaneous order (ch.1.2.1) occurs, exhibits also highest ‘structural stability’ which is useful for adaptive evolution and ‘homeostatic stability’ which should be an effect of evolution. The maximum of spontaneous order is one of the most important features of this area, especially when spontaneous order is taken to be the real one (ch.1.2.1). For Boolean networks it is the case of $K=2$: “If the stability of each state cycle attractor is probed by transient reversing of the activity of each element in each state of the state cycle, then 80-90% of all such perturbations, the system flows back to the same state cycle. Thus state cycles are inherently stable to most minimal transient perturbation” [@ooKauf].
Evolution using random walk in the space of network parameters should tend to this area because there effects of evolution are larger. This expectation is known as ‘life on the edge of chaos’. Kauffman even expects of living systems in solid regime in his “bold hypothesis: Living systems exist in solid regime near the edge of chaos, and natural selection achieves and sustains such a poised state” - page 232 in [@ooKauf].
A large change - large damage avalanche in chaotic systems is obviously taken as improbable in adaptive evolution but conclusion that adaptive evolution is improbable in chaotic systems would be too quick. As noted above (in description of liquid region) in chaotic systems small change only typically causes large effective change. There can also happen small effective changes which may be used by adaptive evolution. It happens more frequently if parameters of network are nearer phase transition between order and chaos. This possibility is discussed in ch.3.3 in fig.6.2 as degree of order of different networks and is one of more important theme of this paper.
In fig.8 emergence of chaos during network ($er$ 4,2 see ch.3.1.2. & ch.3.2.2.) growth is depicted. This network has 64 inputs and 64 outputs, which better describes living systems, it is not autonomous network as in ch.3. As mentioned above, chaos is a high probability of large damage avalanche after small disturbance. Damage $d$ (horizontal axis) is measured as a fraction of all $N$ nodes which have different state than in the not disturbed network. Small network where $N=50$ behave like at the edge of chaos and order. Its distribution is near power law, as for the avalanches in the pile of sand which Per Bak’s self-organized criticality [@Bak96] controls. But living objects have a method against such control - it is self-multiplying. In the effect typically a part of objects avoid large avalanche and can grow. This way they enter chaotic regime and can stay there. Now distribution consists of two peaks: The right one stay in the equilibrium point after large avalanche. Such large avalanches occur with high probability which define system as chaotic. Here ‘matured chaos’ starts about $N=600$, better $N=1000$ where very low probability occurs between peaks, which I discuss in Ref. [@bgec]. Large avalanches change most of node states which means that after such avalanche system works in totally different way. For living objects it means death, i.e. elimination. It cannot be a new evolutionary change (compare to [@Farmer86]) - it is much more than even Lysenko proposed. Can you imagine that more than half of all mechanisms changed and living object survived?
But there remains left peak which is similar to that for $N=50$. Now, however, with a real level of probability it covers only a small fraction of available nodes (each time a different set) but it is enough for adaptive evolution and life, which use only small changes. Without right peak there is no qualitative difference between new state of still living object and dead object. It is possible to investigate such simple model without such difference and with power law distribution, but it is huge simplification. This qualitative difference is important to differentiate between elimination and surviving, which is a base of Darwinian mechanism. Therefore edge of chaos with one left peak of power law distribution cannot be adequate area for living objects on which Darwinian mechanism works. When network grows, or its other parameters change causing growth of chaos level and decreasing order level, then probability of large avalanche also grows and probability of surviving decreases. Therefore life cannot quickly run deep into chaos, however, Darwinian mechanism collects certain mechanisms which can neutralize certain types of initialising changes. Now these known type (ch.1.3.1.) initialisations does not lead to large avalanche, i.e. elimination, and chance of survival goes up. Kauffman (and I too) does not model these ‘known type initialisations’ nor connected to them homeostatic answers of system, based typically on negative feedbacks, but this large simplification must be taken under consideration when stability of living objects is discussed. Authors of Ref. [@Wagner01; @Serra04; @Ramo06; @Serrajtb07] compare real biological stability measured by @Hughes00 to structural stability near the edge of chaos and they obtain similarity. They treat this similarity as the main evidence that living objects stay on the edge of chaos but they neglect taking this simplification into account.
![Damage size $d$ distribution during network $er$ 4,2 (64 inputs and outputs) growth. Emergence of chaos and death (right peak with large avalanche). Damage $d$ is measured as a fraction of all $N$ nodes which have different state than in the no disturbed network. Small network ($N=50$) behave like on the edge of chaos and order (near power law). This peak approximately remains as left one in chaotic area and there is a place for life evolution.[]{data-label="fig:8"}](a8.eps){width="8.5cm"}
R. Albert, A.-L. Barabási, 2000. Dynamics of Complex Systems: Scaling Laws for the Period of Boolean Networks. Phys. Rev. Lett. **84,24**, pp 5660–5663
R. Albert, A.-L. Barabási, 2002. Statistical mechanics of complex networks. Rev. Mod. Phys., **74,1**, pp 47–97
R. Albert, H.G. Othmer, 2003. The topology of the regulatory interactions predicts the expression pattern of the segment polarity genes in Drosophila melanogaster. J. Theor. Biol. **223**. pp 1–18
M. Aldana, 2003. Dynamics of Boolean Networks with Scale Free Topology. Physica D **185**, pp 45-66
M. Aldana, S. Coppersmith, and L. P. Kadanoff, 2003. Boolean Dynamics with Random Couplings.in *Perspectives and Problems in Nonlinear Science*, Applied Mathematical Sciences Series, edited by E. Kaplan, J. E. Marsden, and K. R. Sreenivasan (Springer-Verleg, Berlin, 2003). http://arXiv.org/pdf/nlin.AO/0204062
W. R. Ashby, 1960. *Design for a Brain* 2nd edn. (Wiley, New York )
N. Ay, E. Olbrich, N. Bertschinger, J. Jost, 2006. A unifying framework for complexity measures of finite systems. Proceedings of ECCS06 $http:// cssociety.org/tiki$-$index.php?page$=$ECCS\%2706+Programme$
P. Bak, 1996. *How Nature Works: The Science of Self-Organized Criticality*, New York: Copernicus.
F. J. Ballesteros, B. Luque, 2005. Phase transitions in random networks: Simple analytic determination of critical points. Phys. Rev. E **71**, 031104
W. Banzhaf, 2003. On the Dynamics of an Artificial Regulatory Network. In *Advances in Artificial Life, 7th European Conference, ECAL’03, Lecture Notes in Artificial Intelligence*, **2801**, 217–227 Springer.
A.-L. Barabási, R. Albert, H. Jeong, 1999. Mean-field theory for scale-free random networks. Physica A **272**, pp 173–187
A.-L. Barabási, E. Bonabeau, 2003. Scale-Free Networks. Scientific American, www.sciam.com pp 50–59
S. Bilke, F. Sjunnesson, 2001. Stability of the Kauffman Model. Phys. Rev. E **65**, 016129..
S. Bornholdt, T. Rohlf, 2000. Topological Evolution of Dynamical Networks: Global Criticality From Local Dynamics, Phys. Rev. Letters **84**, 6114–6117.
P. Crucitti, V. Latora, M. Marchiori, A. Rapisarda, 2004. Error and attacktolerance of complex networks Physica A **340** pp 388–394
B. Derrida, Y. Pomeau, 1986. Random Networks of Automata: A Simple Annealed Approximation. Europhys. Lett., **1(2)**, pp 45–49
B. Derrida, G.Weisbuch, 1986. Evolution of Overlaps Between Configurations in Random Boolean Networks. Journal De Physique **47**, 1297– 1303.
P. Erdős and A. Rényi, 1960. Random graphs. Publication of the Mathematical Institute of the Hungarian Academy of Science, **5**, 17–61
J. D. Farmer, S. Kauffman, N. H. Packard, 1986. Autocatalytic Replication of Polymers. Physica D **22(1-3)**, 50–67.
S. Fortunato, 2005. Damage spreading and opinion dynamics on scale-free networks. Physica A **348**, 683–690
P. Fronczak, A. Fronczak, J.A. Ho³yst, 2008. Kauffman Boolean model in undirected scale-free networks. Phys. Rev. E **77**, 036119
L.K. Gallos, P. Argyrakis, A. Bunde, R. Cohen, S. Havlin, 2004. Tolerance of scale-free networks: from friendly to intentional attack strategies. Physica A 344, 504 - 509
A. Gecow, 1975. A cybernetic model of improving and its application to the evolution and ontogenesis description. In: *Proceedings of Fifth International Congress of Biomathematics* Paris
A. Gecow, A. Hoffman, 1983. Self-improvement in a complex cybernetic system and its implication for biology. Acta Biotheoretica **32**, pp 61–71
A. Gecow, M. Nowostawski, M. Purvis, 2005. Structural tendencies in complex systems development and their implication for software systems. Journal of Universal Computer Science, **11** pp 327–356 $http://www.jucs.org/$ $jucs\_11\_2/structural\_tendencies\_in\_complex$
A. Gecow, 2005. From a “Fossil” Problem of Recapitulation Existence to Computer Simulation and Answer. Neural Network World. **3/2005**, pp 189–201 $http:// www.cs.cas.cz/nnw/contents2005/number3.shtml$
A. Gecow, 2008. Structural Tendencies - effects of adaptive evolution of complex (chaotic) systems. Int.J Mod.Phys.C, **19, 4**, pp 647–664.
A. Gecow, 2008a. The purposeful information. On the difference between natural and artificial life. Dialogue & Universalism **11-12**, 191–206.
A. Gecow, 2010. Complexity Threshold for Functioning Directed Networks in Damage Size Distribution. arXiv:1004.3795v1
A. Gecow, 2010a. The differences between natural and artificial life Towards a definition of life arXiv:1012.2889
A. Grabowski, R.A. Kosiński, 2004. Epidemic spreading in a hierarchical social network. Phys.Rev.E, **70**, 031908
A. Grabowski, R.A. Kosiński, 2006. Ising-based model of opinion formation in a complex network of nterpersonal interactions. Physica A, **361**, 651–664
S. E.Harris, B.K. Sawhill, A. Wuensche, S.A. Kauffman, 2002. A model of transcriptional regulatory networks based on biases in the observedregulation rules. Complexity **7**, 23–40
J. H. Holland, 1998. *Emergence: from chaos to order*. Oxford University Press
T. R. Hughes et al., 2000. Functional discovery via a compendium of expression profiles. Cell textbf[102]{}, 109–126.
K. Iguchi, S.-i. Kinoshita, H. Yamada, 2007. Boolean dynamics of Kauffman models with a scale-free network. J. Theor. Biol. **247**, pp 138–151
D. Jacobmeier, 2005. Multidimensional Consensus Model on a Barabasi-Albert Network. Int. J. Mod. Phys. C **16,4**, 633–646
N. Jan, L. de Arcangelis, 1994. Computational Aspects of Damage Spreading. In: *Annual Reviews of Computational Physics I*, ed by D. Stauffer (World Scientific, Singapore) pp 1–16
S.A. Kauffman, 1969. Metabolic stability and epigenesis in randomly constructed genetic nets. J. Theor. Biol. **22**, 437-467
S.A. Kauffman, 1971. Gene regulation networks: a theory for their global structure and behaviour. Current topics in dev. biol. 6, 145.
S. A. Kauffman, 1990. Requirements for Evolvability in Complex Systems - Orderly Dynamics and Frozen Components, Physica D **42**, 135–152
S.A. Kauffman, 1993. *The Origins of Order: Self-Organization and Selection in Evolution* (Oxford University Press New York)
S.A. Kauffman, C. Peterson, B. Samuelsson, C. Troein, 2004. Genetic networks with canalyzing Boolean rules are always stable. PNAS **101,49**, pp 17102–17107
J. F. Knabe, C. L. Nehaniv, M. J. Schilstra, T. Quick, 2006. Evolving Biological Clocks using Genetic Regulatory Networks. In *Proceedings of the Artificial Life X Conference (Alife 10)*, 15–21, MIT Press
B. Korzeniewski, 2001. Cybernetic formulation of the definition of life. J.Theor.Biol. **209**, pp 275–286
B. Korzeniewski, 2005. Confrontation of the cybernetic definition of living individual with the real word. Acta Biotheoretica **53**, pp 1–28
B-O Küppers, 1986. Der Usprung biologischer Information. Zur Naturphilosophie der Lebensentstelung. R.Piper Gmbh & KG., München
B. Luque, R.V. Solé, 1997. Phase transitions in random networks: Simple analytic determination of critical points. Phys. Rev. E 55(1), pp 257–260
B. Luque, F.J. Ballesteros, 2004. Random walk networks. Physica A **342** pp 207–213
M. Nowostawski, M. Purvis, 2007. Evolution and Hypercomputing in Global Distributed Evolvable Virtual Machines Environment. In: *Engineering Self-Organising Systems*, ed by S.A. Brueckner, S. Hassas, M. Jelasity, D. Yamins (Springer-Verlag, Berlin Heidelberg) pp 176–191
P. Rämö, J. Kesseli, O. Yli-Harja, 2006. Perturbation avalanches and criticality in gene regulatory networks. J. Theor. Biol. **242** 164–170
H. Schuster, 1984. Deterministic Chaos: An Introduction. Physik-Verlag
R. Serra, M. Villani, A. Semeria, 2004. Genetic network models and statistical properties of gene expression data in knock-out experiments. J. Theor. Biol. **227**, pp 149–157
R. Serra, M. Villani, L. Agostini, 2004. On the dynamics of random Boolean networks with scale-free outgoing connections. Physica A **339** 665 – 673
R. Serra, M. Villani, A. Graudenzi, S. A. Kauffman, 2007. Why a simple model of genetic regulatory networks describes the distribution of avalanches in gene expression data. J. Theor. Biol. **246** pp 449–460 R. Serra, M. Villani, A. Barbieri, S. A. Kauffman, A. Colacci, 2010. On the dynamics of random Boolean networks subject to noise: attractors, ergodic sets and cell types. J. Theor. Biol. **265**, 185–193
I. Shmulevich, S. A. Kauffman, M. Aldana, 2005. Eukaryotic cells are dynamically ordered or critical but not chaotic. PNAS **102 (38)**, 13439–13444
R.V. Sole, B. Luque, S. Kauffman, 2000. Phase transitions in random networks with multiple states. Technical Report 00-02-011, Santa Fe Institute
A.O. Sousa, 2005. Consensus formation on a triad scale-free network. Physica A **348** pp 701–710
D. Stauffer, A. Sousa, Ch. Schulze, 2004. Discretized Opinion Dynamics of The Deffuant Model on Scale-Free Networks. Journal of Artificial Societies and Social Simulation **7** $http://jasss.soc.surrey.ac.uk/7/3/7.html$
D. Stauffer, S. Moss de Oliveira, P.M.C. de Oliveira, J.S. Sa Martins, 2006. *Biology, Sociology, Geology by Computational Physicists* (Elsevier, Amsterdam), 276 + IX pages.
A. Wagner, 2001. Estimating coarse gene network structure from large-scale gene perturbation data. Santa Fe Institute Working Paper, 01-09-051.
[^1]: It is the original example which Kauffman used (p.201) to explain a base of his hypothesis about spontaneous order. I discuss here this base. Later finding of Samuelsson and Troein (Phys. Rev. Lett. 90, 098701 (2003)) that used here the expected median state cycle length scales faster than any power law, changes only parameters of this base making them less extreme, but doesn’t change the main view. Approximated number like ‘317’ will change significantly but nor regarding this problem.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new single-ion endcap trap for high precision spectroscopy that has been designed to minimize ion-environment interactions. We describe the design in detail and then characterize the working trap using a single trapped $^{171}$Yb$^{+}$ ion. Excess micromotion has been eliminated to the resolution of the detection method and the trap exhibits an anomalous phonon heating rate of $d\langle n\rangle /dt = 24 ^{+30}_{-24}$s$^{-1}$. The thermal properties of the trap structure have also been measured with an effective temperature rise at the ion’s position of $\Delta T_{\rm{(ion)}} = 0.14 \pm 0.14$K. The small perturbations to the ion caused by this trap make it suitable to be used for an optical frequency standard with fractional uncertainties below the $10^{-18}$ level.'
author:
- 'P.B.R. Nisbet-Jones'
- 'S.A.King'
- 'J.M.Jones'
- 'R.M.Godun'
- 'C.F.A.Baynham'
- 'K.Bongs'
- 'M.Doležal'
- 'P.Balling'
- 'P.Gill'
bibliography:
- 'ECTrap\_20102015\_arXiv.bib'
date: 21 October 2015
title: 'A Single-Ion Trap with Minimized Ion-Environment Interactions '
---
Introduction
============
Trapped single atomic ions can provide a near-idealized system in which a quantum object is decoupled from its environment, and any interactions can be precisely tailored. Many uses for such traps have been found, from frequency metrology and tests of fundamental physics [@Godun2014a; @Huntemann2014a; @Wansbeek2008], to simulations of quantum systems [@Gerritsma2010; @Menicucci2007] and quantum information processing [@Hucul2015; @Gessner2013], however the imperfections inherent with any physical system introduce perturbations and uncertainties to the experiments being performed. In many cases these can limit the operational capability of the trapping system.
This paper presents a new single-ion endcap trap specifically designed to reduce the ion’s interaction with its environment. Particular care has been taken to minimize dc and rf phase-offset induced micromotion, thermal effects, and ion heating rates, and also to maximize optical access, resulting in significant improvements over existing designs. Although this trap was developed with a particular interest in precision spectroscopy of the $^{2}$S$_{1/2}-^{2}$F$_{7/2}$ electric octupole (E3) transition in $^{171}$Yb$^{+}$ for frequency metrology and tests of fundamental physics, its performance is applicable to ion trapping for a wide variety of applications.
The requirements on an ion trap for fundamental frequency metrology with fractional uncertainty at the $10^{-18}$ level are presented in Section\[sec:req\]. Details of the design and fabrication of the trap which meets these requirements are given in Section\[sec:design\]. Characterization of the trap using a single $^{171}$Yb$^{+}$ ion is performed in Section\[sec:charac\].
Trap Requirements {#sec:req}
=================
An ion trap and vacuum system for high precision frequency spectroscopy must:
1. [Minimize the thermal black-body radiation (BBR) emissions from the trap, resulting in the BBR environment at the ion’s position being dominated by the well controlled ambient field;]{}
2. [Minimize the anomalous phonon heating rate, $\Gamma_{\rm{h}}$ of the ion;]{}
3. [Minimize stray electric fields which result in rf-induced micromotion;]{}
4. [Minimize collisions with background gas;]{}
5. [Maximize fluorescence collection from the ion to improve the state detection signal-to-noise ratio (SNR).]{}
The first two of these requirements are in conflict. The dominant thermal heating mechanism of the ion trap structure is dielectric loss in the insulators which isolate the rf and ground sections of the trap. This loss takes the form $P = V\epsilon'_{r}\epsilon_{0}\tan\delta\Omega_{\rm{rf}} E^{2}$, where the power $P$ dissipated into a dielectric of volume $V$ depends on the dielectric’s relative permittivity $\epsilon'_{r}$ and loss tangent $\tan\delta$ when an electric field $E$ of frequency $\Omega_{\rm{rf}}$ is applied across it. Whilst selecting materials which have a low loss tangent and relative permittivity is important, the easiest way to reduce the thermal heating of the trap structure is to operate the trap with the lowest possible values for $\Omega_{\rm{rf}}$ and $E$. However, to achieve a low phonon heating rate $\Gamma_{\rm{h}}$ the trap should be operated at the highest possible drive frequency and electric field to provide tight ion confinement and hence high secular frequencies $\omega_{r,z}$.
The anomalous heating rate also has a $z_{0}^{-\alpha}$ dependence on the ion-electrode separation $z_{0}$. The exponent $\alpha$ depends on the noise process involved but comparison between traps produced in different laboratories indicate a range of $2<$$\alpha$$<4$ [@Brownnutt2014] making it highly beneficial to have a large electrode spacing. To maintain a given secular frequency $\omega_{r,z}$ when changing $z_{0}$, $E$ must be increased as $z_{0}^{2}$. These effects make it clear that careful design is required to find a suitable compromise configuration in which the trap can operate successfully within the above constraints.
The other requirements are more straightforward to meet. Relative phase delays between the rf voltage on the two endcaps will result in micromotion that cannot be compensated using only dc fields. The trap structure must therefore be totally symmetric to ensure that the effective path length between the endcaps and the rf feed, including all possible capacitive and inductive delays, is equal. Likewise the ion must be shielded from any rf field that does not come from the endcaps. The ion trap must be contained within an UHV/XHV environment to reduce pressure-dependent shifts on the order of $\sim10^5$HzPa$^{-1}$ [@Dube2013]. Finally, any apertures in the imaging setup should be placed as close to the ion as possible to increase the numerical aperture (NA) and hence fluorescence collection efficiency. Fast state detection is necessary for transitions with short excited state lifetimes to avoid losing excitation probability [@Peik2006], increasing the clock instability via reduced live time, and introducing the Dick effect from the resulting under-sampling of the local oscillator noise. A high SNR also allows easy use of photon correlation spectroscopy to minimize the excess micromotion.
Trap Design and Fabrication {#sec:design}
===========================
The trap presented in this paper is based on the endcap design [@Schrama1993], and consists of two cylindrical electrodes that are held at an equal rf potential facing one another whilst contained within a pair of independent dc electrodes. This basic design is in use in multiple laboratories, [@Godun2014a; @Madej2012; @Takahashi2013] and has the benefit of excellent optical access in the radial plane. Recent experiments with fiber-optics contained within the endcap electrodes have extended this access into the axial direction [@Takahashi2013].
![Vertical cross-section through the trap structure. The rf endcaps are secured into the OFHC copper with M2 set-screws and sit within hollow dc endcap electrodes that are supported by fused silica spacers and recessed alumina bolts.[]{data-label="fig:trap"}](Figure1.pdf){width="\columnwidth"}
A schematic of the trap is shown in Figure\[fig:trap\]. The support structure for the trap, which also acts as the high voltage rf feed, is made of oxygen-free, high thermal conductivity (OFHC) copper. OFHC copper has both high electrical and thermal conductivity, and exhibits low out-gassing which will reduce ion loss via chemical reactions of the ion with oxygen. The two 99.9% purity molybdenum rf endcap electrodes are secured into the copper mount using an interference fit and M2 set-screws. Molybdenum has good electrical and thermal conductivity, is non-magnetic, and can be polished to a high surface quality. Material properties are also thought to contribute to $\Gamma_{\rm{h}}$ and whilst the mechanism behind this is unclear, traps with cleaned Mo electrodes have demonstrated excellent results [@Turchette2000]. The faces of the rf electrodes which subtend the ion were mechanically polished to achieve a mirror quality finish as shown in Figure\[fig:profile\]. The mean surface roughness was measured via focal-variation microscopy to be $\rm{R}_{\rm{a}}=20$nm, resulting in an emissivity of $\epsilon\approx0.02$ at 300K. As the inner endcap electrodes represent the largest solid angle contribution to the ion’s field of view, the low emissivity reduces the thermal radiation that the ion will experience from the temperature of these endcaps. The polished surface also has the potential to reduce $\Gamma_{\rm{h}}$ as it removes any possible points for field emission of electrons and ensures a uniform surface free from adsorbates which have been proposed as a source of electric field noise [@Brownnutt2014]. The electrode tips have a diameter of $750\mu$m and a separation of $1000\mu$m.
![Focal variation microscopy image of the polished surface of the upper rf endcap electrode. The inset plot shows the surface profile along the red line. The 150nm pit highlighted by the yellow marker (1) is typical of the other imperfections which are visible as black dots in this image, although on a larger scale. []{data-label="fig:profile"}](Figure2.pdf){width="\columnwidth"}
Two molybdenum dc electrodes surround the rf endcaps and form the ground reference for the rf potential in the trapping region. The gap between the rf and dc electrodes in the trapping region is 135$\mu$m to ensure a good quadrupole potential. This gap increases to 1mm at the base to reduce the capacitance between the endcaps and thus reduce the rf current required to drive the trap. This also allows the rf electrodes to increase in thickness near the base increasing their mechanical ridigity. The dc electrodes are positioned using a pair of fused silica spacers and are secured with 99.9% purity alumina bolts. By placing the dielectrics in a non-critical region away from the trap center, it is possible to have a large separation between the high voltage and dc electrodes greatly reducing $E^{2}$ across the lossy dielectric. The position of the high emissivity dielectric behind the dc endcap also removes any direct line of sight between it and the ion.
The C-shape of the copper mount breaks the trap’s cylindrical symmetry. The asymmetric mounting protrusions on the dc endcaps effectively shield the ion from the copper mount restoring the surrounding ground potential (Figure\[fig:comsol\]). Without the extra shielding the rf minimum would be shifted from the geometric center of the trap, and there would no longer be a true micromotion zero due to the phase delay between the rf potential from the electrode tips and the center of the C.
[0.45]{} ![Illustration of the importance of extending the dc electrodes to shield the trapping region from the rf potential caused by the C-shaped mount. Simulation of the rf electric potential performed using the COMSOL software package. With the copper C-shaped mount unshielded the rf potential asymmetrically distorts the quadrupole potential in the trapping region.[]{data-label="fig:comsol"}](Figure3a.png "fig:"){width="\columnwidth"}
[0.45]{} ![Illustration of the importance of extending the dc electrodes to shield the trapping region from the rf potential caused by the C-shaped mount. Simulation of the rf electric potential performed using the COMSOL software package. With the copper C-shaped mount unshielded the rf potential asymmetrically distorts the quadrupole potential in the trapping region.[]{data-label="fig:comsol"}](Figure3b.png "fig:"){width="\columnwidth"}
The dc electric field in the trapping region is controlled by applying static voltages to the dc electrodes (axial direction) and to two fixed compensation electrodes set back from trap center by 5mm in the radial plane (Figure\[fig:Cartoon\]) with a separation of 60$^{\circ}$ to maximize optical access in the radial plane. The dc electrodes are capacitively shorted to ground via 1$\mu$F capacitors [^1] to prevent any rf pickup on the dc electrodes that could be re-radiated with a phase offset. These capacitors are located 50mm from trap center on the air-side of the dc feedthrough. The applied dc voltages are filtered close to the feedthrough using a combination of RC and LC $\pi$-filters [^2]
Isotopically enriched neutral $^{171}$Yb is contained within a resistively heated tantalum tube oven. The atomic beam is slightly collimated (opening angle of $\theta\approx10^{\circ}$) reducing the amount of Yb deposited on the trap structure during loading. Yb deposits can cause patch potentials leading to changes in the necessary compensation voltages and might also affect the phonon heating rate of the ion.
Vacuum System {#sec:vac}
-------------
{width="2\columnwidth"}
The vacuum system maximizes the optical access to the trap whilst maintaining a small light-weight package. The chamber was machined out of a single aluminum block and is fitted with non-magnetic CF mounting flanges. The trap region of the chamber takes the form of a $70\times70\times20$mm cuboid with a central $\diameter=40$mm cylindrical opening (Figure\[fig:Cartoon\]). This opening is sealed using two $\diameter=50$mm viewports. The laser access and ion imaging window is anti-reflection coated UV-grade fused silica (UVFS). The other window is magnesium fluoride (MgF$_{2}$), which is transparent up to 8$\mu$m to allow direct thermal imaging of the trap structure during operation. Both windows are cold-welded to the chamber with $\diameter=1$mm indium wire, achieving undetectable He leak rates.
The indium sealed windows allow a maximum optical access opening angle of $2\theta=112\,^{\circ}$, compared to the $2\theta=80\,^{\circ}$ from a traditional CF optic, and allows us to probe and cool the ion in two perpendicular radial directions. This also maximizes the effectiveness of the polarization spinning technique [@Berkeland2001] and ensures a high fluorescence rate. Optical access in the axial direction is limited by the trap structure itself at $2\theta=103\,^{\circ}$.
Ultra-high vacuum is maintained by a SAES NexTorr D100-5 combination non-evaporable getter (NEG) and ion-getter pump. The NEG provides a pumping speed of up to 100ls$^{-1}$ for active gases (H$_{2}$, H$_{2}$O, N$_{2}$, O$_{2}$, etc.), while inert gases are pumped by the ion getter pump at a rate of 6ls$^{-1}$. After final assembly a bakeout at 125$^{\circ}$C roughed by a turbo-molecular pump achieved a base pressure of $<1\times10^{-9}$Pa ($10^{-11}$mbar) determined by the current draw of the ion pump. For a Langevin-approximated frequency shift of 10$^5$HzPa$^{-1}$ [@Dube2013] this corresponds to 100$\mu$Hz.
Light delivery and collection
-----------------------------
The ion is illuminated along three optical axes which probe perpendicular radial directions and sample the axial motion of the ion at an angle of $\theta=40\,^{\circ}$ (Figure\[fig:Cartoon\]). The horizontal beam contains all the wavelengths that are used in the experiment which range from the UV (369nm cooling) to the IR (935nm re-pump). These wavelengths are combined using dichroic beamsplitters and coupled into an endlessly-single-mode photonic crystal fibre[^3]. The fibre output is brought to an achromatic focus of waist $w=19\,\mu$m at the ion’s position using a pair of off-axis parabolic mirrors, ensuring perfect overlap of all lasers onto the ion. The two beams that are raised out of the horizontal plane only supply cooling light.
Atomic fluorescence is collected through the UVFS viewport using an aspheric lens pair in a 2.6:1 telescope. After spatial and spectral filtering an on-resonance fluorescence signal of $20,000\,s^{-1}$ is obtained for a single ion at $I_{\rm{sat}}$. For a typical background count rate of $50\,s^{-1}$ this results in a signal to noise ratio of SNR = 400.
Trap Characterization {#sec:charac}
=====================
The trap performance was characterized using single $^{171}$Yb$^{+}$ ions however the performance should be directly transferable to other elements. Ions have been loaded and trapped at rf frequencies from $\Omega_{\rm{rf}}/2\pi=$$3$$-$$21$MHz. The secular frequencies of trapped single ions were measured by ‘tickling’ the ion with a low-power rf field applied to one of the dc endcaps. When resonant with the ion’s secular frequency this tickling field heats the ion, and greatly increases the observed fluorescence from a far red-detuned cooling laser. Secular frequencies have been arbitrarily set from the kHz to the MHz range up to a maximum of $\omega_{(x,y,z)}=2\pi\times(1.02,1.04,2.04)$MHz, obtained using a helical resonator with a loaded resonant frequency of $14$MHz with 1W of forward power. Ion lifetimes on the order of a week are routinely observed at all frequencies. Uncooled ion lifetimes of up to 15hrs have been observed on multiple occasions. Collisions with background gas molecules can populate the long-lived F-state in Yb$^{+}$, and without repumping, is equivalent to loss of the ion. We routinely observe $>24$hr lifetimes without collisional transfer to this state which supports the low vacuum pressure measured in Section\[sec:vac\].
3D Micromotion minimization {#sec:micromotion}
---------------------------
For ion-based atomic clocks, residual excess micromotion from imperfectly compensated stray potentials causes two of the most significant trap-induced systematic shifts: the dc-Stark shift and the time-dilation shift [@Godun2014a; @Huntemann2012; @Chou2010a]. Additionally, when pushed away from the rf potential minimum the ion is exposed to the possibility of electrical noise from the trap drive.
The projection of the ion’s micromotion in the direction of any one of the cooling beams can be measured using the photon correlation technique. The excess micromotion induces a first-order Doppler shift of the cooling laser in the frame of the ion, which can be observed as a modulation of the ion’s fluorescence at the trap drive frequency when the ion is illuminated with light at the cooling transition half-maximum. The observed fluorescence signal $S_{\rm{obs}}$ from each cooling beam can then be written as $$S_{\rm{obs}} = \frac{S_{\rm{max}}}{2}\left[ 1 + S_{\rm{mod}}\sin(\Omega_{\rm{rf}}t+\phi)\right],$$ where the maximum signal $S_{\rm{max}}$ is sinusoidally modulated at the trap drive frequency $\Omega_{\rm{rf}}$ with phase $\phi$ and fractional amplitude $S_{\rm{mod}}$. To minimize the micromotion, voltages are applied to the compensation electrodes and the resulting electric field pushes the ion towards the rf potential minimum reducing the value of $S_{\rm{mod}}$. A complete analysis of this technique, including effects caused by the fluorescence transition’s finite linewidth is given in [@Keller2015].
![Photon correlation signal and sinusoidal fits for an ion with a flourescence modulation of $S_{\rm{mod}}=0.5$ (black triangles), $S_{\rm{mod}}=0.12$ (blue squares), and $S_{\rm{mod}}=0.01$ (green circles). []{data-label="fig:RFphoton"}](Figure5.pdf){width="\columnwidth"}
An example of the resulting correlation plot is shown in Figure\[fig:RFphoton\] for three different levels of micromotion. Micromotion along the two perpendicular axes of the beams in the radial plane is shown in Figure\[fig:C2\] as a function of the voltage applied to one of the radial compensation electrodes. As the laser beams have components along both of the ion-to-compensation-electrode axes, changing one of the compensation voltages will alter the micromotion observed along both beams. The micromotion projection along a given beam is minimized using the more sensitive compensation electrode, and through an iterative process can be eliminated to the measurement resolution of the system, resulting in $S_{\rm{mod}}=0.000(5)$.
After micromotion minimization the trap has been consecutively reloaded up to twenty times over a 5hour period without noticeable changes in the micromotion at the $S_{mod}=0.01$ level being observed. No long term drifts in the micromotion have been observed suggesting that vacuum windows are sufficiently far from the ion that dielectric charging from the UV laser beams is insignificant [@Harlander2010] and that the trap is electrically and mechanically stable.
![Modulation depth of the photon correlation signal as a function of the voltage on one of the compensation electrodes in the radial plane for the beams that are at an angle of $75^{\circ}$ (black) or $15^{\circ}$ (red) in the radial plane. The green shaded area represents the modulation depth where the dc-Stark and time dilation shifts are below $1\times10^{-18}$. A sign change in the modulation depth corresponds to a $\pi$ phase change in the modulation signal.[]{data-label="fig:C2"}](Figure6.pdf){width="\columnwidth"}
Anomalous heating rate of the ion
---------------------------------
In many experiments such as optical atomic clocks or quantum computation based on optical qubits, it is desirable to resolve narrow lineshapes on electric dipole forbidden transitions. The peak excitation (gate fidelity) is limited by the dephasing caused by the thermal distribution of the vibrational quantum number of the ion. Even if an ion is cooled to its vibrational ground state before interrogation, electric field noise will cause the ion to heat during the probe pulse, thus reducing the fringe contrast. As the wavefunction of a hot ion is larger and it experiences more of the trapping potential a high heating rate will also increase the dc Stark and time-dilation shifts caused by the ion’s secular motion.
The phonon heating rate of the trap was measured using the Doppler re-cooling technique. After a prolonged period without laser cooling the cooling laser, tuned close to the unperturbed line center, is re-applied to the ion. The temperature of the ion before re-application of the cooling laser can be determined by analysis of how the fluorescence returns. A full description of this technique can be found in [@Wesenberg2007; @Epstein2007; @Brownnutt2014].
The trap was operated at an rf frequency of $\Omega_{\rm{rf}}/2\pi = 7.62$MHz with a secular frequency $\omega_{r}/2\pi=500$kHz. This is much lower than the natural linewidth of the cooling transition $\Gamma/2\pi = 19.6$MHz, ensuring that the ion is within the weak-binding limit where the semi-classical Doppler cooling model is valid. The ion was allowed to heat for periods up to 50s before the 5MHz red-detuned cooling laser was reapplied and the fluorescence return counted in 15$\mu$s time-bins. The fluorescence return was measured 3000 times to decrease the statistical uncertainty and the results averaged as shown in Figure\[fig:phonon\](a). There was no resolvable change in the fluorescence level throughout the 10ms counting window. The resolution of the measurement was limited to $d\langle n\rangle /dt = 100$s$^{-1}$ by the signal-to-noise achieved in our 15$\mu$s binning of the photon counter signal. This yields an upper limit on the phonon heating rate of $d\langle n\rangle /dt = 50 ^{+100}_{-50}$s$^{-1}$ with $1\sigma$ uncertainty curtailed at zero.
![The top plot shows ion fluorescence as a function of re-cooling time for a Doppler re-cooling measurement of the trap anomalous heating rate. The ion was illuminated at $I=I_{\rm{sat}}$ and at a detuning of 5MHz. Fluorescence return was measured after 30s without cooling. The red line shows the theoretical fluorescence return for a heating rate $d\langle n\rangle /dt = 50$. The inset plot shows the first 300$\mu$s of recooling. The bottom plot shows the Rabi oscillation decay on the E3 transition for post-cooling delay times of 10ms (hollow circles), 50ms (purple triangles), 200ms (orange squares), and 500ms (green circles). Peak excitation of the first Rabi pulse was limited to 85% by noise on the excitation laser.[]{data-label="fig:phonon"}](Figure7a.pdf){width="\columnwidth"}
![The top plot shows ion fluorescence as a function of re-cooling time for a Doppler re-cooling measurement of the trap anomalous heating rate. The ion was illuminated at $I=I_{\rm{sat}}$ and at a detuning of 5MHz. Fluorescence return was measured after 30s without cooling. The red line shows the theoretical fluorescence return for a heating rate $d\langle n\rangle /dt = 50$. The inset plot shows the first 300$\mu$s of recooling. The bottom plot shows the Rabi oscillation decay on the E3 transition for post-cooling delay times of 10ms (hollow circles), 50ms (purple triangles), 200ms (orange squares), and 500ms (green circles). Peak excitation of the first Rabi pulse was limited to 85% by noise on the excitation laser.[]{data-label="fig:phonon"}](Figure7b.pdf){width="\columnwidth"}
An alternative method of measuring the anomalous heating rate is to observe dephasing of the Rabi oscillations on an atomic transition as a function of the time delay between cooling the ion and starting a Rabi transition. The excitation probability $P$ can be calculated from the excitation probability $P_{\rm{n}}$ and Rabi frequency $\Omega_n$ for each motional state $n$ [@Letchumanan2004] along with the probe-laser-induced decoherence over the probe time $t$, caused by its finite coherence time $\tau$: $$\begin{aligned}
P_{}= \frac{1}{4}\left(1+e^{-t/\tau}\right)\left( 1-\sum_{n}P_{\rm{n}}\cos\left(\Omega_{n} t \right) \right)
\label{eq:rabi}\end{aligned}$$ Ion heating during the probe pulse also affects the dephasing rate; however this effect is small for the conditions in this trap. During this measurement the trap was operated at an rf frequency of $\Omega_{\rm{rf}}/2\pi = 20.89$MHz with a secular frequency $\omega_{r}/2\pi=776$kHz. Rabi oscillations were driven on the 467nm: $^{2}$S$_{1/2}-^{2}$F$_{7/2}$ transition following a delay of up to 500ms between the end of the cooling cycle and interrogating the transition. Fitting Equation\[eq:rabi\] to this data (Fig.\[fig:phonon\]) determined the heating rate to be $d\langle n\rangle /dt = 24 ^{+30}_{-24}$s$^{-1}$, which corresponds to an electric field noise density of $S_{\rm{E}} = 5.5\times 10^{-13}$V$^2$m$^{-2}$Hz$^{-1}$. This heating rate is in line with heating rates for other traps of this size. More precise measurements of the heating rate would require alternative techniques such as carrier-to-sideband ratio measurements, which are not possible at this stage due to the necessary probe laser powers. The fundamental limit to the ion heating rate is expected to be the Johnson noise of the LRC circuit connected to the endcap electrodes. [@Brownnutt2014].
Thermal measurements of the trap structure {#sec:heat}
------------------------------------------
One of the dominant causes of uncertainty in state-of-the-art optical atomic clocks is the Stark shift caused by the emission of black-body radiation from the trap and vacuum chamber apparatus. It is therefore essential that any trap is designed with good thermal properties so that any temperature rise during operation is low, can be accurately determined, and can be kept stable.
Finite element models (FEM) were developed by CMI [@Dolezal2015] to calculate the changes in the temperature distribution across the ion trap caused by its operation, and thus the effective temperature of the total BBR environment seen by the ion. Models of the trap show that the maximum temperature rise of any part of the trap structure should be 0.15K at standard operating conditions, and are shown in Fig.\[fig:CMIModel\]. As the solid angle subtended by the hot regions of the trap is very small, and the rf electrode, which has a very large solid angle, has been polished to a mirror finish, the effective temperature rise that should be seen by the ion is much lower than this, $\Delta T_{\left(\rm{ion}\right)} = 0.02 \pm 0.02$K.
![FEM simulation of the thermal profile cross-section for an applied rf potential at $\Omega_{\rm{rf}}/2\pi=15$MHz required to give $\omega_{r}/2\pi=1$MHz . A maximum temperature rise of 0.15K was found in the fine wires which connect to the dc endcap electrodes.[]{data-label="fig:CMIModel"}](Figure8.png){width="\columnwidth"}
The thermal characteristics of the operating trap were measured in situ through the MgF$_{2}$ window using a Cedip Silver thermal infrared camera in the 3-5$\mu$m radiation band Fig.\[fig:ThermalImage\]. The imaging setup was pre-calibrated by measuring target samples at known temperatures. During initial characterization of the trap [@Dolezal2015] it was found that dielectric heating in the UHV feedthrough completely dominated the heating from the trap structure itself, causing a temperature rise of 6.5K when secular frequencies of 1MHz were generated at a drive frequency of 21MHz. A simple heat sink made from a thermally conductive ceramic post (Shapal $\diameter=13$mm) reduced this temperature rise by a factor of two, and a further improvement in the feedthrough temperature was obtained by decreasing the drive frequency by 30%. With these minor modifications a temperature rise of 1.5K was measured both with the thermal camera and calibrated PT100 sensors on the air-side of the feedthrough, when an ion was trapped in a $\Omega_{\rm{rf}}/2\pi=14\,$MHz field with $\omega_{r}/2\pi=1\,$MHz. This results in an effective temperature rise at the ion of $\Delta T_{\left(\rm{ion}\right)} = 0.14 \pm0.14$K.
![IR images of the trap structure taken in the 3-5$\,\mu$m range (a) without an applied rf voltage, and (b) with $\Omega_{\rm{rf}}/2\pi=14\,$MHz field at $\omega_{r}/2\pi=1\,$MHz. The IR images are not corrected for emissivity and thus show the effective temperature that the ion observes rather than the actual temperature of the trap. Reflections belong to ambient laboratory radiation which is shielded during normal operation.[]{data-label="fig:ThermalImage"}](Figure9.png){width="\columnwidth"}
Conclusion
==========
We have designed and constructed a trap for single ions with the endcap topology in order to minimize trap-ion interactions and thus approach an ideal isolated quantum system. The trap has been characterized using a single $^{171}$Yb$^{+}$ ion to investigate the micromotion and anomolous heating rate properties of the trap. The trap was operated over a range of frequencies from $\Omega_{\rm{rf}}/2\pi=3 \rightarrow 21$MHz, and with secular frequencies $\omega_{r,z}$ from $200\,$kHz to the MHz range. Cooled ions remain in the trap almost indefinitely, and the uncooled ion-retention time is in excess of 15hrs. A fluorescence signal-to-noise ratio from a single ion of 400 is achieved.
Excess micromotion has been minimized to the level of $S_{\rm{mod}} = 0.000(5)$ as determined via the rf-photon correlation technique. The anomalous phonon heating rate of the trap was measured using the Doppler recooling method to be $d\langle n\rangle /dt = 50 ^{+100}_{-50}$s$^{-1}$ and via the observation of Rabi flop dephasing to be $d\langle n\rangle /dt = 24 ^{+30}_{-24}$s$^{-1}$ with the uncertainty limited by the experimental methods. Further investigation of the anomalous phonon heating rate using sideband spectroscopy is planned. The effective rise in the BBR temperature at the ion has been determined via a combination of thermal imaging and FEM models to be $\Delta T_{\left(\rm{ion}\right)}=0.14\pm 0.14$K limited by the rf loss in the vacuum feedthrough.
When used for an atomic clock based on the frequency of the E3: $^{2}S_{1/2}-^{2}F_{7/2}$ transition in $^{171}$Yb$^{+}$, perturbations induced by these trap parameters are presented in Table\[tab:ErrBud\]. These contributions, when summed in quadrature, constitute a fractional uncertainty on the transition frequency of $3.5\times10^{-19}$. This should be compared with the uncertainty contribution from trap-based frequency shifts in recently published absolute frequency measurements of $\sim1.5\times10^{-17}$ [@Godun2014a; @Huntemann2012]. The uncertainty could be further improved by reducing the secular shifts through well established techniques for ground-state cooling of the ion, and by fully heat-sinking the feedthrough. If successful, these improvements should reduce the uncertainty in the error budget by more than an order of magnitude.\
------------------------------------------------------------------------------------- -- --
**[Source of Shift]{} & $\delta\nu/\nu_{0}(10^{-18})$ & $\sigma/\nu_{0}(10^{-18})$\
Background gas collisions & 0.15 & 0.15\
Excess $\mu$-motion scalar dc-Stark & 0 & 0.1\
Excess $\mu$-motion time dilation & 0 & 0.1\
Secular motion scalar dc-Stark & -0.3 & 0.2\
Secular motion time dilation & 0.7 & 0.1\
BBR dc-Stark: offset from 293K & 0.2 & 0.2\
**[TOTAL]{} & 1.35 & 0.35\
****
------------------------------------------------------------------------------------- -- --
: Error budget for trap induced shifts and uncertainties for the E3: $^{2}$S$_{1/2}-^{2}$F$_{7/2}$ transition in $^{171}$Yb$^{+}$ when the trap is operated at $\Omega_{\rm{\rm{rf}}}/2\pi=14$MHz, with $\omega_{x,y,z}/2\pi=(1,1,2)$MHz. []{data-label="tab:ErrBud"}
This work was funded by the European Metrology Research Programme (EMRP), the UK National Measurement System, and the European Space Agency (ESA). The EMRP is jointly funded by the EMRP participating countries within EURAMET and the European Union.
[^1]: High SRF SMD components (X7R dielectric) ensures good performance in the 1-10MHz range.
[^2]: 10Hz and 100kHz respectively.
[^3]: NKT LMA-PM-5
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Th. Roudier, M. Rieutord,, V. Prat,, J.M. Malherbe, N. Renon , Z. Frank, M. Švanda, T. Berger, R. Burston, L. Gizon'
bibliography:
- 'biblio.bib'
date: 'Received / Submitted '
title: 'Comparison of solar horizontal velocity fields from SDO/HMI and Hinode data'
---
Introduction
============
The dynamics of the Sun’s surface is one of the major elements in understanding the time evolution of its magnetic activity. It is a real challenge to measure the surface motions at all spatial and temporal scales and compare them to those coming from the simulations. Recently, the HMI instrument aboard the SDO satellite allowed us a new step in that direction. By following the proper motions of the solar granules, representative of solar plasma evolution, it is possible to define the flow field on the solar surface [@Roud2012] from a small spatial scale of 2.5 Mm up to nearly 85% of the solar radius (Fig. 6 of that paper). We measured the velocities in a cartesian coordinate system, where $x$ and $y$ denote the coordinates in the sky plane with $x$ parallel to the direction of the solar rotation and $z$ directed towards the observer along the line of sight. Beyond 0.85 R$_\odot$, the $v_{\rm x}$ and $v_{\rm y}$ components appeared to be noisy, but the $ v_{\rm x}$ component showed a trend indicative of the solar rotation. In order to identify whether improvements could be made in the determination of the horizontal velocities beyond that limit, we used simultaneous observations of the Sun in white light at low (SDO/HMI) and high (Hinode) spatial resolution.
In this paper, we describe a comparison between velocity fields projected on the sky plane that are obtained in the south pole region with Hinode data and SDO/HMI data using the CST code [@RRRD07; @Roud2012]. In the next section we discuss the corrections done in order to get an accurate velocity close to the solar limb. From the $v_{\rm x}$, $v_{\rm y}$ and Doppler observations, we describe the transformation to the local surface velocities $v_r$, $v_\theta$, $v_\varphi$. Finally, we present an application to measure solar differential rotation with a short time sequence (3h) up to high latitudes with low noise. Discussion and conclusions follow.
Observations
============
Hinode observations
-------------------
We used data sets from the Solar Optical Telescope (SOT), onboard the *Hinode* [^1] mission [e.g., @STISO08; @ITSSO04]. The SOT has a 50 cm primary mirror with a spatial resolution of about 0.2 at a wavelength of 550 nm. For our study, we used blue continuum observations at 450.45 nm. from the *Hinode*/SOT BFI (Broadband Filter Imager). The observations were recorded continuously on December 10, 2011, from 16:11:33 to 19:07:05 UT. To get the limb as reference, the south pole was observed at the position reported in the Flexible Image Transport System (FITS) header of Hinode, i.e, $X_{\rm cent}=-10.08\arcsec$ and $Y_{\rm cent}=-969.84\arcsec$. The time step was 45 sec and the field of view was $111.6~\arcsec\times111.6~\arcsec$ with a pixel size of 0 1089. After alignment, the useful field of view reduced to $104\arcsec\times82\arcsec$. To remove the effects of the oscillations, we applied a subsonic Fourier filter. This filter was defined by a cone in $k$-$\omega$ space, where $k$ and $\omega$ are spatial and temporal frequencies. All Fourier components such that $\omega/k\geq
V_{\rm cut-off}=7\,\mathrm{km~s}^{-1}$ were removed so as to keep only convective motions [@TTTFS89].
SDO/HMI observations
--------------------
The HMI [@Scherrer2012; @Schou2012] onboard the Solar SDO provides uninterrupted observations over the entire disk. This gives a unique opportunity for mapping surface flows on various scales (spatial and temporal). Using the SDO/HMI white light data on December 10, 2011, from 16:11:15 to 19:06:45 UT, we derived horizontal velocity fields from image granulation tracking using the newly developed version of the CST code [@Roud2012]. The time step was 45 seconds with a pixel size of 05. The solar differential rotation discussed in Sect. 6 was determined from SDO/HMI white light and Doppler data taken on August 30, 2010 from 8:00:45 to 11:09:45 UT.
Hinode south pole flow field
============================
Measuring the solar rotation near the poles is a difficult task because there are few solar structures at high latitudes suitable for tracking. However, the Hinode images allow us to follow solar granules with good contrast up to the solar limb. First, we compare results of two methods to measure the solar motion close to the south pole using Hinode observations: local correlation tracking (LCT) [@NS88] and CST [@RRRD07; @Roud2012]. Both methods track the horizontal motions of granules in the field of view. More precisely, the LCT method of obtaining the horizontal velocity uses a spatial window that simultaneously accounts for the solar granules and integranular structures and may cover several granules. In contrast, the CST method measures the velocities by following the trajectory of each granule, i.e., solar plasma, during the life of the coherent object, which is defined by its appearance and disappearance if the granule does not split or merge. The data were aligned and the P angle evolution corrected by 0.018/hour in order to get a perfect co-alignment with SDO data used in the following. On December 10, 2011, the $B_0$ angle was $-0.25\degr$, resulting in insignificant projection effects with a negligible evolution during the observation period. Figure \[Vx3h\] shows good agreement of the $v_{\rm x}$ component from the LCT and CST up to latitude 75. Close to the limb, the effect of the spatial window used in the LCT prevents correct velocity determination. Figure \[Omega3h\] gives a solar sidereal rotation rate close to the pole of around 11/day, which corresponds to a polar period of 32.73 days. This is in agreement with previous determinations [@Beck2000].
Comparison of the velocities from SDO and Hinode
================================================
Our main goal is to compare the surface flows measured with a high spatial resolution (Hinode, 0.1) and a low resolution (SDO, 0.5). We extracted the same field of view from the SDO data set as the Hinode one and performed a very precise co-alignment of both sequences. To get the best co-alignment, the south polar region was observed and the solar limb was taken as an absolute reference. An additional check was performed after the alignment process by locating the brightest features (facular points) at the beginning and end of both sequences (Hinode and SDO). A very good match of the structures indicated very good alignment of both sequences during the 3 hours. We then applied the CST to both sequences to get the horizontal velocities ($v_{\rm x}$ and $v_{\rm y}$) over the entire field of view. Figures \[Vx2h\] and \[Omega2h\] show a good agreement up to 78of southern latitude. This indicates that SDO observations can be used to determine the solar rotation in the central meridian region up to significantly high latitudes. However, as shown in Fig. \[vy2h\], the meridional component measured on SDO data exhibits an offset of 0.4 km s$^{-1}$ relative to the Hinode component, which was used as the reference (because the velocity is close to zero at the limb). Figure \[vy2h\] shows a good correspondence between Hinode and SDO velocities when the 0.4 km s$^{-1}$ offset is removed (curve marked as SDO$-$0.4). We tried to elucidate the origin of that offset by determining all possible errors as listed in Table II of [@Strous2000b]. The estimation on the SDO measurement gives a total error on the $v_{\rm y}$ component of around 0.034 km $s^{-1}$, which is ten times smaller than the measured offset.
\
As described above, one of the major differences between the two sequences is the pixel size, which is 0.1 and 0.5 for Hinode and SDO, respectively. In order to analyze the sequences under the same conditions, we degraded the pixel size of the Hinode observation to the SDO one: 0.5. We then applied CST to the Hinode degraded sequence to get the horizontal velocities, which were also $k$-$\omega$ filtered. Figure \[compar\] still shows good agreement for the $v_{\rm
x}$ component and, in the same way, for the siderial angular rotation, at the latitudes up to 78. The offset of 0.4 km s$^{-1}$ observed in $v_{\rm y}$ component of the Hinode data at 0.5 is clearly visible in Fig. \[compar\]. We conclude that the offset is caused by the combination of the lower spatial resolution and decreasing contrast to the limb. One has to bear in mind that we only observe the south pole region in detail, where $v_{\rm y}$ is practically identical to the radial component of the flow. Thus, our observation can be generalized so that the radial component is always affected by the offset and we can correct for it. Indeed, different processes play a role in generating that radial component: at very high heliocentric angles, one pixel covers several granules. The granule part close to the limb is of lower contrast: it tends to be lost in the segmented images. This introduces an artificial radial motion of some granules that we see as the offset in the $v_{\rm y}$ component on the central meridian when comparing high-resolution (Hinode) and low-resolution (SDO) measurements. This radial effect was previously observed in our measurements, but its origin was not identified. Since we now have the origin of that error, we can measure it and correct for it all over the Sun. One way to get the radial offset is to average the radial velocity component over a circle centered on the solar disk. Due to the $B_0$ angle, the radial correction is not necessarily symmetrical in the northern and southern hemispheres of the Sun. This is why we treat the average process in the northern and southern regions of the Sun separately. Figure \[correction2\] shows, for example, the plot of the measured average radial component in the southern part and the overplotted fit obtained by using a polynomial function of the fifth degree. Figure \[correction\] shows the entire profile of the correction of the radial velocity to be applied to the velocities $v_{\rm x}$ and $v_{\rm y}$ all over the Sun (north and south), allowing a good representation of the velocity field around 90% of the solar radius. In the following, this correction is applied systematically to all velocity fields.
.
\
\
\
Determination of the spherical components of the velocity
=========================================================
From the velocity components $v_{\rm x}$ and $v_{\rm
y}$ measured in the sky plane and the simultaneous line of sight velocity $v_D$ measured from the SDO/HMI dopplergrams (e.g., Fig. \[original\]), we can derive the spherical velocity components $v_r$, $v_\theta$, $v_\varphi$, projected onto spherical coordinates $r$, $\theta$, $\varphi$ using
$$\begin{aligned}
v_r(\theta,\varphi) & =& \cos\theta * \sin\varphi * v_{\rm x} \nonumber\\
&+& (\sin \theta * \cos B_0 - \cos \theta * \cos \varphi * \sin B_0)* v_{\rm y}\nonumber\\
&+& (\cos \theta * \cos \varphi * \cos B_0 + \sin \theta * \sin B_0) * v_{\rm Dop}\nonumber\\
%
v_\theta(\theta,\varphi) &=& -\sin\theta * \sin\varphi * v_{\rm x} \nonumber\\
&+& (\sin \theta * \cos \varphi * \sin B_0 + \cos \theta * \cos B_0 ) * v_{\rm y} \nonumber\\
&+& (\cos \theta * \sin B_0 - \sin \theta * \cos \varphi * \cos B_0 ) * v_{\rm Dop} \nonumber\\
%
v_\varphi(\theta,\varphi) &=& \cos \varphi * v_{\rm x} \nonumber\\
&+& \sin \varphi * \sin B_0 * v_{\rm y} \nonumber\\
&-& \sin \varphi * \cos B_0 * v_{\rm Dop}\nonumber\end{aligned}$$
The system of coordinates is also shown in Fig. \[coordonnees\]. We note that $\varphi$ is along the longitude and $\theta$ is along the latitude. The co-alignment of all components has been checked carefully by using the information from FITS headers of the SDO data. Before using the dopplergrams, some data reduction had to be applied. First, the dopplergrams were averaged over the same time interval as the $v_{\rm x}$ and $v_{\rm y}$ sequence, here three hours. The limbshift had to be be corrected [@Ulrich2010] and, in this case, we used the limbshift function determined by the SDO team over the month of August 2010 (P. Scherrer, private communication).
By defining $z=1- \cos \rho$, where $\rho$ is heliocentric angle, the limb-shift correction (in km s$^{-1}$) is given by limb-shift$(z)=
-0.664 z + 0.775 z^2 + 0.284 z^3$. The mean velocity of the central region of the dopplergrams was taken as zero origin. The dopplergrams were resampled to the same size as the $v_{\rm x}$ and $v_{\rm y}$ components.
To convert the $v_{\rm x}$ and $v_{\rm y}$ velocities in the sidereal system $v_x^{\rm sid}$ and $v_y^{\rm sid}$, the Earth’s orbital displacement was taken into account. This correction depends on the P angle as
$$\begin{aligned}
v_x^{\rm sid} &=& v_{\rm x} + v_{\rm e}\cos P \nonumber\\
v_y^{\rm sid} &=& v_{\rm y} + v_{\rm e}\sin P \nonumber\end{aligned}$$
On August 30, 2010, it was 0.13 km s$^{-1}$ on $v_{\rm x}$ and 0.04 km s$^{-1}$ on $v_{\rm y}$, where $v_{\rm e}$ is the Earth velocity component projected on the Sun’s surface.
Figure \[spheric\] shows the resulting velocities $v_r, v_\theta,
v_\varphi$ for the three-hour sequence on August 30, 2010. The $v_r$ map exhibits the radial component where the downflow is visible in the sunspot regions. Close to disk center we observe a lower contrast of $v_r$, due to the low contrast in Doppler velocity in this part of the Sun because of the mostly horizontal flow on the Sun’s surface. The $v_r$ component is not well determined close to the limb because the projection effects are predominant. The $v_\varphi$ component clearly shows the solar differential rotation with a lower velocity close to the pole. To our knowledge, this is the first time that we can measure and visualize the solar differential rotation with a three-hour time sequence. The latitudinal component $v_\theta$ map allows us to essentially observe motion on supergranulation scales with a comparatively short time sequence.
\
\
\
\
Determination of the solar differential rotation
================================================
One of the first scientific applications of the CST algorithm on SDO/HMI data described in [@Roud2012] was to determine the solar rotation from the granule displacements. From the longitudinal velocity component $v_\varphi$, it is possible to calculate the solar differential rotation. In order to reduce the noise, we average $v_\varphi$ over bands in longitudes. Figure \[rotdif\] shows the computed differential rotation averaged over bands limited to longitude of $(-2,2)$, $(-10,10)$, $(-20,20)$, $(-40,40)$ degrees. The overplotted continuous line represents the solar rotation measured by the spectroscopic method [@HH70], which is our reference to evaluate the noise of our measurements. The noise level is found to be 0.94, 0.37, 0.27, 0.267 /day for the set of above-mentioned longitudinal bands. Thus, the average equatorial solar rotation determined for these four bands around the central meridian is 1.99 km s$^{-1}$ $\pm$ (0.133, 0.052, 0.038, 0.037) km s$^{-1}$ for the different longitudinal bands.
Our results show that, for the first time, using the CST we can get a determination of the solar rotation with a time sequence as short as three hours at very high precision, namely, 1.9% in the best case.
Discussion and conclusion
=========================
The comparison between high and low spatial resolution observations from the Hinode and SDO satellites allowed us to quantify the quality of the horizontal velocities determined over the full Sun. We found that the differential rotation along the central meridian is well determined up to latitude 60. With a longer time series, higher latitudes will be reached on depending the $B_0$ angle. Comparison of the meridional components shows an offset of 0.4 km s$^{-1}$ due to a combination of three factors: the low spatial resolution, the limb gradient contrast, and the segmentation process. However, we describe a way to measure and correct for the radial effect all over the Sun and get nearly a full Sun velocity measurement. From the velocities $v_{\rm
x}$ and $v_{\rm y}$ measured in the sky plane and the simultaneous line of sight velocity from SDO/HMI dopplergrams, we derived the spherical velocity components ($v_r, v_\theta, v_\varphi$). From the longitudinal component, it is possible to get the solar differential rotation with high precision ( $\pm 0.037$ km s$^{-1}$) using a temporal sequence of only three hours. This is remarkable because other methods require at least one month of data. That particularity opens a new field of study of the solar rotation and motions over the solar surface. In this way, we can revisit the dynamics of the solar surface at high spatial and temporal resolution from hours to months and years with the SDO data. In particular, it will be of great interest to compare CST convective velocities with numerical simulations and constraints from helioseismology [@Gizon2012] .
We thank the *Hinode* team for letting us use their data. *Hinode* is a Japanese mission developed and launched by ISAS/JAXA, collaborating with NAOJ as a domestic partner and with NASA and STFC (UK) as international partners. Support for the post-launch operation is provided by JAXA and NAOJ (Japan), STFC (U.K.), NASA, ESA, and NSC (Norway). We thank the HMI team members for their hard work.We thank the German Data Center for SDO and BAS2000 for providing SDO/HMI data.We thank F. Rincon for his private comments. This work was granted access to the HPC resources of CALMIP under the allocation 2011-\[P1115\]. L.G. acknowledges support from DFG Collaborative Research Center 963 “Astrophysical Flow Instabilities and Turbulence” (Project A1). This work was supported by the CNRS Programme National Soleil Terre. M. Š is supported by the Czech Science Foundation (grant P209/12/P568).
[^1]: Launched in 2006 the Hinode spacecraft , was designed and is now operated by JAXA (Japan Aerospace Exploration Agency) in cooperation with NASA (National Aeronautics and Space Administration) and ESA (European Space Agency).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a very recent work, G. E. Andrews defined the combinatorial objects which he called [*singular overpartitions*]{} with the goal of presenting a general theorem for overpartitions which is analogous to theorems of Rogers–Ramanujan type for ordinary partitions with restricted successive ranks. As a small part of his work, Andrews noted two congruences modulo 3 which followed from elementary generating function manipulations. In this work, we show that Andrews’ results modulo 3 are two examples of an infinite family of congruences modulo 3 which hold for that particular function. We also expand the consideration of such arithmetic properties to other functions which are part of Andrews’ framework for singular overpartitions.'
address:
- 'Institute of Contemporary Mathematics, Department of Mathematics and Information Sciences, Henan University, Kaifeng, 475001, China, [email protected]'
- 'School of Mathematics and Statistics, UNSW, Sydney 2052, Australia, [email protected]'
- 'Department of Mathematics, Penn State University, University Park, PA 16802, USA, [email protected]'
author:
- 'Shi-Chao Chen'
- 'Michael D. Hirschhorn'
- 'James A. Sellers'
title: 'Arithmetic Properties of Andrews’ Singular Overpartitions'
---
[^1]
2010 Mathematics Subject Classification: 05A17, 11P83
Keywords: singular overpartition, congruence, generating function, sums of squares
Introduction {#intro}
============
In a very recent work, Andrews [@A_sing] defined the combinatorial objects which he called [*singular overpartitions*]{} with the goal of presenting a general theorem for overpartitions which is analogous to theorems of Rogers–Ramanujan type for ordinary partitions with restricted successive ranks. In the process, Andrews proves that these singular overpartitions, which depend on two parameters $k$ and $i,$ can be enumerated by the function $\overline{C}_{k,i}(n)$ which gives the number of overpartitions of $n$ in which no part is divisible by $k$ and only parts $\equiv \pm i \pmod{k}$ may be overlined. Andrews also notes that, for all $n\geq 0,$ $\overline{C}_{3,1}(n) = \overline{A}_3(n)$ where $\overline{A}_3(n)$ is the number of overpartitions of $n$ into parts not divisible by 3. The function $\overline{A}_k(n),$ which counts the number of overpartitions of $n$ into parts not divisible by $k$, plays a key role in the work of Lovejoy [@L].
As part of his work, Andrews [@A_sing] uses elementary generating function manipulations to prove that, for all $n\geq 0,$ $$\label{And_congs_mod3}
\overline{C}_{3,1}(9n+3) \equiv \overline{C}_{3,1}(9n+6) \equiv 0 \pmod{3}.$$ In Section \[C31\_section\], we prove (\[And\_congs\_mod3\]) as part of an infinite family of mod 3 congruences satisfied by $\overline{C}_{3,1}(n).$ We also prove a number of arithmetic properties modulo powers of 2 satisfied by $\overline{C}_{3,1}(n).$ In Section \[C41\_section\], we prove similar results for $\overline{C}_{4,1}(n)$ while in Section \[C6\_section\], we prove a wide variety of results for $\overline{C}_{6,1}(n)$ and $\overline{C}_{6,2}(n),$ respectively. All of the proofs will follow from elementary generating function considerations and $q$–series manipulations.
Before we transition to our proofs, we note that, for $k\geq 3$ and $1\leq i\leq \lfloor \frac{k}{2}\rfloor,$ the generating function for $\overline{C}_{k,i}(n)$ is given by $$\label{main_genfn}
\sum_{n=0}^\infty \overline{C}_{k,i}(n)q^n = \frac{(q^k;q^k)_\infty(-q^i;q^k)_\infty(-q^{k-i};q^k)_\infty}{(q;q)_\infty}$$ where $$(A;q)_n = (1-A)(1-Aq)\dots (1-Aq^{n-1})$$ and $$(A;q)_\infty = \lim_{n\to\infty} (A;q)_n.$$ For certain values of $k$ and $i,$ (\[main\_genfn\]) can be manipulated in elementary ways to generate the Ramanujan–like congruences which appear in this paper.
Results for $\overline{C}_{3,1}(n)$ {#C31_section}
===================================
Motivated by Andrews, we first focus on the generating function for $\overline{C}_{3,1}(n)$ which, according to (\[main\_genfn\]), is given by $$\begin{aligned}
\sum_{n=0}^\infty\overline{C}_{3,1}(n)q^n
&=&
\frac{(q^3;q^3)_\infty(-q;q^3)_\infty(-q^2;q^3)_\infty}{(q;q)_\infty}\\
&=&
\frac{(q^3;q^3)_\infty(-q;q)_\infty}{(q;q)_\infty(-q^3;q^3)_\infty}\\
&=&
\frac{(q^3;q^3)_\infty^2(q^2;q^2)_\infty}{(q;q)_\infty^2(q^6;q^6)_\infty}\\
&=&
\frac{(q^3;q^3)_\infty^2}{(q^6;q^6)_\infty}\Big{/}\frac{(q;q)_\infty^2}{(q^2;q^2)_\infty}\\
&=& \frac{\varphi(-q^3)}{\varphi(-q)},\end{aligned}$$ where $\varphi(q)$ is Ramanujan’s theta function given by $$\label{phi_defn}
\varphi(q) = 1+2\sum_{n\geq 1}q^{n^2}.$$ Given that $$\label{genfn31phis}
\sum_{n=0}^\infty\overline{C}_{3,1}(n)q^n = \frac{\varphi(-q^3)}{\varphi(-q)},$$ we can see rather quickly how one might develop congruences modulo 3 which are satisfied by $\overline{C}_{3,1}(n).$
\[C31mod3general\] Let $N$ be a positive integer which is not expressible as the sum of two nonnegative squares. Then $$\overline{C}_{3,1}(N) \equiv 0 \pmod{3}.$$
Note that $$\begin{aligned}
\sum_{n=0}^\infty(-1)^n\overline{C}_{3,1}(n)q^n
&=&
\frac{\varphi(q^3)}{\varphi(q)}\\
&\equiv &
\frac{\varphi(q)^3}{\varphi(q)}\pmod3\\
&=&
\varphi(q)^2\\
&=&
\left (1+2\sum_{n=1}^\infty q^{n^2}\right )^2\\
&\equiv &
1+\sum_{n=1}^\infty q^{n^2}+\sum_{m,n=1}^\infty q^{m^2+n^2} \pmod{3}.\end{aligned}$$ The result follows.
Let $r_2(n)$ be the number of representations of $n$ as the sum of two squares. From the proof of Theorem \[C31mod3general\], we know $$\overline{C}_{3,1}(n)\equiv(-1)^nr_2(n)\pmod3.$$ Recall the well–known formula for $r_2(n),$ as noted in [@by], which states $$r_2(n)=4\prod_{\substack{p|n\\p\equiv1\pmod4}}(1+\nu_p(n))\prod_{\substack{p|n\\p\equiv3\pmod4}}\frac{1+(-1)^{\nu_p(n)}}{2},$$ where $p$ is prime and $\nu_p(n)$ is the exponent of $p$ dividing $n$. In light of this formula for $r_2(n),$ we have the following corollaries of Theorem \[C31mod3general\].
\[cor1mod3\] For all $k,m\geq 0,$ $$\overline{C}_{3,1}(2^k(4m+3)) \equiv 0 \pmod{3}.$$
Let $p\equiv1\pmod4$ be prime. Then for all $k,m\ge0$ with $p\nmid m,$ $$\overline{C}_{3,1}(p^{3k+2}m)\equiv0\pmod3.$$
\[cor2mod3\] Let $p\equiv3\pmod4$ be prime. Then for all $k,m\ge0$ with $p\nmid m,$ $$\overline{C}_{3,1}(p^{2k+1}m)\equiv0\pmod3.$$
Note that Andrews’ original congruences modulo 3, as given in (\[And\_congs\_mod3\]), are the $p = 3, k = 0$ and $m\equiv 1,2\pmod3$ cases of Corollary \[cor2mod3\].
We now transition to a consideration of congruence results satisfied by $\overline{C}_{3,1}(n)$ modulo small powers of 2. We begin with a lemma which will allow us to obtain an alternate form of the generating function for $\overline{C}_{3,1}(n)$ from which we can obtain such results.
$\varphi(-q^2)^2 = \varphi(q)\varphi(-q)$ where $\varphi(q)$ is defined in (\[phi\_defn\]).
$$\varphi(q)\varphi(-q)=\frac{(q^2;q^2)_\infty^5}{(q;q)_\infty^2(q^4;q^4)_\infty^2}\cdot\frac{(q;q)_\infty^2}{(q^2;q^2)_\infty}
=\frac{(q^2;q^2)_\infty^4}{(q^4;q^4)_\infty^2}=\varphi(-q^2)^2.$$
\[corA\] $$\frac{1}{\varphi(-q)} = \frac{\varphi(q)}{\varphi(-q^2)^2}.$$
This result is obvious based on the previous lemma.
\[alternate\_genfn31\] $$\sum_{n= 0}^\infty \overline{C}_{3,1}(n)q^n = \varphi(-q^3)\prod_{i=0}^\infty \varphi(q^{2^i})^{2^i}.$$
We simply use (\[genfn31phis\]) and iterate Corollary \[corA\] ad infinitum.
We can now state a few characterization theorems for $\overline{C}_{3,1}(n)$ modulo small powers of 2.
For all $n\geq 1,$ $\overline{C}_{3,1}(n) \equiv 0\pmod{2}.$
This follows from (\[genfn31phis\]) and (\[phi\_defn\]).
\[C31mod4\] For all $n\geq 1,$ $$\overline{C}_{3,1}(n) \equiv
\begin{cases}
2\pmod{4} & \text{if } n = k^2 \text{ or } n = 3k^2,\\
0\pmod{4} & \text{otherwise. }
\end{cases}$$
Thanks to Corollary \[alternate\_genfn31\] and (\[phi\_defn\]) we know $$\begin{aligned}
\sum_{n= 0}^\infty \overline{C}_{3,1}(n) q^n
&=&
\varphi(-q^3)\prod_{i=0}^\infty \varphi(q^{2^i})^{2^i} \\
&\equiv &
\varphi(-q^3)\varphi(q) \pmod{4} \\
&=&
\left( 1+2\sum_{k\geq 1}(-1)^kq^{3k^2} \right)\left( 1+2\sum_{k\geq 1}q^{k^2} \right) \\
&\equiv &
1+2\sum_{k\geq 1}q^{k^2}+2\sum_{k\geq 1}q^{3k^2} \pmod{4}.\end{aligned}$$ The result follows.
It is clear that one can write down numerous Ramanujan–like congruences modulo 4 satisfied by $\overline{C}_{3,1}(n)$ thanks to Theorem \[C31mod4\]. We refrain from doing so here.
Note that it is also possible to write a relatively clean characterization modulo 8 for $\overline{C}_{3,1}(n)$ using the same strategy as that employed above. This is because $$\begin{aligned}
\sum_{n= 0}^\infty \overline{C}_{3,1}(n) q^n
&=&
\varphi(-q^3)\prod_{i=0}^\infty \varphi(q^{2^i})^{2^i} \\
&\equiv &
\varphi(-q^3)\varphi(q)\varphi(q^2)^2 \pmod{8}.\end{aligned}$$ With that said, we consider a slight variant, namely obtaining a clean characterization modulo 8 for $(-1)^n\overline{C}_{3,1}(n),$ which can then be used rather quickly to prove numerous Ramanujan–like congruences modulo 8.
\[C31mod8\] $$\begin{aligned}
&&\sum_{n=0}^\infty (-1)^n\overline{C}_{3,1}(n)q^n \\
&\equiv &
1+6\sum_{k\geq 1}q^{k^2} + 4\sum_{k \geq 1}q^{2k^2} + 2\sum_{k\geq 1}q^{3k^2} + 4\sum_{k,\ell \geq 1}q^{k^2+3\ell^2} \pmod{8}.\end{aligned}$$
Since $\displaystyle{\varphi(q) = 1+2\sum_{n\geq 1}q^{n^2}},$ it is clear that $\varphi(q)^4 \equiv 1 \pmod{8}.$ It follows that $$\begin{aligned}
&&
\sum_{n=0}^\infty (-1)^n\overline{C}_{3,1}(n)q^n \\
&=&
\frac{\varphi(q^3)}{\varphi(q)} \\
&\equiv &
\varphi(q^3)\varphi(q)^3 \pmod{8} \\
&=&
\left( 1+2\sum_{k\geq 1}q^{k^2} \right)^3 \left( 1+2\sum_{k\geq 1}q^{3k^2} \right) \\
&\equiv &
\left( 1+6\sum_{k\geq 1}q^{k^2} + 4\sum_{k,\ell \geq 1}q^{k^2+\ell^2} \right) \left( 1+2\sum_{k\geq 1}q^{3k^2} \right) \pmod{8} \\
&\equiv &
1+6\sum_{k\geq 1}q^{k^2} + 2\sum_{k\geq 1}q^{3k^2} + 4\sum_{k,\ell \geq 1}q^{k^2+\ell^2} + 4\sum_{k,\ell \geq 1}q^{k^2+3\ell^2} \pmod{8} \\
&\equiv &
1+6\sum_{k\geq 1}q^{k^2} + 4\sum_{k \geq 1}q^{2k^2} + 2\sum_{k\geq 1}q^{3k^2} + 4\sum_{k,\ell \geq 1}q^{k^2+3\ell^2} \pmod{8}\end{aligned}$$ since solutions of $n=k^2+\ell^2$ with $k\not=\ell$ come in pairs.
We close this section by briefly noting a few corollaries of Theorem \[C31mod8\].
For all $k, m\geq 0,$ $$\begin{aligned}
\overline{C}_{3,1}(4^k(16m+6)) &\equiv & 0 \pmod{8}, \\
\overline{C}_{3,1}(4^k(16m+10)) &\equiv & 0 \pmod{8}, \text{\ \ and}\\
\overline{C}_{3,1}(4^k(16m+14)) &\equiv & 0 \pmod{8}.\end{aligned}$$
For all $k, m\geq 0,$ $$\begin{aligned}
\overline{C}_{3,1}(2^k(6m+5)) &\equiv & 0 \pmod{8}.\end{aligned}$$
Let $p$ be prime, $p\equiv 5, 11\pmod{12}.$ For all $k, m\geq 0$ with $p\nmid m,$ $$\overline{C}_{3,1}(p^{2k+1}m) \equiv 0 \pmod{8}.$$
Note that $n=p^{2k+1}m$ with $p\nmid m$ is neither a square, twice a square, three times a square, nor of the form $x^2+3y^2$ (since $\left(\frac{-3}{p}\right)=-1$ and $\nu_p(n)$ is odd).
Results for $\overline{C}_{4,1}(n)$ {#C41_section}
===================================
We now wish to consider other examples of the functions $\overline{C}_{k,i}(n)$ where arithmetic properties can be proven using elementary means. In this section, we concentrate on the function $\overline{C}_{4,1}(n).$
\[genfn\_C41\] $$\sum_{n= 0}^\infty \overline{C}_{4,1}(n)q^n = \frac{(q^2;q^2)_\infty^2}{(q;q)_\infty^2}.$$
Beginning with (\[main\_genfn\]), we see that $$\begin{aligned}
\sum_{n= 0}^\infty \overline{C}_{4,1}(n)q^n
&=&
\frac{(q^4;q^4)_\infty(-q;q^4)_\infty(-q^3;q^4)_\infty}{(q;q)_\infty} \\
&=&
\frac{(q^4;q^4)_\infty(-q;q^2)_\infty}{(q;q)_\infty} \\
&=&
\frac{(q^4;q^4)_\infty(q^2;q^4)_\infty}{(q;q)_\infty(q;q^2)_\infty} \\
&=&
\frac{(q^2;q^2)_\infty^2}{(q;q)_\infty^2}.\end{aligned}$$
Theorem \[genfn\_C41\] provides the following characterization of $\overline{C}_{4,1}(n)$ modulo 2.
\[C41mod2\] For all $n\geq 1,$ $$\overline{C}_{4,1}(n) \equiv
\begin{cases}
1\pmod{2} & \text{if } n = k(3k-1),\\
0\pmod{2} & \text{otherwise. }
\end{cases}$$
Thanks to Theorem \[genfn\_C41\], we have $$\begin{aligned}
\sum_{n= 0}^\infty \overline{C}_{4,1}(n) q^n
&=&
\frac{(q^2;q^2)_\infty^2}{(q;q)_\infty^2}\\
&\equiv &
\frac{(q^2;q^2)_\infty^2}{(q^2;q^2)_\infty} \pmod{2}\\
&=&
(q^2;q^2)_\infty \\
&\equiv &
\sum_{k=-\infty}^\infty q^{k(3k-1)} \pmod{2}\end{aligned}$$ thanks to Euler’s Pentagonal Number Theorem [@A Corollary 1.7].
Two corollaries follow immediately from the above.
For all $n\geq 0,$ $$\overline{C}_{4,1}(2n+1) \equiv 0\pmod{2}.$$
Note that $k(3k-1)$ is even for all integers $k.$
Let $p$ be prime and let $1\leq r\leq p-1$ with $12r+1$ a quadratic nonresidue modulo $p.$ Then, for all $m\geq 0,$ $$\overline{C}_{4,1}(pm+r) \equiv 0\pmod{2}.$$
We have $$\sum_{n=0}^\infty\overline{C}_{4,1}(n)q^{12n+1}\equiv\sum_{k= -\infty}^\infty q^{(6k-1)^2}\pmod2.$$ Here $n=pm+r$, so $12n+1=12pm+12r+1\equiv12r+1\pmod{p}$ is not a square modulo $p.$ Thus, $12n+1$ is not a square, and $\overline{C}_{4,1}(n)\equiv0\pmod2$.
From Theorem \[genfn\_C41\], we can also obtain results modulo 4 satisfied by $\overline{C}_{4,1}(n).$
\[C41mod4\] If $n$ cannot be represented as the sum of two pentagonal numbers, or if $n$ cannot be represented as the sum of a square and four times a pentagonal number, then $\overline{C}_{4,1}(n) \equiv 0 \pmod{4}.$
From the fact that $(1-q^2)^2\equiv(1-q)^4\pmod4$, we find $$\begin{aligned}
\sum_{n=0}^\infty\overline{C}_{4,1}(n)q^n
&=&
\frac{(q^2;q^2)_\infty^2}{(q;q)_\infty^2}\\
&\equiv &
\frac{(q;q)_\infty^4}{(q;q)_\infty^2} \pmod{4}\\
&=&
(q;q)_\infty^2\\
&=&\sum_{k,l=-\infty}^\infty (-1)^{k+l}q^{\frac{k(3k-1)}{2}+\frac{l(3l-1)}{2}},\end{aligned}$$ and $$\begin{aligned}
\sum_{n=0}^\infty(-1)^n\overline{C}_{4,1}(n)q^n
&\equiv &
(-q;q^2)_\infty^2(q^2;q^2)_\infty^2 \pmod{4}\\
&=&
\frac{(q^2;q^4)_\infty^2(q^2;q^2)_\infty^2}{(q;q^2)_\infty^2}\\
&=&
\frac{(q^2;q^2)_\infty^6}{(q;q)_\infty^2(q^4;q^4)_\infty^2}\\
&=&
(q^4;q^4)_\infty\varphi(q)\\
&=&
\sum_{k,l=-\infty}^\infty(-1)^kq^{2k(3k-1)+\ell^2}.\end{aligned}$$ The result follows.
Using Theorem \[C41mod4\], we can explicitly write infinitely many Ramanujan–like congruences modulo 4 satisfied by $\overline{C}_{4,1}(n).$
\[C41mod4primes3\] Let $p\ge5$ be a prime and $p\not\equiv1\pmod{12}.$ Then for all $k, m\ge0$ with $p\nmid m$, $$\overline{C}_{4,1}\left (p^{2k+1}m+\frac{p^{2k+2}-1}{12}\right )\equiv0\pmod4.$$
First suppose $p$ is prime, $p\equiv 7$ or $11\pmod{12}.$ We have $$\sum_{n=0}^\infty\overline{C}_{4,1}(n)q^{24n+2}
\equiv
\sum_{k,l=-\infty}^\infty (-1)^{k+l}q^{(6k-1)^2+(6l-1)^2}\pmod4.$$ Thus, if $24n+2$ is not the sum of two squares, $\overline{C}_{4,1}(n)\equiv0\pmod4$.
We have $n=p^{2k+1}m+\displaystyle\frac{p^{2k+2}-1}{12}$, so $$24n+2=24p^{2k+1}m+2p^{2k+2}=p^{2k+1}\left (24m+2p\right )$$ and $\nu_p(24n+2)$ is odd. By Fermat’s two–squares theorem, $24n+2$ is not the sum of two squares, so $\overline{C}_{4,1}(n)\equiv0\pmod4$.
Now suppose $p$ is prime, $p\equiv 5\pmod{12}.$ We have $$\sum_{n=0}^\infty(-1)^n\overline{C}_{4,1}(n)q^{12n+1}\equiv
\sum_{k,l=-\infty}^\infty(-1)^kq^{(6k-1)^2+3(2l)^2} \pmod{4}.$$ If $N$ is of the form $x^2+3y^2$, then it follows by a standard argument that $\nu_p(N)$ is even since $\displaystyle{\left (\frac{-3}{p}\right )=-1}.$
However, here $n=p^{2k+1}m+\displaystyle\frac{p^{2k+2}-1}{12}$ and $\nu_p(12n+1)$ is odd. So $12n+1$ is not of the form $x^2+3y^2$, and $\overline{C}_{4,1}(n)\equiv0\pmod4.$
As we close this section, we note that the generating function for $\overline{C}_{4,1}(n)$ is a modular function on $\Gamma_0(2).$ As such, we can slightly modify the proof of Theorem 1 of [@SCC] to obtain the following:
Let $p\geq 5$ be prime and let $\delta_p$ be the least positive residue of $p$ modulo 12. Then, for all $m\geq 0$ with $p\nmid m,$ $$\overline{C}_{4,1}\left(pm + \frac{p^2-1}{12}\right) \equiv 0\pmod{2^{\delta_p - 1}}.$$
Hence, for example, we have the following:
For all $m\geq 0,$ $$\begin{aligned}
\overline{C}_{4,1}(5m+2) &\equiv & 0\pmod{2^4} \text{\ \ if\ \ } 5\nmid m, \\
\overline{C}_{4,1}(7m+4) &\equiv & 0\pmod{2^6} \text{\ \ if\ \ } 7\nmid m, \\
\overline{C}_{4,1}(11m+10) &\equiv & 0\pmod{2^{10}} \text{\ \ if\ \ } 11\nmid m.\end{aligned}$$
Results for $\overline{C}_{6,1}(n)$ and $\overline{C}_{6,2}(n)$ {#C6_section}
===============================================================
Next, we consider the two functions $\overline{C}_{6,1}(n)$ and $\overline{C}_{6,2}(n).$ We begin by proving an elementary result for $\overline{C}_{6,1}(n)$ modulo 3.
\[C61mod3\] If $n$ cannot be represented as the sum of a pentagonal number and twice a triangular number, or if $n$ cannot be represented as the sum of a triangular number and four times a pentagonal number, then $\overline{C}_{6,1}(n)\equiv0\pmod3$.
Beginning with (\[main\_genfn\]), we see that $$\begin{aligned}
\sum_{n=0}^\infty\overline{C}_{6,1}(n)q^n
&=&
\frac{(q^6;q^6)_\infty(-q;q^6)_\infty(-q^5;q^6)_\infty}{(q;q)_\infty}\\
&=&
\frac{(q^6;q^6)_\infty(-q;q^2)_\infty}{(q;q)_\infty(-q^3;q^6)_\infty}\\
&=&
\frac{(q^6;q^6)_\infty(q^2;q^4)_\infty(q^3;q^6)_\infty}{(q;q)_\infty(q;q^2)_\infty(q^6;q^{12})_\infty}\\
&=&
\frac{(q^2;q^2)_\infty^2(q^3;q^3)_\infty(q^{12};q^{12})_\infty}
{(q;q)_\infty^2(q^4;q^4)_\infty(q^6;q^6)_\infty}\\
&\equiv &
\frac{(q^2;q^2)_\infty^2(q;q)_\infty^3(q^4;q^4)_\infty^3}{(q;q)_\infty^2(q^4;q^4)_\infty(q^2;q^2)_\infty^3}\pmod3\\
&=&
\frac{(q;q)_\infty(q^4;q^4)_\infty^2}{(q^2;q^2)_\infty}\\
&=&
(q;q)_\infty\psi(q^2)\\
&=&
\sum_{k=-\infty}^\infty\sum_{\ell=1}^\infty (-1)^kq^{\frac{k(3k-1)}{2}+\ell(\ell-1)},\end{aligned}$$ and $$\begin{aligned}
\sum_{n=0}^\infty (-1)^n\overline{C}_{6,1}(n)q^n
&\equiv &
\frac{(-q;q^2)_\infty(q^2;q^2)_\infty(q^4;q^4)_\infty^2}{(q^2;q^2)_\infty}\pmod{3}\\
&=&
\frac{(q^2;q^4)_\infty(q^4;q^4)_\infty^2}{(q;q^2)_\infty}\\
&=&
\frac{(q^2;q^2)_\infty^2(q^4;q^4)_\infty}{(q;q)_\infty}\\
&=&
(q^4;q^4)_\infty\psi(q)\\
&=&
\sum_{k=-\infty}^\infty\sum_{\ell=1}^\infty (-1)^kq^{2k(3k-1)+\frac{\ell(\ell-1)}{2}}.\end{aligned}$$ Here we have used Ramanujan’s theta function $\psi(q)$ which satisfies $$\psi(q)=\sum_{\ell=1}^\infty q^{\frac{\ell(\ell-1)}{2}}=\frac{(q^2;q^2)^2_\infty}{(q;q)_\infty}.$$ The result follows.
From Theorem \[C61mod3\], we can obtain the following set of Ramanujan–like congruences modulo 3 for $\overline{C}_{6,1}(n).$
\[corC61mod3\] Let $p\geq 5$ be prime and $p\not\equiv 1,\ 7 \pmod{24}.$ Then for $k, m\ge0$ with $p\nmid m$, $$\overline{C}_{6,1}\left (p^{2k+1}m+7\times\frac{p^{2k+2}-1}{24}\right )\equiv0\pmod3.$$
First suppose $p$ is prime, $p\equiv13,\ 17,\ 19$ or $23\pmod{24}.$ We have $$\sum_{n=0}^\infty\overline{C}_{6,1}(n)q^{24n+7}\equiv\sum_{k=-\infty}^\infty\sum_{\ell=1}^\infty
q^{(6k-1)^2+6(2\ell-1)^2}\pmod3.$$ So, if $24n+7$ is not of the form $(6k-1)^2+6(2\ell-1)^2$, then $\overline{C}_{6,1}(n)\equiv0\pmod3$.
If $N$ is of the form $x^2+6y^2$, then $\nu_p(N)$ is even since $\displaystyle\left (\frac{-6}{p}\right )=-1.$ However, here $
n=p^{2k+1}m+7\times\displaystyle\frac{p^{2k+2}-1}{24},
$ and $\nu_p(24n+7)$ is odd. So $24n+7$ is not of the form $x^2+6y^2$, and $\overline{C}_{6,1}(n)\equiv0\pmod3$.
Now suppose $p$ is prime, $p\equiv 5$ or $11\pmod{24}.$ We have $$\sum_{n\ge0}(-1)^n\overline{C}_{6,1}(n)q^{24n+7}\equiv\sum_{k=-\infty}^\infty\sum_{\ell=1}^\infty (-1)^kq^{(12k-2)^2+3(2\ell-1)^2}.$$ If $N$ is of the form $x^2+3y^2$, then $\nu_p(N)$ is even since $\displaystyle\left (\frac{-3}{p}\right )=-1$. However, $\nu_p(24n+7)$ is odd, so $24n+7$ is not of the form $x^2+3y^2.$ Therefore, we can conclude that $\overline{C}_{6,1}(n)\equiv0\pmod3$.
We now transition to a similar analysis of $\overline{C}_{6,2}(n).$
\[C62mod2\] For all $n\geq 1,$ $$\overline{C}_{6,2}(n) \equiv
\begin{cases}
1\pmod{2} & \text{if } n \text{\ is a pentagonal number},\\
0\pmod{2} & \text{otherwise. }
\end{cases}$$
Beginning with (\[main\_genfn\]), we have $$\begin{aligned}
\sum_{n=0}^\infty\overline{C}_{6,2}(n)q^n
&=&
\frac{(q^6;q^6)_\infty(-q^2;q^6)_\infty(-q^4;q^6)_\infty}{(q;q)_\infty}\\
&=&
\frac{(q^6;q^6)_\infty(-q^2;q^2)_\infty}{(q;q)_\infty(-q^6;q^6)_\infty}\\
&=&
\frac{(q^6;q^6)_\infty(q^4;q^4)_\infty(q^6;q^6)_\infty}{(q;q)_\infty(q^2;q^2)_\infty(q^{12};q^{12})_\infty}\\
&=&
\frac{(q^4;q^4)_\infty(q^6;q^6)_\infty^2}{(q;q)_\infty(q^2;q^2)_\infty(q^{12};q^{12})_\infty}\\
&\equiv &
\frac{(q;q)_\infty^4(q^{12};q^{12})_\infty}{(q;q)_\infty(q;q)_\infty^2(q^{12};q^{12})_\infty}\pmod2\\
&=&
(q;q)_\infty\\
&\equiv &
\sum_{k= -\infty}^\infty q^{\frac{k(3k-1)}{2}}\pmod2.\end{aligned}$$
Let $p\geq 5$ be prime and let $1\leq r\leq p-1$ with $24r+1$ a quadratic nonresidue modulo $p.$ Then, for all $m\geq 0,$ $$\overline{C}_{6,2}(pm+r) \equiv 0\pmod{2}.$$
We have $$\sum_{n=0}^\infty\overline{C}_{6,2}(n)q^{24n+1}\equiv\sum_{k= -\infty}^\infty q^{(6k-1)^2}.$$ Here, $n=pm+r$, so $24n+1=24pm+24r+1\equiv24r+1\pmod{p}$ is not a square modulo $p.$ Thus, $24n+1$ is not a square, and $\overline{C}_{6,2}(n)\equiv0\pmod2.$
We close this section by considering $\overline{C}_{6,2}(n)$ modulo 3.
\[C62mod3\] If $n$ cannot be represented as the sum of a pentagonal number and a square, or if $n$ cannot be written as the sum of a pentagonal number and twice a square, then $$\overline{C}_{6,2}(n)\equiv0\pmod3.$$
We have $$\begin{aligned}
\sum_{n=0}^\infty\overline{C}_{6,2}(n)q^n
&=&
\frac{(q^4;q^4)_\infty(q^6;q^6)_\infty^2}{(q;q)_\infty(q^2;q^2)_\infty(q^{12};q^{12})_\infty}\\
&\equiv &
\frac{(q^4;q^4)_\infty(q^2;q^2)_\infty^6}{(q;q)_\infty(q^2;q^2)_\infty(q^4;q^4)_\infty^3}\pmod3\\
&=&
\frac{(q^2;q^2)_\infty^5}{(q;q)_\infty(q^4;q^4)_\infty^2}\\
&=&
(q;q)_\infty\varphi(q)\\
&=&
\sum_{k,\ell=-\infty}^\infty (-1)^kq^{\frac{k(3k-1)}{2}+\ell^2}\end{aligned}$$ and $$\begin{aligned}
\sum_{n=0}^\infty(-1)^n\overline{C}_{6,2}(n)q^n
&\equiv &
(-q;q^2)_\infty(q^2;q^2)_\infty\varphi(-q) \pmod{3}\\
&=&
(-q;q^2)_\infty(q^2;q^2)_\infty\frac{(q;q)_\infty^2}{(q^2;q^2)_\infty}\\
&=&
\frac{(q;q)_\infty^2(q^2;q^4)_\infty}{(q;q^2)_\infty}\\
&=&
\frac{(q;q)_\infty^2(q^2;q^2)_\infty^2}{(q;q)_\infty(q^4;q^4)_\infty}\\
&=&
\frac{(q;q)_\infty(q^2;q^2)_\infty^2}{(q^4;q^4)_\infty}\\
&=&
(q;q)_\infty\varphi(-q^2)\\
&=&
\sum_{k,\ell=-\infty}^\infty (-1)^l q^{\frac{k(3k-1)}{2}+2\ell^2}.\end{aligned}$$ The result follows.
We close our paper by demonstrating an infinite family of Ramanujan–like congruences satisfied by $\overline{C}_{6,2}(n)$ modulo 3.
\[corC62mod3\] Let $p\geq 5$ be prime, $p\not\equiv1$ or $7\pmod{24}$, then for all $k,m\ge0$ with $p\nmid m$, $$\overline{C}_{6,2}\left (p^{2k+1}m+\frac{p^{2k+2}-1}{24}\right )\equiv0\pmod3.$$
First suppose that $p$ is prime, $p\equiv 13,\ 17,\ 19$ or $23\pmod{24}.$ This means $\displaystyle\left (\frac{-6}{p}\right )=-1$. We have $$\sum_{n=0}^\infty\overline{C}_{6,2}(n)q^{24n+1}\equiv\sum_{k,\ell=-\infty}^\infty(-1)^kq^{(6k-1)^2+6(2\ell)^2}.$$ The proof now goes through just as the proof of the first half of Corollary \[corC61mod3\].
Next, suppose $p$ is prime, $p\equiv 5$ or $11\pmod{24}.$ Then $\displaystyle\left (\frac{-3}{p}\right )=-1$. We have $$\sum_{n=0}^\infty(-1)^n\overline{C}_{6,2}(n)q^{24n+1}\equiv\sum_{k,\ell=-\infty}^\infty(-1)^\ell
q^{(6k-1)^2+3(4\ell)^2}.$$ The proof now goes through just as the proof of the second half of Corollary \[corC61mod3\].
[99]{}
G. E. Andrews, Singular overpartitions, in progress
G. E. Andrews, [*The Theory of Partitions*]{}, Addison-Wesley, Reading 1976; reprinted, Cambridge University Press, Cambridge, 1984, 1998
A. Berkovich and H. Yesilyurt, Ramanujan’s identities and representation of integers by certain binary and quaternary quadratic forms, [*Ramanujan J.*]{} [**20**]{} (2009), 375–408
S.-C. Chen, Arithmetic properties of a partition pair function, [*Int. J. Num. Thy.*]{}, to appear
J. Lovejoy, Gordon’s theorem for overpartitions, [*J. Comb. Thy. A*]{} [**103**]{} (2003), 393–401
[^1]: The first author was supported by the NSF of China (No.11101123).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present new measurements of the temperature-density ($T-\rho$) relation for neutral hydrogen in the $2.0 < z < 2.8$ intergalactic medium (IGM) using a sample of $\sim$6000 individual absorbers fitted with Voigt profiles constrained in all cases by multiple Lyman series transitions. We find model independent evidence for a positive correlation between the column density of ([$N_{\rm HI}$]{}) and the minimum observed velocity width of absorbers ([${b_{\rm min}}$]{}). With minimal interpretation, this implies that the $T-\rho$ relation in the IGM is not “inverted”, contrary to many recent studies. Fitting [${b_{\rm min}}$]{} as a function of [$N_{\rm HI}$]{} results in line width - column density dependence of the form ${\ensuremath{{b_{\rm min}}}}= b_0 \left( {\ensuremath{N_{\rm HI}}}/N_{\rm HI,0} \right)^{\Gamma -1}$ with a minimum line width at mean density ($\rho/\bar\rho = 1, N_{\rm HI, 0} = 10^{13.6}$ cm$^{-2}$) of $b_0= {\ensuremath{17.9 \pm 0.2}}$ km s$^{-1}$ and a power-law index of $(\Gamma -1) = {\ensuremath{0.15 \pm 0.02}}$. Using analytic arguments, these measurements imply an “equation of state” for the IGM at $\langle z \rangle = 2.4$ of the form $T=T_0 \left(\rho/\bar\rho\right)^{\gamma-1}$ with a temperature at mean density of $T_0= [{\ensuremath{1.94 \pm 0.05}}] \times 10^4 $ K and a power-law index $(\gamma -1) = {\ensuremath{0.46 \pm 0.05}}$.'
author:
- 'Gwen C. Rudie, Charles C. Steidel, & Max Pettini'
bibliography:
- 'ms.bib'
title: 'The Temperature-Density Relation in the Intergalactic Medium at redshift $\langle z \rangle =2.4$ '
---
Introduction
============
The “equation of state” of the low-density intergalactic medium is believed to be controlled by two principle processes: photo-heating and adiabatic cooling. The cooling is most directly tied to the expansion of the Universe, while the heating is expected to be a complicated mixture of relic effects from the reionization of hydrogen and helium plus the current heating, predominantly from the UV background [@hui97; @sch99]. This naturally imposes a relationship between the temperature and density of intergalactic gas. Denser gas is expected to trace larger over-densities, for which adiabatic cooling is suppressed because such regions are more bound against the expansion of the Universe. At the same time, denser gas has a much faster recombination time scale and thus presents a larger cross-section for photo-ionization which leads denser gas to experience greater heating.
Recently, this simple picture has been called into question on the basis of several statistical studies of the transmitted flux observed in individual pixels within the Lyman-$\alpha$ forest [@bec07; @bol08; @vie09; @lid10; @cal12; @gar12]. These studies require comparison with large numerical simulations in order to interpret the physical implications of the detailed shape of these flux distributions. Further, the shape of the distribution may be sensitive to the full thermal history of the gas, not just the present temperature-density ($T-\rho$) relation [@pee10].
@bec07 considered the transmitted flux probability distribution function (PDF) and compared it to the numerical models of @mir00, finding that none of the models could explain the observations. They found that a $T-\rho$ relation in which lower density regions were *hotter* (an inverted $T-\rho$ relation) provided a better fit to the data. Since this study, several other authors have obtained similar results using transmitted flux PDF analysis [@bol08; @vie09; @cal12], wavelet analysis [@lid10], or a combination of both [@gar12].
In contrast, early work on the $T-\rho$ relation and its evolution with redshift was performed via Voigt profile fitting of individual absorption lines. These early studies found a monotonic $T-\rho$ relation with a positive power-law index [@sch00; @ric00; @bry00; @mcd01].
In this letter, we return to the more direct test of the $T-\rho$ relation in the IGM using Voigt profile fits to individual absorbers. This method relies on the expected relationship between the column density of neutral hydrogen of an absorber, [$N_{\rm HI}$]{}, and its local overdensity, $\rho/\bar{\rho}$ [@dav99; @sch01]. If such a relationship exists, then one expects to observe a correlation between [$N_{\rm HI}$]{} and the velocity widths of individual thermally broadened absorbers, [${b_{\rm d}}$]{}. Absorbers having only thermal broadening[^1] are expected to have the smallest [${b_{\rm d}}$]{} at a given [$N_{\rm HI}$]{}, [${b_{\rm min}}$]{}. Thus, by observing the behavior of the low-[${b_{\rm d}}$]{} “edge” of the distribution of [${b_{\rm d}}$]{} versus [$N_{\rm HI}$]{}, it is possible to infer the relationship between $T$ and [$N_{\rm HI}$]{} and through theory, the IGM $T-\rho$ relationship.
Here we present analysis of individual absorbers fitted in 15 high-resolution, high-S/N spectra of luminous QSOs at $2.5 < z < 2.9$ from the Keck Baryonic Structure Survey (KBSS; @gcr12). We discuss the data set used in this study as well as the line-fitting procedure in Section \[data\]. In Section \[b-NHI\], the fit to the minimum [${b_{\rm d}}$]{} ([${b_{\rm min}}$]{}) as a function of [$N_{\rm HI}$]{} is presented, followed by Section \[results\] in which the physical implications of the results in the context of the $T-\rho$ relation are discussed. The results and their implications are summarized in Section \[con\].
Throughout this paper we assume a $\Lambda$-CDM cosmology with $H_{0} = 70$ Mpc$^{-1}$, $\Omega_{\rm m} = 0.3$, and $\Omega_{\Lambda} = 0.7$.
Data and Analysis {#data}
=================
The data analyzed in this paper are taken from the Keck Baryonic Structure Survey (KBSS). The KBSS is a large spectroscopic survey designed to study the gaseous surroundings of high-$z$ star-forming galaxies by combining redshift surveys for galaxies at $2 < z < 3$ in the fields of bright QSOs with absorption line data from those same QSOs. The data include high-resolution, high signal-to-noise echelle spectra of 15 luminous ($m_V \simeq 15.5-17$) QSOs located near the centers of the KBSS fields. The KBSS will be described fully by Steidel et al. (in prep); however, the data of relevance to this paper as well as the analysis of the absorption-line sample have been presented in @gcr12. Here we give a brief summary.
The QSO spectra were obtained with the High Resolution Echelle Spectrometer [HIRES; @vog94] on the Keck I telescope. The HIRES spectra have $R\simeq 45,000$ (FWHM$\simeq 7$ ) and S/N $\sim 50-200$ per pixel. They cover the wavelength range 3100 – 6000 Å with no spectral gaps, allowing for the observation of Ly$\beta$ $\lambda 1025.7$ down to at least $z= 2.2$ in all 15 of our sightlines. The additional constraints provided by Ly$\beta$ (and in many cases, additional Lyman series transitions) allow for highly accurate measurements of column densities and (most importantly for this paper) line widths for the absorbers in our sample.
[llcrc]{} Q0100+130 (PHL957) & 2.721 & 2.0617– 2.6838 & 77 & 50\
HS0105+1619 & 2.652 & 2.1561– 2.6153 & 127 & 89\
Q0142$-$09 (UM673a) & 2.743 & 2.0260– 2.7060 & 71 & 45\
Q0207$-$003 (UM402) & 2.872 & 2.1532– 2.8339 & 82 & 55\
Q0449$-$1645 & 2.684 & 2.0792– 2.6470 & 73 & 41\
Q0821+3107 & 2.616 & 2.1650– 2.5794 & 50 & 33\
Q1009+29 (CSO 38) & 2.652 & 2.1132– 2.6031& 99 & 58\
SBS1217+499 & 2.704 & 2.0273– 2.6669 & 68 & 38\
HS1442+2931 & 2.660 & 2.0798– 2.6237 & 99 & 47\
HS1549+1919 & 2.843 & 2.0926– 2.8048 & 173 & 74\
HS1603+3820 & 2.551 & 2.1087– 2.5066 & 108 & 58\
Q1623+268 (KP77) & 2.535 & 2.0544– 2.4999 & 48 & 28\
HS1700+64 & 2.751 & 2.0668– 2.7138 & 98 & 42\
Q2206$-$199 & 2.573 & 2.0133– 2.5373 & 88 & 46\
Q2343+125 & 2.573 & 2.0884– 2.5373 & 71 & 45 \[field\]
The reduction of the QSO spectra and a detailed description of the process of fitting the forest are discussed in @gcr12. The final Voigt profiles fits to the full Ly$\alpha$ and Ly$\beta$ forests of the 15 KBSS QSOs were completed using VPFIT[^2] written by R.F. Carswell and J.K. Webb. The redshift range included in the fit is given in Table \[field\]. In this paper, we use the pathlength weighted mean redshift of the sample $\langle z \rangle = 2.37$ as the fiducial redshift [@bah69].[^3]
The final absorber catalog includes 5758 absorbers with $12.0 < \log({\ensuremath{N_{\rm HI}}}/ \textrm{cm}^{-2}) < 17.2$ and $ 2.02 < z < 2.84$ over a total redshift path length of $\Delta z = 8.27$. This sample is the largest ever compiled at these redshifts and is more than an order of magnitude larger than previous samples that included constraints from higher-order Lyman lines.
The Temperature-Density Relation in the IGM {#b-NHI}
===========================================
The equation of state of the IGM is expected to have the form $$T=T_0 \left( \frac{\rho}{\bar{\rho}} \right)^{\gamma-1}$$ where $T_0$ is the temperature at the mean mass density ($\overline \rho$) [@hui97].
Under the assumption of a relatively uniform ionizing radiation field, a power-law relationship between [$N_{\rm HI}$]{} and $\rho$ is also expected. @sch01 presented a model for the low-density IGM in which ‘clouds’ are in local hydrostatic equilibrium and therefore typically have sizes comparable to the local Jeans length. Employing this assumption, @sch01 derived a relationship between [$N_{\rm HI}$]{} and local overdensity, $\rho/\overline \rho$. Using updated cosmology and evaluating at the path-length weighted mean redshift of the sample: $$\label{jeans}
\rho_b/\overline \rho_b \approx \left(\frac{N_{\textrm{\footnotesize{HI}}}} {10^{13.6}}\right)^{2/3}T_4^{0.17} \left( \frac{\Gamma_{12}}{0.5} \right)^{2/3} \left( \frac{1+z}{3.4} \right)^{-3},
\label{schaye}$$ where $T_4$ is the gas temperature in units of $10^4$ K and $\Gamma_{12}$ is the hydrogen photoionization rate in units of $10^{-12}$ s$^{-1}$ with the normalization suggested by @fau08. Assuming this scaling[^4], absorbers with $\log({\ensuremath{N_{\rm HI}}}/ \textrm{cm}^{-2}) \approx 13.6$ are expected to trace gas at the mean density of the universe at $z=2.4$.
![The doppler widths of absorbers ([${b_{\rm d}}$]{}) versus their column density ([$N_{\rm HI}$]{}) for all the absorbers in the HI sample with relative errors in [$N_{\rm HI}$]{} and [${b_{\rm d}}$]{} less than 50%. Point-density contours are over plotted to guide the eye in the bottom panel. Note that the minimum value of [${b_{\rm d}}$]{} at each [$N_{\rm HI}$]{} increases with increasing [$N_{\rm HI}$]{} suggesting a normal $T-\rho$ relation.[]{data-label="b_vs_N"}](f1.png){width="45.00000%"}
Thermally broadened absorbers are also expected to follow a power law relation between [${b_{\rm d}}$]{} and temperature: ${\ensuremath{{b_{\rm d}}}}~\propto~T^{1/2}$. Combining with the expected $T-\rho$ relation and the conversion between [$N_{\rm HI}$]{} and $\rho$, the relationship between [${b_{\rm min}}$]{} and [$N_{\rm HI}$]{} would be a power law of the form: $${\ensuremath{{b_{\rm min}}}}= b_0 \left(\frac{{\ensuremath{N_{\rm HI}}}}{N_{\rm{HI, 0}}}\right)^{\Gamma -1}
\label{eqn_b_n}$$ where $b_0$ is the minimum line width of absorbers with ${\ensuremath{N_{\rm HI}}}= N_{\rm HI,0}$. With this formalism, $(\gamma -1 )$ is proportional to $(\Gamma -1)$ [see eg. @sch99].
More explicitly, for pure thermal broadening: $$b= (2 k_B T/m_{\rm p})^{1/2}
\label{bT}$$ where $k_B$ is the Boltzmann constant and $m_{\rm p}$ is the mass of the proton. This suggests: $$\log \left( T \right)= C + 2 \log\left( \frac{b}{\textrm{km s}^{-1}} \right)
\label{eqn_Tb}$$ where $$C=\log \left( \frac{m_{\rm p}}{2 k_B} \frac{(\textrm{km s}^{-1})^2}{\textrm{K}}\right) = 1.78.$$
Rearranging the above equations, we expect the conversion between the index of ${\ensuremath{{b_{\rm min}}}}({\ensuremath{N_{\rm HI}}})$ and the $T - \rho$ relationship to be roughly:[^5] $$\gamma-1\approx 3 \left( \Gamma -1\right)
\label{eqn_gamma}$$
In the sections that follow, we fit to the trend of the [${b_{\rm min}}$]{} as a function of [$N_{\rm HI}$]{} for the data sample. We then use equations \[eqn\_Tb\] and \[eqn\_gamma\] to estimate the $T-\rho$ relation in the $\langle z \rangle =2.4$ IGM.
{width="40.00000%"}{width="65.00000%"}
{width="40.00000%"}{width="65.00000%"}
A “normal” $T-\rho$ relation
----------------------------
The measured values of [${b_{\rm d}}$]{} and [$N_{\rm HI}$]{} considered in this letter are presented in Figure \[b\_vs\_N\]. Examination of Figure \[b\_vs\_N\] already suggests the main result of this paper; focusing on the low-[${b_{\rm d}}$]{} edge of the distribution, one can easily see that the lower envelope increases monotonically with [$N_{\rm HI}$]{}. Such behavior is expected for a *normal* temperature-density relationship in which denser regions are hotter.
Outlier Rejection {#outlier}
-----------------
Before fitting to the “ridge-line” of [${b_{\rm min}}$]{} versus [$N_{\rm HI}$]{}, we exclude those absorbers that have large errors in their measured parameters. @sch99 suggest a simple algorithm for the outlier rejection in which absorbers with relative error in [${b_{\rm d}}$]{} or [$N_{\rm HI}$]{} larger than 50% are excluded.[^6] This method is expected to primarily reject lines which originate in blends and are thus unlikely to have accurate line-width measurements. Additionally, we exclude those absorbers with [${b_{\rm d}}$]{}= 8 and [${b_{\rm d}}$]{}= 100 . These line widths correspond to the allowed line-width limits input to VPFIT and as such are artificial. The absorbers which remain after those exclusions are shown in Figure \[b\_vs\_N\].
Considering Figure \[b\_vs\_N\], we note a small set of points at very low [${b_{\rm d}}$]{} which seem to lie significantly below the main locus of points. These absorbers are likely to be unidentified metal line contaminants, and thus we consider a $\sigma$-rejection algorithm to exclude them.
The $\sigma$-rejection algorithm is applied as follows. Only absorbers with ${\ensuremath{{b_{\rm d}}}}< 40$ are considered. The data are sorted by their [$N_{\rm HI}$]{}. The absorbers are placed into [$N_{\rm HI}$]{} bins of width 0.25 dex and the mean and standard deviation of [${b_{\rm d}}$]{} are computed for the bin. Absorbers with [${b_{\rm d}}$]{} 2$\sigma$ or larger from the mean are flagged and excluded. This process is iterated; the set of absorbers excluded via this method with [${b_{\rm d}}$]{} lower than the mean are surrounded by (blue) boxes in Figure \[bN\] and are excluded from the fitting procedure. We have checked the majority of $\sigma$-rejected absorbers by hand and find that they are either previously unidentified metal lines or parts of blended saturated systems whose individual properties are poorly constrained.
In the following section, we proceed with the measurement of the [${b_{\rm min}}$]{}– [$N_{\rm HI}$]{} relation. We show that the application of the $\sigma$-rejection algorithm does not produce significant changes to the results. In the conclusions of the paper, however, we prefer the combined error and $\sigma$-rejection method as it excludes the majority of the poorly measured absorbers as well as those which lie far below the [${b_{\rm min}}$]{} – [$N_{\rm HI}$]{} trend.
Fitting [${b_{\rm min}}$]{} versus [$N_{\rm HI}$]{} {#Schaye_power_law}
---------------------------------------------------
In this section we measure the [${b_{\rm min}}$]{} – [$N_{\rm HI}$]{} relation using the two data sets resulting from the outlier rejection methods described above. To fit the trend between [${b_{\rm min}}$]{} and [$N_{\rm HI}$]{}, we follow the iterative power-law method proposed by @sch99. A power-law relationship of the form shown in equation \[eqn\_b\_n\] with $\log(N_{\rm HI,0}/$2)$ = 13.6$ is fit to the data. Data points which have [${b_{\rm d}}$]{} larger than one mean absolute deviation above the fit are discarded. The power law is refit to the remaining absorbers and the rejection and refitting are repeated until the power law converges. Then, absorbers more than one mean absolute deviation below the fit are removed *once*, and the power law is refit. This fit is taken as the final form of the minimum [${b_{\rm min}}$]{} – [$N_{\rm HI}$]{} relation. Errors in the parameters of the fit are derived by applying the bootstrap resampling method to the data sample and following the above outlined procedure.
[lccccccccc]{} default & $2.0-2.8$ & 2.4 &13.6 & 0.156 $\pm$ 0.032 & 0.47 $\pm$ 0.10 & 17.56 $\pm$ 0.40 & 1.87 $\pm$ 0.08\
$2\sigma$-rej. & $2.0-2.8$ & 2.4 & 13.6 & 0.152 $\pm$ 0.015 & 0.46 $\pm$ 0.05 & 17.90 $\pm$ 0.21 & 1.94 $\pm$ 0.05\
\
\
$2\sigma$-rej. & $2.0 - 2.6$ & 2.3 & 13.7 & 0.156 $\pm$ 0.016 & 0.47 $\pm$ 0.05 & 18.50 $\pm$ 0.22 & 2.07 $\pm$ 0.05\
$2\sigma$-rej. & $2.6 - 2.8$ & 2.7 & 13.4 & 0.171 $\pm$ 0.032 & 0.51 $\pm$ 0.10 & 17.39 $\pm$ 0.44 & 1.83 $\pm$ 0.09 \[igm\_temp\]
[llcccl]{} & Line Fitting & 2.3 & $1.17$ & 0.27 &\
& Line Fitting & 2.5 & $1.38$ & 0.56 &\
\
& Line Fitting & 2.75 & $2.52$ & $0.22 \pm 0.10$ &\
& Line Fitting & 1.90 & $ 1.77 $ & $0.32 \pm 0.30$ &\
\
@bry00 & Line fitting & 2.7 & $1.65 $& $0.29$ &\
@mcd01 & Line Fitting & 2.41 & $1.74 \pm 0.19 $ & $ 0.52 \pm 0.14$ & Their Table 5, $z_{\rm sim}=3$\
@zal01 & Transmission Power Spectra & 2.4 & $1-3 $ & 0 to 0.6 & Their Figure 3\
@bec07 & Flux PDF & $2-6$ & & $\approx -0.5$ &\
\
& Flux PDF & 2.07& & $-0.33$ &\
& Flux PDF & 2.52 & & $-0.46$ &\
& Flux PDF & 2.94 & & $-0.56$ &\
\
@vie09 & Flux PDF & 3 & $1.9 \pm 0.6$ & $-0.25 \pm 0.21$ &\
\
& Wavelet Analysis & 2.6 & $1.6 \pm .6 $ & &\
& Wavelet Analysis & 2.2 & $2.1 \pm .7 $ & &\
\
@bec11 & Curvature Analysis & $2.4$ & $1.11 \pm 0.06$& &Their Table 3\
\
& Wavelet Analysis + Flux PDF & 2.1 & $1.7 \pm 0.2 $ &$0.11 \pm 0.11$ &\
& Wavelet Analysis & 2.1 & $1.6 \pm 0.5 $ &$> -0.14$ &\
& Flux PDF & 2.1 & $1.5 \pm 0.3 $ &$-0.01 \pm 0.14$ &\
& Wavelet Analysis + Flux PDF & 2.5 & $1.3 \pm 0.4 $ & $ >-0.05$ &\
& Wavelet Analysis & 2.5 & $1.6 \pm 0.4$& $ >-0.08$ &\
& Flux PDF & 2.5 & $1.4 \pm 0.9 $ & $>-0.31$ &\
\
& Flux PDF & 3 & $1.93 \pm 0.48$ & $-0.10\pm0.21$ &\
& Flux PDF & $3$ & $1.79 \pm 0.35$ & $-0.30\pm 0.12$ &\
\
\
& & & & & Results from data set\
& & & & &using both relative error\
& & & & &and $\sigma$-rejection \[prev\_res\]
Figure \[bN\] shows the results of the iterative power-law fit for both data sets formed via the outlier rejection algorithms. The red curves correspond to the fit including the $\sigma$-rejected points, while the blue curves correspond to those without. The center and right panels show the distribution of fit parameters obtained through bootstrap resampling. As expected, those fits which exclude the low-[${b_{\rm d}}$]{} outliers via $\sigma$-rejection (blue histograms) result in tighter distributions for the fitted parameters; however, the best fit values are nearly indistinguishable.
The best fit parameters $b_0$ and $(\Gamma - 1)$ and their relative errors, as well as their physical counterparts $T_0$ and $(\gamma-1)$, are presented in Table \[igm\_temp\]. $T_0$ and $(\gamma-1)$ are calculated assuming equations \[eqn\_Tb\] and \[eqn\_gamma\] respectively. Notably, both outlier rejection algorithms produce statistically consistent results suggesting that the measurements are robust to changes in the rejection algorithm; however, (as expected) slightly higher $b_0$ are preferred by the algorithms with $\sigma$-rejection.
Henceforth, we consider the results from the data set formed with both the error rejection and $\sigma$-rejection algorithms applied prior to the power-law fit to [${b_{\rm min}}$]{}([$N_{\rm HI}$]{}).
Results
=======
As discussed above, the mapping of [${b_{\rm min}}$]{}([$N_{\rm HI}$]{}) to the equation of state for intergalactic gas requires a theoretical framework. @sch00 considered simulations of the IGM and found that the minimum values of [${b_{\rm d}}$]{} at each [$N_{\rm HI}$]{} were larger than expected given purely thermal broadening. They suggested that all absorbers in the [$N_{\rm HI}$]{} range used to measure the $T-\rho$ relationship experienced mild Hubble broadening (also expected to scale as $T^{1/2}$ for clouds having sizes similar to the local Jeans scale.) In this case, estimates made using the above equations would mildly over-predict the temperature at mean density. We prefer to report physical values tied closely to our measurements rather than rely directly on a specific set of simulations; however, our measurements of $b_0$ and $\Gamma-1$ can be converted to physical parameters using a different set of assumptions. Our simplifying assumption of pure thermal line broadening for absorbers near [${b_{\rm min}}$]{}([$N_{\rm HI}$]{}) results in $T_0$ values slightly higher than some previous studies which used simulations as a reference (see Table \[prev\_res\]).
In Table \[prev\_res\], we compare our results with those of previous studies. We find generally good agreement between our results and those of past studies which used line fitting. We further emphasize that our results differ significantly from those obtained using the transmitted flux PDF and similar statistical methods, calling into question the interpretation of such techniques.
Reionization
-------------
One of the expected results of reionization, believed to occur at $z\approx3$, is a significant change in the $T-\rho$ relation due to excess photoionization heating [@hui97]. These changes are not long lived, and the exact effects depend on the speed and patchiness with which reionization proceeds and the spectral hardness of the ionizing sources (see @mcq09). While detailed discussion is beyond the scope of this letter, we briefly mention that the absorbers in the KBSS sample show no strong evidence for a change in the slope of ${\ensuremath{{b_{\rm min}}}}({\ensuremath{N_{\rm HI}}})$ for higher-redshift absorbers, which (presumably) are temporally closer to the reionization epoch. The bottom panels of Figure \[bN\] show the iterative power law fit to absorbers with, respectively, $z < 2.6$ (green) and $z > 2.6$ (orange) from the sample with $\sigma$-rejection. Notably, the ${\ensuremath{{b_{\rm min}}}}-{\ensuremath{N_{\rm HI}}}$ relation appears to be independent of redshift (left panel); however, the expected evolution in the value of [$N_{\rm HI}$]{} at the mean density (equation \[schaye\]) results in differing values of $N_{\rm HI,0}$ for the two redshift bins. This in turn leads to expected variation in the values of $b_0$ and $T_0$ with $z$. The implied parameters from the fit are listed in Table \[igm\_temp\].
Conclusions {#con}
===========
We have inferred the $T-\rho$ relationship in the $\langle z \rangle = 2.4$ IGM using Voigt profile fitting of individual Lyman lines. Fitting the trend of the minimum line width ([${b_{\rm min}}$]{}) as a function of column density ([$N_{\rm HI}$]{}) with a power law, we find best fit values of [${b_{\rm min}}$]{} at mean density (which at $\langle z \rangle = 2.4$ corresponds to $N_{\rm HI, 0} = 10^{13.6}$ 2) $b_0={\ensuremath{17.9 \pm 0.2}}$ km s$^{-1}$ and a power-law index $(\Gamma -1) = {\ensuremath{0.15 \pm 0.02}}$. Assuming a monotonic relation between [$N_{\rm HI}$]{} and $\rho/\bar{\rho}$, these data support the conclusion that the temperature-density ($T-\rho$) relationship in the IGM is not inverted at $\langle z \rangle = 2.4$ but instead has a power-law index $(\gamma -1) = {\ensuremath{0.46 \pm 0.05}}$. Further, our results suggest a temperature at mean density of $T_0=[{\ensuremath{1.94 \pm 0.05}}] \times 10^4 $K. Within our sample spanning the redshift range $2.0 \lesssim z \lesssim2.8$, there is no evidence for significant evolution in the ${\ensuremath{{b_{\rm min}}}}- {\ensuremath{N_{\rm HI}}}$ relation or in the power-law index ($\gamma - 1$) of the $T-\rho$ relation.
We thank George Becker for his early interest in the $T-\rho$ relation results from the KBSS sample, Joop Schaye who provided helpful advice and insightful comments, and Allison Strom and Ryan Trainor for their careful reading of early drafts. Thanks also to Olivera Rakic for her contributions to the reduction of the QSO data set and for her pertinent advice. The authors wish to acknowledge Ryan Cooke who contributed the fits to the damped profiles in our QSO spectra.
We wish to acknowledge the staff of the the W.M. Keck Observatory whose efforts insure the telescopes and instruments perform reliably. We thank those of Hawaiian ancestry on whose sacred mountain we are privileged to be guests.
This work has been supported by the US National Science Foundation through grants AST-0606912 and AST- 0908805. CCS acknowledges additional support from the John D. and Catherine T. MacArthur Foundation and the Peter and Patricia Gruber Foundation. This research has made use of the Keck Observatory Archive (KOA).
*Facilities:*
[^1]: Other sources of broadening of absorbers include bulk motions of the gas (generally parametrized as a turbulent component of [${b_{\rm d}}$]{}) and differential Hubble flow which broadens the absorption lines originating from the most diffuse and physically extended absorbers.
[^2]: http://www.ast.cam.ac.uk/$\sim$rfc/vpfit.html; © 2007 R.F. Carswell, J.K. Webb
[^3]: The comoving pathlength, $dX$, was defined by @bah69 such that absorbers with constant physical size and comoving number density would have a contant number per $dX$.
[^4]: We note here that with $T_4 = 2$ as we find in this paper, the [$N_{\rm HI}$]{} corresponding to $\bar{\rho}$ is $\log({\ensuremath{N_{\rm HI}}}/$2$)=13.7$, a change of 0.1 dex. However, the uncertainty in other parameters (e.g. $\Gamma_{12}$) is large enough, that we do not consider this small effect in this paper.
[^5]: In the following equation, we have neglected the minor dependence of $\rho_b/\overline \rho_b$ on $T$ from equation \[schaye\].
[^6]: The relative error considered is computed from the formal error bars output by VPFIT for both [${b_{\rm d}}$]{} and [$N_{\rm HI}$]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report a measurement of intergalactic magnetic fields using combined data from Atmospheric Cherenkov Telescopes and Fermi Gamma-Ray Space Telescope, based on the spectral data alone. If blazars are assumed to produce both gamma rays and cosmic rays, the observed spectra are not sensitive to the intrinsic spectrum of the source, because, for a distant blazar, secondary photons produced along the line of sight dominate the signal. In this case, we find $1\times 10^{-17}~{\rm G} < B < 3\times 10^{-14}~{\rm G}$. If one excludes the cosmic-ray component, the $10^{-17}$ G lower limit remains, but the upper limit depends on the spectral properties of the source. We present the allowed ranges for a variety of model parameters.'
address:
- 'Department of Physics and Astronomy, University of California, Los Angeles, CA 90095-1547, USA'
- 'California Institute of Technology, Mail Code 350-17, Pasadena, CA 91125, USA'
- 'IPMU, University of Tokyo, Kashiwa, Chiba 277-8568, Japan'
author:
- Warren Essey
- 'Shin’ichiro Ando'
- Alexander Kusenko
title: Determination of intergalactic magnetic fields from gamma ray data
---
Intergalactic magnetic fields (IGMFs) play an important role in many astrophysical processes and may offer a new window on the cosmology in the early universe, but the size and origin of these fields are still poorly understood [@Kronberg:1993vk]. Until recently, only the upper limits of $10^{-9}$ G were inferred from the observational data [@Barrow:1997mj]. One can measure IGMFs using gamma-ray observations, for example, using time delays [@1995Natur.374..430P], or by searching extended halos around the point objects [@Aharonian:1993vz; @Ando:2010rb]. Here we employ an independent and complementary approach to obtain new upper and lower limits on IGMFs based on the spectra of three blazars observed by HESS, for which there also exist upper limits from Fermi Gamma-Ray Space Telescope. We have calculated the spectra numerically, and we use the goodness of fit to the data as the means to infer the average values of IGMFs. Blazar spectra have been used to set [*lower*]{} bounds on the IGMF [@Neronov:1900zz; @Tavecchio:2010mk; @Dolag:2010ni; @Dermer:2010mm]. However, in each of Refs. [@Neronov:1900zz; @Tavecchio:2010mk; @Dolag:2010ni], only one EBL model and a single power law injection spectrum were considered, and the cosmic ray contribution was not included in Refs. [@Neronov:1900zz; @Tavecchio:2010mk; @Dolag:2010ni; @Dermer:2010mm], while it can be the dominant contribution for distant sources [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. Our analysis is different from previous work in that we have scanned over a very wide range of model parameters, and we included the cosmic-ray component.
AGN are known to be sources of very high energy (VHE) gamma rays; they are also believed to produce cosmic rays [@Biermann:2009cz; @stanev]. There is growing evidence that both primary gamma rays emitted at the source and secondary gamma rays produced by cosmic rays along the line of sight contribute to the observed signals [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. We, therefore, consider two cases: (i) pure gamma-ray emission at the source, and (ii) mixed gamma-ray and cosmic ray emission. In the case (i) the intrinsic source spectra determine the observed spectra. If, however, the cosmic-ray contribution (ii) is included, the observed gamma-ray spectra are remarkably independent of the intrinsic gamma-ray or cosmic-ray spectra at the source. The observed spectrum is determined almost entirely by the shape (but not the overall normalization) of extragalactic background light (EBL) spectrum in case (ii), and the overall agreement of such a prediction with observed spectra of distant blazars can be considered evidence of cosmic ray contribution, and, therefore, of cosmic ray acceleration in AGN [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. Nevertheless, in what follows, we will consider both possibilities (i) and (ii). In the case (i), most of the observed secondary gamma rays are produced within tens of megaparsecs from the source, and our measurement reflects IGMFs closer to the source. In the case (ii), the field strength we infer corresponds to equally weighted averaging along the line of sight.
To probe IGMF we constructed a detailed Monte Carlo simulation that accurately tracks the effects of both electromagnetic showers and cosmic-ray propagation and interactions. Included in the simulation are both pion photo-production and proton pair production for cosmic rays and pair production and inverse Compton scattering for gamma rays and electrons. Any relevant secondary decays such as neutron and pion decays were also included. Deflections due to the IGMF are tracked with both momentum and position vectors stored at all points of the propagation. To construct observed spectra, the position and momentum vectors of gamma rays at the $z=0$ surface are used to calculate the image size. Individual gamma-rays are tracked to the instrument, and their arrival direction further scattered by the instrumental PSF. Any events with arrival directions remaining with the PSF radius are then retained. (A study of the source morphology as a function of energy can be a good diagnostic technique, but in this paper we concentrate only on the spectral analysis.) For $E_\gamma<100$ GeV the publicly available Fermi PSF was used for comparison and for $E_\gamma>100$ GeV HESS PSF was used.
Suppression of the low-energy spectra of point images depends on the mean field along the line of sight more than on the (poorly known) distribution of IGMFs. Indeed, one can use the results of Ref. [@Dolag:2010ni] to compare the effects of uniform magnetic fields with those of more realistic simulations, which take into account the distribution of matter in the universe [@Dolag:2004kp]. The spectrum for “MHD(x)” (“Model 3x”) model, with the mean field of 100 fG (0.1 fG), is close to that of the uniform 100 fG (0.1 fG) field. One can also understand this from general arguments [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. Thus, we modeled the IGMF as a characteristic strength organized into cubes of a characteristic correlation length each with a random direction. Of course, if the distribution of matter along the line of sight is known, e.g., from Lyman-$ \alpha $ forest in the direction to some blazars, one can further improve our inferences of IGMFs. From the available sources, we chose a subset based on three conditions: (a) the source must be observed at energies above 1 TeV, (b) the distance to the source must be relatively large, and (c) the source must show no temporal variability above 1 TeV. The latter condition is an necessary but not sufficient condition that secondary gamma rays dominate the observed signal, since primary gamma rays from blazars are predicted to show temporal variability. Based on these criteria, we chose three sources: 1ES 0229+200 ($z=0.14$), 1ES1101-232 ($z=0.186$), and 1ES0347-121 ($z=0.188$), each of which has been observed by HESS [@Aharonian:2007wc; @Aharonian:2007nq; @Aharonian:2007tc]. Temporal variability has been measured above the 200 GeV energy threshold of the HESS array for 1ES1101 and 1ES0347 [@Zech:ICRC].
[c]{}\
\
We consider a range of viable EBL models. A lower limit on the EBL density is set by galaxy counts [@Xu:2000ss]. The upper limit is less known. Although gamma-ray data have been used to exclude higher EBL models such as Stecker et al. [@Stecker:2005qs], it was recently pointed out that the inclusion of secondary gamma rays from VHE gamma rays and cosmic rays significantly improves the fits to the data for these “high EBL” models [@Essey:2010er]. Therefore, we include the model of Stecker et al. [@Stecker:2005qs] as the upper limit (for UV EBL). A number of other models fall between these two limits [@Salamon:1997ac]. EBL models with a higher level of UV background radiation are difficult to reconcile with a limit based on gamma-ray burst data [@LAT:2010kz].
We shall first discuss the case of primary gamma rays only. Following most recent papers on IGMFs, we will assume that the blazar spectra follow a power law $dN/dE \sim \rm E^{-\Gamma}$ with $\Gamma \geq 1.5$ [@1981MNRAS.196..135P]. However harder spectra are also possible [@Stecker:2007zj]. Furthermore, it is possible that the intrinsic spectra do not conform to a single power law. Deviations from single power law may affect the limits on IGMFs derived under the assumption of a power-law spectrum. We consider the range $\Gamma = 0.5-2$, which encompasses a broad variety of models. We also allow the high-energy cutoff of the power law spectrum to vary in a wide range, $10-10^4$ TeV.
We will consider magnetic fields ranging from $10^{-18}$ G to $10^{-11}$ G. The IGMF correlation lengths are also not known; we assume it is in the range 0.01–1 Mpc and use 1 Mpc as a reference point in Fig. 1.
We fit the calculated spectrum to the data obtained by HESS Atmospheric Cherenkov Telescope (ACT) at high energy, perform statistical analysis to calculate 95% confidence intervals, and we demand that the low-energy prediction not overshoot the Fermi upper limit. The spectra obtained for 1ES 0229+200 ($z=0.14$), 1ES 1101-232 ($z=0.186$), and 1ES 0347-121 ($z=0.188$), are shown in Fig. 1 for IGMFs in the range $10^{-18}-10^{-13}$ G, with a correlation length $\lambda_c = 1$ Mpc. The results for a source with intrinsic spectrum $\Gamma=1.75$, a 100 TeV high-energy cutoff and the “low” EBL model [@Xu:2000ss], shown in Fig. 1, illustrate the effects of IGMF on the spectrum. For larger IGMFs, more of the secondary gamma rays arrive outside the cone determined by the angular resolution of the instrument, with lower energy gamma rays arriving at larger angles. This causes an energy dependent drop in the spectrum.
As one can see from comparing $10^{-15}$ G and $10^{-13}$ G curves in Fig. 1, the wash-out of the signal due to magnetic deflections causes a break in the spectrum at some energy, which is a function of the magnetic field. The following semi-quantitative analytical estimate may be helpful. The magnetic deflections depend on the energy [@Neronov:2009gh]: $$\Theta_{\rm defl}\simeq 0.1^\circ (1+z)^{-2}(\tau_\theta / 5)^{-1}(E_\gamma / 0.1~TeV)^{-1}(B / 10^{-15}~G),$$ for $\rm \lambda_c\gg D_e$, where $\tau_\theta$ is the optical depth of the primary gamma ray and $\rm D_e$ the electron cooling distance. For the energies considered here, $\tau_{\theta} \propto E_\gamma^{-1/2}$, and for inverse-Compton (IC) cooling $E_\gamma \propto E_{\gamma 0}^2$, where $E_{\gamma 0}$ is the energy of the primary photon, and $n_\gamma \propto E_\gamma^{-1/2}$, where $n_{\gamma}$ is the number of photons produced from IC cooling. Thus we expect $\Theta_{\rm defl} \propto B/E_{\gamma 0}^{3}$. Let us consider the flux of gamma rays that contribute to the point image (determined by a fixed instrumental resolution): $\Theta_{\rm defl} < \Theta_{PSF}$ implies $ E_{\gamma 0} > {\rm const} \times B^{1/3}$. Thus for an intrinsic power-law spectrum with index $\Gamma$ we expect the point flux $$F
\propto \frac{1}{E_{\gamma}^{1/2}}\int^\infty_{B^{1/3}}\frac{dN}{dE_{\gamma 0}}\sim E_{\gamma}^{-1/2}\int^\infty_{B^{1/3}}E_{\gamma 0}^{-\Gamma}\sim \frac{1}{E_{\gamma}^{1/2}B^{(\Gamma-1)/3}}.$$ This relation implies the approximate scaling $B^{2(\Gamma-1)} E^3 = {\rm const}$ for the position of the break in the spectrum. By comparing the $10^{-15}$ G and $10^{-13}$ G curves in Fig. \[fig:spechi\] we see that this approximation works reasonably well. Of course, this discussion is admittedly simplistic: the relative contributions of primary and secondary gamma rays, and some other effects were neglected. In deriving the actual limits we use the results of detailed numerical Monte Carlo for propagation, shower development, and the effects of magnetic fields.
For angular resolution of about $0.1^\circ$ we expect magnetic fields larger than about 1 fG to wash out the signal observed by the ACTs, while fields lower than a fG would boost the signal in the Fermi energy range. This allows us to use the combination of ACT and Fermi data to set both upper and lower limits on the IGMF. A lower limit can be found by requiring that the IGMF is large enough to wash out enough of the GeV signal so that the observed spectrum is below the Fermi 95% CL upper limits for the source. An upper limit can be found by requiring that the IGMF be small enough to avoid washing out the TeV signal. More specifically, we combined the ACT data for these three blazars and used 25 data points shown in Fig. 1. Then for each model with given $B$ and $\Gamma$, we have three independent parameters, i.e., normalizations (or luminosities) of the spectra of three blazars. The number of degrees of freedom of the fit is 22. If the value of $\chi_{\rm min}^2$ for each model specified with fixed $B$ and $\Gamma$ is larger than 33.9, then such a model is acceptable at the level of only 5% probability. This way, we set excluded regions (or upper limits on $B$) at 95% CL.
Our results for pure gamma-ray sources are summarized in Fig. 2. The results depend on the choice of EBL model and high energy cutoff. We show both the “low” EBL model, based on lower limits from galaxy counts, and the “high” EBL model [@Stecker:2005qs], as well as high-energy cutoff values of 20 TeV and 100 TeV.
Let us now include the the cosmic ray component [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. Now the observed gamma-ray spectrum does not depend on the intrinsic spectrum of gamma rays, and it is also practically independent of the spectral properties of cosmic rays, such as the high energy cutoff and the cosmic ray spectral index (under some very mild assumptions) [@Essey:2009zg; @Essey:2009ju; @Essey:2010er]. Therefore, instead of the exclusion regions shown in Fig. 2, we find the following bounds (for $\lambda_c=1$ Mpc): $$\begin{aligned}
3\times 10^{-16}~{\rm G} & < B < &\rm 3\times 10^{-14}~{\rm G \ (High \ EBL) }\nonumber\\
1\times 10^{-17}~{\rm G} & < B < &\rm 8\times 10^{-16}~{\rm G \ (Low \ EBL)} \label{IGMF_limits}\end{aligned}$$ Of course it is possible, and likely, that the observed spectrum is a combination of secondary particles from gamma rays and cosmic rays emitted by the source. In this case the most conservative allowed range for IGMF would be the union of the gamma-ray and cosmic-ray allowed ranges.
-------------------------------------------------------- -------------------------------------------------------
{width="45.00000%"} {width="45.00000%"}
{width="45.00000%"} {width="45.00000%"}
-------------------------------------------------------- -------------------------------------------------------
The limits in Fig. 2 imply $\Gamma \le 1.8$ for the set of blazars we chose, which is in good agreement with mean spectral index of 1.9 reported by Fermi for nearby sources.
The limits presented in Fig. 2 are subject to some caveats. First, we have assumed that our three sources had similar intrinsic spectra, which need not be the case. Second, as discussed above, the likely contribution of protons along the line of sight [@Essey:2009zg; @Essey:2009ju; @Essey:2010er] can reduce the limits to the model-independent form shown in Eq. \[IGMF\_limits\]. Finally, time variability of the primary gamma ray sources can affect our conclusions [@Dermer:2010mm].
Following the insightful discussion of Dermer et al. [@Dermer:2010mm], it can be shown that the typical time delays for secondary photons are $$\Delta t \simeq 2\times 10^6 \lambda_{100} \left(\frac{B_{-15}}{E_\gamma/10~\rm{GeV}}\right)^2~\rm{yrs},
\label{eq:timedelays}$$ where $\lambda_{100}\sim1$ for the Stecker et al EBL model, $B_{-15}$ is the IGMF in units of $10^{-15}$ G. If the source is active for a timescale that is shorter than these delays, the observed signal is suppressed by a factor $(t_{\rm active}/\Delta t)$ where $t_{\rm active}$ is the time that the source was active for. This suppression mimics the effects of a stronger magnetic field, and, therefore, the predicted limit on the IGMF would be suppressed by the a factor $$\kappa \simeq \left(\frac{t_{\rm active}}{2\times 10^6~\rm{yrs}}\right)^{1/2}\left(\frac{10^{-15}~\rm{G}}{B_{\rm limit}}\right)\left(\frac{E_\gamma}{10~\rm{GeV}}\right),
\label{eq:timedamp}$$ where $B_{\rm limit}$ is either the upper or lower limit shown in Fig. 2.
The observations of 1ES 0229+100 by HESS [@Aharonian:2007wc] and VERITAS [@Perkins:VERITAS] were separated by roughly 3 years and show no evidence of significant variability. Variability may be present during the times when the source was not observed, or at a level below the instrumental sensitivity. We assume $t_{\rm active}\sim 3$ yrs as a lower limit on the time the source was active. The most restrictive energy for the lower limits is near $E_\gamma \sim 10$ GeV and most of the lower limits fall in the range $10^{-15}-10^{-17}$ G which means the most conservative lower limits would be suppressed by a factor of $10-10^3$. For upper limits, the most restrictive energy is at $E_\gamma\sim 1$ TeV and since most upper limits fell into the range $10^{-13}-10^{-15}$ G we expect the upper limits to also be lowered by roughly $10-10^3$. Thus, the overall effect of including a short activity time is to lower the 95% confidence intervals shown in Fig. 2 and our robust $10^{-17}$ G lower limit would be lowered to $10^{-18}$ G in the most conservative case.
It is important to distinguish between the time scale of the source activity and short-term variability. A variability of the source would not lead to a suppression discussed above because signals of variable activity would undergo significant delays in the IGMF and would be observed as a constant source with the time averaged luminosity of the flares. Thus, if the source is active on a timescale of $10^6$ yrs, the limits in Fig. 2 are valid, even in the case of variability on shorter time scales.
The effect the source activity–dormancy time scale is much weaker in the case of the cosmic ray limits, because the delays in the stronger magnetic fields near the source can wash out variability on even longer time scales. The magnetic fields within the host galaxies containing source are likely to be of the order of $1~\mu G$ which can lead to significant delays. Fig. \[prosource1e6\] shows the delays for a 100 kpc wide source of $10^{10}$ GeV protons with $B=1~\mu \rm{G}$ and $l_c=0.1$ kpc. It can be seen that the delays are significantly longer than the observed variability of blazars and thus we expect any intrinsic variability to be washed out. Such a source appears as a constant, long-lived source, and any time delays incurred in the host galaxy are not measurable. Of course, there is a significant uncertainty in the magnetic field distribution around the source, in particular, in the direction of the jet, and the values assumed here are taken solely to illustrate the idea.
The effect of the source variability would be to damp the observed power of the source by a factor of $$f_{damp}\sim N_{\rm active}\frac{t_{\rm active}}{t_{\rm delay}},$$ where $t_{\rm delay}$ is the typical proton delay at the source, $t_{\rm active}$ is the typical time the source is active or flaring and $N_{\rm active}$ is the number of times the source is active in the time period $t_{\rm delay}$. This damping will not be a significant effect, especially since the typical deflections at the source will not be enough to affect the beaming factors.
![ Arrival time probability distribution in arbitrary units for primary cosmic rays including only delays at the source. A cosmic ray source of $10^{10}$ GeV protons was used in a source 100 kpc wide with $B=10^{-6}$ G and $l_c=0.1$ kpc.[]{data-label="prosource1e6"}](Arrival_pr_1e10_1e6.eps){width="\textwidth"}
Using Eq. (\[eq:timedamp\]) and these typical time delays we see that the model independent lower limit of $10^{-17}$ G would remain unchanged and the upper limit of $3\times 10^{-14}$ G would be lowered by roughly one order of magnitude. Thus the limits we provided for cosmic rays are still the most conservative limits even including all timing uncertainties. Thus, for the case of cosmic rays dominating the signal, we can report a model-independent 95% CL interval for the IGMF of $$1\times 10^{-17}~\rm G < B < 3\times 10^{-14}~G.$$
Our results represent the first 95% confidence interval measurements of the IGMF based on the spectral data alone. The method used here to probe the IGMF will become more powerful in the future. Future observations of distant blazars will not only improve the limits on the IGMF, but also constrain AGN and EBL models. The measured values of IGMFs can be used to distinguish between the cosmological and astrophysical scenarios for the origin of all astrophysical magnetic fields. In addition, this information will lead to a better understanding of gamma rays and cosmic rays, as well as the properties of AGN and of universal photon backgrounds.
The authors thank F. Aharonian, J. Beacom, P. Blasi, C. Dermer, S. Razzaque, and E. Waxman for helpful comments. The work of W.E. and A.K. was supported by DOE grant DE-FG03-91ER40662 and NASA ATP grant NNX08AL48G. The work of S.A. was supported by Japan Society for Promotion of Science.
[99]{}
P. P. Kronberg, Rept. Prog. Phys. [**57**]{} (1994) 325; A. Kandus, K. E. Kunze and C. G. Tsagas, \[arXiv:1007.3891\]. J. D. Barrow, P. G. Ferreira and J. Silk, Phys. Rev. Lett. [**78**]{} (1997) 3610; K. Subramanian and J. D. Barrow, Phys. Rev. Lett. [**81**]{} (1998) 3575; P. Blasi, S. Burles and A. V. Olinto, Astrophys. J. [**514**]{} (1999) L79; K. Jedamzik, V. Katalinic and A. V. Olinto, Phys. Rev. Lett. [**85**]{} (2000) 700. R. Plaga, Nature, 374 (1995) 430.
F. A. Aharonian, P. S. Coppi, H. J. Volk, Astrophys. J. [**423**]{}, L5-L8 (1994).
S. Ando and A. Kusenko, Astrophys. J. [**722**]{} (2010) L39; A. Neronov [*et al.*]{}, Astron. and Astrophys., [**526**]{} (2010) 90.
A. Neronov, I. Vovk, Science [**328**]{} (2010) 73. F. Tavecchio [*et al.*]{}, Mon. Not. Roy. Astron. Soc. [**406**]{} (2010) L70. K. Dolag [*et al.*]{}, Astrophys. J. [**727**]{} (2011) L4. C. D. Dermer [*et al.*]{}, \[arXiv:1011.6660\]. W. Essey and A. Kusenko, Astropart. Phys. [**33**]{} (2010) 81. W. Essey, O. E. Kalashev, A. Kusenko and J. F. Beacom, Phys. Rev. Lett. [**104**]{} (2010) 141102. W. Essey, O. Kalashev, A. Kusenko and J. F. Beacom, Astrophys. J., 731 (2011) 51. P. L. Biermann [*et al.*]{}, Int. J. Mod. Phys. D [**18**]{} (2009) 1577; arXiv:1012.0204 \[astro-ph.GA\]. For recent reviews, see, [*e.g.*]{}, T. Stanev, Mod. Phys. Lett. A, 25, (2010) 1467; T. K. Gaisser, arXiv:1010.5996 \[astro-ph.HE\]. K. Dolag [*et al.*]{}, JCAP [**0501**]{} (2005) 009; D. Ryu, H. Kang, J. Cho, S. Das, Science, 320 (2008) 909.
F. Aharonian [*et al.*]{}, Astron. Astrophys. [**475**]{} (2007) L9. F. Aharonian [*et al.*]{}, Astron. Astrophys. [**470**]{} (2007) 475. F. Aharonian [*et al.*]{}, Astron. Astrophys. [**473**]{} (2007) L25. A. Zech [*et al.*]{}, in Proceedings of 31st ICRC (Lodz), 2009.
J. S. [Perkins]{} and [VERITAS Collaboration]{}, [Bulletin of the American Astronomical Society]{},[**42**]{} (2010) 708.
C. Xu [*et al.*]{}, Astrophys. J., [**562**]{} (2001) 179.
F. W. Stecker, M. A. Malkan and S. T. Scully, Astrophys. J. [**648**]{} (2006) 774.
M. H. Salamon and F. W. Stecker, Astrophys. J. [**493**]{} (1998) 547; T. M. Kneiske et al., Astron. Astrophys. [**386**]{}(2002) 1; [*ibid.*]{}, [**413**]{} (2004) 807; A. Franceschini, G. Rodighiero and M. Vaccari, Astron. Astrophys. [**487**]{} (2008) 837; S. Horiuchi, J. F. Beacom and E. Dwek, Phys. Rev. D [**79**]{} (2009) 083013; J. R. Primack, R. C. Gilmore and R. S. Somerville, AIP Conf. Proc. [**1085**]{} (2009) 71; R. C. Gilmore [*et al.*]{}, MNRAS, [**399**]{} (2009) 1694; S. Razzaque, C. D. Dermer and J. D. Finke, Astrophys. J. [**697**]{} (2009) 483; J. D. Finke, S. Razzaque and C. D. Dermer, Astrophys. J. [**712**]{} (2010) 238. A. A. Abdo [*et al.*]{}, Astrophys. J. [**723**]{} (2010) 1082. J. A. Peacock, MNRAS [**196**]{} (1981) 135; J. G. Kirk and P. Schneider, Astrophys. J., [**315**]{} (1987) 425; A. F. [Heavens]{} and L. O. [Drury]{}, MNRAS 235 (1988) 997; J. Bednarz and M. Ostrowski, Phys. Rev. Lett. [**80**]{} (1998) 3911; M. A. Malkov and L. O’C Drury, Reports Prog. Phys., 64 (2001) 429.
F. W. Stecker, M. G. Baring and E. J. Summerlin, Astrophys. J. [**667**]{} (2007) L29; F. Aharonian, D. Khangulyan and L. Costamante, MNRAS, 387, (2008) 1206.
A. Neronov and D. V. Semikoz, Phys. Rev. D [**80**]{} (2009) 123012.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Rhodonea curves are classical planar curves in the unit disk with the characteristic shape of a rose. In this work, we use point samples along such rose curves as node sets for a novel spectral interpolation scheme on the disk. By deriving a discrete orthogonality structure on these rhodonea nodes, we will show that the spectral interpolation problem is unisolvent. The underlying interpolation space is generated by a parity-modified Chebyshev-Fourier basis on the disk. This allows us to compute the spectral interpolant in an efficient way. Properties as continuity, convergence and numerical condition of the scheme depend on the spectral structure of the interpolation space. For rectangular spectral index sets, we show that the interpolant is continuous at the center, the Lebesgue constant grows logarithmically and that the scheme converges fast if the function under consideration is smooth. Finally, we derive a Clenshaw-Curtis quadrature rule using function evaluations at the rhodonea nodes and conduct some numerical experiments to compare different parameters of the scheme.'
address: |
University of Padova\
Department of Mathematics “Tullio Levi-Civita”\
Via Trieste 63, 35121 Padova, Italy
author:
- Wolfgang Erb
title: Rhodonea curves as sampling trajectories for spectral interpolation on the unit disk
---
Spectral interpolation on the disk ,rhodonea curves ,intersection and boundary nodes of rhodonea curves ,parity-modified Chebyshev-Fourier series ,Clenshaw-Curtis quadrature on the disk ,numerical condition and convergence of interpolation schemes
Introduction {#sec:introduction}
============
Rose curves are classical planar curves in a disk that have the shape of a patelled rose. Guido Grandi, studying these curves profoundly in the beginning of the 18th century [@Grandi1728], used the corresponding greek name for them: rhodonea curves. These algebraic curves have a particular simple parametric and polar form. This makes them to interesting trajectories for data sampling in imaging. Examples of scanning systems using rhodonea curves are, for instance, Magnetic Particle Imaging [@Knopp2009PhysMedBio; @Knopp2017; @Szwargulski2015b] and laser scanners based on rotating Risley prisms [@DumaSchitea2018; @Lu2014]. Further, rose curves are also very popular in calculus text books to teach parametrization and integration in polar coordinates.
In this work, we study rhodonea curves as sampling trajectories for new and promising sets of interpolation nodes on the unit disk. If the samples are taken in a time-equidistant way along the curve, the nodes form a pair of interlacing polar grids. This structure of the so called rhodonea nodes together with an accordingly chosen basis system allows us to construct a simple and efficient spectral interpolation and quadrature scheme on the disk.
As a suitable basis system for the spectral interpolation on the rhodonea nodes we use a parity-modified Chebyshev-Fourier basis. Among other well-known basis systems as the Logan-Shepp ridge polynomials or the Zernike polynomials, the Chebyshev-Fourier basis is a very popular choice for spectral methods on the unit disk [@Boyd2000; @BoydYu2011; @Fornberg1995; @Fornberg1996; @Shen2011; @Trefethen2000; @TownsendWilberWright2017]. One main advantage of the Chebyshev-Fourier basis is the possibility to compute the interpolating function very efficiently using fast Fourier methods. In relation to other systems, this basis system performs however not so well at the center of the unit disk. For a detailed comparison of the different spectral methods on the unit disk we refer to the profound discussion in [@BoydYu2011].
[**1.1. Main contributions.**]{}
- *Characterization of the rhodonea interpolation nodes.* We provide new descriptions of the intersection and boundary points of the rhodonea curves and show how they can be used as nodes for a spectral interpolation scheme on the disk.
- *Unisolvence of interpolation scheme on rhodonea nodes.* We will prove the unisolvence of the spectral interpolation problem on the rhodonea nodes. The interpolation spaces are spanned by a general spectral set of Chebyshev-Fourier basis functions.
- *Efficient implementation.* We show that the spectral interpolation on the rhodonea nodes can be performed efficiently using a two-dimensional fast Fourier transform.
- *Numerical condition and convergence analysis.* The main interpolation space considered in this work is based on a rectangular spectral index set. For this space we show that the numerical condition of the interpolation is growing only logarithmically in the number of nodes and that the scheme converges fast if the interpolated function is smooth.
- *Continuity and quadrature.* For the rectangular spectral index set we can guarantee that the interpolant is continuous at the center of the disk. Further, we show how the interpolation scheme can be used to define a Clenshaw-Curtis quadrature rule on the disk.
[**1.2. Comparison to existing work.**]{}
*Comparison to standard tensor-product schemes on the disk.* Spectral methods based on a Chebyshev-Fourier basis can be implemented efficiently by fast Fourier algorithms. In many common implementations, the calculation of the coefficients in the Chebyshev-Fourier series is performed on a tensor-product polar grid [@BoydYu2011; @Fornberg1995; @Shen2011; @Trefethen2000; @TownsendWilberWright2017]. The rhodonea nodes used in this work allow a similar computation of the Chebyshev-Fourier coefficients with equivalent efficiency and convergence rates. Compared to the tensor-product case, the new scheme provides the following additional features:
- The data can be collected by sampling along one or several rhodonea curves. This is particularly interesting for the applications in which rose curves are used as scanning trajectories. In this perspective, rhodonea nodes can be interpreted as polar analogs of rank-$1$ trigonometric lattices [@KKP2012; @KPV2015] or rank-$1$ Chebyshev lattices [@CoolsPoppe2011; @PottsVolkmer2015].
- The presented interpolation scheme on the rhodonea curve is more flexible in terms of the underlying interpolation space. The unisolvence of the interpolation problem is guaranteed for a large class of spectral index sets. This is a polar version of a bivariate result in which a similar flexibility is known for polynomial interpolation on interlacing grids [@Floater2017].
*Complementation of work on Lissajous nodes.* Rhodonea curves can be regarded as polar counterparts of bivariate Lissajous curves on the square $[-1,1]^2$ and of spherical Lissajous curves. This article is a continuation of the work on polynomial interpolation on Lissajous curves [@DenckerErb2017a; @DenckerErb2015a; @Erb2015; @ErbKaethnerAhlborgBuzug2015] and on spherical Lissajous nodes [@ErbSphere2017] and extends it to the polar setting. The differences between the actual work on the disk and the previous works on the hypercube and the unit sphere arise naturally from the differing geometries. In all three settings, the generating curves and the interpolation nodes have own characteristic properties and the interpolation spaces have to be set up according to the given symmetries. Nevertheless, the core ideas in all three setups are similar and many of the ideas used for Lissajous curves can be carried over to the setting of rhodonea curves. In particular, as for multivariate Lissajous-Chebyshev points in the hypercube [@DenckerErb2017a; @DenckerErb2015a], a main step in the proof of the quadrature and interpolation formulas is a discrete orthogonality structure linked to the structure of the rhodonea nodes. Compared to previous works, a major progress in this article is the larger flexibility in the choice of the interpolation space.
[**1.3. Organization.**]{} After a short introduction, we provide three different characterizations of the rhodonea nodes: 1) by time equidistant samples along the rhodonea curve (Section \[sec:rhodonea\]), 2) in terms of a union of two interlacing polar grids (Section \[sec:nodesphere\]), and 3) by using the algebraic description of the rhodonea varieties (Section \[sec:rhodvar\]).
The technical background for the interpolation results in form of a discrete orthogonal structure and spectral index sets is given in Section \[1507091240\]. The main results providing the unisolvence of the spectral interpolation on the rhodonea nodes are proven in Section \[sec:interpolation\].
In Section \[sec:implementation\], we describe an efficient implementation of the interpolation scheme using the fast Fourier transform. We conclude this work with Section \[sec:convergence\] and a mathematical description of various properties of the interpolation scheme including: 1) the behavior of the interpolant at the center of the disk, 2) the numerical condition of the scheme, 3) convergence rates, and 4) the application to a Clenshaw-Curtis quadrature rule on the disk. The proofs of all results are collected in Section \[sec:proof\].
Rhodonea curves on the unit disk {#sec:rhodonea}
================================
#### **2.1. General properties**
For a frequency vector ${\boldsymbol{m}} = (m_1, m_2) \in{{\mathbb N}}^{2}$ and a rotation parameter $\alpha \in {{\mathbb R}}$, we define the *rhodonea curves* in parametric form as $$\label{201509161237}
{{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t) = \Big( \cos(m_2 t) \cos (m_1 t - \alpha \pi), \ \cos(m_2 t) \sin (m_1 t - {\textstyle}\alpha \pi) \Big), \quad t \in \mathbb{R}.$$ The rhodonea curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ is contained in the unit disk ${\mathbb{D}}= \{{\boldsymbol{x}} \in {{\mathbb R}}^2 \ : \ |{\boldsymbol{x}}| \leq 1\}$. Because of its characteristic shape of a patelled rose, these curves are also referred to as *rose curves* or *roses of Grandi*, after the monk and mathematician Guido Grandi who studied them intensively in [@Grandi1728]. Two typical examples of rose curves are illustrated in Figure \[fig:LS-1\].
The frequency parameters $m_1$ and $m_2$ in the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ determine a superposition of a radial and an angular harmonic motion. For this reason, rose curves can also be regarded as polar variants of bivariate Lissajous curves [@DenckerErb2017a; @DenckerErb2015a; @Erb2015; @ErbKaethnerAhlborgBuzug2015]. If the numbers $m_1$ and $m_2$ are relatively prime, the minimal period $P$ of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ is given by $P=2\pi$ if $m_1+m_2$ is odd, and $P = \pi$ if $m_1+m_2$ is even (see Proposition \[prop-11\]). Depending on these two cases, the properties of the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ vary slightly and we will have to distinguish them at several occasions.
For general ${\boldsymbol{m}} \in {{\mathbb N}}^2$ and $g = \mathrm{gcd}({\boldsymbol{m}}) \geq 1$, we can write ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t) = {\boldsymbol{\varrho}}^{({\boldsymbol{m}}/g)}_{\alpha}( g t)$. In this case, the minimal period of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ is given by $P/g$. In particular, all properties of a rose ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ with general ${\boldsymbol{m}} \in {{\mathbb N}}^2$ can be obtained from the curve ${\boldsymbol{\varrho}}^{({\boldsymbol{m}}/g)}_{\alpha}$ with the relatively prime parameter ${\boldsymbol{m}}/g$. When analyzing the properties of a single rose curve it is therefore enough to restrict the considerations to relatively prime frequency numbers $m_1$ and $m_2$. However, if more than one rhodonea curve is used to generate the interpolation nodes, also the general case will play an important role later on.
#### **2.2. The self-intersection points of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$**
To extract all self-intersection points of the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$, we consider for $t \in [0,2\pi)$ the sets $\mathcal{S}^{({\boldsymbol{m}})}(t) = \{ s \in [0,2\pi): \ {{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(s) = {{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t) \}$ and the sampling points $$\begin{aligned}
t^{({\boldsymbol{m}})}_{l} &= \frac{l \pi}{2 m_1 m_2}, \quad l \in \{0,1, \ldots, 4m_1m_2-1\}. \label{eq-samples1}\end{aligned}$$
\[prop-11\] Let $m_1$ and $m_2$ be relatively prime numbers.\
If $m_1+m_2$ is odd, the minimal period of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ is $2 \pi$ and $$\begin{array}{lll}
(i) & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 2 m_2 & \text{if}\quad t \in \{\,t^{({\boldsymbol{m}})}_{l}\,| \ l\in \{0,\ldots,4m_1m_2-1\}, \ l \equiv m_1 \mod 2 m_1 \},\\
(ii) & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 2 & \text{if}\quad t \in \{\,t^{({\boldsymbol{m}})}_{l}\,| \ l\in \{0,\ldots,4m_1m_2-1\}, \ l \not\equiv 0 \mod m_1 \}, \\
(iii) & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 1 & \text{for all other $t \in [0,2\pi)$.}
\end{array}$$ If $m_1+m_2$ is even, then the minimal period of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ is $\pi$ and $$\begin{array}{lll}
(i)' & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 2 m_2 & \text{if}\quad t \in \{\,t^{({\boldsymbol{m}})}_{l}\,| \ l\in \{0,\ldots,4m_1m_2-1\}, \ l \equiv m_1 \mod 2 m_1 \}, \\
(ii)' & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 4 & \text{if}\quad t \in \{\,t^{({\boldsymbol{m}})}_{2l}\,| \ l\in \{0,\ldots,2m_1m_2-1\}, \ l \not\equiv 0 \mod m_1 \}, \\
(iii)' & \# \mathcal{S}^{({\boldsymbol{m}})}(t) = 2 & \text{for all other $t \in [0,2\pi)$.}
\end{array}$$
We can extract a series of properties from this result. The nodes ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)$ in $(i)$ and $(i)'$ with $\# \mathcal{S}^{({\boldsymbol{m}})}(t) = 2 m_2$ correspond to the center $(0,0)$ of the unit disk ${\mathbb{D}}$. As $t$ varies from $0$ to $P$, the center is traversed $2m_2$ times in the case that $m_1+m_2$ is odd and $m_2$ times if $m_1+m_2$ is even. All the points ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)$ in $(ii)$ and $(ii)'$ are doubly traversed in one period $P$. Therefore, if $m_1+m_2$ is odd, Proposition \[prop-11\] ensures that the set $$\label{1709171731}
{{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha} = \left\{\,{{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t^{({\boldsymbol{m}})}_{l})\,|\, l\in \{0,\ldots,4m_1m_2-1\} \,\right\}$$ contains all self-intersection points of the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$. The additional nodes ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t^{({\boldsymbol{m}})}_{l})$ with $l \equiv 0 \mod 2 m_1$ describe precisely the set of all points at which the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ touches the boundary of the unit disk ${\mathbb{D}}$ (i.e. the unit circle). If $m_1+m_2$ is even, the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha}$ is larger than the union of self-intersection and boundary points of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$. Nevertheless, also in this case the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha}$ will play an important role in our considerations. We summarize all important properties of the rhodonea curves in the following Corollary \[cor-1\].
\[cor-1\] Let $m_1$ and $m_2$ be relatively prime natural numbers.
- If $m_1 + m_2$ is odd, then ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$ is the union of all self-intersection and all boundary points of the closed curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$. ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$ contains $2 m_1 m_2 + 1$ points in ${\mathbb{D}}$. It includes the center $(0,0)$ that is traversed $2 m_2$ times in one period $P = 2\pi$, $2 (m_1-1) m_2$ ordinary double points distinct from $(0,0)$ and $2m_2$ points on the boundary of ${\mathbb{D}}$.
- If $m_1 + m_2$ is even, the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ contains $\frac12(m_1-1) m_2$ ordinary double points distinct from $(0,0)$ and $m_2$ points on the boundary of ${\mathbb{D}}$. The center $(0,0)$ is traversed $m_2$ times in one period $P = \pi$.
Various properties of the rhodonea curves described in this section are known for a long time. The number and type of intersection points are, for instance, originally derived in [@Himstedt1888]. A general historic overview with a lot of additional features of rhodonea curves can be found in [@Loria1902 p. 297-306]. Further graphical illustrations of rhodonea curves are given in [@Gorjanc2010]. The novel aspects of this article are the different characterizations of the rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$. In Corollary \[cor-1\] i) we could describe the union of intersection and boundary points as the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$ of time-equidistant samples along the rhodonea curve. Further characterizations of ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$ are now obtained in the next part.
The interpolation nodes generated by rhodonea curves {#sec:nodesphere}
====================================================
The nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha}$ of the rhodonea curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ given in Corollary \[cor-1\] i) can be described as the union of two interlacing rectangular grids in polar coordinates. Without restriction to generality, we set $\alpha = 0$ and consider the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{0}$. Further, we will use general frequencies ${\boldsymbol{m}} = (m_1, m_2) \in{{\mathbb N}}^{2}$ for this second description. If $m_1$ and $m_2$ are not relatively prime, the so obtained nodes contain ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{0}$ as a subset and can be interpreted as sampling nodes of more than one rhodonea curve. Similarly, if $m_1+m_2$ is even the given description will contain the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{0}$ as a subset. First examples are given in Figure \[fig:LS-2a\] and \[fig:LS-2b\].
To obtain this more general characterization, we introduce the nodal index set $$\label{eq:0911}
{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}= \left\{ \,(i_1, i_2) \in {{\mathbb Z}}^{2}\ \left|\begin {array}{ll}
& 0\leq i_{1}\leq m_{1}, \; -2 m_2 < i_2 \leq 2 m_2, \\
& \text{$i_2 \leq 0$ \, if $i_{1} = m_1$}, \\
& \text{$i_{1} + i_{2}$ is even}
\end{array}\right. \, \right\}.$$ The set ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ can be decomposed into the two disjoint finite grids ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0$ and ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1$ given by $$\label{eq:0901}
{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0 = \{(i_1, i_2) \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}\ | \ \text{$i_1$, $i_2$ are even} \, \}, \quad
{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1 = \{(i_1, i_2) \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}\ | \ \text{$i_1$, $i_2$ are odd} \, \}.$$ A tuple ${\boldsymbol{i}} = (i_1, i_2)$ in ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ has a one to one relation to a point in polar coordinates by introducing a radial and an angular component $$r^{(m_1)}_{i_1} = \cos \left( \frac{i_1}{2 m_1} \pi \right) \in [0,1], \qquad \theta^{(m_2)}_{i_2} = \frac{i_2}{2 m_2} \pi \in (-\pi,\pi].$$ The general *rhodonea nodes* on the unit disk are then defined as the point set $$\label{eq:09172}
{{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}=\left\{\, {\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}\,\left|\,{\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}\right.\right\},$$ with the nodes ${\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} \in {\mathbb{D}}$ given in polar coordinates $(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2})$ by $$\label{eq:0917234} {\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} = \left(r^{(m_1)}_{i_1} \cos (\theta^{(m_2)}_{i_2}), r^{(m_1)}_{i_1} \sin (\theta^{(m_2)}_{i_2})\right).$$ From the almost rectangular form of ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0$ and ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1$ in (see also Figure \[fig:LS-2a\] (a) and \[fig:LS-2b\] (a)), the cardinalities $\# {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0$ and $\# {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1$ can be determined by a simple counting argument: $$\label{eq:0917}
\#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0 = (m_1+1) m_2, \quad \#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1 = m_1 m_2, \quad \#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}=\#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0 +\#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1 = (2 m_1+1) m_2.$$ Since the $m_2$ points ${\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}$ with coordinate $i_1 = m_1$ all describe the center $(0,0)$ of ${\mathbb{D}}$, the cardinality of ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ is smaller campared to $\#{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$: $$\#{{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}= 2 m_1 m_2 + 1.$$ In the setting of Corollary \[cor-1\] i), the cardinality of the node set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0$ corresponds exactly to the cardinality of the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$. For general ${\boldsymbol{m}} \in {{\mathbb N}}^{2}$ with $g = \gcd({\boldsymbol{m}}) \geq 1$, we have the following relation between ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ and the node points ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha}$ defined in .
\[cor-111\] Let ${\boldsymbol{m}} \in {{\mathbb N}}^2$. Then $${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}= \bigcup_{\rho = 0}^{2 g-1} {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\rho/m_2} =
\bigcup_{\rho = 0}^{2 g-1}\left\{\,{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{\rho/m_2} (t^{({\boldsymbol{m}})}_{l})\ \Big| \ l\in \{0,\ldots,4m_1m_2/g-1\} \,\right\},$$ where ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ are the rhodonea curves introduced in , $t^{({\boldsymbol{m}})}_{l}$ the equidistant sampling points given in , and ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{\alpha}$ the node points defined in .
\[ex:1\] If $m_1$ and $m_2$ are relatively prime, we have $g = \gcd ({\boldsymbol{m}}) = 1$ and Theorem \[cor-111\] states that ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ is generated by the two rhodonea curves ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ and ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{1/m_2}$. Let $a$ and $b$ be two integers from Bézout’s lemma such that $a m_1 + b m_2 = 1$. Then $$\label{eq:111123421324}
{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}(t - a \pi / m_2) = (-1)^{a+b} {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{1/m_2}(t).$$
1. If $m_1+m_2$ is odd, then the rhodonea curve ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ satisfies ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}(t) = -{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}(t-\pi)$ and ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}({{\mathbb R}})$ is point symmetric with respect to the origin. The same holds true for the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0$ given in . Therefore, the identity implies that ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}({{\mathbb R}}) = {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{1/m_2}({{\mathbb R}})$ and $${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}= {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0 = {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{1/m_2}.$$ Thus, according to Corollary \[cor-1\] i), the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ corresponds to the union of self-intersection and boundary points of the curve ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$. Rotating the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ by an angle $\alpha \pi$ gives a corresponding identity for the points ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_\alpha$.
2. If $m_1+m_2$ is even, then $a + b$ is odd and gives ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}(t - a \pi / m_2) = - {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{1/m_2}(t)$. Further, in this case the sets ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}({{\mathbb R}})$ and ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0$ are not point symmetric with respect to the center $(0,0)$. This implies that ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_{1/m_2} = -{{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0 \neq {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0$ and $${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}= {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0 \cup -{{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}_0.$$ In particular, ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ is generated by the samples of two distinct rhodonea curves.
A link to rhodonea varieties {#sec:rhodvar}
============================
The union of rhodenea curves used to generate the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ in Theorem \[cor-111\] can be identified as an algebraic variety. For $r \in [-1,1]$, we denote by $T_{m_1}(r) = \cos (m_1 \arccos r)$ the univariate Chebyshev polynomial of degree $m_1$ and by $H_{m_2}(x_1,x_2)$ the bivariate polynomial $$H_{m_2}(x_1,x_2) = \sum_{k=0}^{\lfloor m_2/2 \rfloor} \binom{m_2}{2k} (-1)^k x_1^{m_2-2k} x_2^{2k}.$$ $H_{m_2}$ is a bivariate homogeneous polynomial of total degree $m_2$. The *rhodonea variety* $\mathcal{R}^{{\boldsymbol{m}}}$ on the unit disk ${\mathbb{D}}$ is defined as $$\label{1509222010}
\mathcal{R}^{({\boldsymbol{m}})}
= \left\{\,{\boldsymbol{x}} \in {\mathbb{D}}\,\Big| \, (x_1^2+x_2^2)^{m_2} T_{m_1}\left(\sqrt{x_1^2+x_2^2}\right)^2 = H_{m_2}(x_1,x_2)^2 \, \right\}.$$ This affine real algebraic variety is of order $2 m_1 + 2 m_2$. In polar coordinates, we get a simpler description of this variety. With the substitution $x_1(r,\theta) = r \cos \theta$ and $x_2(r,\theta) = r \sin \theta$ and the trigonometric formula $$\cos (m_2 \theta) = \sum_{k=0}^{\lfloor m_2/2 \rfloor} \binom{m_2}{2k} (-1)^k \cos(\theta)^{m_2-2k} \sin(\theta)^{2k}$$ we can rewrite as $$\label{1509222011}
\mathcal{R}^{({\boldsymbol{m}})}
= \left\{\,\left.{\boldsymbol{x}}(r,\theta) \in {\mathbb{D}}\,\right|\, T_{m_1}(r)^2 = \cos^2(m_2 \theta) \,\right\}.$$ Since $H_{m_2}(\cos \theta, \sin \theta) = \cos (m_2 \theta)$ we see that $H_{m_2}(x_1,x_2)$ is in fact a harmonic homogeneous polynomial of degree $m_2$.
\[thm:decompositionrhodonea\]
1. The variety $\mathcal{R}^{({\boldsymbol{m}})}$ can be decomposed as $\displaystyle\mathcal{R}^{({\boldsymbol{m}})} = \bigcup_{\rho = 0}^{2 g-1} {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{\rho/m_2}([0,P)).$
2. The rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ can be written as $${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}= \left\{\,\left.{\boldsymbol{x}}(r,\theta) \in {\mathbb{D}}\,\right|\, T_{m_1}(r)^2 = \cos^2(m_2 \theta) \in \{0,1\} \,\right\},$$ i.e., the set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ consists of those points of the variety $\mathcal{R}^{({\boldsymbol{m}})}$ for which $T_{m_1}(r)^2$ and $\cos^2(m_2 \theta)$ get maximal or minimal.
We mention (without explicit proof) that in addition to the statements of Theorem \[thm:decompositionrhodonea\], the points in ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ can also be categorized in terms of singularity theory. This yields a description similar to the one given in Corollary \[cor-1\] i). Namely, the elements of ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ in the interior of the unit disk ${\mathbb{D}}$ are precisely the singular points of the algebraic variety $\mathcal{R}^{({\boldsymbol{m}})}$. The singular points distinct from the center $(0,0)$ are all ordinary double points while the center itself is a singular point with multiplicity $2 m_2$.
1. We consider the setting of Corollary \[cor-1\] i), i.e., $m_1$ and $m_2$ are relatively prime and $m_1 + m_2$ is odd. Then, according to Example \[ex:1\] i) and Theorem \[thm:decompositionrhodonea\] a), we have $\mathcal{R}^{({\boldsymbol{m}})} = {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}([0,2 \pi))$, i.e. $\mathcal{R}^{({\boldsymbol{m}})}$ provides the algebraic equation of the curve ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ given in parametric form in . This characterization of ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ is well known in the literature. Slightly less compact variants of the definition in can be found in [@Gorjanc2010; @Himstedt1888; @Loria1902].
2. If $m_1$ and $m_2$ are relatively prime and $m_1+m_2$ is even, we get a different scenario. In this case, Example \[ex:1\] ii) and Theorem \[thm:decompositionrhodonea\] a) imply that $\mathcal{R}^{({\boldsymbol{m}})} = {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}([0,\pi)) \cup -{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}([0,\pi))$, i.e. the algebraic variety $\mathcal{R}^{({\boldsymbol{m}})}$ is the union of two distinct rhodonea curves. The algebraic varieties $\mathcal{R}^{({\boldsymbol{m}})}_+$ and $\mathcal{R}^{({\boldsymbol{m}})}_-$ describing the single curves $\pm {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ are given as $$\mathcal{R}^{({\boldsymbol{m}})}_{\pm} = \left\{\,\left.{\boldsymbol{x}} \in {\mathbb{D}}\,\right|\, (x_1^2+x_2^2)^{\frac{m_2}{2}}
T_{m_1}\left(\sqrt{x_1^2+x_2^2}\right) = \pm H_{m_2}(x_1,x_2) \,\right\}.$$ $\mathcal{R}^{({\boldsymbol{m}})}_{\pm}$ are algebraic varieties only in the given particular case that both $m_1$ and $m_2$ are odd. The description of ${\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{0}$ as the algebraic variety $\mathcal{R}^{({\boldsymbol{m}})}_+$ is also usually provided in the literature, see [@Gorjanc2010; @Himstedt1888; @Loria1902]. The particular variety $\mathcal{R}^{(5,3)}$ and its two subvarieties $\mathcal{R}^{(5,3)}_{\pm}$ are illustrated in Figure \[fig:LS-2a\] (b).
3. For ${\boldsymbol{m}} = (1,1)$, the points of the rhodonea variety $\mathcal{R}^{(1,1)}$ satisfy the equation $$(x_1^2 + x_2^2)^2 = x_1^2 \quad \Leftrightarrow \quad x_1^2 + x_2^2 = \pm x_1.$$ In this case, the variety consists of two circles with diameter $1$ and radius $1/2$ centered at $\pm 1/2$. The two circles correspond to the two rhodonea curves ${\boldsymbol{\varrho}}^{(1,1)}_0$ and ${\boldsymbol{\varrho}}^{(1,1)}_1$. The variety $\mathcal{R}^{(1,1)}$ is part of the larger variety $\mathcal{R}^{(4,4)}$ illustrated in Figure \[fig:LS-2b\] (b).
4. For ${\boldsymbol{m}} = (1,2)$, the rhodonea variety $\mathcal{R}^{(1,2)}$ is determined by the equation $$(x_1^2 + x_2^2)^3 = (x_1^2-x_2^2)^2.$$ The corresponding curve ${\boldsymbol{\varrho}}^{(1,2)}_0$ gives the so called four leave rose, a curve having the form of a rose with four petals. In general, the curve ${\boldsymbol{\varrho}}^{(1,m_2)}_0$ has the shape of a rose with $2m_2$ leaves if $m_2$ is even, and $m_2$ leaves if $m_2$ is odd. The description of ${\boldsymbol{\varrho}}^{(1,m_2)}_0$ in polar coordinates is given by $r = \cos (m_2 \theta)$. For some illustrations of these roses we refer to [@Gorjanc2010].
Spectral index sets and discrete orthogonality on ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ {#1507091240}
==========================================================================================================
#### **5.1. Discrete function space on ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$**
We denote by $\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ the space of all discrete functions $f: {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}\to {{\mathbb C}}$ on the index set ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. In $\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, we consider further the family of functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\in \mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, ${\boldsymbol{\gamma}} \in {{\mathbb Z}}^2$, given by $$\label{A1508291531}
{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) = \cos(\gamma_{1} i_1 \pi/(2m_{1})) \mathrm{e}^{\imath \gamma_{2} i_2 \pi/(2m_{2}) }.$$ In the following, our objective is to derive a discrete orthogonality structure for the functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ on ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. As the functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ are a discretization of the Chebyshev-Fourier basis ${{X}_{{\boldsymbol{\gamma}}}}$ AT the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ (this will be derived in Section \[sec:interpolation\]), this discrete orthogonality is the key ingredient for the proof of the main Theorems \[201512131945\] and \[201512131946\] on spectral interpolation on the rhodonea nodes.
To introduce an inner product on the space $\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ we define for ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ the weights $$\label{1507091748}
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}= \frac{1}{4m_1m_2}\left\{ \begin{array}{ll} 1 & \text{if}\ \ {\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}, \; i_1 = 0\\
2 & \text{if}\ \ {\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}, \; 0 < i_1 \leq m_1.
\end{array} \right.$$ The corresponding discrete measure $\mathrm{w}$ on the power set of ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ is defined by $\mathrm{w}(\{{\boldsymbol{i}}\})=\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}$. Then, the inner product $$\langle f,g \rangle_{\mathrm{w}} = \int f \, \overline{g} \, \mathrm{d}\rule{1pt}{0pt}\mathrm{w}
= \sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} \mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} f({\boldsymbol{i}}) \overline{g({\boldsymbol{i}})}$$ turns $\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ into a Hilbert space. We denote the corresponding norm by $\|\cdot\|_{\mathrm{w}}$.
\[def:spectralindex\] We call ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\subset {{\mathbb Z}}^2$ a *spectral index set for ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$* if the system $\{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\ | \ {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\}$ forms an orthogonal basis of the inner product space $(\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}),\langle\,\cdot,\cdot\,\rangle_{\mathrm{w}})$. We additionally assume that the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ is a subset of $${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}= \{ {\boldsymbol{\gamma}} \in {{\mathbb Z}}^2 \ | \ 0 \leq \gamma_1 \leq 2 m_1, \ -2 m_2 < \gamma_2 \leq 2 m_2 \}.$$ and that $\gamma_1 + \gamma_2$ is even for all ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$. The last condition is referred to as parity condition.
#### **5.2. Rectangular spectral index sets**
For our purpose, the most important example of a spectral index set is the rectangular set $$\label{1508222042base}
{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}= \left\{\,{\boldsymbol{\gamma}}\in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\left| \ - m_2 < \gamma_{2} \leq m_{2}, \
\gamma_1 + \gamma_2 \ \text{is even}
\right.
\right\}.$$ The set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ contains $\# {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}= (2m_1 + 1) m_2$ elements. This corresponds exactly to the cardinality of ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. In fact, we obtain:
\[1507091911\] The set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ is a spectral index set for ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, i.e. the system $\{ {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\ | \ {\boldsymbol{\gamma}}\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\}$ is an orthogonal basis of the $(2m_1+1) m_2$ dimensional space $(\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}),\langle\,\cdot,\cdot\,\rangle_{\mathrm{w}})$. The basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ are normed by $$\label{1508221825}
\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2 =
\left\{ \begin{array}{rl} 1,\; & \text{if}\quad {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}, \ \gamma_1 \in \{0,2m_1\},\\
\frac12,\; & \text{if}\quad {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}, \ \gamma_1 \notin \{0,2m_1\}.
\end{array} \right.$$
#### **5.3. General spectral index sets**
Based on the rectangular index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, we can characterize all further spectral index sets ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ contained in Definition \[def:spectralindex\]. On ${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$, we define a flip operator ${}^*: {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\to {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$, ${\boldsymbol{\gamma}} \to
{\boldsymbol{\gamma}}^*$ by $$\label{eq:201807201118} {\boldsymbol{\gamma}}^* = (2 m_1 - \gamma_1 , \gamma_2 + 2m_2 \mod 4 m_2).$$ This flip operator combines a reflection at $\gamma_1 = m_1$ with a glide operation along the $\gamma_2$ coordinate. If $\gamma_1 +\gamma_2$ is even, then the basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ are invariant under this glide-reflection operation, i.e we have for all ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$: $$\label{eq:glidereflection}
\chi_{{\boldsymbol{\gamma}}^*}^{({\boldsymbol{m}})}({\boldsymbol{i}}) = {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) \quad \text{if $\gamma_1 + \gamma_2$ is even}.$$ The flip operator on ${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$ is an involution, i.e. ${\boldsymbol{\gamma}}^{**} = {\boldsymbol{\gamma}}$. Further, if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, then ${\boldsymbol{\gamma}}^* \in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\setminus{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ and $\gamma_1^* + \gamma_2^*$ is also an even number.
Now, for an arbitrary subset $\Omega$ of ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, we define the index set $$\label{eq:201807211209}
{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\Omega}}= \{ {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\setminus \Omega\} \ \cup \ \{{\boldsymbol{\gamma}} \in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\ | \ {\boldsymbol{\gamma}}^* \in \Omega \}.$$ By the considerations above, we have $\# {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\Omega}}= \# {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}= (2 m_1 +1 ) m_2$ and Theorem \[1507091911\] combined with the glide-reflection symmetry of the basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ implies that also ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\Omega}}$ is a spectral index set for ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$.
On the other hand, every spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\subset {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$ given in Definition \[def:spectralindex\], contains $\# {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}= (2 m_1 +1 ) m_2$ elements and the glide-reflection symmetry implies for ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ that ${\boldsymbol{\gamma}}^* \in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ and $\gamma_1^* + \gamma_2^*$ is even. By setting $\Omega = {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\cap {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, the set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ is therefore identical to the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\Omega}}$ given in . We summarize these findings:
\[1507091911BB\] Every spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ in Definition \[def:spectralindex\] can be written in the form , i.e., ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}= {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\Omega}}$ with $\Omega \subseteq {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$. The basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, are normed by $$\label{1508221825BB}
\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2 =
\left\{ \begin{array}{rl} 1,\; & \text{if}\quad {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \ \gamma_1 \in \{0,2m_1\},\\
\frac12,\; & \text{if}\quad {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \ \gamma_1 \notin \{0,2m_1\}.
\end{array} \right.$$
The choice $\Omega = \emptyset$ gives exactly the rectangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$. The choice $\Omega = \{ {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\ | \ \gamma_1/m_1 + \gamma_2/m_2 > 1, \ \gamma_1/m_1 - \gamma_2/m_2 > 1 \}$ gives a triangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$ of the form $$\label{1508222042triangle}
{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}= \left\{\,{\boldsymbol{\gamma}}\in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\left| \ \begin{array}{l} \gamma_1/m_1 + |\gamma_2|/m_2 < 1 \\ \gamma_1 + \gamma_2 \ \text{is even} \end{array} \right. \right\} \cup
\left\{\,{\boldsymbol{\gamma}}\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\left| \ \gamma_1/m_1 + |\gamma_2|/m_2 = 1 \right. \right\}.$$ In the forthcoming applications we will mostly use the rectangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ or the triangular set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$. For the frequency parameter ${\boldsymbol{m}} = (5,3)$, these two spectral index sets are shown in Figure \[fig:LS-3\].
#### **5.4. Real basis systems**
For computational issues it is convenient to have also a real orthogonal basis for the space ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ at disposition. For this, we define the subset $$\label{eq:201807211209b}
{{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}= \left\{ {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\ | \ (\gamma_1,-\gamma_2) \notin {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\right\}$$ and the real valued discrete functions $$\label{1702291124}
{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}= \left\{ \begin{array}{lll}
\cos(\gamma_{1} i_1 \pi/(2m_{1})) \cos(\gamma_{2} i_2 \pi/(2m_{2})),
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 \geq 0$}, \; & (i)\\
\cos(\gamma_{1} i_1 \pi/(2m_{1})) \sin(\gamma_{2} i_2 \pi/(2m_{2})),
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 < 0$}, & (ii)\\
\cos(\gamma_{1} i_1 \pi/(2m_{1})) \cos(\gamma_{2} i_2 \pi/(2m_{2})),
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 \leq m_1$}, & (iii)\\
\cos(\gamma_{1} i_1 \pi/(2m_{1})) \sin(\gamma_{2} i_2 \pi/(2m_{2})),
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 > m_1$}. & (iv)
\end{array} \right.$$
\[1507091912\] Let ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ be a spectral index set according to Definition \[def:spectralindex\]. Then, the functions ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}}\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, form a real orthogonal basis of the inner product space $({\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}),\langle\,\cdot,\cdot\,\rangle_{\mathrm{w}})$. The norms of the basis functions ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}$ are given as $$\label{1508221826}
\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2 =
\left\{ \begin{array}{rl} 1\; & \text{if}\quad {\boldsymbol{\gamma}} \in \{(0,0), (2 m_{1},0), (0,2m_2) \}, \\[2mm]
\frac12\; & \text{if} \quad \begin{array}{l} {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus \{(0,0), (2 m_{1},0), (0,2m_2) \} \ \text{and} \\
\text{$\gamma_1 = 0$ or $\gamma_1 = 2 m_1$ or $\gamma_2 = 0$ or $\gamma_2 = 2m_2$}, \end{array} \\[4mm]
\frac12\; & \text{if}\quad {\boldsymbol{\gamma}} \in \{(m_1,m_2), (m_1,-m_2) \}, \\[2mm]
\frac14\; & \text{for all other ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$}.
\end{array} \right.$$
1. For the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, we have ${{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}_{\square} = \{ {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\ | \ \gamma_2 = m_2\}$.
2. For the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$, the set ${{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}_{\triangle}$ is empty if $m_1 + m_2$ is odd. If $m_1+m_2$ is even, then ${{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}_{\triangle} = \{(m_1,m_2)\}$.
Spectral interpolation on the rhodonea nodes {#sec:interpolation}
============================================
#### **6.1. Formulation of the interpolation problem**
We formulate now the spectral interpolation problem on the disk ${\mathbb{D}}$ based on the rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ as interpolation nodes. As coordinate system, we will use polar coordinates $(r,\theta)$ in the domain $[0,1] \times (-\pi,\pi]$. According to the definitions in and , the rhodonea nodes in polar coordinates are given as $(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2})$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$.
We generate the interpolation spaces using the *Chebyshev-Fourier basis functions* ${{X}_{{\boldsymbol{\gamma}}}}$. For ${\boldsymbol{\gamma}} \in {{\mathbb N}}_0 \times {{\mathbb Z}}$ and $(r,\theta) \in [0,1] \times (-\pi,\pi]$, these basis functions ${{X}_{{\boldsymbol{\gamma}}}}(r,\theta)$ are given by $$\label{A1508291123}
{{X}_{{\boldsymbol{\gamma}}}}(r,\theta) = T_{\gamma_1}(r) \mathrm{e}^{\imath \gamma_{2} \theta },$$ where $T_{\gamma_1}(r) = \cos (\gamma_1 \arccos r)$ is the Chebyshev polynomial of first kind of degree $\gamma_1 \in {{\mathbb N}}_0$. The space of all linear combinations of the functions ${{X}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\mathbb N}}_0 \times {{\mathbb Z}}$, is denoted by $\Pi$.
Our aim is to solve the following interpolation problem: for given data values $f \in {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ we want to obtain a spectral interpolant $ P^{({\boldsymbol{m}})}_f \in \Pi$ such that $$\label{1508220011}
P^{({\boldsymbol{m}})}_f (r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}) = f({{\boldsymbol{i}}}) \quad \text{for all} \ {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}.$$
#### **6.2. Unisolvence of spectral interpolation**
To obtain uniqueness in , we have to specify a proper subspace of $\Pi$. For this, we will use the key relation $$\label{1508201411}
{{X}_{{\boldsymbol{\gamma}}}}(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}) = {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}), \qquad {\boldsymbol{\gamma}}\in{{\mathbb N}}_0 \times {{\mathbb Z}}, \quad {\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},$$ between the Chebyshev-Fourier basis ${{X}_{{\boldsymbol{\gamma}}}}$ and the discrete basis ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ for the space ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. From the previous section we know that ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, is an orthogonal basis of ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ if ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ is a spectral index set for ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. This turns spectral index sets ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ also to ideal index sets for the construction of the interpolation spaces. We define: $${\Pi^{({\boldsymbol{m}})}}= \mathrm{span} \left\{\, {{X}_{{\boldsymbol{\gamma}}}}\,\left|\, {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\right. \right\}.$$ By Definition \[def:spectralindex\] of the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ the sum $\gamma_1 + \gamma_2$ is even. This parity condition ensures that the functions $P \in {\Pi^{({\boldsymbol{m}})}}$ can be extended naturally onto $[-1,1] \times (-\pi,\pi]$ such that the continuous glide-reflection symmetry $P(-r,\theta) = P(r,\theta+\pi)$ is satisfied. For this reason, the basis $X_{{\boldsymbol{\gamma}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, is also referred to as *parity-modified Chebyshev-Fourier basis* of the space ${\Pi^{({\boldsymbol{m}})}}$. Two such basis systems are illustrated in Figure \[fig:LS-3B\].
Data values $f \in {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ obtained by sampling a continuous function on the disk are constant at all coordinates ${\boldsymbol{i}}$ with $i_1 = m_1$ representing the center of the unit disk. To take this fact into account, we additionally define the subspaces $${\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}) = \left\{f \in \mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}): \, f({\boldsymbol{i}}) = f_{\mathrm{C}} \ \text{is constant at $i_1 = m_1$} \right\}$$ and $${\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}= \left\{P \in {\Pi^{({\boldsymbol{m}})}}\ | \ P(\theta^{(m_1)}_{i_1}, {\varphi}^{(m_2)}_{i_2})) \equiv P(\theta^{(m_1)}_{j_1}, {\varphi}^{(m_2)}_{j_2}))
\quad \text{if} \; i_1 = j_1 = m_1 \ \right\}. \label{eq:186181745}$$ We have ${\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}) \subset {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, ${\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}\subset {\Pi^{({\boldsymbol{m}})}}$ and $\dim {\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}) = \dim {\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}= \# {{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$. The space ${\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ can naturally be used to describe all given data functions on the rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$. On the other side, the space ${\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}\subset {\Pi^{({\boldsymbol{m}})}}$ contains exactly all $P \in {\Pi^{({\boldsymbol{m}})}}$ such that the discrete data set $p({\boldsymbol{i}}) = P(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}))$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, is contained in ${\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. Although $P \in {\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}$ satisfies this discrete consistency condition at the center of ${\mathbb{D}}$, the function $P \in {\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}$ is in general not constant on the entire line $r = 0$ describing the center. We will show in Section \[sec:convergence\] that for the particular interpolation space ${\Pi^{({\boldsymbol{m}})}_{\square,\mathrm{D}}}$ based on the index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ we can guarantee the continuity of $P$ at the center.
In the literature on spectral spectral methods a tensor-product grid in polar coordinates is usually used in place of the rhodonea points to build up collocation schemes on the unit disk. The underlying interpolation spaces spanned by a parity-modified Chebyshev-Fourier basis in a rectangular spectral set are similar to the spaces ${\Pi^{({\boldsymbol{m}})}_{\square}}$, see [@BoydYu2011; @Fornberg1995; @Shen2011; @Trefethen2000; @TownsendWilberWright2017].
In our considerations, the spaces ${\Pi^{({\boldsymbol{m}})}_{\square}}$ play a dominant role as well. Nevertheless, also ${\Pi^{({\boldsymbol{m}})}_{\triangle}}$ has some interesting resemblances to spaces in other works. For $m \in {{\mathbb N}}$ odd and ${\boldsymbol{m}} = (m-1,m)$, the interpolation space ${\Pi^{({\boldsymbol{m}})}_{\triangle}}$ is spanned by all parity-modified Chebyshev-Fourier basis functions ${{X}_{{\boldsymbol{\gamma}}}}$ with total degree $|\gamma_1| + |\gamma_2| \leq m-1$. Because of this, the rhodonea nodes ${\boldsymbol{\operatorname{RD}}}^{(m-1,m)}$ can be regarded as polar version of the Padua points studied in [@BosDeMarchiVianelloXu2006; @CaliariDeMarchiVianello2005]. If ${\boldsymbol{m}} = (m,m)$, the points ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ and the interpolation space ${\Pi^{({\boldsymbol{m}})}_{\triangle}}$ provide a setup that is very similar to the one provided by the Morrow-Patterson-Xu points [@Xu1996]. For general ${\boldsymbol{m}} \in {{\mathbb N}}^2$, the results obtained for the nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ can be regarded as a polar version of the theory on polynomial interpolation on Lissajous nodes [@DenckerErb2017a; @DenckerErb2015a; @Erb2015; @ErbKaethnerAhlborgBuzug2015] and on spherical Lissajous nodes [@ErbSphere2017].
Our main result on spectral interpolation on the disk reads as follows:
\[201512131945\] Let $f\in {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ and ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ be a spectral index set according to Definition \[def:spectralindex\]. Then, the interpolation problem has a unique solution in the space ${\Pi^{({\boldsymbol{m}})}}$ given by $$\label{201513121708}
P^{({\boldsymbol{m}})}_{f}(r,\theta) = \sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} f({{\boldsymbol{i}}}) \, L^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}(r,\theta)$$ with the Lagrange functions $$\label{1508220009}
L^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}(r,\theta) = \mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{\overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}})}}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}
{{X}_{{\boldsymbol{\gamma}}}}(r,\theta), \qquad {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},$$ forming a basis of the vector space ${\Pi^{({\boldsymbol{m}})}}$. For $f\in {\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, the interpolant $P^{({\boldsymbol{m}})}_{f} \in {\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}$ is of the form $$\label{201513121700}
P^{({\boldsymbol{m}})}_{f}(r,\theta) = \underset{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}: i_1 \neq m_1}{\sum} f({{\boldsymbol{i}}}) \, L^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}(r,\theta) + f_{\mathrm{C}} \underset{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}: i_1 = m_1}{\sum} L^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}(r,\theta).$$
#### **6.3. Real valued interpolation spaces**
In order to establish a similar interpolation result for real vector spaces, we define for ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ the real basis functions $${X_{\mathcal{R},{\boldsymbol{\gamma}}}}(r,\theta) = \left\{ \begin{array}{ll} T_{\gamma_{1}} (r) \cos (\gamma_{2} \theta ) & \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 \geq 0$}, \\
T_{\gamma_{1}} (r) \sin (\gamma_{2} \theta ) & \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 < 0$}, \\
T_{\gamma_{1}} (r) \cos (\gamma_{2} \theta ) & \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 \leq m_1$}, \\
T_{\gamma_{1}} (r) \sin (\gamma_{2} \theta ) & \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 > m_1$}. \\
\end{array} \right.$$
Evaluating the functions ${X_{\mathcal{R},{\boldsymbol{\gamma}}}}$ at the polar nodes $(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2})$, we obtain precisely the discrete basis functions of the space ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ given in , i.e. $$\label{1508201413}
{X_{\mathcal{R},{\boldsymbol{\gamma}}}}(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}) = {\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}({\boldsymbol{i}}), \qquad \text{${\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$}.$$ In the same way as for the basis function ${{X}_{{\boldsymbol{\gamma}}}}$ and the space ${\Pi^{({\boldsymbol{m}})}}$, we can now introduce the real valued interpolation space $${\Pi^{({\boldsymbol{m}})}_{\mathcal{R}}}= \mathrm{span} \left\{\, {X_{\mathcal{R},{\boldsymbol{\gamma}}}}\,\left|\, {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\right. \right\}$$ and obtain in analogy to Theorem \[201512131945\] the following result:
\[201512131946\] Let $f\in {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ and ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ be a spectral index set according to Definition \[def:spectralindex\]. Then, the interpolation problem has a unique solution in the real space ${\Pi^{({\boldsymbol{m}})}_{\mathcal{R}}}$ given by $$\label{201513121708B}
P^{({\boldsymbol{m}})}_{\mathcal{R},f}(r,\theta) = \sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} f({{\boldsymbol{i}}}) \, L^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{i}}}(r,\theta).$$ with the Lagrange functions $$\label{1508220009B}
L^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{i}}}(r,\theta) = \mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}
\sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}({\boldsymbol{i}})}{\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}
{X_{\mathcal{R},{\boldsymbol{\gamma}}}}(r,\theta), \qquad {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},$$ forming a basis of the vector space ${\Pi^{({\boldsymbol{m}})}_{\mathcal{R}}}$. If $f\in {\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, the solution of the interpolation problem can be written as $$\label{201513121700B}
P^{({\boldsymbol{m}})}_{\mathcal{R},f}(r,\theta) = \underset{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}: i_1 \neq m_1}{\sum} f({{\boldsymbol{i}}}) \, L^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{i}}}(r,\theta) + f_{\mathrm{C}} \underset{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}: i_1 = m_1}{\sum} L^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{i}}}(r,\theta).$$
In the discrete setting the basis systems ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ and ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ generate the same vector space ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. In the continuous setup this is no longer true and we generally have ${\Pi^{({\boldsymbol{m}})}}\neq {\Pi^{({\boldsymbol{m}})}_{\mathcal{R}}}$. An example in which ${\Pi^{({\boldsymbol{m}})}}$ coincides with ${\Pi^{({\boldsymbol{m}})}_{\mathcal{R}}}$ can be obtained for the triangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$. We get ${\Pi^{({\boldsymbol{m}})}_{\triangle}}= {\Pi^{({\boldsymbol{m}})}_{\triangle,\mathcal{R}}}$ in the case that $m_1 + m_2$ is odd. This follows from the fact that in this case ${{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}_{\triangle}$ is empty.
Efficient implementation of the interpolation algorithm {#sec:implementation}
=======================================================
#### **\[sec:implementation\].1. Calculation of the expansion coefficients**
An efficient way to calculate the interpolation polynomial $P^{({\boldsymbol{m}})}_{f} \in {\Pi^{({\boldsymbol{m}})}}$ from given data values $f \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ is based on the expansion $$\label{20170303146} P^{({\boldsymbol{m}})}_{f}(r,\theta) = \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} c_{{\boldsymbol{\gamma}}}(f) {{X}_{{\boldsymbol{\gamma}}}}(r,\theta).$$ Using this series expansion, $P^{({\boldsymbol{m}})}_{f}$ can be evaluated once the coefficients $c_{{\boldsymbol{\gamma}}}(f)$ are calculated. Both steps, the calculation of the coefficients and the evaluation of the sum can be implemented by applying a discrete Fourier transform. Theorem \[201512131945\] and the definition $\eqref{1508220009}$ of the Lagrange basis functions provide us with the representation $$P^{({\boldsymbol{m}})}_{f}(r,\theta) = \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{1}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2} \left(\sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}}
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} f({{\boldsymbol{i}}}) \overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}})} \right) {{X}_{{\boldsymbol{\gamma}}}}(r,\theta).$$ Since the set $\{ {{X}_{{\boldsymbol{\gamma}}}}\ | \ {\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\}$ is a basis for ${\Pi^{({\boldsymbol{m}})}}$, we have the identity $$c_{{\boldsymbol{\gamma}}}(f) = \frac{1}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2} \left(\sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}}
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} f({{\boldsymbol{i}}}) \overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}})} \right) = \frac{1}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}\,{\displaystyle}\langle\;\! f,{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\rangle_{\mathrm{w}}.$$ This formula enables us to calculate the expansion coefficients $c_{{\boldsymbol{\gamma}}}(f)$ by a double Fourier transform on the finite abelian group $( {{\mathbb Z}}/ 4 m_1 {{\mathbb Z}}) \times ( {{\mathbb Z}}/ 4 m_2 {{\mathbb Z}})$ identified with $${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}= \{ {\boldsymbol{i}} \in {{\mathbb Z}}^2 \ | \ -2m_1 < i_1 \leq 2 m_1, \ -2 m_2 < i_2 \leq 2 m_2 \}.$$ We consider ${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$ as a subset of ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ and the flip operator $\phantom{}^{*}$ introduced in on ${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$. Further, we introduce a second reflection operator on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ by setting ${\boldsymbol{i}}^\dag = (- i_1 \mod 4m_1 , i_2 )$ for ${\boldsymbol{i}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$. The two mappings allow us to extend $f \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ symmetrically to ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$. We set $$\label{eq:g} g({\boldsymbol{i}}) = \frac{1}{8 m_1m_2} \left\{ \begin{array}{rl} f({\boldsymbol{i}}), \quad & \text{if}\; {\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}, \rule[-0.65em]{0pt}{1em}\\
f({\boldsymbol{i}}^*), \quad & \text{if}\; {\boldsymbol{i}}\in{{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}, \; {\boldsymbol{i}}^*\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},
\rule[-0.65em]{0pt}{1em}\\
f({\boldsymbol{i}}^\dag), \quad & \text{if}\; {\boldsymbol{i}}\in{{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}, \; {\boldsymbol{i}}^\dag\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},
\rule[-0.65em]{0pt}{1em}\\
f({\boldsymbol{i}}^{\dag *} ), \quad & \text{if}\; {\boldsymbol{i}}\in{{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}, \; {\boldsymbol{i}}^{\dag *}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}},
\rule[-0.65em]{0pt}{1em}\\
0, \quad
& \text{otherwise}. \end{array} \right.$$ The coefficients $c_{{\boldsymbol{\gamma}}}(f)$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ can now be obtained directly from the Fourier transform $$\hat{g}({\boldsymbol{\gamma}}) =\sum_{{\boldsymbol{i}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}} g({\boldsymbol{i}}) \mathrm{e}^{-\imath \gamma_{1} i_1 \pi/m_{1} } \mathrm{e}^{-\imath \gamma_{2} i_2 \pi/m_{2} } \quad \text{of the function $g$ on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$.}$$ The relation between $\hat{g}({\boldsymbol{\gamma}})$ and the coefficients $c_{{\boldsymbol{\gamma}}}(f)$ is herein given by $$\begin{aligned}
c_{{\boldsymbol{\gamma}}}(f)&={\textstyle}\frac{1}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}\,{\displaystyle}\sum_{{\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} \mathrm{w}^{\!({\boldsymbol{m}})}_{{\boldsymbol{i}}} f({\boldsymbol{i}}) \overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}})}
= {\textstyle}\frac{1}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}\,{\displaystyle}\sum_{{\boldsymbol{i}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}} g({\boldsymbol{i}}) \mathrm{e}^{-\imath \gamma_{1} i_1 \pi/m_{1} } \mathrm{e}^{-\imath \gamma_{2} i_2 \pi/m_{2} } \notag \\
&= \left\{ \begin{array}{rl} 2 \hat{g}({\boldsymbol{\gamma}}), & \text{if}\, {\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \, \gamma_1 \notin \{0,2m_1\},
\rule[-0.65em]{0pt}{1em}\\ \hat{g}({\boldsymbol{\gamma}}), & \text{if}\, {\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \, \gamma_1 \in \{0,2m_1\}. \end{array} \right. \label{20170304}\end{aligned}$$ The entire calculation of the coefficients is summarized in Algorithm \[algorithm1\]. The main computational step in Algorithm \[algorithm1\] consists in the calculation of the Fourier transform $\hat{g}$. By using a fast Fourier algorithm this step can be executed in ${\mathcal{O}}(m_1m_2 \log (m_1m_2))$ arithmetic operations. The values of $\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2$ used in are known from .
The symmetry of the function $g$ can be seen as a combination of a reflection and a glide reflection symmetry. The reflection symmetry corresponds to the invariance of $g$ under the reflection operator ${\boldsymbol{i}}^{\dag}$ on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$, while the glide reflection symmetry is described by the operator ${\boldsymbol{i}}^{*}$ on the subset ${{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}$. In [@TownsendWilberWright2016; @TownsendWilberWright2017], this glide reflection symmetry is referred to as block-mirror centrosymmetric (BMC) structure.
[2]{}
\[algorithm1\]
**Calculate** $g \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ from $f$ by
**Calculate** Fourier transform $\hat{g}$ of $g$ on the group ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$
For ${\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, **set** $$c_{{\boldsymbol{\gamma}}}(f)
= \left\{ \begin{array}{rl} 2 \hat{g}({\boldsymbol{\gamma}}), & \text{if} \, \gamma_1 \notin \{0,2m_1\},
\rule[-0.65em]{0pt}{1em}\\ \hat{g}({\boldsymbol{\gamma}}), & \text{if}\, \gamma_1 \in \{0,2m_1\}. \end{array} \right.$$
\[algorithm2\]
**Calculate** $h \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ from $c_{{\boldsymbol{\gamma}}}(f)$ by
**Calculate** adjoint Fourier transform $\check{h}$ of $h$ on the (dual) group ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$
For ${\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, **set** $$f({\boldsymbol{i}}) = \check{h}({\boldsymbol{i}}).$$
#### **\[sec:implementation\].2. The inverse transform**
From a known set of coefficients $c_{{\boldsymbol{\gamma}}}(f)$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, we can reversely reconstruct the function values $f \in \mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. This inverse transform is also determined by a discrete Fourier transform. Combining the interpolation condition with the expansion , we get $$f({{\boldsymbol{i}}}) = P^{({\boldsymbol{m}})}_f (r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}) =
\sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} c_{{\boldsymbol{\gamma}}}(f) {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) \quad \text{for all}\quad {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}.$$ As in the previous section, we extend the coefficients $c_{{\boldsymbol{\gamma}}}(f)$ first symmetrically to the (dual) group ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ and define the function $h$ on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$ as $$\label{eq:h} h({\boldsymbol{\gamma}}) = \left\{ \begin{array}{rl} 2 c_{{\boldsymbol{\gamma}}}(f), \quad & \text{if}\; {\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \, \gamma_1 \notin \{0,2 m_1\},
\rule[-0.65em]{0pt}{1em}\\
2 c_{{\boldsymbol{\gamma}}}(f), \quad & \text{if}\; (-\gamma_1,\gamma_2)\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \, \gamma_1 \notin \{0,2 m_1\},
\rule[-0.65em]{0pt}{1em}\\
c_{{\boldsymbol{\gamma}}}(f), \quad & \text{if}\; {\boldsymbol{\gamma}}\in{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}, \, \gamma_1 \in \{0,2m_1\},
\rule[-0.65em]{0pt}{1em}\\
0, \quad
& \text{otherwise}. \end{array} \right.$$ This definition together with the definition of the discrete basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ yields $$f({{\boldsymbol{i}}}) = \sum_{{\boldsymbol{\gamma}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}} h({\boldsymbol{\gamma}}) \mathrm{e}^{\imath \gamma_{1} i_1 \pi/m_{1} } \mathrm{e}^{\imath \gamma_{2} i_2 \pi/m_{2} } = \check{h}({\boldsymbol{i}})
\quad \text{for}\; {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}.$$ Therefore, the function $f \in \mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ can be recovered from the coefficients $c_{{\boldsymbol{\gamma}}}(f)$ by computing the adjoint Fourier transform $\check{h}$ of $h$ on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$. The single steps of the calculation are summarized in Algorithm \[algorithm2\]. As for Algorithm \[algorithm1\], the entire inverse transform can be computed in ${\mathcal{O}}(m_1m_2 \log (m_1m_2))$ arithmetic operations.
#### **\[sec:implementation\].3. Calculation of real expansion coefficients**
When working with the real basis ${X_{\mathcal{R},{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, the expansion coefficients $c_{\mathcal{R},{\boldsymbol{\gamma}}}(f)$ of the interpolant $P^{({\boldsymbol{m}})}_{\mathcal{R},f}$ can be computed in a similar way. Using the formula $$c_{\mathcal{R},{\boldsymbol{\gamma}}}(f)={\textstyle}\frac{1}{\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}\,{\displaystyle}\langle\;\! f,{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\rangle_{\mathrm{w}},$$ the coefficients can be rewritten as $$c_{\mathcal{R},{\boldsymbol{\gamma}}}(f) = \frac{1}{\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2} \left\{ \! \begin{array}{ll}
{\operatorname{Re}}\hat{g}({\boldsymbol{\gamma}})
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 \geq 0$}, \\
- {\operatorname{Im}}\hat{g}({\boldsymbol{\gamma}})
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}\setminus {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_2 < 0$}, \\
{\operatorname{Re}}\hat{g}({\boldsymbol{\gamma}})
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 \leq m_1/2$},\\
- {\operatorname{Im}}\hat{g}({\boldsymbol{\gamma}})
& \text{if ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Upsilon}}^{({\boldsymbol{m}})}}$, $\gamma_1 > m_1/2$}.
\end{array} \right.$$ This formula can be derived as in by using the real basis ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}$ in instead of the complex-valued functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$. The values $\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2$ are explicitly known from and the calculation of the Fourier transform $\hat{g}({\boldsymbol{\gamma}})$ is the same as in Section \[sec:implementation\].1.
#### **\[sec:implementation\].4. Averaged expansion coefficients**
A further option to alter the structure of the interpolation spaces is to use, for some of the indices ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$ , the averaged basis functions $X_{\mathcal{A},{\boldsymbol{\gamma}}} = \lambda X_{{\boldsymbol{\gamma}}} + (1 - \lambda) X_{{\boldsymbol{\gamma^*}}}$ with $\lambda \in {{\mathbb R}}$ instead of the standard basis functions $X_{{\boldsymbol{\gamma}}}$. Since the discrete basis functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, are invariant under the flip operator $\phantom{}^{*}$, we have $$X_{{\boldsymbol{\gamma}}}(r_{i_1}^{(m_1)},\theta_{i_2}^{(m_2)}) = {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) = X_{\mathcal{A},{\boldsymbol{\gamma}}}(r_{i_1}^{(m_1)},\theta_{i_2}^{(m_2)}).$$ Therefore, while both basis systems define different interpolation spaces, they both lead to the same interpolation problem on the rhodonea nodes and the expansion coefficients are identical. Interpolation spaces with such an averaging for some of the boundary elements of the spectral index set were originally studied in the bivariate setting for the Morrow-Patterson-Xu points in [@Harris2010; @Xu1996]. For multivariate interpolation on Lissajous-Chebyshev nodes, this averaging process is studied in more detail in [@DenckerErb2017a].
An alternative to the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ in the implementation of the interpolation scheme is to use the more symmetric set (compare the definition of ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$) $${\overline{{\boldsymbol{\Gamma}}}^{({\boldsymbol{m}})}_{\square}}= \left\{\,{\boldsymbol{\gamma}}\in {{\boldsymbol{\operatorname{K}}}^{({\boldsymbol{m}})}}\left| \ - m_2 \leq \gamma_{2} \leq m_{2}, \
\gamma_1 + \gamma_2 \ \text{is even}
\right.
\right\}$$ and to average the basis functions for the coefficients ${\boldsymbol{\gamma}}$, $|\gamma_2| = m_2$ at the upper and lower boundary of ${\overline{{\boldsymbol{\Gamma}}}^{({\boldsymbol{m}})}_{\square}}$ with $\lambda = 1/2$. In this way we get an interpolation function of the form $$P_{\mathcal{A},f}^{({\boldsymbol{m}})}(r,\theta) = \sum_{{\boldsymbol{\gamma}} \in {\overline{{\boldsymbol{\Gamma}}}^{({\boldsymbol{m}})}_{\square}}} c_{\mathcal{A},{\boldsymbol{\gamma}}}(f) X_{{\boldsymbol{\gamma}}},$$ in which the coefficients are given as $$c_{\mathcal{A},{\boldsymbol{\gamma}}}(f) = \left\{ \begin{array}{ll} c_{{\boldsymbol{\gamma}}}(f)/2 & \text{if} \ {\boldsymbol{\gamma}} \in {\overline{{\boldsymbol{\Gamma}}}^{({\boldsymbol{m}})}_{\square}}, \ |\gamma_2| = m_2, \\
c_{{\boldsymbol{\gamma}}}(f) & \text{if} \ {\boldsymbol{\gamma}} \in {\overline{{\boldsymbol{\Gamma}}}^{({\boldsymbol{m}})}_{\square}}, \ |\gamma_2| \neq m_2. \end{array}\right.$$ The coefficients $c_{{\boldsymbol{\gamma}}}(f)$ itself are calculated as in . For the real basis functions $X_{\mathcal{R},{\boldsymbol{\gamma}}}$ similar averaging strategies are of course also possible.
Properties of the spectral interpolation scheme {#sec:convergence}
===============================================
Goal of this section is to provide a convergence analysis of the presented interpolation scheme and to answer questions typically considered in approximation theory and numerical analysis. This includes quadrature possibilities, the behavior of the interpolating functions at the center of the unit disk, as well as the numerical condition and the convergence of the interpolation scheme if the number of nodes gets large. In general, these properties depend on the geometric form of the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$. To obtain more concrete results, we will restrict our studies mainly to the two particular spectral sets ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ and ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$.
The interpolating functions $P^{({\boldsymbol{m}})}_{f}$ considered in this section are determined by data values $f$ that are obtained from the samples of a continuous function on the disk. For a continuous function ${\mathscr{f}}(r,\theta)$ on ${\mathbb{D}}$ in polar coordinates, we have the relation $$\label{eq:186181822}
f({\boldsymbol{i}}) = {\mathscr{f}}(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2}) \quad \text{for}\quad {\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}.$$ In particular, $f \in {\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ and Theorem \[201512131945\] ensures that we obtain a unique interpolant $P^{({\boldsymbol{m}})}_{f}$ in ${\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}\subset {\Pi^{({\boldsymbol{m}})}}$ that interpolates the function ${\mathscr{f}}$ at the rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$.
#### **\[sec:convergence\].1. Continuity at the center**
When using polar coordinates $(r,\theta) \in [0,1] \times [-\pi,\pi]$ to describe a continuous function ${\mathscr{f}}$ on ${\mathbb{D}}$ we have to add the usual topological identifications for the polar coordinates. We can describe the space of all continuous functions on ${\mathbb{D}}$ by $$C({\mathbb{D}}) = \left\{{\mathscr{f}}\in C([0,1]\times[-\pi,\pi]) \ \left| \ \begin{array}{lll} \mathrm{(i)} & {\mathscr{f}}(r,-\pi) = {\mathscr{f}}(r,\pi), & 0 \leq r \leq 1, \\
\mathrm{(ii)} & {\mathscr{f}}(0,\theta_1) = {\mathscr{f}}(0,\theta_2), & -\pi \leq \theta_1,\theta_2 \leq \pi. \end{array}
\right. \right\}$$ Not all basis functions ${{X}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, are contained in $C({\mathbb{D}})$. While ${{X}_{{\boldsymbol{\gamma}}}}\in C([0,1]\times[-\pi,\pi])$ and the periodicity $\mathrm{(i)}$ are always satisfied, the continuity $\mathrm{(ii)}$ at the center holds only true if $\gamma_1$ and $\gamma_2$ are odd. Therefore, also for the interpolant $P^{({\boldsymbol{m}})}_{f}$ we can in general not expect that the continuity $\mathrm{(ii)}$ is satisfied. However, if the rectangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ is used for the interpolation space, we can guarantee that $P^{({\boldsymbol{m}})}_{f}$ is continuous also at the center.
\[th:201807231722\] Let ${\mathscr{f}}\in C({\mathbb{D}})$ and $P^{({\boldsymbol{m}})}_{f}$ be the unique interpolant in the space ${\Pi^{({\boldsymbol{m}})}_{\square}}$ according to Theorem \[201512131945\]. Then, $P^{({\boldsymbol{m}})}_{f} \in C({\mathbb{D}})$.
\[rem:AB1\] Note that Theorem \[th:201807231722\] does in general not hold true if ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ is replaced by a different spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$. In general, the trigonometric polynomial $P^{({\boldsymbol{m}})}_{f}(0,\theta)$ (see also in the proof of Theorem \[th:201807231722\]) is contained in a space of dimension larger than $m_2$ and the $m_2$ given boundary conditions can not guarantee that $P^{({\boldsymbol{m}})}_{f}(0,\theta)$ is constant. Also for real interpolation spaces ${\Pi^{({\boldsymbol{m}})}_{\square,\mathcal{R}}}$ a careful view at the boundary conditions is necessary. For the interpolation space ${\Pi^{({\boldsymbol{m}})}_{\square,\mathcal{R}}}$, the statement of Theorem \[th:201807231722\] holds only true if $m_2$ is odd.
Although Theorem \[th:201807231722\] establishes that, at least in the case of the rectangular set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, the interpolant $P^{({\boldsymbol{m}})}_{f}$ is continuous, we can not expect higher order smoothness of $P^{({\boldsymbol{m}})}_{f}$ at the center of ${\mathbb{D}}$. In Section \[sec:convergence\].3. we will see that these singularities of $P^{({\boldsymbol{m}})}_{f}$ at the center have no influence on the global convergence of the interpolation scheme. If the function ${\mathscr{f}}$ is sufficiently smooth, the derivatives of $P^{({\boldsymbol{m}})}_{f}$ will approximately satisfy the continuity condition $\mathrm{(ii)}$ as soon as the node set ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$ gets large. In [@Boyd1978], such a property is referred to as natural boundary condition. For a lot of applications it is sufficient if such a natural boundary condition is satisfied. A deeper discussion about the behavior of spectral methods at coordinate singularities can be found in [@Boyd2000; @HMS2002; @Shen2011].
#### **\[sec:convergence\].2. Numerical condition of the interpolation scheme**
The Lebesgue constant of the interpolation problem is defined as the operator norm $$\Lambda^{({\boldsymbol{m}})}_{\Omega} = \sup_{\|{\mathscr{f}}\|_{\infty} \leq 1} \|P^{({\boldsymbol{m}})}_{f}\|_{\infty}, \quad \text{with} \;\| {\mathscr{f}}\|_{\infty} = \sup_{(r,\theta)} |{\mathscr{f}}(r,\theta)|.$$ In numerical analysis, this constant is interpreted as the absolute condition number of the interpolation problem . It gives an upper bound on how a small error in the function ${\mathscr{f}}$ affects the corresponding interpolant $P^{({\boldsymbol{m}})}_{f}$ in the uniform norm. Beside the distribution of nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$, the Lebesgue constant depends on the geometric structure of the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$. For the rectangular spectral set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, we can guarantee that this Lebesgue constant grows only slowly if $m_1$ and $m_2$ get large.
\[thm:lebesgueconstant\] For the interpolation space ${\Pi^{({\boldsymbol{m}})}_{\square}}$, the Lebesgue constant $\Lambda^{({\boldsymbol{m}})}$ is bounded by $$\Lambda^{({\boldsymbol{m}})}_{\square} \leq C_{\square} \ln (m_1+1) \ln (m_2+1)$$ with a constant $C_{\square}$ independent of ${\boldsymbol{m}}$.
A similar logarithmic estimate holds true for the Lebesgue constant $\Gamma^{({\boldsymbol{m}})}_{\triangle}$ when the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ is replaced by the triangular set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$. In this case, techniques developed in [@DEKL2017 Section 2] can be used to obtain the estimates of the resulting double integrals. For a similar setting on the unit sphere, the respective proof can be found in [@ErbSphere2017].
#### **\[sec:convergence\].3. Convergence of the interpolation scheme**
Once an estimate for the Lebesgue constant is known, the convergence of $P^{({\boldsymbol{m}})}_{f}$ towards ${\mathscr{f}}\in C({\mathbb{D}})$ can be established easily if the underlying function ${\mathscr{f}}$ is smooth. Further, we obtain better rates of convergence the smoother the function ${\mathscr{f}}$ is. This is a general principle for spectral methods in a variety of settings [@Boyd2000; @Trefethen2000]. For multivariate polynomial interpolation in the hypercube $[-1,1]^{\mathsf{d}}$ similar results for Lissajous sampling nodes can be found in [@DEKL2017; @Erb2015]. For spherical Lissajous nodes a respective derivation is given in [@ErbSphere2017]. Similar error estimates for a tensor product spectral collocation scheme on the sphere can also be found in [@Ganesh1998].
We consider $P^{({\boldsymbol{m}})}_{f}$ in the interpolation space ${\Pi^{({\boldsymbol{m}})}_{\square}}$. If $P^*$ denotes the best possible approximation of ${\mathscr{f}}$ in ${\Pi^{({\boldsymbol{m}})}_{\square}}$ the uniform error $\|{\mathscr{f}}- P^{({\boldsymbol{m}})}_{f} \|_\infty$ can be bounded by $$\begin{aligned}
\|{\mathscr{f}}- P^{({\boldsymbol{m}})}_{f} \|_\infty &\leq \|{\mathscr{f}}- P^* \|_\infty + \| P^* - P^{({\boldsymbol{m}})}_{f} \|_\infty \\ & \leq (\Lambda^{({\boldsymbol{m}})}_{\square} + 1) \|f - P^* \|_\infty
= (C_{\square}+1) \ln (m_1+1) \ln (m_2+1) \|{\mathscr{f}}- P^* \|_\infty.\end{aligned}$$ In the second estimate, we used the fact that the interpolation operator ${\mathscr{f}}\to P^{({\boldsymbol{m}})}_{f}$ reproduces $P^* \in {\Pi^{({\boldsymbol{m}})}_{\square}}$ together with the estimate of the Lebesgue constant in Theorem \[thm:lebesgueconstant\]. If ${\mathscr{f}}$ is $s$ times continuously differentiable in ${\mathbb{D}}$, we can estimate the best error $\|{\mathscr{f}}- P^* \|_\infty$ with help of a multivariate Jackson inequality for trigonometric functions (see [@Timan1960 Section 5.3]). In this way, we get the error bound $$\label{eq:1806211617} \|{\mathscr{f}}- P^{({\boldsymbol{m}})}_{f}\|_\infty \leq C_{{\mathscr{f}},s} \ln(m_1+1) \ln(m_2+1) \left( \frac{1}{m_1^s} + \frac{1}{m_2^s}\right).$$ The constant $C_{{\mathscr{f}},s}$ is independent of ${\boldsymbol{m}}$. Thus, if ${\mathscr{f}}$ is smooth, this estimate guarantees a fast uniform convergence of the interpolant towards ${\mathscr{f}}$ if $m_1$ and $m_2$ get large.
#### **\[sec:convergence\].4. Quadrature formula on the rhodonea nodes** {#sec:cc}
In order to formulate a Clenshaw-Curtis quadrature rule for the rhodonea points ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$, only the expansion and the explicit integration of the basis functions ${{X}_{{\boldsymbol{\gamma}}}}\in {\Pi^{({\boldsymbol{m}})}}$ over the disk ${\mathbb{D}}$ are necessary. In polar coordinates the area element on ${\mathbb{D}}$ is given by $r \, \mathrm{d} r \mathrm{d} \theta$. The tensor product structure of the Chebyshev-Fourier basis functions ${{X}_{{\boldsymbol{\gamma}}}}$ then yields the formula $$\frac{1}{\pi} \int_0^{2\pi} \int_0^{1} {{X}_{{\boldsymbol{\gamma}}}}(r,\theta) r \, \mathrm{d} r \mathrm{d} \theta = \left\{ \begin{array}{ll}
\frac{1}{1-\gamma_1^2/4} & \text{if $\gamma_2 = 0$ and $\gamma_1 \in 4 {{\mathbb Z}}$,} \\ 0 & \text{otherwise}.\end{array} \right.$$ For this formula, we used the identities $\int_{0}^{2\pi} e^{\imath \gamma_2 \theta} \mathrm{d} \theta = 2 \pi \delta_{\gamma_2,0}$ and $\int_0^{1} T_{\gamma_1} (r) r \mathrm{d} r = \frac12 \frac{1}{1-\gamma_1^2/4}$ if $\gamma_1$ is an element of $4 {{\mathbb Z}}$ and zero otherwise. With the expansion of $P^{({\boldsymbol{m}})}_{f}$ in the space ${\Pi^{({\boldsymbol{m}})}_{\square}}$, we obtain the Clenshaw-Curtis quadrature formula $$Q(f) = \frac{1}{\pi} \int_0^{2\pi} \int_0^{1} P^{({\boldsymbol{m}})}_{f}(r,\theta) r \, \mathrm{d} r \mathrm{d} \theta = \sum_{k = 0}^{\lfloor
m_1 /2 \rfloor} \frac{c_{(4k,0)}(f)}{1 - 4 k^2}.$$ The coefficients $c_{(4k,0)}(f)$ on the right hand side depend only on the data values $f({\boldsymbol{i}})$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, and, therefore by , on the function samples given at the rhodonea nodes ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$. The quadrature rule $Q(f)$ is exact for all functions in ${\Pi^{({\boldsymbol{m}})}_{\square}}$. Since $c_{(k,0)}(f) = c_{\mathcal{R},(k,0)}(f)$, the same formula holds also true using $P^{({\boldsymbol{m}})}_{\mathcal{R},f}$ as an interpolation function. The quadrature formula $Q(f)$ remains also the same if we use the triangular spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$ instead of ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$.
#### **\[sec:convergence\].5. A numerical example**
As a final numerical experiment we test the developed interpolation scheme and the Clenshaw-Curtis quadrature formula for the function $$\label{eq:testfunction}
f({\boldsymbol{x}}) = e^{- 2 ((1.6 x_1-0.1)^2+(2.4 x_2-0.2)^2)} \cos((4x_1-0.25)^2+(6x_2-0.5)^2).$$ The results of this test for different frequency parameters ${\boldsymbol{m}} = (m,m+1)$, $m \in {{\mathbb N}}$, are illustrated in Figure \[fig:LS-4\]. We observe a fast convergence of the interpolant $P_{\mathcal{R},f}^{(m,m+1)}$ and the quadrature value $Q(f)$ towards $f$ and $I(f)$, respectively, as the parameter $m \in {{\mathbb N}}$ gets large. This fast spectral convergence is not surprising since $f$ is analytic. The error estimate provides a convergence rate of the interpolation scheme faster than any polynomial.
We also compare the interpolation scheme for the two spectral index sets ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ and ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\triangle}}$. For smaller values of $m$ the discontinuity of the interpolant in the space ${\Pi^{({\boldsymbol{m}})}_{\triangle,\mathcal{R}}}$ is clearly visible at the center $(0,0)$ of ${\mathbb{D}}$ while, according to Theorem \[th:201807231722\] and Remark \[rem:AB1\], the considered interpolants $P^{({\boldsymbol{m}})}_{\mathcal{R},f}$ have no discontinuities in ${\Pi^{({\boldsymbol{m}})}_{\square,\mathcal{R}}}$. For increasing values of $m$ the differences between the interpolant $P^{({\boldsymbol{m}})}_{\mathcal{R},f}$ in ${\Pi^{({\boldsymbol{m}})}_{\triangle,\mathcal{R}}}$ and ${\Pi^{({\boldsymbol{m}})}_{\square,\mathcal{R}}}$ almost vanish. Since the Clenshaw-Curtis quadrature formula $Q(f)$ is the same for the interpolation spaces ${\Pi^{({\boldsymbol{m}})}_{\square,\mathcal{R}}}$ and ${\Pi^{({\boldsymbol{m}})}_{\triangle,\mathcal{R}}}$ we observe no differences in the evaluations of $Q(f)$ for the two spectral index sets. A [Matlab]{} code for this numerical example with an implementation of the spectral interpolation scheme on the rhodonea nodes can be found at [*https://github.com/WolfgangErb/RDisk*]{}.
Proofs {#sec:proof}
======
#### **9.1. Proof of Proposition \[prop-11\]**
[Proposition \[prop-11\]]{} For $s,t \in {{\mathbb R}}$, we write $s\eqsim t$ if $s$ and $t$ satisfy the equivalence relation $t-s\in 2\pi\mathbb{Z}$. In a first step, we determine all points $t \in [0,2\pi)$ so that ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)=(0,0)$, i.e., ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)$ is the center of the unit disk. By the definition of the rhodonea curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$, we have ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)=(0,0)$ if and only if $$\label{170210-1}
m_2 t \eqsim \pm \pi/2,$$ i.e., if and only if $t = t^{({\boldsymbol{m}})}_{l}$ for some $l\in \{m_1, 3m_1, \ldots ,(4m_2-1)m_1\}$. This provides the first statements (i) and (i)’ of Proposition \[prop-11\].
In a second step, we consider now for fixed $t \in [0,2\pi)$ the case ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t) \neq (0,0)$. By the definition of ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$, we get $s\in \mathcal{S}(t)$ if and only if $$\cos(m_2 s) = v \cos(m_2 t) \quad \text{and} \quad \left( \begin{array}{l} \cos(m_1 s - \alpha \pi) \\ \sin(m_1 s - \alpha \pi) \end{array} \right) = v
\left( \begin{array}{l} \cos(m_1 t - \alpha \pi) \\ \sin(m_1 t - \alpha \pi) \end{array} \right),$$ for some $v \in \{-1,1\}$. In the left equation, we have equality exactly if $m_2 s \eqsim u m_2 t + \frac{1-v}{2} \pi$. In the right equation, equality is obtained if $m_1 s \eqsim m_1 t + \frac{1-v}{2} \pi$ holds true. Combining these observations, we get $s\in \mathcal{S}(t)$ if and only if $$\label{170210-2}
m_2( s - u t) + \frac{1-v}{2} \pi \eqsim 0 \quad \text{and} \quad m_1 ( s - t) + \frac{1-v}{2} \pi \eqsim 0 \qquad \text{for some $u,v\in \{-1,1\}$}.$$ We characterize now all $s \in [0,2\pi)$ that satisfy the conditions in . As $m_1$ and $m_2$ are relatively prime, Bézout’s lemma provides two integers $a, b\in\mathbb{Z}$ such that $a m_1 + b m_2 = 1$. Multiplying the first and the second identity in with $b$ and $a$, respectively, and adding them up, we obtain $$\label{170210}
s \eqsim t - b m_2 (1-u) t - (a+b) \frac{1-v}{2} \pi$$ as a description for all possible solutions $s$ of . In particular, this implies that $s \eqsim t$ or $s \eqsim t + \pi$ are the only possible solutions for $u = 1$. If $u = -1$, we multiply the first and the second identity in with $m_1$ and $m_2$, respectively, and obtain $$\begin{aligned}
\notag 2 m_1 m_2 s\eqsim m_1 m_2 (s + t) + m_2 m_1 ( s - t) \eqsim (m_1 + m_2) \frac{1-v}{2} \pi, \\
2 m_1 m_2 t \eqsim m_1 m_2 (s + t) - m_2 m_1 ( s - t) \eqsim (m_1 - m_2) \frac{1-v}{2} \pi. \label{eq:00101} \end{aligned}$$ Therefore, for $u = -1$ we can conclude that $s = t_{l'}^{({\boldsymbol{m}})}$ and $t = t_{l}^{({\boldsymbol{m}})}$ for some $l,l' \in {{\mathbb Z}}$.\
Based on these deductions we can now derive the remaining properties. We distinguish the two cases $m_1 + m_2$ odd and $m_1 + m_2$ even:
If we suppose that $m_1 + m_2$ is odd and ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t) \neq (0,0)$, $s \eqsim t + \pi$ can not satisfy both identities in , and is therefore not a solution of . Hence, if $t \neq t_{l}^{({\boldsymbol{m}})}$ for some $l \in \{0, \ldots, 4 m_1 m_2 -1\}$, then $t$ is the only element of $[0,2 \pi)$ in $\mathcal{S}^{({\boldsymbol{m}})}(t)$ (corresponding to the sole solution of with the values $u = 1$ and $v = 1$) and the largest part of statement (iii) is proven.\
If $t \in [0, 2\pi)$ and $t = t_{l}^{({\boldsymbol{m}})}$ for some $l\in \{0,\ldots,4m_1m_2-1\}$, then further solutions of for the value $u = -1$ are possible. The corresponding possibilities given by are $s \eqsim t - 2 b m_2 t$ or $s' \eqsim s + \pi$. Since we are in the case $m_1+m_2$ odd, only one of the two solutions $s$ and $s'$ is possible. Also, $s$ or $s'$ do not depend on the particular value of the integer $b$ from Bézout’s lemma (since $b$ is uniquely determined modulo $m_1$). The identity therefore gives in this case exactly one solution of for $u = -1$ (we denote this solution as $s \in [0,2\pi)$). If $t = t_{l}^{({\boldsymbol{m}})}$ with $l \equiv 0 \mod 2 m_1$ we obtain $s \eqsim t$ and $s \not\eqsim t$ if $l \not \equiv 0 \mod m_1$. This yields the remaining assertions (ii) and (iii) of the proposition. Note that the case $l \equiv m_1 \mod 2 m_1$ is already treated in (i).
Finally, we shortly discuss the case when both integers $m_1$ and $m_2$ are odd, or in other words, when $m_1 + m_2$ is even. The statements (ii)’ and (iii)’ can be deduced in a similar way as before with one cardinal difference: in this case the curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}(t)$ is traversed twice as $t$ varies from $0$ to $2\pi$. In , we consequently see that if $s \in \mathcal{S}^{({\boldsymbol{m}})}(t)$ then also $s + \pi \in \mathcal{S}^{({\boldsymbol{m}})}(t)$. This yields additional solutions which double the value of $\# \mathcal{S}^{({\boldsymbol{m}})}(t)$ in (ii)’ and (iii)’. Further, from we can deduce that the double points of the curve are given at the positions $t = t_{2l}^{({\boldsymbol{m}})}$ for some $l \in {{\mathbb Z}}$.
#### **9.2. Proof of Theorem \[cor-111\]**
To prove Theorem \[cor-111\], it is necessary to establish a relation between the nodal index sets ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and the sampling points along the rhodonea curves ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$. This relation can be extracted from the following auxiliary result:
\[1509221521\] Let ${\boldsymbol{m}} \in {{\mathbb N}}^2$ and $g = \gcd({\boldsymbol{m}})$ For all $l \in \{0,\ldots, 4 m_{1} m_{2}/g -1\}$ and $\rho \in \{0,\ldots,2 g-1\}$, there exists an ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and $u,v \in\{-1,1\}$ such that $$\begin{aligned}
i_{1} &\equiv u (v l + (1-v) m_1) \mod 4 m_1, \label{1509221526} \\
i_{2} &\equiv l - 2\rho - (1-v) m_2 \mod 4 m_2. \label{1509221526B}\end{aligned}$$ The index ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ is uniquely determined by and and provides a well-defined surjective mapping ${\boldsymbol{i}}^{({\boldsymbol{m}})}: \{0,\ldots, 4 m_{1} m_{2}/g -1\} \times \{0,\ldots,2 g-1\} \to {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ by ${\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)={\boldsymbol{i}}$. Further, ${\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_{0}$ ( ${\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_{1}$) holds true if and only if $l$ is even ($l$ is odd).\
With the additional convention $$\label{1509221526C}
\text{$u = 1$ if $l \equiv 0 \mod 4 m_1$,\hspace{1cm} $u = -1$ if $l \equiv 2 m_1 \mod 4 m_1$,}$$ the numbers $u,v \in \{-1,1\}$ are uniquely determined by and . This gives for ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ the cardinalities $$\label{1509221526D} \#\{\,(l,\rho) \,|\,{\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)={\boldsymbol{i}}\,\}=\left\{ \begin{array}{ll} 4 \quad & \text{if \ $0 < i_1 \leq m_1$,} \\
2 \quad & \text{if \ $i_1 = 0$.} \end{array}\right.$$
For $l \in \{0,\ldots, 4 m_{1} m_{2}/g -1\}$ we can find an integer $0\leq i_{1}\leq m_{1}$ and $u, v \in \{-1,1\}$ such that $u i_1 + (1-v) m_1 \equiv l \mod 4 m_1$ holds true, i.e. that is satisfied. The number $i_1$ in this equation is uniquely determined by $l$, whereas, with the convention , the numbers $u$ and $v$ are uniquely determined by exactly if $l \not \equiv m_1 \mod 4 m_1$ and $l \not \equiv - m_1 \mod 4 m_1$. In this case, the tuple $(l,\rho)$ and the number $v$ given by determine an unique integer $-2m_2 < i_{2} \leq 2m_{2}$ such that is satisfied. In the remaining case when $l \equiv m_1 \mod 4 m_1$ or $l \equiv - m_1 \mod 4 m_1$, the tuple $(l,\rho)$ yields an unique $-2m_2 < i_{2} \leq 0$ and $v \in \{-1,1\}$ such that is satisfied. Furthermore, in this case, the so determined $v \in \{-1,1\}$ fixes also the number $u \in \{-1,1\}$ in . Since $i_{1} \equiv l\equiv i_{2} \mod 2$, we can finally state that the index ${\boldsymbol{i}}$ determined in this way from and is an element of ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. Further, looking at the definition in we also see that ${\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_{0}$ or ${\boldsymbol{i}}^{({\boldsymbol{m}})}(l,\rho)\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_{1}$ holds if and only if $l\equiv 0 \mod 2$ or $l\equiv 1 \mod 2$, respectively.
We finally prove . Let ${\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and $u,v\in \{-1,1\}$. Then for $a_{1} = u (v i_1 + (1-v) m_1)$ there is a uniquely determined $\rho\in \{0,\ldots,2 g-1\}$ such that $a_2 = i_2 + (1-v) m_2 + 2 \rho$ satisfies $a_1 \equiv a_2 \mod 4g$. The Chinese remainder theorem now yields a unique number $l\in \{0,\ldots,4m_1m_2/g-1\}$ satisfying $$a_1 \equiv l \mod 4m_1, \qquad a_2 \equiv l \mod 4m_2.$$ With the convention the numbers $a_1$ and $a_2$ are uniquely determined by ${\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and $u,v\in \{-1,1\}$. Thus, the numbers $(l,\rho)$ satisfying and are uniquely determined by ${\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and $u,v\in \{-1,1\}$. In the case $0 < i_1 = m_1$ both choices of $u$ and $v$ give distinct elements $(l,\rho)$, whereas, according to the convention , in the case $i_1 = 0$ only the parameter $v$ can be chosen freely.
[Theorem \[cor-111\]]{} The definitions and of the rhodonea curve ${{\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\alpha}}$ and the sampling points $t^{({\boldsymbol{m}})}_{l}$ give us directly the identity $${\textstyle}{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{\rho/m_2} (t^{({\boldsymbol{m}})}_{ l}) =
\left( \cos \left( \frac{ l \pi}{2 m_1} \right) \cos \left( \frac{l \pi}{2 m_2} - \frac{2 \rho}{2 m_2} \pi \right), \
\cos \left(\frac{ l \pi}{2 m_1} \right)
\sin \left(\frac{l \pi}{2 m_2} - \frac{2 \rho}{2 m_2} \pi \right) \right).$$ Now, by Proposition \[1509221521\], we can find an index ${\boldsymbol{i}}\in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$ and $u,v \in\{-1,1\}$ such that and are satisfied. This implies $$\begin{aligned}
{\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{\rho/m_2} (t^{({\boldsymbol{m}})}_{ l}) &= {\textstyle}\left( \cos \left( \frac{ v i_1 \pi}{2m_1} + \frac{1-v}{2} \pi \right) \cos \left( \frac{ i_2 \pi}{2m_2} + \frac{1-v}{2}\pi \right), \
\cos \left( \frac{ v i_1 \pi}{2m_1} + \frac{1-v}{2} \pi \right) \sin \left( \frac{ i_2 \pi}{2m_2} + \frac{1-v}{2} \pi \right) \right) \\
&= {\textstyle}\left( \cos \left( \frac{ i_1 \pi}{2m_1} \right) \cos \left( \frac{ i_2 \pi}{2m_2} \right), \ \cos \left( \frac{ i_1 \pi}{2m_1} \right)
\sin \left(\frac{ i_2 \pi}{2m_2} \pi \right) \right) = {\boldsymbol{x}}_{{\boldsymbol{i}}}^{({\boldsymbol{m}})}.\end{aligned}$$ The reverse implication is obtained by inverting these steps: for given ${\boldsymbol{x}}_{{\boldsymbol{i}}}^{({\boldsymbol{m}})}$, we can fix $u,v \in\{-1,1\}$ and Proposition \[1509221521\] yields a unique tuple $(l,\rho)$ with ${\boldsymbol{x}}_{{\boldsymbol{i}}}^{({\boldsymbol{m}})} = {\boldsymbol{\rho}}^{({\boldsymbol{m}})}_{\rho/m_2} (t^{({\boldsymbol{m}})}_{ l})$.
#### **9.3. Proof of Theorem \[thm:decompositionrhodonea\]**
[Theorem \[thm:decompositionrhodonea\]]{} (a) Let ${\boldsymbol{x}}(r,\theta)\in \mathcal{R}^{({\boldsymbol{m}})}$. We choose $s \in [0,\pi/2]$ and $t'\in \mathbb{R}$ such that $r=\cos(s)$ and $T_{m_{1}}(r) = \cos (m_1 s) = \cos (m_2 \theta) = \cos(\frac{m_1 m_2}{g} t')$ . Then, we can find $v_1,v_2 \in \{-1,1\}$ and $h_1,h_2 \in\mathbb{Z}$ such that $$\begin{aligned}
m_1 s &= v_1 \left(\frac{m_1 m_2}{g} t' + h_1 \pi\right), \qquad m_2 \theta = v_2 \left(\frac{m_1 m_2}{g} t' + h_2 \pi\right), \end{aligned}$$ and therefore $$r = \cos \left(\frac{m_2}{g} t' + \frac{h_1}{m_1} \pi\right) \quad \text{and} \quad \theta = v_2 \left(\frac{m_1}{g} t' + \frac{h_2}{m_2} \pi \right).$$ Further, there is a unique $\rho \in \{0, \ldots, 2 g -1\}$ such that $h_1 \equiv h_2 + v_2 \rho \mod 2g$. Then, by the Chinese remainder theorem we can find an $l \in {{\mathbb Z}}$ such that $l\equiv h_1 \mod 2 m_1$ and $l\equiv h_2 + v_2 \rho \mod 2 m_1$. This gives (we assume that the angle $\theta$ is an element in ${{\mathbb R}}/ (2 \pi {{\mathbb Z}})$) $$r = \cos \left(\frac{m_2}{g} t' + \frac{l}{m_1} \pi \right) \quad \text{and} \quad \theta = v_2 \left(\frac{m_1}{g} t' + \frac{l- v_2\rho}{m_2} \pi \right).$$ Then, introducing $t= v_2(t'+l g \pi/ (m_1 m_2))$ we obtain $$r = \cos \left(\frac{m_2}{g} t \right) \quad \text{and} \quad \theta = \frac{m_1}{g} t - \frac{\rho}{m_2} \pi.$$ Therefore ${\boldsymbol{x}} \in {\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\rho/m_2}([0,P))$ for some $\rho \in \{0, \ldots, 2g-1\}$. The implication ${\boldsymbol{\varrho}}^{({\boldsymbol{m}})}_{\rho/m_2} \subseteq \mathcal{R}^{({\boldsymbol{m}})}$ is easily verified by inserting the curve in the description of the rhodonea variety $\mathcal{R}^{({\boldsymbol{m}})}$.
\(b) We have a look at the definition of the points ${\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}$ in ${{\boldsymbol{\operatorname{RD}}}^{({\boldsymbol{m}})}}$. Plugging the points ${\boldsymbol{x}}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}$ into the polar equation of the rhodonea variety, we obtain $T_{m_1}(r^{(m_1)}_{i_1}) = \cos (m_2 \theta^{(m_2)}_{i_2}) = 1$ if ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0$ and $T_{m_1}(r^{(m_1)}_{i_1}) = \cos (m_2 \theta^{(m_2)}_{i_2}) = 0$ if ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1$. On the other hand, it is well-known that the extrema of the univariate functions $T_{m_1}(r)$ and $\cos (m_2 \theta)$ are attained at $r=r^{(m_1)}_{i_1}$ and $\theta = \theta^{(m_2)}_{i_2}$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_0$, respectively. Also, it is well-known that all roots of $T_{m_1}(r)$ and $\cos (m_2 \theta)$ are given by $r=r^{(m_1)}_{i_1}$ and $\theta = \theta^{(m_2)}_{i_2}$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}_1$, respectively.
#### **9.4. Proofs of Section \[1507091240\]**
The proofs of Section \[1507091240\] base on the following technical result:
\[1507211320\] Let ${\boldsymbol{\gamma}}\in{{\mathbb Z}}^{2}$ and $\gamma_1 \equiv \gamma_2 \mod 2$. If $\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w}\neq 0$, then $$\label{1507201132}
\text{there exist $(h_1,h_2) \in {{\mathbb Z}}^{2}$ with $\gamma_{1}=2 h_{1}m_{1}$, $\gamma_{2}=2h_{2}m_{2}$, and
$h_1 + h_2 \equiv 0 \mod 2$}.$$ If is satisfied, then $\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w}=1$.
In the proof of Proposition \[1507211320\], we use the well-known trigonometric identity $$\label{1506171253}
\sum_{l=0}^N \mathrm{e}^{\imath l \vartheta} = \left\{ \begin{array}{ll} \frac{\mathrm{e}^{\imath (N+1) \vartheta}-1}{\mathrm{e}^{\imath \vartheta}-1}
\quad & \vartheta\notin 2\pi\mathbb{Z},\\ N+1 & \vartheta \in 2\pi\mathbb{Z}, \end{array} \right.
\qquad N\in{{\mathbb N}}_0.$$
Using Proposition \[1509221521\], we can manipulate the discrete integral $\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w}$ as follows: $$\begin{aligned}
\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w}
&= \sum_{{\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}}
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}} \cos(\gamma_{1} i_1 \pi/(2m_{1})) \mathrm{e}^{\imath \gamma_{2} i_2 \pi/(2m_{2}) } \\
&= \frac14 \sum_{u,v \in \{\pm 1\}}\sum_{{\boldsymbol{i}}\in{{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} \mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{i}}}
\mathrm{e}^{\imath \,( \gamma_{1} u v i_1/(2m_1) \pi + \gamma_1 (1-v)/2 \pi) + \gamma_{2} i_2/(2m_2) \pi + \gamma_2 (1-v)/2 \pi))}\\
&= \frac{1}{8m_1m_2} \sum_{l = 0}^{4 m_1 m_2/g} \sum_{\rho = 0}^{2g - 1}
\mathrm{e}^{\imath \,( \gamma_{1} l \pi/(2m_{1}) + \gamma_{2} l \pi/(2m_{2}) + 2 \gamma_2 \rho \pi / (2m_2))}\end{aligned}$$ The trigonometric identity implies that the last sum is different from zero if and only if $\gamma_{1} /(2m_{1}) + \gamma_{2} /(2m_{2}) \in 2 {{\mathbb Z}}$ and $\gamma_2 /(2m_2) \in {{\mathbb Z}}$ are satisfied. Therefore, if we assume that $\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w} \neq 0$ then $\gamma_2 = 2 h_2 m_2$ with some integer $h_2 \in {{\mathbb Z}}$ and $\gamma_{1} /(2m_{1}) + \gamma_2 / (2m_2) \in 2 {{\mathbb Z}}$. In particular, also $\gamma_1 = 2 h_1 m_1$ with some $h_1 \in {{\mathbb Z}}$. Further, we have $h_1 + h_2 \in 2 {{\mathbb Z}}$. This proves the identity . If is satisfied then the trigonometric identity yields $$\begin{aligned}
\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\mathrm{d}\rule{1pt}{0pt}\mathrm{w}
&= \frac{1}{8m_1m_2} \sum_{l = 0}^{4 m_1 m_2 / g} \mathrm{e}^{\imath \,l ( \gamma_{1} \pi/(2m_{1}) + \gamma_{2} \pi/(2m_{2}))} \sum_{\rho = 0}^{2 g -1}
\mathrm{e}^{\imath \,( 2 \gamma_2 \pi / 2 m_2) \rho } = 1.\end{aligned}$$
In order to prove Theorem \[1507091911\], i.e., to show that the rectangular set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ is a spectral index set, we will use the two identities $$\begin{aligned}
\label{1507222159}
{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\overline{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}'}}
&= \frac{1}{2} \left(\chi^{({\boldsymbol{m}})}_{(\gamma_1+\gamma_1',\gamma_2 - \gamma_2')}
+ \chi^{({\boldsymbol{m}})}_{{(\gamma_1-\gamma_1',\gamma_2 - \gamma_2')}} \right),\\
\overline{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}
&= \chi^{({\boldsymbol{m}})}_{(\gamma_1,-\gamma_2)}, \label{1507222159C}\end{aligned}$$ which are satisfied for all ${\boldsymbol{\gamma}},{\boldsymbol{\gamma}}' \in {{\mathbb Z}}^2$. Formulas and are a direct consequence of the definition of the discrete functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$ as well as the cosine product formula.
[Theorem \[1507091911\]]{} We will constantly use condition and Proposition \[1507211320\] to derive the values of the integrals. For a simpler notation, we denote the index vectors on the right hand side of by $${\boldsymbol{\gamma}}^+ = (\gamma_1+\gamma_1',\gamma_2-\gamma_2') \quad \text{and} \quad {\boldsymbol{\gamma}}^- = (\gamma_1-\gamma_1',\gamma_2-\gamma_2').$$ We assume first that ${\boldsymbol{\gamma}},{\boldsymbol{\gamma}}'\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ and ${\boldsymbol{\gamma}} \neq {\boldsymbol{\gamma}}'$: Since $-2 m_2 < \gamma_2-\gamma_2' < 2 m_2$, the condition can be satisfied for ${\boldsymbol{\gamma}}^+$ and ${\boldsymbol{\gamma}}^-$ only if $\gamma_2 = \gamma_2'$. Since we assume that ${\boldsymbol{\gamma}} \neq {\boldsymbol{\gamma}}'$, we get in this case $\gamma_1 \neq \gamma_1'$. This, on the other hand, implies that $\gamma_1-\gamma_1' \in 4 {{\mathbb Z}}$, $\gamma_1 + \gamma_1' \in 4 {{\mathbb Z}}$ is not possible, and therefore that ${\boldsymbol{\gamma}}^+$ and ${\boldsymbol{\gamma}}^-$ can not satisfy the condition . The product formula now yields the orthogonality $\int{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\overline{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}'}} \mathrm{d}\rule{1pt}{0pt}\mathrm{w} = 0$.\
Now, consider ${\boldsymbol{\gamma}},{\boldsymbol{\gamma}}'\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ and ${\boldsymbol{\gamma}} = {\boldsymbol{\gamma}}'$: In this case, we have ${\boldsymbol{\gamma}}^+ = (2 \gamma_1,0)$ and ${\boldsymbol{\gamma}}^- = (0,0)$. Since ${\boldsymbol{\gamma}}\in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, we have $0 \leq 2 \gamma_1 \leq 4 m_1$. Therefore, condition is always satisfied for ${\boldsymbol{\gamma}}^-$ and satisfied for ${\boldsymbol{\gamma}}^+$ precisely if $\gamma_1 \in \{0,2m_1\}$. Proposition \[1507211320\] therefore implies .\
Finally, since the functions ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$, are pairwise orthogonal, they are in particular linearly independent and span a subspace of dimension $\# {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}= 2 m_1 m_2 + m_2$ in $\mathcal{L}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. Since $\dim {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}) = \# {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}= (2 m_1 +1) m_2$ is of the same complexity, this subspace coincides with ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$.
[Theorem \[1507091912\]]{} The functions ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}$ are real and satisfy $${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}= \left\{ \begin{array}{ll}
{\operatorname{Re}}\chi^{({\boldsymbol{m}})}_{({\boldsymbol{\gamma}})} = \frac12
(\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}} + \overline{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}),
& \text{if ${\boldsymbol{\gamma}}$ is in the sets (i) or (iii) of \eqref{1702291124}}, \\
{\operatorname{Im}}\chi^{({\boldsymbol{m}})}_{({\boldsymbol{\gamma}})} = \frac1{2 \imath}
(\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}} - \overline{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}),
& \text{if ${\boldsymbol{\gamma}}$ is in the sets (ii) or (iv) of \eqref{1702291124}}.
\end{array} \right.$$ Using the condition of Proposition \[1507211320\] in combination with the trigonometric identities and , the orthogonality of the basis functions can be derived in the same way as in the proof of Theorem \[1507091911\]. In the following, we provide the calculation of the norms $\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2$. Using and for ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}= {\operatorname{Re}}\chi^{({\boldsymbol{m}})}_{({\boldsymbol{\gamma}})}$, we get $$\begin{aligned}
\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2 &= \frac14 \int \left|\chi^{({\boldsymbol{m}})}_{(\gamma_1, \gamma_2)}
+ \chi^{({\boldsymbol{m}})}_{(\gamma_1,- \gamma_2)}\right|^2 \mathrm{d}\rule{1pt}{0pt}\mathrm{w} \\
&= \frac{1}{8} \int \left( 2 \chi^{({\boldsymbol{m}})}_{(0,0)} + 2 \chi^{({\boldsymbol{m}})}_{(2 \gamma_1,0)} + \chi^{({\boldsymbol{m}})}_{(2\gamma_1, 2\gamma_2)} + \chi^{({\boldsymbol{m}})}_{(0, 2\gamma_2)}
+ \chi^{({\boldsymbol{m}})}_{(2\gamma_1, - 2\gamma_2)} + \chi^{({\boldsymbol{m}})}_{(0, -2\gamma_2)} \right) \mathrm{d}\rule{1pt}{0pt}\mathrm{w}.\end{aligned}$$ On the other hand, if ${\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}= {\operatorname{Im}}\chi^{({\boldsymbol{m}})}_{({\boldsymbol{\gamma}})}$, we have $$\begin{aligned}
\|{\chi^{({\boldsymbol{m}})}_{\mathcal{R},{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2 &= \frac14 \int \left|\chi^{({\boldsymbol{m}})}_{(\gamma_1, \gamma_2)}
- \chi^{({\boldsymbol{m}})}_{(\gamma_1,- \gamma_2)}\right|^2 \mathrm{d}\rule{1pt}{0pt}\mathrm{w} \\
&= \frac{1}{8} \int \left( 2 \chi^{({\boldsymbol{m}})}_{(0,0)} + 2 \chi^{({\boldsymbol{m}})}_{(2 \gamma_1,0)} - \chi^{({\boldsymbol{m}})}_{(2\gamma_1, 2\gamma_2)} - \chi^{({\boldsymbol{m}})}_{(0, 2\gamma_2)}
- \chi^{({\boldsymbol{m}})}_{(2\gamma_1, - 2\gamma_2)} - \chi^{({\boldsymbol{m}})}_{(0, -2\gamma_2)} \right) \mathrm{d}\rule{1pt}{0pt}\mathrm{w}.\end{aligned}$$ In both cases, using Proposition \[1507211320\], we can explicitly evaluate the integrals on the right hand side. Depending on the different cases given in , the corresponding values for the norm in can be obtained directly.
#### **9.5. Proofs of Section \[sec:interpolation\]**
[Theorem \[201512131945\] and Theorem \[201512131946\]]{} The proof of the two theorems differs only in the choice of the basis system. We will therefore restrict our attention to Theorem \[201512131945\].
We denote by $\delta_{{\boldsymbol{j}}}({\boldsymbol{i}}) = \delta_{{\boldsymbol{i}}{\boldsymbol{j}}}$, ${\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, the set of all Dirac functions on ${{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. They clearly form an orthogonal basis of ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. By Definition \[def:spectralindex\] of the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, also the function system ${\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}$, ${\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}$, is an orthogonal basis for the space ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$. We can therefore expand the Dirac functions $\delta_{{\boldsymbol{j}}}$, ${\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, as $$\delta_{{\boldsymbol{j}}}({\boldsymbol{i}}) = \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{\langle\;\! \delta_{{\boldsymbol{j}}},{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\rangle_{\mathrm{w}}}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2} {\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) =
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{j}}} \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) \overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{j}})}}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}.$$
Evaluating the Lagrange function $L^{({\boldsymbol{m}})}_{{\boldsymbol{j}}}$, ${\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, given in at the nodes $(r^{(m_1)}_{i_1}, \theta^{(m_2)}_{i_2})$, ${\boldsymbol{i}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, and using the relation , we get the identity $$L^{({\boldsymbol{m}})}_{{\boldsymbol{j}}}(r^{(m_1)}_{i_1},\theta^{(m_2)}_{i_2}) =
\mathrm{w}^{({\boldsymbol{m}})}_{{\boldsymbol{j}}} \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}}} \frac{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{i}}) \overline{{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}({\boldsymbol{j}})}}{\|{\chi^{({\boldsymbol{m}})}_{{\boldsymbol{\gamma}}}}\|_{\mathrm{w}}^2}.$$ Therefore, $L^{({\boldsymbol{m}})}_{{\boldsymbol{j}}}(r^{(m_1)}_{i_1},\theta^{(m_2)}_{i_2}) = \delta_{{\boldsymbol{j}}}({\boldsymbol{i}})$ and for $f \in {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ we have $$f({\boldsymbol{i}}) = \sum_{{\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} f({\boldsymbol{i}}) \delta_{{\boldsymbol{j}}}({\boldsymbol{i}}) =
\sum_{{\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}} f({\boldsymbol{i}}) L^{({\boldsymbol{m}})}_{{\boldsymbol{j}}}(r^{(m_1)}_{i_1},\theta^{(m_2)}_{i_2})
= P^{({\boldsymbol{m}})}_{f}(r^{(m_1)}_{i_1},\theta^{(m_2)}_{i_2}).$$ Thus, the function $P^{({\boldsymbol{m}})}_{f}$ solves the interpolation problem $\eqref{1508220011}$, and the mapping $f \to P^{({\boldsymbol{m}})}_{f}$ is an injective linear mapping from ${\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$ into ${\Pi^{({\boldsymbol{m}})}}$. Further, since $\dim {\mathcal{L}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}) = \dim {\Pi^{({\boldsymbol{m}})}}$ this mapping is indeed an automorphism. This implies that the interpolant $P^{({\boldsymbol{m}})}_{f}$ is unique and that the system $L^{({\boldsymbol{m}})}_{{\boldsymbol{j}}}$, ${\boldsymbol{j}} \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$, forms a basis of ${\Pi^{({\boldsymbol{m}})}}$. Finally, if $f$ is in the subspace ${\mathcal{L}_{\mathrm{D}}}({{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}})$, we have $f(m_1,i_2) = f_C$ for all tuples $(m_1,i_2) \in {{\boldsymbol{\operatorname{I}}}^{({\boldsymbol{m}})}}$. In this way, the interpolating function $P^{({\boldsymbol{m}})}_{f}$ has the form and is contained in the subspace ${\Pi^{({\boldsymbol{m}})}_{\mathrm{D}}}$.
#### **9.6. Proofs of Section \[sec:convergence\]**
[Theorem \[th:201807231722\]]{} By the discussion in front of Theorem \[th:201807231722\], we only have to show that the continuity condition (ii) in the definition of $C({\mathbb{D}})$ is satisfied. The interpolant $P^{({\boldsymbol{m}})}_{f} \in {\Pi^{({\boldsymbol{m}})}_{\square}}$ is of the form $$\label{eq:1807221940} P^{({\boldsymbol{m}})}_{f}(r,\theta) = \sum_{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}} c_{{\boldsymbol{\gamma}}}(f)
{{X}_{{\boldsymbol{\gamma}}}}(r,\theta) = \sum_{\gamma_2 = -m_2 + 1}^{m_2} p_{\gamma_2}(r) e^{\imath \gamma_2 \theta},$$ with univariate polynomials $p_{\gamma_2}(r)$, $-m_2 < \gamma_2 \leq m_2$ of degree $2 m_1$. Further, the polynomials $p_{\gamma_2}(r)$ are even if $\gamma_2$ is even and odd otherwise. Thus, for the center $r = 0$ we obtain $$\label{eq:1807221941} P^{({\boldsymbol{m}})}_{f}(0,\theta) = \sum_{\gamma_2 \in \{-m_2 + 1, \ldots, m_2\} \atop \gamma_2 \ \text{even}} p_{\gamma_2}(0) e^{\imath \gamma_2 \theta}.$$ In particular, $P^{({\boldsymbol{m}})}_{f}(0,\theta)$ is a $\pi$-periodic trigonometric polynomial in the $m_2$ dimensional space spanned by the functions $\{e^{\imath \gamma_2 \theta} \ | \ \gamma_2 \in \{-m_2+1, \ldots, m_2\}, \ \gamma_2 \; \text{even}\}$. Moreover, we have exactly $m_2$ different points $(0,\theta_{i_2}^{(m_2)})$, $i_2 \in \{0,2,2m_2-2\}$ in $[0,\pi)$ at which $P^{({\boldsymbol{m}})}_{f}(0,\theta)$ is equal to the constant value $f_{\mathrm{C}} = {\mathscr{f}}(0,0)$ given at the center of ${\mathbb{D}}$. These $m_2$ conditions determine the trigonometric polynomial $P^{({\boldsymbol{m}})}_{f}(0,{\varphi})$ uniquely such that $P^{({\boldsymbol{m}})}_{f}(0,\theta) = f_{\mathrm{C}}$ is constant for $\theta \in [-\pi,\pi]$.
[Theorem \[thm:lebesgueconstant\]]{} We split the spectral index set ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}$ into the two canonical parts $${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square,0}}= \left\{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\ | \ \gamma_1, \gamma_2 \ \text{are even}\ \right\}, \quad
{{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square,1}}= \left\{{\boldsymbol{\gamma}} \in {{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square}}\ | \ \gamma_1, \gamma_2 \ \text{are odd}\ \right\}.$$ Then, we have $\Lambda^{({\boldsymbol{m}})}_{\square} \leq \Lambda^{({\boldsymbol{m}})}_{\square,0} +
\Lambda^{({\boldsymbol{m}})}_{\square,1}$, where $$\Lambda^{({\boldsymbol{m}})}_{\square,0} = \sup_{\|{\mathscr{f}}\|_{\infty} \leq 1} \|P^{({\boldsymbol{m}})}_{f}|_{\Pi^{({\boldsymbol{m}})}_{\square,0}}\|_{\infty}, \quad
\Lambda^{({\boldsymbol{m}})}_{\square,1} = \sup_{\|{\mathscr{f}}\|_{\infty} \leq 1} \|P^{({\boldsymbol{m}})}_{f}|_{\Pi^{({\boldsymbol{m}})}_{\square,1}}\|_{\infty},$$ and $\Pi^{({\boldsymbol{m}})}_{\square,0}$ and $\Pi^{({\boldsymbol{m}})}_{\square,0}$ are the subspaces of ${\Pi^{({\boldsymbol{m}})}_{\square}}$ with respect to the spectral index sets ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square,0}}$ and ${{\boldsymbol{\Gamma}}^{({\boldsymbol{m}})}_{\square,1}}$, respectively. As the estimates for $\Lambda^{({\boldsymbol{m}})}_{\square,0}$ and $\Lambda^{({\boldsymbol{m}})}_{\square,1}$ are very similar, we will restrict all upcoming considerations to the number $\Lambda^{({\boldsymbol{m}})}_{\square,0}$.
We use and to reformulate $P^{({\boldsymbol{m}})}_{f}|_{\Pi^{({\boldsymbol{m}})}_{\square,0}}$ in terms of a double sum. We get $$\begin{aligned}
P^{({\boldsymbol{m}})}_{f}|_{\Pi^{({\boldsymbol{m}})}_{\square,0}}(r,\theta) &=
\sum_{\gamma_1 = 0}^{m_1} \sum_{\gamma_2 = - \lceil m_2/2 \rceil + 1}^{\lfloor m_2/2 \rfloor} \frac{\hat{g}(2 {\boldsymbol{\gamma}})}{\|\chi^{({\boldsymbol{m}})}_{2 {\boldsymbol{\gamma}}}\|_{\mathrm{w}}^2}
T_{2 \gamma_1}(r) \mathrm{e}^{\imath 2 \gamma_{2} \theta}.\end{aligned}$$ Using the reflection symmetry $\hat{g}({\boldsymbol{\gamma}}) = \hat{g}(- \gamma_1 \mod 4m_1, \gamma_2)$ of $g$ on ${{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}$, we further get $$\begin{aligned}
P^{({\boldsymbol{m}})}_{f}|_{\Pi^{({\boldsymbol{m}})}_{\square,0}}(r,\theta) &=
\sum_{\gamma_1 = -m_1}^{m_1} \sum_{\gamma_2 = - \lceil m_2/2 \rceil + 1}^{\lfloor m_2/2 \rfloor}
(1 - {\textstyle}\frac12 \delta_{|\gamma_1|,m_1})\hat{g}(2 {\boldsymbol{\gamma}})
\mathrm{e}^{\imath (2 \gamma_1 \arccos(r) + 2 \gamma_{2} \theta)}.\end{aligned}$$ For $\Lambda^{({\boldsymbol{m}})}_{\square,0}$ we get in this way the bound $$\begin{aligned}
\Lambda^{({\boldsymbol{m}})}_{\square,0} & \leq \sup_{\|{\mathscr{f}}\|_{\infty} \leq 1} \sup_{(r,\theta)} \left|
\sum_{{\boldsymbol{i}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}} \sum_{\gamma_1 = -m_1}^{m_1} \!\!\! (1 - {\textstyle}\frac12 \delta_{|\gamma_1|,m_1}) \!\!\!\!\!\!\!\!\! {\displaystyle}\sum_{\gamma_2 = - \lceil m_2/2 \rceil + 1}^{\lfloor m_2/2 \rfloor} \!\!\!\!\!\! g({\boldsymbol{i}}) \mathrm{e}^{-\imath 2 \gamma_{1} (i_1 \pi/m_{1} - \arccos(r)) }
\mathrm{e}^{-\imath 2 \gamma_{2} (i_2 \pi/m_{2}-\theta) }\right| \\
&\leq \sup_{(\theta,{\varphi})} \frac{1}{8 m_1 m_2} \sum_{{\boldsymbol{i}} \in {{{\boldsymbol{\operatorname{J}}}}^{({\boldsymbol{m}})}}} \left|
\sum_{\gamma_1 = -m_1}^{m_1} \!\!\! (1 - {\textstyle}\frac12 \delta_{|\gamma_1|,m_1}) \!\!\!\!\!\!\!\!\! {\displaystyle}\sum_{\gamma_2 = - \lceil m_2/2 \rceil + 1}^{\lfloor m_2/2 \rfloor} \mathrm{e}^{-\imath 2 \gamma_{1} (i_1 \pi/m_{1} - \arccos(r)) }
\mathrm{e}^{-\imath 2 \gamma_{2} (i_2 \pi/m_{2}-\theta) }\right| \\
&\leq C \int_{0}^{2 \pi} \left| \sum_{\gamma_1 = -m_1}^{m_1} \!\!\! (1 - {\textstyle}\frac12 \delta_{|\gamma_1|,m_1})
\mathrm{e}^{- \imath 2 \gamma_{1} \rho } \right| {\mathrm{d}\rho} \int_0^{2\pi} \left|
\sum_{\gamma_2 = - \lceil m_2/2 \rceil + 1}^{\lfloor m_2/2 \rfloor} \mathrm{e}^{-\imath 2 \gamma_{2} \theta' } \right| {\mathrm{d}\theta'}.\end{aligned}$$ The last transition from the two discrete sums to the continuous integrals with a constant $C >0$ independent of ${\boldsymbol{m}}$ is a twofold application of a Marcinkiewicz-Zygmund inequality, see [@Zygmund X, Theorem 7.10]. The two univariate integrals in the last line can be considered, up to minor modifications, as the classical univariate Lebesgue constants in the trigonometric setting [@Zygmund II, §12]. Both can be estimated in terms of a log term such that $$\begin{aligned}
\Lambda^{({\boldsymbol{m}})}_{\square,0} & \leq C_{\square,0} \ln (m_1+1) \ln (m_2+1).\end{aligned}$$ This, together with a respective very similar estimate for $\Lambda^{({\boldsymbol{m}})}_{\square,1}$ gives the statement.
Conclusion
==========
In this manuscript, we derived a novel spectral interpolation scheme for the unit disk in which samples along rhodonea curves form the set of interpolation nodes. We derived three characterizations of the rhodonea nodes. In particular, the possibility to describe these nodes as the union of two interlacing polar grids allowed us to implement the interpolation scheme in an efficient way using fast Fourier algorithms.
The interpolation spaces are determined by a spectral index set selecting the Chebyshev-Fourier basis. While uniqueness of the interpolation scheme can be shown for a general class of interpolation spaces, the restriction to a rectangular spectral index set turned out to be advantageous for several reasons: in this case the numerical condition number is growing slowly in the number of nodes, the interpolation scheme converges fast if the interpolated function is smooth and continuity of the interpolant can be guaranteed. This could be verified theoretically and also in a numerical experiment. Further, the interpolation scheme was applied to obtain a Clenshaw-Curtis quadrature rule on the disk.
[10]{}
, 1 (2006), 15–25.
. (1978), 1184–1191.
Dover Publications Inc., New York, 2000.
. (2011), 1408–1438.
, 2 (2005), 261–274.
. , 2 (2011), 275–288.
. (2017).
. (2017), 15–45.
. (2017), 1–27.
. (2018), 53–60.
. (2016), 409–425.
. (2018).
. , 1 (2016), 685–705.
Polynomial interpolation on interlacing rectangular grids. (2017), 64–73.
. (1995), 1071–1081.
. Cambridge University Press, 1996.
A new spectral boundary integral collocation method for three-dimensional potential problems. (1998), 778–804.
Rose surfaces and their visualizations. , 1 (2010), 1–9.
, 1728.
, 12 (2010), 4447–4453.
. (1888).
. , 6 (2003), 2333–2349.
Interpolation lattices for hyperbolic cross trigonometric polynomials. (2012), 76–92.
. (2015), 543–576.
Trajectory analysis for magnetic particle imaging. , 2 (2009), 385–397.
Magnetic particle imaging: From proof of principle to preclinical applications. , 14 (2017), R124.
. , 1902.
Frame frequency prediction for risley-prism-based imaging laser radar. (2014), 3556–3564.
Fast and exact reconstruction of arbitrary multivariate algebraic polynomials in [C]{}hebyshev form. In [*Proceedings of the 11th International Conference on Sampling Theory and Applications*]{} (2015), pp. 392–396.
. Springer Series in Computational Mathematics 41, Springer-Verlag Berlin Heidelberg, 2011.
A radial lissajous trajectory for magnetic particle imaging. In [*International Workshop on Magnetic Particle Imaging 2015*]{} (2015).
. Pergamon Press, Oxford, 1963.
. , 4 (2016).
. SIAM, Philadelphia, 2000.
. , 3 (2017).
, 2 (1996), 220–238.
. Cambridge University Press, Cambridge, 2002.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This paper contains an account of arbitrary cubic function fields of characteristic three. We define a standard form for an arbitrary cubic curve and consider its function field. By considering an integral basis for the maximal order of these function fields, we are able to calculate the field discriminant and the genus. We also describe the splitting behavior of any place, and give composition and reduction algorithms for arithmetic in the ideal class group.\
address:
- 'University of Calgary, Calgary AB T2L 1N4, CANADA'
- 'Bates College, Lewiston ME 04240, USA'
author:
- Mark Bauer
- Jonathan Webster
bibliography:
- 'mcom-l.bib'
title: Computations in Cubic Function Fields of Characteristic Three
---
Introduction
============
Calculating invariants of a field and its maximal order remains one of the central problems in computational number theory. Motivated by hyperelliptic curve cryptography and well-studied cubic number fields, a host of authors have researched computational properties of cubic function fields. From calculating fundamental units [@fun_unit], to computing in the ideal class group [@bauer], to tabulating [@rozen], to describing and classifying arbitrary cubic function fields [@exp_cff; @aacff], the results (mostly) exclude characteristic three.
In the case in which characteristic three is considered, it is often through generic methods. The function field analogue of the Round 2 algorithm, algebraic methods involving desingularization, or using Groebner basis to do ideal arithmetic, all may be applied to the problems considered in this paper. However, these methods are often impractical and can make it difficult to understand how the basic invariants of a function field arise from the defining curve. For elliptic curves and hyperelliptic curves, we may compute the desired quantities directly from the defining curve and the underlying finite field. For cubic function fields in characteristic greater than three, much progress has been made in this regard; our goal is to extend these computations to characteristic three. It is important to mention that work to this end has also been undertaken in [@tobias]. However, the aim of our project is slightly different — as opposed to developing a coherent theory for signatures across different characteristics, we have chosen to completely analyze all cubic function fields in characteristic three and develop the associated algorithms for computations.
We begin by developing the basic invariants of cubic function fields. Section 2 defines function fields and states the standard model that will be used to define the field. Section 3 contains the calculation of the integral basis and the field discriminant for the fixed model. The following section describes the splitting behavior for places and this is used in Section 5 to calculate the genus. This will conclude the calculation of basic invariants. We review the relationship between the ideal class group and the Jacobian to motivate an explicit means of doing computations in the ideal class group. We state integral basis for the prime ideals and their powers in Section 7. Using this basis motivates arbitrary ideal arithmetic which is given over the next two sections. Finally, we state an algorithm to do composition and reduction in the ideal class group and conclude with an example computation.
Standard Form
=============
As there are many good introductions to algebraic function fields (for example [@sticht; @dino]), we will only seek to clarify the notation used in this paper. As usual, let $\mathbb{F}_q$ be a finite field and $\mathbb{F}_q[x]$ and $\mathbb{F}_q(x)$ be the ring of polynomials and the field of rational functions, respectively, in $x$ over $\mathbb{F}_q$. An algebraic function field is a finite extension $\mathcal{F}$ of $\mathbb{F}_q(x)$; it thus may be written as $\mathcal{F} = \mathbb{F}_q(x,y)$ with $y$ a root of $H(T)$, where $H(T)$ is an irreducible monic polynomial in $({\mathbb{F}_q}[x])[T]$ of degree $n = [\mathcal{F}:\mathbb{F}_q(x)]$. Hence to study cubic function fields, we consider affine planar curves $H(x,T)$, where $H \in \mathbb{F}_q[x,T]$ is a bivariate polynomial that is absolutely irreducible and of degree three in $T$.
When $\textrm{char}(\mathbb{F}_q) \neq 3$, cubic function fields may be studied by examining the standard form for the defining polynomial which is given by $T^3 - AT + B = 0$ with $A, B \in \mathbb{F}_q[x]$ provided there is no non-constant $Q \in \mathbb{F}_q[x]$ such that $Q^2|A$ and $Q^3|B$ (see [@aacff] for details). By considering these birationally equivalent curves, it is possible to study arbitrary curves as a two-parameter family. Our goal will be to find a model which gives a similar two-parameter family in characteristic three. Henceforth, let $\textrm{char}(\mathbb{F}_q) = 3$ unless explicitly stated otherwise.
Write $H(x,T) = ST^3 + UT^2 + VT + W$ with $S,U,V,W \in \mathbb{F}_q[x]$ and $SW \neq 0$. If $U=V=0$, then the function field associated with this curve is purely inseparable, and hence isomorphic to the rational function field (see Proposition III.9.2 of [@sticht]). We thus require $U \neq 0$ or $V \neq 0$ to avoid this degenerate case.
If $U = 0$ then making the polynomial monic yields a curve in of the form $T^3 - AT + B = 0$. In terms of the original parameters, $A = -SV$ and $B = S^2W$. Otherwise, $U \neq 0$ and transform via $T \rightarrow (T + V)/U$ to eliminate the linear term. Considering the monic, integral, reciprocal polynomial, we get a curve in the form $T^3 - AT + B = 0$. If $N = S - U^2V+U^3W$, then $A = -N^2U^2$ and $B = N^2S$ in terms of the parameters of the original curve.
Henceforth, we will restrict our attention to curves of the form $T^3 - AT + B = 0$. In what follows we use the fact that $T \rightarrow T + i$ yields the birationally equivalent curve $T^3 -AT + (i^3 - iA + B) = 0$. Our goal will be two-fold — to minimize both the repeated factors dividing $A$ and the degree of $B$.
If there is a polynomial $Q \in \mathbb{F}_q[x]$ such that $Q^2|A$ and $Q^3|(i^3 -iA + B)$ for some $i \in \mathbb{F}_q[x]$ then it is possible to consider the curve given by $$\label{sing_remove}T^3 - \left(\frac{A}{Q^2} \right)T +\left(\frac{i^3 - iA + B}{Q^3}\right) = 0.$$ If the characteristic is not three, the existence of a $Q$ such that $Q^2|A$ and $Q^3|B$ implies that $y/Q$ is integral. Likewise, in characteristic three the existence of $Q$ and $i$ implies $(y+i)/Q$ is integral and has a minimal polynomial (\[sing\_remove\]). Hence this is the appropriate generalization from the case where the characteristic is different from three. Also, in each case $Q$ corresponds to removable singularities that preserve the shape of the model for the given curve.
To find $Q$ and $i$, it is sufficient to check irreducible polynomials $P$ such that $P^2|A$. Begin by writing $i = i_0 + i_1P + i_2P^2$ with $i_0, i_1, i_2 \in {\mathbb{F}_q}[x]$ with degree less than that of $P$. Since only $i_0$ affects the congruence $i^3 -iA + B \equiv 0 \pmod{P^3}$, we solve $i_0^3 + B \equiv 0 \pmod{P}$. It then becomes a matter of checking whether $i_0^3 - i_0A + B \equiv 0 \pmod{P^3}$. If the congruence holds, redefine $A$ as $A/P^2$ and $B$ as $(i^3 - iA + B)/P^3$. This process may be repeated as needed.
We now turn our focus to reducing the degree of $B$. If $3 | \deg B$ and $ 2\deg B > 3\deg A $, we can write $B(x) = b_{3n}x^{3n} + b_{3n-1}x^{3n-1} + \cdots + b_0$ with $b_{3n} \neq 0$. Then consider the linear transformation $T \rightarrow T - (b_{3n})^{1/3}x^n$ (note that $b_{3n}^{1/3} \in \mathbb{F}_q$ because $\mathbb{F}_q$ is perfect). Under this map we get a new curve $$T^3 - A(x)T + b_{3n-1}x^{3n-1} + \ldots + b_0 + A(x)b_{3n}^{1/3}x^n = 0,$$ where the polynomial $ b_{3n-1}x^{3n-1} + \ldots + b_0 + A(x)(b_{3n})^{1/3}x^n$ has a lower degree than $B(x)$. By repeating this procedure it is possible to force the curve to satisfy one of the following two distinct criteria: $$\label{wildram} 3 \nmid \deg B \mbox{ \quad and \quad } 2\deg B > 3\deg A$$ or $$\label{tame} 2\deg B \leq 3\deg A.$$
A curve is said to be a *standard model* (or in *standard form*) for a cubic function field if it is of the form $T^3 - AT + B = 0$ with no $Q, i\in \mathbb{F}_q[x]$ such that $Q^2|A$ and $Q^3 | i^3 - iA + B$ and satisfies either (\[wildram\]) or (\[tame\]).
It will also be useful to have a simple criterion to detect singularities.
The curve $T^3 - A(x)T + B(x) = 0$ is nonsingular if and only if $\deg d=0$ where $d = {\rm gcd}(A(x), A'(x)^3B (x)+ B'(x)^3)$.
A singular point $(a,b) \in \overline{\mathbb{F}}_q^2$ satisfies the following three equations. $$\begin{aligned}
\label{sing1} b^3 - A(a)b + B(a) = 0. \\
\label{sing2} A(a) = 0. \\
\label{sing3} - A'(a)b + B'(a) = 0.\end{aligned}$$
From (\[sing2\]), $a$ is a root of $A(x)$, and combined with (\[sing1\]) we see that $-b^3 = B(a)$. Cubing (\[sing3\]) we get $-A'(a)^3b^3 + B'(a)^3 = 0$ which implies $A'(a)^3B(a) + B'(a)^3 = 0$. Thus $a$ is a common root of $A(x)$ and $A'(x)^3B(x) + B'(x)^3$.
For the converse let $a$ be a common root of $A(x)$ and $A'(x)^3B (x)+ B'(x)^3$. Since $a$ is a root of $A(x)$, (\[sing2\]) is satisfied. Since $\overline{\mathbb{F}_q}$ is perfect, we can find $b$ such that $b^3 = -B(a)$ in order to satisfy (\[sing1\]). With (\[sing1\]) and (\[sing2\]) satisfied, it is clear that (\[sing3\]) is also satisfied by the above construction.
Note that for large $q$ we do not expect a curve selected in standard form to be singular. That is, if singularity is detected by $\deg d$ not being $0$, then it is a question of when two “random" polynomials are relatively prime. This happens with probability roughly $1 - 1/q$.
Calculating the standard form and the integral basis, as well as finding the field discriminant and the genus are all closely related to singularity. The square factors removed from $A$ in the conversion to standard form correspond to singular points, which simplifies future calculations. In fact, if the standard form is nonsingular then $\{1, y, y^2\}$ is an integral basis for the maximal order (see Prop 5.10 of [@dino]). We now know that $D = {\rm disc}(y) = A^3$ (for the reader who is unfamiliar with this concept, it will be defined more formally below). In the next section we will show that the square-free factorization of $d = {\rm gcd}(A(x), A'(x)^3B (x)+ B'(x)^3)$ is $I = {\rm ind}(y)$. With $D$ and $I$ in hand, we will have $\Delta = {\rm disc}(\mathcal{F})$.
Knowing that $D = A^3$ and that $\Delta$ differs from $D$ by square factors is enough to determine when $\mathcal{F}$ is an Artin-Schreier extension.
$\mathcal{F}$ is an Artin-Schreier extension if and only if $A(x)$ is a square.
Cubic extensions are Galois (which is to say an Artin-Schreier extension in characteristic 3) if and only if their discriminant is a square. In order to have a square discriminant, $A(x)$ must be a square. Conversely, if $T^3 - T = f/g$ with $f,g \in \mathbb{F}_q[x]$ is an Artin-Schreier extension then $T^3 - g^2T = fg^2$ is an integral model for this equation. By renaming, we have $A(x) = g(x)^2$ a square.
Integral basis and field discriminant
=====================================
We will follow Chapter 2 Section 17 of [@dandf] to find an integral basis for $\mathcal{O}_\mathcal{F}$, the integral closure of $\mathbb{F}_q[x]$ in $\mathcal{F}$. Recall that the powers $1,\ y,\ y^2$ form a basis of the $\mathbb{F}_q(x)$-vector space $\mathcal{F}$. An $\mathbb{F}_q(x)$-basis given by $\{\alpha_0, \alpha_1, \alpha_2 \}$ is triangular if $\alpha_0$ and $\alpha_1$ are an $\mathbb{F}_q(x)$-linear combination of $1$ and $1, y$, respectively. The three conjugate mappings taking $y$ to the three roots $ y = y^{(0)}, y^{(1)}, y^{(2)}$ defines for every $\alpha \in \mathcal{F}$ its three conjugates $\alpha = \alpha^{(0)}, \alpha^{(1)},\alpha^{(2)}$, and allows for the following definition of the discriminant of three elements: $$\rm{disc}(\alpha_0, \alpha_1, \alpha_{2}) = \rm{det}(\alpha_i^{(j)})^2_{0\leq i, j, \leq 2} \in \mathbb{F}_q(x).$$
The ring $\mathcal{O}_\mathcal{F}$ always admits a triangular basis, one element of which is (obviously) in $\mathbb{F}_q^*$. The discriminant of $\mathcal{F}/\mathbb{F}_q(x)$ is $\rm{disc}(\mathcal{F}) = \rm{disc}(\alpha_0, \alpha_1, \alpha_{2})$ where $\{\alpha_0, \alpha_1,\alpha_{2} \}$ is an integral basis of $\mathcal{F}/\mathbb{F}_q(x)$, i.e. a basis for $\mathcal{O}_\mathcal{F}$. For any element $\alpha \in \mathcal{F}$, the index of $\alpha$ satisfies $\rm{disc}(\alpha) = \rm{ind}(\alpha)^2\rm{disc}(\mathcal{F})$, which will be crucial in determining a basis for $\mathcal{O}_\mathcal{F}$.
Writing down a basis in triangular form, we will be able to deduce restrictions on the elements of the basis simply by using the fact that they are integral. These restrictions arise naturally by examining the minimal polynomial of each element. Following [@dandf], we choose the product of the latter two basis elements to be in $\mathbb{F}_q[x]$. Consider the integral basis given by $$\left[1, \frac{y - i}{I_1}, \frac{ c + by + y^2}{I_2} \right] = [1, \rho, \omega ]$$ with $I_1,I_2, i, c, b \in \mathbb{F}_q[x]$ (the choice to reuse $i$ will become clear). As mentioned before, the integral basis construction was a motivation for the choice of the standard model; in particular, the minimal polynomial of $\rho$ is given by $$\rho^3 -\frac{A}{I_1^2}\rho + \frac{i^3 - iA + B}{I_1^3} = 0 .$$ Since this is an integral equation in $\rho$, it must be that $I_1^2 |A$ and $I_1^3 |i^3 -iA +B$. This is the same criterion as (\[sing\_remove\]). Thus the reduction to standard form forces $I_1 = 1$. Now consider $\rho\omega \in \mathbb{F}_q[x]$ to get additional criteria on $i,b,c, $ and $I_2$: $$\rho\omega = \frac{(b-i)y^2 + (A-ib+c)y - (ic + B)}{I_2}.$$ This implies $i=b$, $c = i^2 - A$, and $I_2|ic + B$. Combining the last two statements, $I_2|i^3 - iA + B$. Rewrite $\omega$ as $(y^2 + iy + i^2 - A)/I_2$ and consider its minimal polynomial to get our final criterion: $$\omega^3 + \frac{A}{I_2}\omega^2 - \frac{(i^3 -iA + B)^2}{I_2^3} = 0.$$
This gives $I_2 | A$ and $I_2^3|(i^3 -iA + B)^2$. Choosing $i$ such that $I_2$ is of maximal degree yields the basis. This observation will in fact force $I_2$, which is the index of $y$, to be square-free. From this point forward the subscript of $I_2$ will be dropped and the index of $y$ will be denoted $I$.
\[t:squarefree\] A curve in standard from has $I = {\rm ind}(y)$ being square-free.
Let $P \in {\mathbb{F}_q}(x)$ be irreducible such that $P|I$, the index. If $v_P(A) = 1$ then $v_P(I) = 1$. So assume $v_P(A) \geq 2$ and consider $i$ such that $I^3 |( i^3 -iA + B)^2$. If $v_P( i^3 -iA + B ) = 2$ then $v_P(I) = 1$. However, if $v_P(A) \geq 2$ and $v_P( i^3 -iA + B ) \geq 3$ then the curve is not in standard form.
Having established that the index is square free, it is possible to calculate $i$. Since $i$ is unique modulo $I$, $i$ is determined by its residue class modulo each distinct prime dividing $I$. For each irreducible polynomial $P|I$, we solve $(i^3 -iA + B)^2 \equiv 0 \pmod{P^3}$ and constuct the solution using the Chinese Remainder Theorem. As we did when removing singularities, we write $ i = i_0 + i_1P + i_2P^2$ and solve congruence equations modulo $P$, $P^2$, and $P^3$.
With the index of $y$ calculated, it is straightforward to determine the discriminant of the function field simply by noting that $D=A^3$ and hence $\Delta=A^3/I^2$. It is well known that the field discriminant is closely tied to ramification. This is developed more fully in the next section.
Letting $A = EI$ and $FI^2 = i^3 - iA + B$, we have the following identities for various products of integral basis elements: $$\rho^2 = I\omega + A, \quad \omega^2= - E\omega - F\rho, \quad \rho\omega = -FI .$$
Splitting of Places
===================
The places of $\mathbb{F}_q(x)$ consist of finite places, identified with the monic irreducible polynomials in $\mathbb{F}_q[x]$, and the place at infinty $P_{\infty}$, identified with $1/x$. Every place $P$ has a corresponding discrete valuation on $\mathbb{F}_q(x)$ denoted $v_P$ and a discrete valuation ring $\mathcal{O}_P = \{G \in \mathbb{F}_q(x) | v_P(G) \geq 0 \}$. These definitions may be naturally extended to the field $\mathcal{F}$. That is, the finite places are associated with the non-zero prime ideals in $\mathcal{O}_\mathcal{F}$ and the infinite places are associated to the non-zero prime ideals in the integral closure of $\mathcal{O}_{P_{\infty}}$. If $\mathfrak{p}$ is a place of $\mathcal{F}$ then let $v_\mathfrak{p}$ denote its associated discrete valuation and $\mathcal{O}_\mathfrak{p} = \{ \alpha \in \mathcal{F} | v_\mathfrak{p}(\alpha) \geq 0 \}$ its discrete valuation ring. There exists a place $P \in \mathbb{F}_q(x)$ with $v_\mathfrak{p}(P) > 0$; we say $\mathfrak{p}$ lies above $P$ and write $\mathfrak{p} | P$. The positive integer $e(\mathfrak{p}|P) = v_\mathfrak{p}(P)$ is the ramification index and we say $P$ is ramified if $e(\mathfrak{p}|P) > 1$ and unramified otherwise. Further, if $\gcd (e(\mathfrak{p}|P), q) = 1$ a place is called tamely ramified and wildly ramified otherwise. The inertial degree of a place is denoted $f(\mathfrak{p}|P)$ and has value $[ \mathcal{O}_\mathfrak{p}/\mathfrak{p} : \mathcal{O}_P/(P)]$ if $P$ is a finite place and $[\mathcal{O}_\mathfrak{p}/\mathfrak{p}:\mathbb{F}_q]$ for the infinite place.
Knowing the splitting behavior of places is a key component to determine the genus of the function field $\mathcal{F}$. We now turn our attention to characterizing the splitting behavior of all the places, starting with the finite places and concluding with the infinite place.
\[t:n\_finiteplaces\] Let $P \in \mathbb{F}_q[x]$ be an irreducible polynomial and let $q_1 = q^{deg(P)}$. Also let $a$ and $b$ be defined by $T^3 - aT + b \equiv T^3 - AT + B \pmod{P}$. Then the principal ideal $(P)$ splits into prime ideals in $\mathcal{O}_\mathcal{F}$ as follows:
1. If $v_P(\Delta) > 2 $ then $(P) = \mathfrak{p}^3$.
2. If $v_P(\Delta) = 1 $ then $(P) = \mathfrak{qp}^2$.
3. Otherwise $P\nmid A$, $d = gcd( T^{q_1} - T, T^3 - aT + b)$, and we consider three cases:
1. If $\deg d = 0$ then $(P) = \mathfrak{p}$.
2. If $\deg d = 1$ then $(P) = \mathfrak{pq}$.
3. If $\deg d = 3$ then $(P) = \mathfrak{pqr}$.
For primes not dividing $A$, $\{1, y, y^2\}$ is an integral basis of $\mathcal{O}_P[y]/\mathcal{O}_P$ and thus Kummer’s Theorem (see Theorem III.3.7 of [@sticht]) may be applied to get the desired result. By Dedekind’s Different Theorem, ramified primes are distinguished by the multiplicity with which they divide the field discriminant and thus the the two ramified cases are as claimed (see Theorem III.5.1 in [@sticht]).
While we could consider a transformation to bring the infinite place to a finite place and invoke Kummer’s Theorem as above, there is no guarantee that the infinite place is nonsingular. We will avoid this approach and appeal to completions using Theorem 3.1 of [@exp_cff]. This theorem implies that there will be a root in $\mathbb{F} \langle x^{-1} \rangle$, where $\mathbb{F}$ is some finite extension of $\mathbb{F}_q$, if and only if the infinite place is not wildly ramified. We will show that a curve in the form of (\[tame\]) characterizes the infinite place being tamely ramified or unramified by constructing just such a root in $\mathbb{F} \langle x^{-1} \rangle$. From the construction, it will then be a matter of counting the number of roots, and hence finding $[\mathbb{F}: \mathbb{F}_q]$ as this corresponds to the inertial degree. After dealing with the four cases that arise in this situation, it will be clear why (\[wildram\]) implies that the place at infinity is wildly ramified.
Assume the curve is in standard form and satisfies (\[tame\]). Consider constructing a root $y \in \mathbb{F} \langle x^{-1} \rangle$ of $\phi(T)$. We can write $$y = y_nx^n + y_{n-1}x^{n-1} + \ldots$$ where $y_i \in \mathbb{F}$. Let $A(x) = a_{2n}x^{2n} + \ldots + a_0$ and $B(x) = b_{3n}x^{3n} + \ldots b_0$ with $a_i, b_i \in \mathbb{F}_q$. By writing the polynomials this way, we only assume that either $a_{2n}$ or $a_{2n-1}$ is nonzero. If $a_{2n}=0$ then $b_{3n} = 0$ and $b_{3n-1}= 0$ in order to satisfy (\[tame\]). The coefficients of the powers of $x$ in the equation $y^3 - A(x)y + B(x) = 0$ are as follows:
$$\begin{aligned}
x^{3n} \ : & \ y_n^3 - a_{2n}y_n + b_{3n} \\
x^{3n-1} \ : &\ -a_{2n-1}y_n - a_{2n}y_{n-1} + b_{3n-1}\\
x^{3n-2} \ : &\-a_{2n-2}y_n - a_{2n-1}y_{n-1} - a_{2n}y_{n-2} + b_{3n-2}\\
x^{3n-3} \ : &\ y_{n-1}^3 -a_{2n-3}y_n - a_{2n-2}y_{n-1} - a_{2n-1}y_{n-2} -a_{2n}y_{n-3} + b_{3n-3}\\
\vdots \quad \ : & \ \vdots \end{aligned}$$
The equation associated with $x^{3n}$ is cubic in $y_n$. After the initial cubic equation, we have an equation associated to $x^{3n-i}$ that is linear in $y_{n-i}$ for $i > 0$. That is, the values for $y_{n-i}$ are uniquely determined by the initial choice for $y_n$. Therefore, we examine the solutions to $Y^3 - a_{2n}Y + b_{3n} = 0$.
If $\deg(A)$ is odd then $a_{2n} = 0$ and $Y^3 + b_{3n} = 0$ has exactly one solution. This occurs in the partially ramified case because there is exactly one root in $\mathbb{F}_q \langle x^{-1} \rangle$ and another distinct root in $\mathbb{F}_q \langle x^{-1/2} \rangle$. Otherwise when $a_{2n} \neq 0$, the finite field extension which contains $y_n$ determines the splitting type. The completely split case has three distinct roots in $\mathbb{F}_q \langle x^{-1} \rangle$. In this case $Y^3 - a_{2n}Y + b_{3n}$ splits completely in $\mathbb{F}_q[Y]$. The inert case will have a root in $\mathbb{F}_{q^3} \langle x^{-1} \rangle$ and therefore $Y^3 - a_{2n}Y + b_{3n}$ is irreducible over $\mathbb{F}_q[Y]$. The last case corresponds to the polynomial factoring as a linear and an irreducible quadratic in $\mathbb{F}_q[Y]$. Therefore the place is partially split.
Suppose the curve in standard form satisfies (\[wildram\]). If we let $b_k$ be the leading coefficient of $B(x)$ (with $3 \not | k$, and $k>3n$), then the first equation is $b_k = 0$ which is a contradiction. This forces the above system of equations to be inconsistent. Therefore a curve in standard form satisfying (\[wildram\]) must have a wildly ramified place at infinity. We summarize the above discussion in following theorem.
\[infinity\_splitting\] The place at infinity splits as follows.
1. If $\phi(T)$ satisfies (\[wildram\]) then $(\infty) = \mathfrak{p}^3$.
2. If $\phi(T)$ satisfies (\[tame\]) and $\deg(A)$ is odd then $(\infty) = \mathfrak{p}\mathfrak{q}^2$.
3. If $\phi(T)$ satisfies (\[tame\]) and $\deg(A)$ is even then $ d = {\rm gcd}(T^q - T, T^3 - a_{2n}T + b_{3n})$ determines the splitting type.
1. If $\deg d = 0$ then $(\infty) = \mathfrak{p}$.
2. If $\deg d = 1$ then $(\infty) = \mathfrak{p}\mathfrak{q}$.
3. If $\deg d = 3$ then $(\infty) = \mathfrak{p}\mathfrak{q}\mathfrak{r}$.
This gives a classification of the signature of the function field that involves a simple computation using very basic parameters of the curve. We can now turn our attention to calculating the genus of the function field.
Genus
=====
We will calculate the genus with the Hurwitz Genus Formula (see Theorem III.4.12 in [@sticht]), which requires knowledge of the degree of the different. Having the field discriminant, Dedekind’s Different Theorem gives the different exponents for the finite places. The infinite place is more complicated. However, since the problematic case is when wild ramifiation occurs, we are fortunate in that wild ramification is also total ramification. For totally ramified places, determining the different exponent is a matter of finding a uniformizer for the place and evaluating a particular valuation (see Theorem III.5.12 of [@sticht]).
If the place at infinity is totally ramified then it has different exponent $\delta_{\infty} = 2\deg B - 3\deg A + 2$.
Let $\mathfrak{p}$ be the place at infinity in ${\mathcal{F}}$. Since it is totally ramified and must lie above the unique infinite place in ${\mathbb{F}_q}(x)$ with uniformizer $1/x$, $v_{\mathfrak{p}}(x) = -3$. By examining the equation $y^3 - A(x)y + B(x) = 0$, we can determine $v_{\mathfrak{p}}(y) = -\deg B$. We will apply Theorem III.5.12 of [@sticht]. The uniformizer used in the theorem will depend on which residue class $\deg B$ resides in modulo $3$.
If $\deg B = 3m - 1$ then a uniformizer of ${\mathfrak{p}}$ is given by $t = y/x^m$. The minimal polynomial for $t$ is $f(t) = t^3 - Atx^{-2m} + Bx^{-3m}$. Applying the theorem we see $$\delta_{\infty} = v_{\mathfrak{p}}(f'(t)) = v_{\mathfrak{p}}(Ax^{-2m}) = -3\deg A + 6m = 2\deg B -3\deg A + 2.$$ The case $\deg B = 3m + 1$ follows in a similar manner.
If $\mathcal{F}$ has a totally ramified place at infinity then the genus of $\mathcal{F}$ is $g = \deg B - \deg I - 1.$
The Hurwitz Genus Formula gives $$2g - 2 = -2[\mathcal{F}:\mathbb{F}_q(x)] + \sum_{\mathfrak{p} \in \mathbb{P}_{\mathcal{F}} } d(\mathfrak{p}|P).$$ This yields $$2g - 2 = -6 + (3\deg A - 2 \deg I ) + (2\deg B - 3 \deg A + 2),$$ which upon simplification gives the desired result.
If $\mathcal{F}$ has a place at infinity that is tamely ramified or unramified, then $g = (3\deg{A} - 2\deg I + \delta_{\infty} - 4 )/2$ where $\delta_{\infty} = 0$ if $\deg A$ is even and $\delta_{\infty} = 1$ if $\deg A$ is odd.
The proof follows as before except now that the infinite place is not wildly ramified, and hence its different exponent $\delta_{\infty}$ can only take the values 0 or 1. Since the genus is an integer, the parity of $\deg A$ determines the value of $\delta_{\infty}$.
We have described the basic invariants of cubic function fields in characteristic three. The focal point for the remainder of this paper is to develop the arithmetic of ideals. As in the previous sections, one can appeal to generic algorithms to solve this problem. However, these algorithms typically require operations on large matrices or an appeal to Groebner basis. We desire, like elliptic curves and hyperelliptic curves (using Cantor’s algorithm), a method to do computations that depends only on the underlying curve parameters and the finite field.
Divisor Class Groups and Ideal Class Groups
===========================================
This section provides an overview of the relationship between the Jacobian of a curve and the ideal class group of a function field. As there are many sources for this material (see e.g. [@bauer], [@landquist], [@hasse] [@aacff], [@idealarithmetic]), we will be relatively brief and only provide the relevant definition and results where needed. Once this is completed, it will be possible to develop arithmetic on ideals and, for a certain class of curves, fully realize arithmetic in the ideal class group.
A divisor is a finite formal sum of places in $\mathcal{F}$. The set of all divisors forms a free abelian group. We will work in a specific finite subgroup of this group. Let $S$ be the set of finite places in $\mathcal{F}$. There is an isomorphism between the divisors with support in $S$, $\mathcal{D}_F(S)$, and the fractional ideals in $\mathcal{O}_\mathcal{F}$, $\mathcal{I}(\mathcal{O}_\mathcal{F})$. The *Fundamental theorem of ideal theory in an algebraic function field* [@hasse p 401] gives the isomorphism as $$\label{fund_iso} \Phi \ : \ \mathcal{D}^0_\mathcal{F}(S) \rightarrow \mathcal{I}(\mathcal{O}_\mathcal{F}), \quad D \longmapsto \left\{ \alpha \in F^{\times} \ \vrule \ \sum_{P \in S}v_P(\alpha)P \geq D \right\} \cup \{ 0 \}.$$
This may also be defined by $$\sum n_PP \longmapsto \prod_{P \in S} (P \cap \mathcal{O}_\mathcal{F})^{n_P}.$$ In general the ideal class group is related to the Jacobian by the following exact sequence (see Theorem 14.1 of [@rosen]) $$(0) \rightarrow \mathcal{D}_\mathcal{F}(S^c)/\mathcal{P}_\mathcal{F}(S^c) \rightarrow \mathcal{J}_\mathcal{F} \rightarrow \mathcal{C}l(\mathcal{O}_F) \rightarrow \mathbb{Z}/f\mathbb{Z} \rightarrow (0),$$ where $S^c$ is the set of infinite places (the set compliment of $S$ in $\mathbb{P}_\mathcal{F}$). Specifically, if a function field has a unique place at infinity of degree 1, the points on the Jacobian will be isomorphic to the ideal class group.
We use the hierarchy of divisors (and hence ideals) defined in [@bauer] so that there is a way to represent elements of the divisor class group of degree zero in a unique way with minimal information. A divisor $D$ is *effective* if $D>0$ (that is, $n_P \geq 0$ for all $P \in \mathbb{P}_F$) and denote its effective part as $D^+$, i.e. $$D= \sum_{P\in \mathbb{P}_\mathcal{F}} n_PP \quad \Rightarrow \quad D^+ = \sum_{P\in \mathbb{P}_\mathcal{F}, n_P > 0} n_PP .$$ A degree zero divisor is called *finitely effective* if its finite part is effective; it can be shown that every divisor $D \in \mathcal{D}_\mathcal{F}^0$ is equivalent to a finitely effective divisor. This is the first step in the hierarchy.
A finitely effective divisor is *semi-reduced* if there does not exist a non-empty sub-sum of the form $(\alpha)$ where $\alpha \in \mathbb{F}_q[x] \backslash \mathbb{F}_q$. Again, it is straightforward to show that every divisor is equivalent to a semi-reduced divisor, extending the hierarchy. A *semi-reduced* divisor $D$ is *reduced* if $\deg D^+ \leq g$ where $g$ is the genus of the curve. Using the Riemann-Roch Theorem, it is possible to prove that every divisor class also contains a reduced divisor.
To complete the hierarchy, we define a *distinguished* divisor to be a divisor $D$ such that for all other equivalent finitely effective divisors $D_1$, we have that $\deg D_1^+ \leq \deg D^+$ implies $D = D_1$. If a divisor is distinguished, it is reduced [@bauer Lemma 1.12]. Unfortunately, we have no apriori way of knowing if such a divisor exists or of verifying that a divisor is distinguished.
The above definitions for a divisor $D$ can immediately be transferred to fractional ideals by first considering $D^+$ and then applying the isomorphism . Finitely effective divisors map to integral ideals, and hence we can do computations in this context. Note that in the ideal class group we will mostly work with primitive ideals, that is: $\mathfrak{a}$ is primitive if and only if there is no non-constant polynomial $a(x) \in \mathbb{F}_q[x]$ such that $\langle a(x) \rangle \ | \ \mathfrak{a}$ where $\langle a(x) \rangle$ represents $a(x) \mathcal{O}_F$. Under the above correspondence, we see that primitive integral ideals give an equivalent notion to semi-reduced divisors. We will call an ideal reduced (resp. distinguished) if it is the image under the above correspondence of a reduced (resp. distinguished) divisor. We now turn our attention to determining when it is possible to show that each divisor class, or equivalently, ideal class, contains a distinguished element.
Let $\alpha = a + b\rho + c\omega \in \mathcal{F}$ with $a,b,c \in \mathbb{F}_q(x)$. Then the norm of $\alpha$ is given by $$N_{\mathcal{F}/\mathbb{F}_q(x)}(\alpha) = N(a + b\rho + c\omega)=$$ $$a^3 - a^2cE + abcIF -ab^2A + b^2cAE + bc^2AF - bc^2EFI -c^3F^2I -b^3FI^2.$$
\[t:n\_norm\] Let $\alpha = a + b\rho + c\omega \in \mathcal{O}_\mathcal{F}$, $2 \deg B > 3 \deg A$, and $3\nmid \deg FI^2$ . Then $ \deg{N(\alpha)} = \max\{\deg{a^3}, \deg{b^3FI^2}, \deg{c^3F^2I}\}$.
The proof follows from a careful analysis of the degrees of the relative terms in the norm expression, and noting that the criterion that $3\nmid \deg FI^2$ actually forces $\deg FI^2 = \deg B$ and thus the curve satisfies . A detailed version of the proof can be found in [@webster].
It is natural to wonder if implies $3 \nmid \deg{FI^2}$. Unfortunately, a class of curves exists for which this implication is not true. We can construct a curve such that $3\nmid \deg{B}$ and $3|\deg{FI^2}$. In general we do not expect to deal with such curves; it requires a very special sort of singularity. An example of this type of singularity is given in the following construction.
Consider the function field given with parameters $A = (x^2 + x - 1)(x^2 + 1)$ and $B = - x^8 + x^6 + x^5 + x^4 + x^2 + 1$. These parameters define a curve that is in standard form and satisfies . Both divisors of $A$ are singular, $I = (x^2 + x - 1)(x^2 + 1)$, and $i = x^3 + x^2$. Thus $\deg{FI^2} =\deg{( i^3 - iA + B)} = 9$.
Having established this property of the norm, we can now return to the specifics of distinguished ideals. In particular, Theorem \[t:n\_norm\] is exactly what is needed to extend Theorem 5.1 of [@bauer] to this case.
\[t:minnormexists\] If $2 \deg B > 3 \deg A$, and $3\nmid \deg FI^2$, then every nonzero ideal contains a nonzero element of minimal norm which is unique up to multiplication by an element in $\mathbb{F}_q^{\times}$.
The proof is identical to that in [@bauer] but we sketch the key points. The validity is established using Theorem \[t:n\_norm\]. Assume there are two elements $\alpha_i = a_i + b_i\rho + c_i\omega$ for $i = 1,2$ whose norm has the same degree and suppose $\deg N(\alpha_i) = \deg a_i^3$. Let $k$ be the quotient of the leading term of $a_1$ divided by the leading term of $a_2$ then $\alpha_3 = \alpha_1 - k\alpha_2$ has smaller norm. A similar argument works when the degree of the norm is determined by $b_i$ or $c_i$.
\[distinguished\] If $2 \deg B > 3 \deg A$, and $3\nmid \deg FI^2$, then every ideal class contains a unique distinguished ideal.
By Theorem \[t:minnormexists\] there exists $\alpha_1 \in J$ of minimal norm from some ideal $J$. We consider the primitive integral ideal $J_1 = \langle \alpha_1 \rangle J^{-1}$ . Assume there is some integral primitive ideal $J_2$ equivalent to $J_1$ with $\deg N(J_2) \leq \deg N(J_1)$. Then $J_2J = \langle \alpha_2 \rangle$ with $\alpha_2 \in J$. This gives $\deg N(J_1) = \deg N(\alpha_1) - \deg N(J) \geq \deg N(J_2) = \deg N(\alpha_2) - \deg N(J)$ which implies $\deg N(\alpha_1) \geq \deg N(\alpha_2)$. By assumption $\alpha_1$ is an element of minimal norm; therefore $\alpha_2 = k\alpha_1$ for $k \in \mathbb{F}_q^{\times}$ and $J_1 = J_2$. Thus every ideal class contains a unique distinguished ideal.
All the theoretical pieces are in place to develop arithmetic in the ideal class group. Ideal inversion and multiplication pose no major theoretical obstacles, and the above establishes a unique way to find a distinguished ideal in a given class. Combining all of the pieces will allow composition and reduction in the ideal class group. The remaining sections make the above explicit for the considered function fields. We start by describing an integral basis for primes and their powers which gives insight to how inversion and multiplication will work as explained afterward. Finally, we give explicit algorithms that produce the element of minimal norm and its corresponding distinguished ideal.
Triangular basis for prime ideals
=================================
Having described how the finite places split, it will be helpful to have a concrete description of generators for the prime ideals in terms of the basis elements developed in Section 4. Scheidler provided a comparable statement in Theorem 3.1 of [@idealarithmetic] for all prime ideals in a purely cubic function field of characteristic not 3 that was an analog of the theorem of Voronoi [@vor] for number fields. Having classified the splitting type of prime ideals, we follow their lead and give the triangulr bases along with basic products and powers of the prime ideals.
Throughout the following sections, proofs will occasionally be omitted for the sake of brevity. In particular, when a particular technique may be used successfully to compute the basis in multiple cases, it will only be included once. The interested reader may always refer to [@webster] for complete proofs.
Ramified primes
---------------
There are three cases to consider for the ramified primes. When calculating powers of primes, ramification tends to make the treatment here a little easier for a given prime. A ramified prime is expected to be totally ramified so $\mathfrak{p}^3 = (P)\mathcal{O}_\mathcal{F} = (P)[1, \rho, \omega]$ for some irreducible polynomial $P \in {\mathbb{F}_q}[x]$. This leaves only the calculation of the basis for $\mathfrak{p}$ and $\mathfrak{p}^2$.
\[t:wildrambasis\] Let $v_P(A) \geq 1$ and $v_P(I) = 0$ so that $(P) = \mathfrak{p^3}$. Then $$\mathfrak{p} = [P, f + \rho,- I^{-1}f^2 + \omega]
\mbox{ \ and \ }
\mathfrak{p}^2 = [P, P\rho, I^{-1}f^2 - I^{-1}f\rho + \omega]$$ where $f^3 \equiv FI^2 \pmod{P}$, and $I^{-1}I \equiv 1 \pmod{P}$.
Apply Kummer’s theorem to the minimal polynomial of $\rho$ to see $$\rho^3 -A\rho + FI^2 \equiv \rho^3 + FI^2 \equiv (\rho + f)^3 \pmod{P}.$$ This implies $\mathfrak{p} = \langle P, f + \rho \rangle$. To get the last element we consider $\rho (f + \rho) = f\rho + I \omega + A \in \mathfrak{p}$. Using that $I$ is relatively prime to $P$ we get the last element as claimed.
For $\mathfrak{p}^2$, consider $(f + \rho)^2 = f^2 - f\rho + I\omega + A \in \mathfrak{p}^2$ to see that the third term in the basis has the form claimed. Note $v_{\mathfrak{p}}(P) = 3$, so $P \in \mathfrak{p}^2$. Since the ideal has to have norm $P^2$ this forces the second element of the basis to be as stated.
The next two primes we consider both lie over primes dividing the index. Unlike the above primes and the unramified primes (as we will see in the next subsection), these primes can give rise to ideals that will have $\omega$ with a polynomial coefficient.
\[sing\_wildram\] Let $v_P(A) > 1$ and $v_P(I) = 1$ so that $(P) = \mathfrak{p}^3$. Then $$\mathfrak{p} = [P, \rho, \omega]
\mbox{ \ and \ }
\mathfrak{p}^2 = [P, \rho, P\omega].$$
By applying Kummer’s theorem to the minimal polynomials of $\rho$ and $\omega$, we see that both are in $\mathfrak{p}$. Squaring $\mathfrak{p}$ and considering the 9 resulting elements leads to the latter result.
Tamely ramified primes are all that remain to consider. We state their bases here, but we will treat powers of $\mathfrak{p}$ and $\mathfrak{q}^2$ with unramified primes.
\[splitramified\]\[splitram\] Let $v_P(A) = 1$ and $v_P(I) = 1$ so that $(P) = \mathfrak{pq}^2$. Then $$\mathfrak{p} = [P, \rho, E + \omega], \quad \mathfrak{q} = [P, \rho, \omega],$$ $$\mathfrak{q}^2 = [P, P\rho, E^{-1}F\rho + \omega],\mbox{ and }\ \mathfrak{pq} = [P, \rho, P\omega]$$ where $E^{-1}$ is the inverse of $E$ modulo $P$.
Apply Kummer’s theorem to the minimal polynomials of $\rho$ and $\omega$ and proceed as before.
Unramified primes
-----------------
The unramified primes are the primes that we expect to deal with most often in the course of doing computations. There are several different cases to be accounted for depending on the inertial degree and the primes involved in the product. First we deal with the case in which $\mathfrak{p}$ has inertia degree 1, and then when $\mathfrak{p}$ has inertia degree 2. This case will also handle the product of two primes of inertia degree 1 and the powers of $\mathfrak{q}^2$ for the split ramified case. The last case will be where $\mathfrak{p}$, $\mathfrak{q}$ both have inertial degree 1 and we need to consider products of the form $\mathfrak{p}^i\mathfrak{q}^j$ with $i \neq j$.
\[unramified\] Let $\mathfrak{p}|P$ have inertial degree 1 and ramification index 1. Then $$\mathfrak{p} = [ P, - \alpha + \rho, - I^{-1}(\alpha^2 - A) + \omega]$$ where $\alpha$ is a root of the minimal polynomial of $\rho$ modulo $P$ and $I^{-1}I \equiv 1 \pmod{P}$.
The ideal $\mathfrak{p}$ is generated by $\langle P, -\alpha + \rho \rangle$, and the rest follows.
The proposition below deals with primes that have ramification index 1 and inertia degree 1. It therefore also handles the unramified prime lying over the tamely ramified primes.
\[t:n\_single\] For $\mathfrak{p}$ with ramification index 1 and inertial degree 1, we have $$\mathfrak{p}^i= [P^i, -X_i + \rho, -Z_i + \omega]$$ where
- $Z_{i+1} = Z_i + kP^i$,
- $k \equiv -C_i(EZ_i)^{-1} \pmod{P}$,
- $C_i = -(Z_i^3 - EZ_i^2 + F^2I)/P^i$,
- $X_{i+1} \equiv -FIZ_{i+1}^{-1} \pmod{P}$,
and $X_1$ and $Z_1$ are defined and given in Propositions \[splitramified\] and \[unramified\].
The definitions in this proposition make it important that $Z_1$ be invertible modulo $P$. In Proposition \[splitramified\], $E$ is invertible modulo $P$. For Proposition \[unramified\] the element $Z_1$ is invertible because it is a nonzero root of the minimal polynomial of $\omega$ modulo $P$, that is to say only ramified primes correspond to $0$ being a root modulo $P$.
Since $P^i|N(\omega - Z_i)$, that basis element can be written as $\omega - (Z_i + kP^i)$. We now describe how to choose $k$ so that the element is correct for $\mathfrak{p}^{i+1}$. $$\begin{aligned}
N(- (Z_i + kP^i) + \omega ) & = -[(Z_i + kP^i)^3 + E(Z_i + kP^i)^2 +F^2I] \\
&\equiv -(Z_i^3 - EZ_i^2 + F^2I) - EZ_ikP^i \pmod{P^{i+1}}\\
&\equiv -(C_iP^i - EZ_ikP^i) \pmod{P^{i+1}}\end{aligned}$$ Since we want $C_iP^i - EZ_ikP^i \equiv 0 \pmod{P^{i+1}}$, we can choose $k \equiv C_i(EZ_i)^{-1} \pmod{P}$. Such an inverse exists because $P$ is relatively prime to both $E$ and $Z_i$. Now that $ - Z_{i+1} + \omega \in \mathfrak{p}^{i+1}$ we can see that $(Z_{i+1} - \omega)\rho = FI + Z_{i+1}\rho \in \mathfrak{p}^{i+1}$. This gives the term with $-X_{i+1} + \rho$ as claimed.
\[partsplit\] Let $\mathfrak{q}$ be a prime with inertia degree 2. Then $$\mathfrak{q} = [ P , P\rho, I^{-1}(W + A) - I^{-1}M\rho + \omega]$$ where $\rho^3 - A\rho + FI^2 \equiv (\rho-\alpha)(\rho^2 - M\rho + W) \pmod{P}$.
Kummer’s theorem gives $\mathfrak{q} = \langle P, \rho^2 - M\rho + W \rangle$, and similar techniques complete the proof.
Notice the form of the product of two distinct unramified primes lying over a completely split prime: $$\begin{aligned}
\mathfrak{pq} &= \langle P,- \alpha_1 + \rho \rangle \langle P, - \alpha_2 + \rho \rangle \\
&= \langle P^2, P(-\alpha _ 1 + \rho), P(- \alpha_2 + \rho), A +\alpha_1\alpha_2 -(\alpha_1 + \alpha_2)\rho +I\omega \rangle \\
&= [ P, P\rho, I^{-1}(A +\alpha_1\alpha_2) -I^{-1}(\alpha_1 + \alpha_2)\rho + \omega ].\end{aligned}$$ Here the last line is justified by the fact that $(\alpha_1 - \alpha_2)$ is relatively prime to $P$. Thus, the greatest common divisor of $P(\alpha_1 - \alpha_2)$ and $P^2$ is $P$. There are three types of ideals that can have the form $[P, P\rho, -N_1 - M_1\rho + \omega]$:
- $\mathfrak{q} = [P, P\rho, \omega - M\rho - W]$ from Proposition \[partsplit\],
- $\mathfrak{q}^2 = [P, P\rho, E^{-1}F\rho + \omega ]$ from Proposition \[splitramified\], and
- $\mathfrak{pq} =[ P, P\rho, I^{-1}(A +\alpha_1\alpha_2)-I^{-1}(\alpha_1 + \alpha_2)\rho + \omega]$ from the exposition above.
\[t:n\_double\] Let $\mathfrak{r}$ represent any of the three ideals above. Then $$\mathfrak{r}^i = [P^i, P^i\rho, N_i - M_i\rho + \omega]$$ where
- $L(M_{i-1}M_1I + N_{i-1} + N_1 - E) \equiv 1 \pmod{P^i}$,
- $M_i \equiv -L(F + M_{i-1}N_1 + M_1N_{i-1}) \pmod{P^i}$, and
- $N_i \equiv L( M_1FI + M_{i-1}FI M_{i-1}M_1A N_{i-1}N_1) \pmod{P^i}$.
The previous work establishes the base case $i = 1$ and we argue by induction.
$$\mathfrak{r^{i-1}}\mathfrak{r} = [ P^{i-1}, P^{i-1}\rho, \omega - M_i\rho + N_i][P, P\rho, \omega - M_1\rho + N_1].$$
In the nine possible products of the basis elements only $ (\omega - M_i\rho + N_i)(\omega - M_1\rho + N_1)$ does not contain a factor of $P$. Thus the coefficient of $\omega$ has to be relatively prime to $P$. If it were not, the product would not be primitive. Multiplying through by its inverse modulo $P^i$ gives the desired basis element.
The product contains $P^i$ and $P^i\rho$. A norm argument shows that $\mathfrak{r}^i$ cannot contain $P^{i-1}$ or $P^{i-1}\rho$. Thus the ideal has the desired norm and the elements stated form a basis.
All that remains is to handle $\mathfrak{p}^i\mathfrak{q}^{i + j}$ where $j>0$ and each prime has inertia degree 1. By Proposition \[t:n\_double\] we know $$\begin{aligned}
\label{prod1} ( \mathfrak{pq})^i& = [ P^i, P^i\rho, N_i - M_i\rho + \omega] \\
\intertext{and by Proposition \ref{t:n_single}}
\label{prod2} \mathfrak{q}^j &= [P^j, - X_{\mathfrak{q}_j} + \rho , - Z_{\mathfrak{q}_j} + \omega ], \\
\label{prod3} \mathfrak{q}^{i+j} &= [P^{i+j}, - X_{\mathfrak{q}_{i+j}} + \rho, - Z_{\mathfrak{q}_{i+j}} + \omega ], \mbox{\quad and } \\
\label{prod4} \mathfrak{p}^i &= [P^i, - X_{\mathfrak{p}_i} + \rho, - Z_{\mathfrak{p}_i} + \omega ]. \end{aligned}$$ Combinations of the above products will help determine the proper basis of $\mathfrak{p}^i\mathfrak{q}^{i + j}$.
\[partial\_split\] Using notation as above $$\mathfrak{p}^i\mathfrak{q}^{i + j} = [ P^{i+j}, P^i( - X_{\mathfrak{q}_j}+ \rho ), H + G\rho + \omega ]$$ where we let $N$ be defined by $NX_{\mathfrak{p}_i} \equiv 1 \pmod{P^{i+j}}$ and $$G \equiv NZ_{\mathfrak{q}_{i+j}} \pmod{P^i} \mbox{ \quad and \quad } H \equiv N(-FI - X_{\mathfrak{p}_i}Z_{\mathfrak{q}_{i+j}}) \pmod{P^{i+j}}.$$
Considering the product of and , we see that $P^{i+j}$ and $P^i( - X_{\mathfrak{q}_j} + \rho )$ are in $\mathfrak{p}^i\mathfrak{q}^{i + j} $. By considering the product of and , we can see that $$( - X_{\mathfrak{p}_i} + \rho) (- Z_{\mathfrak{q}_{i+j}} + \omega ) \in \mathfrak{p}^i\mathfrak{q}^{i + j}$$ Since $X_{\mathfrak{p}_i}$ is relatively prime to $P$ it is invertible modulo $P^{i+j}$. Multiplying through by its modular inverse gives the third element of the basis. The other two elements are in the ideal by construction. It remains to establish that they are indeed basis elements, which is easily accomplished by a norm argument.
We have dealt with all of the prime ideals and their possible powers and products. We now turn to arbitrary ideal arithmetic. Any given ideal factors into the product of four ideals: $$J = [s_1, s_1'(u_1 + \rho), v_1 + w_1\rho + \omega][s_2, s_2(u_2 + \rho), v_2 + w_2\rho + \omega]$$ $$[s_3, \rho, s_3''\omega][s_4, s_4'(u_4 + \rho), s_4''(v_4 + w_4\rho + \omega)] = J_1J_2J_3J_4.$$ Each of the ideals is relatively prime to the others and is determined by which type of prime appears in the factorization. For $\mathfrak{p}|P$, we have four criteria:
- $\mathfrak{p}$ divides $J_1$ iff $P$ is unramified,
- $\mathfrak{p}$ divides $J_2$ iff $P$ is totally ramified and does not divide the index,
- $\mathfrak{p}$ divides $J_3$ iff $P$ is totally ramified and divides the index, and
- $\mathfrak{p}$ divides $J_4$ if and only if $P$ is split ramified.
We call these primes Type I, Type II, Type III and Type IV, respectively. Recombining ideals factored in this way is a straightforward application of the Chinese Remainder Theorem, while finding the factorization for a given ideal is an application of polynomial factorization. There are a few reasons for this approach. The first is for simplicity as the propositions are easier to state for a given type. The combined proposition would look much like each constituent part where the final result is an application of the Chinese Remainder Theorem. The second reason is that the difficulty often lies in a particular case and this allows the exposition to highlight the trouble.
Furthermore, from a computational perspective, we are also drawn to this approach. Two of the four cases involve curves that have singularities, and hence we can choose to avoid them. We could also easily choose a curve with no finite ramification and ignore three of the four cases. Even in the worst case scenario where all types of primes are possible, we still do not expect to deal with three of the four products in the course of doing arithmetic. A rough heuristic argument shows that probability of two randomly chosen ideals with degree less than $g$ contain a ramified prime is $4g/q$, which will be small if $q$ is large. Thus from a computational point of view, three of the four cases will rarely occur even when a curve is singular.
Inversion and Division
======================
Some basic properties of the structure of ideals in cubic function fields developed in [@idealarithmetic] remain true even in characteristic three. We cite without proof the containment criterion for ideals written with a triangular basis.
[**(Lemma 4.1 of [@idealarithmetic])**]{} Let $I_i = [s_i, s_i'(u_i + \rho), s_i''(v_i + w_i\rho + \omega)]$ for $i = 1,2$ be two ideals. Then $I_1 \subseteq I_2$ if and only if $$s_2 | s_1, \quad s_2' | s_1', \quad s_2''|s_1'', \quad s_1'u_1 \equiv s_1'u_2 \pmod{s_2},$$ $$s_1''w_1 \equiv s_1''w_2 \pmod{s_2'},\mbox{ \ and \ } s_1''v_1 \equiv s_1''(v_2 + u_2(w_1-w_2)) \pmod{s_2}.$$
Our first goal is to develop ideal inversion. As we only wish to work with integral ideals, we compute a primitive ideal that is in the ideal class of the inverse of a given ideal. As a reminder, the notation for such an inverse will be $\overline{J}$ and the notation for division will be $J^{-1}$. We label the propositions depending on which of the four ideal types the proposition covers.
\[t:inverse\_s\_1\][**(Inversion for Type I and II primes)** ]{} If $I_1 = [s, s'(u + \rho), v + w\rho + \omega]$, then $I_2 = \overline{I_1} = \langle s \rangle I_1^{-1}$ is given by $I_2 = [S, S'(U+ \rho), V + W\rho +\omega]$, where $$S = s, \quad S' = s/s', \quad U \equiv -Iw \pmod{s'},$$ $$W \equiv -uI^{-1} \pmod{s/s'}, \ \mbox{ and } \ V \equiv E - v - WIw \pmod{s}.$$
Since $s \in I_1$, it is clear that $\langle s \rangle I_1^{-1}$ is an integral ideal. We show that the above choices provide a correct ${\mathbb{F}_q}[x]$ basis for $I_2$. The fact that $I_1I_2 = \langle s \rangle$ will be used extensively in this proof (and the proofs to follow). Since $s \in I_2$, $S|s$. Examining $S(v + w\rho + \omega) \in \langle s \rangle $, we conclude $s|S$ and hence $s = S$. Consider the norm of the ideal $\langle s \rangle$ to determine $S'$: $$s^3 = N( \langle s \rangle ) = N(I_1)N(I_2) = ss'sS'.$$ Therefore $S' = s/s'$ as claimed. The other products of the two ideals give the remaining congruences. We start with $S'(U + \rho)(v + w\rho + \omega)$ and examine the coefficient of $\omega$: $$S'(U + \rho)(v + w\rho + \omega) \in \langle s \rangle \Rightarrow s \ \vrule \ \frac{s}{s'}(U + Iw) \Rightarrow U \equiv -Iw \pmod{s'}.$$ The congruence for $W$ (resp. $V$) follows by considering the coefficient of $\omega$ in the product $s'(u + \rho)(V + W\rho + \omega)$ (resp. $(V + W\rho + \omega)(v + w\rho + \omega)$ ) and arguing as above.
We note that the above proposition is simpler for nonsingular curves because $I = 1$ and most of the congruences can be replaced by equalities. It was fortunate that we could deal with Type I and Type II primes with a single proposition without an appeal to the Chinese Remainder Theorem. We will often be able to deal with these two types of primes together.
\[t:inverse\_s\_3\][**(Inversion for Type III primes)** ]{} If $I_1 = [s, \rho, s''\omega]$, then $I_2 = \overline{I_1} = \langle s \rangle I_1^{-1}$ is given by $I_2 = [s, \rho, (s/s'')\omega]$.
This follows immediately from Proposition \[sing\_wildram\].
Here the index divisors do not seem too complicated, but for Type IV ideals they will prove very troublesome. The approach taken for this case will become familiar. Just as factorization helps simplify the four cases, factorization within this case will prove useful. Our approach is to consider a factorization that most closely resembles some of the basic prime powers of Section 7. The treatment of Proposition \[partial\_split\], that dealt with products of the form $\mathfrak{p}^i\mathfrak{q}^{i+j}$, is enlightening here. In the proof we appeal to the fact that the product can be viewed as $(\mathfrak{pq})^i$ and $\mathfrak{q}^j$. This is the sort of factorization that we will use in many of the following propositions.
\[t:inverse\_s\_4\][**(Inversion for Type IV)**]{} If $I_1 = [s, s'(u + \rho), s''(v + w\rho + \omega)]$, then $I_2 = \overline{I_1} = \langle s \rangle I_1^{-1}$ is given by $I_2 = [S, S'(U+ \rho), S''(V + W\rho +\omega)]$, where $$S = s, \quad S' = \frac{s}{s's''s_I}, \quad S'' = s_I = \gcd\left( \frac{s}{s's''} , v \right)$$ $$U \equiv \left\{ \begin{array}{ll}
0 & \pmod{s''s_I} \\
-Iw & \pmod{s'}
\end{array}
\right. , \quad \quad
V \equiv \left\{ \begin{array}{ll}
0 & \pmod{ s'' }\\
0 & \pmod {s/s's''s_I}\\
E & \pmod{ s' }
\end{array}
\right. ,$$ $$W \equiv E^{-1}F \pmod{ s/s's''s_I},$$ and $s''$, $s/s's''s_I$, and $s'$ are pairwise coprime.
Factor $I_1$ as $$I_1 = [s'', \rho, s''\omega][s', s'\rho, w\rho + \omega] \left[ \frac{s}{s's''}, u + \rho, v + w\rho + \omega \right].$$ By factoring the ideal in this fashion, we can find the inverse of each factor. The inverse of the first two factors is an immediate consequence of Proposition \[splitram\]. The inverse of the last ideal in the above factorization has two factors since it could contain either ramified primes or powers of unramified primes, which is determined by the term associated with $\omega$ and $s_I$. For the ramified primes in this product, the inverse is $[s_I, \rho, s_I\omega]$ and this gives $V \equiv 0 \pmod{s_I}$. The remaining factor of the inverse has the form $$\left[ \frac{s}{s's''s_I}, \frac{s}{s's''s_I}\rho, E^{-1}F\rho + \omega \right],$$ yielding the only congruence for $W$ and the remaining congruence for $V$. The above immediately shows that the choices for $S'$, and $S''$ are correct. A quick norm argument shows that $S = s$ as claimed.
The remaining portion of this section leads to arbitrary ideal division. We begin with a series of lemmata that will handle the simplest case of division, and will later be used to handle the general case. Consider two primes $\mathfrak{p}$ and $ \mathfrak{q}$ lying over a completely split place $P$. Given $ (\mathfrak{pq})^i$ and $\mathfrak{p}^i$, we find (somewhat trivially) $(\mathfrak{pq})^i\mathfrak{p}^{-i} = \mathfrak{q}^i$. Thus the product $(\mathfrak{pq})^i$ is split into two ideals of equal norm, $\mathfrak{p}^i$ and $\mathfrak{q}^i$.
\[l:splitting\_s\_1\][**(Splitting for Type I and II primes)**]{} Let $I_2 = [s, s\rho, v_2 + w_2\rho + \omega]$ and $I_1 = [s, u_1 + \rho, v_1 + \omega ]$ be two ideals such that $I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = [s, U + \rho, V + \omega]$, where $$U \equiv Iw_2 - u_1 \pmod{s}, \quad V \equiv v_2 - Iw_2^2 +u_1w_2 \pmod{s}.$$
By Proposition \[t:inverse\_s\_1\], $I_2 = \langle s \rangle [s, Iw_2 + \rho, E - v_2 + \omega]^{-1}$. Therefore we can write $$J [s, u_1 + \rho, v_1 + \omega][s, Iw_2 +\rho, u_2w_2 - v_2 + \omega ] = \langle s \rangle.$$ Since $J = [s, \rho + U, \omega + V]$, it is only a matter of finding the correct congruences for $V$ and $U$. Using $(U + \rho)(u_1 + \rho)(Iw_2 +\rho) \in \langle s \rangle$ and the coefficient of $\omega$, we find $U \equiv Iw_2 - u_1 \pmod{s}$. To find $V$, we note that $v_2 + w_2\rho + \omega \in J$ and subtract $w_2(U + \rho)$.
\[l:splitting\_s\_3\][**(Splitting for Type III primes)**]{} Let $I_2 = [s, \rho, s\omega]$ and $I_1 = [s, \rho, \omega ]$ be two ideals such that $I_1 \subseteq I_2$. Then $J = I_2I_1^{-1} = [s, \rho, \omega]$.
This follows immediately from Proposition \[sing\_wildram\].
\[l:splitting\_s\_4\][**(Splitting for Type IV primes)**]{} Let $I_2 = [s's'', s'\rho, s''(v_2 + w_2\rho + \omega)]$ and $I_1 = [s's'', \rho, v_1 + \omega ]$ be two ideals such that $I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = [s's'', \rho, V + \omega]$, where $$V \equiv E \pmod{d}, \quad V \equiv 0 \pmod{s's''/d}, \ \mbox{ and } \ d=\gcd(s'', v_1).$$
By Proposition \[splitram\], $J = [s's'', \rho, V + \omega]$ for some $V$. For a given prime $P$, $I_2$ contains either $\mathfrak{pq}$ or $\mathfrak{q}^2$ and no higher powers, and the ideal $I_1$ contains either $\mathfrak{p}$ or $\mathfrak{q}$. The quantity $d$ corresponds to the ramified primes in $I_1$. For these primes the unramified conjugate is the inverse, and hence justifies the choice for $V$ modulo $d$.
Rather than proceed straight to the division propositions, we illustrate the method behind the division in Figure 1. The hardest part of division is tracking the various products lying over completely split primes. The figure illustrates the order of operations (as described in the proof) used to complete ideal division. For $\mathfrak{p}$ and $\mathfrak{q}$ lying over a completely split prime $P$ we will walk through the division process in the case that the dividend is $\mathfrak{p}^8\mathfrak{q}^6$ and the divisor is $\mathfrak{p}^5\mathfrak{q}$.
\[div\_walk\]
@C=.3pc@R=.4pc[ & & \^8\^6 @[-]{}\[d\] @[-]{}\[dr\] & & \^5 @[-]{}\[d\] @[-]{}\[dr\] &\
& & \^6\^6 @[-]{}\[d\]& \^2 @[-]{}\[d\] & @[-]{}\[d\] &\^4 @[-]{}\[d\]\
& & \^5\^5 @[-]{}\[d\] & \^2 @[-]{}\[d\] & 1 @[-]{}\[d\] & \^4 @[-]{}\[d\]\
& & \^5\^5 @[-]{}\[d\] @[-]{}\[dl\] & 1 @[-]{}\[d\] & 1 @[-]{}\[d\] & \^2 @[-]{}\[d\]\
& \^3\^3 @[-]{}\[d\] & \^2\^2 @[-]{}\[d\] & 1 @[-]{}\[d\] & 1 @[-]{}\[d\] & \^2 @[-]{}\[d\]\
& \^3\^3 & \^2 & 1 &1& 1\
]{}
The tree for the dividend ends with three branches. It should be noted that the last two nodes on the tree are relatively prime; more specifically, at least one of them is one. This will be key for the next proof because it relies on the product of the those two nodes being relatively prime.
\[t:division\_s\_1\][**(Division for Type I and II primes)**]{} Let $I_i = [s_i, s_i'(u_i + \rho), v_i + w_i\rho + \omega]$ for $i = 1,2$ be such that $I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = [S, S'(U+ \rho), V + W \rho + \omega]$, where $$S = \frac{s_2}{s_1'd}, \quad S' = \frac{s_2'd}{s_1}, \quad
U \equiv \left\{ \begin{array}{ll}
Iw_2 - u_1 & \pmod{ s_1/(s_1'd) }\\
u_2 & \pmod{ s_2/(s_2'd) }
\end{array}
\right. ,$$ $$\quad V \equiv (W - w_2)U + v_2 \pmod{S},\quad W \equiv w_2 \pmod{S'},$$ $$\mbox{ and} \ d = \gcd\left(\frac{s_2}{s_2'}, \frac{s_1}{s_1'}, u_1 - u_2 \right).$$
We begin by factoring both $I_1$ and $I_2$ into two different ideals: $$I_i = I_{i,1}I_{i,2} = [ s_i', s_i'\rho, v_i + w_i\rho + \omega] \left[ \frac{s_i}{s_i'}, u_i + \rho, v_i - u_iw_i + \omega \right].$$ The first division is $$\label{s_1_evendeg} I_{2,1}I_{1,1}^{-1}= \left[ \frac{s_2'}{s_1'}, \frac{s_2'}{s_1'}\rho, v_2 + w_2\rho + \omega \right].$$ All that remains of the divisor is $I_{1,2}= \left[s_1/s_1', u_1 + \rho, v_1 - u_1w_1 + \omega \right].$ We consider the greatest common divisor of this ideal with the corresponding ideal arising from $I_2$. This is the justification for $d$ in the proposition statement. We perform the following division: $$\left[\frac{s_2}{s_2'}, u_2 + \rho, v_2 - u_2w_2 + \omega \right] \left[ d , u_1 + \rho, v_1 - u_1w_1 + \omega \right]^{-1} =$$ $$\left[\frac{s_2}{s_2'd}, u_2 + \rho, v_2 - u_2w_2 + \omega \right],$$ which justifies one of the two congruences for $U$. We factor out of the ideal in the part that matches the remaining divisor. That is, $$\left[ \frac{s_2'}{s_1'}, \frac{s_2'}{s_1'}\rho, v_2 + w_2\rho + \omega \right] =$$ $$\label{s_1_split} \left[ \frac{s_2'd}{s_1}, \frac{s_2'd}{s_1}\rho, v_2 + w_2\rho + \omega \right] \left[ \frac{s_1}{s_1'd}, \frac{s_1}{s_1'd}\rho, v_2 + w_2\rho + \omega \right].$$ We apply Lemma \[l:splitting\_s\_1\] to the right hand ideal of and the remainder of the divisor to get $$\left[ \frac{s_1}{s_1'd}, \frac{s_1}{s_1'd}\rho, v_2 + w_2\rho + \omega \right] \left[ \frac{s_1}{s_1'd}, u_1 + \rho, v_1 -w_1u_1 + \omega \right]^{-1}$$ $$= \left[ \frac{s_1}{s_1'd}, Iw_2 - u_1 + \rho, v_2 - Iw_2^2 + u_1w_2 + \omega \right].$$ This ideal gives the other congruence for $U$ and the division is complete at this step. The choice for $S$ is justified by looking at the first term in the three ideals that remain; likewise $S'$ is the product of the coefficients of $\rho$: $$S = \left(\frac{s_1}{s_1'd} \right)\left(\frac{s_2'd}{s_1} \right)\left(\frac{s_2}{s_2'd} \right)= \frac{s_2}{s_1'd} \quad \mbox{and} \quad S' = \frac{s_2'd}{s_1}.$$ Since $v_2 + w_2\rho + \omega \in J$, it just remains to modify this element so that it is canonical. This justifies the choice for $V$ and $W$.
\[t:division\_s\_3\][**(Division for Type III primes)**]{} Let $I_i = [s_i, \rho, s_i''\omega]$ for $i = 1,2$ be such that $I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = [S, \rho, S'' \omega]$, where $$S = \frac{s_2}{s_1''d}, \quad S'' = \frac{s_2''d}{s_1}, \ \mbox{ and } \ d = \gcd\left( \frac{s_1}{s_1''}, \frac{s_2}{s_2''} \right) .$$
This follows by using the same arguments presented in the proof of Proposition \[t:division\_s\_1\]. The key distinction is how the ideals are factored: $$I_i = [s_i'', \rho, s_i''\omega] \left[ \frac{s_i}{s_i''},\rho, \omega \right].$$ The rest of the arguments are simplified given that these are products of totally ramified primes.
\[t:division\_s\_4\][**(Division for Type IV primes)**]{} Let $I_i = [s_i, s_i'(u_i + \rho), s_i''(v_i + w_i\rho + \omega)]$ for $i = 1,2$ be such that $I_2 \subseteq I_1$. Then $J= I_2I_1^{-1} = [S, S'(U+ \rho), S''(V + W \rho + \omega)]$, where $$S = \frac{s_2}{s_1's_1''d}, \quad S' = \gcd\left( \frac{ds_2's_2''}{s_1}, \frac{s_2'}{s_1'} \right), \quad S'' = \gcd\left( \frac{ds_2's_2''}{s_1}, \frac{s_2''}{s_1''} \right),$$ $$U \equiv \left\{ \begin{array}{ll}
0 & \pmod{ s_1/(s_1's_1''d) }\\
u_2 & \pmod{ s_2/(s_2's_2''d) }
\end{array}
\right. ,$$ $$d = \gcd\left(\frac{s_2}{s_2's_2''}, \frac{s_1}{s_1's_1''}, v_1 - w_1u_1 - v_2 + w_2u_2 \right),$$ $$S''V \equiv s_2''((W - w_2)U + v_2) \pmod{S}, \mbox{ and } S''W \equiv s_2''w_2 \pmod{S'}.$$
We begin by factoring both $I_1$ and $I_2$ into the ideals $I_{i,1}$ and $I_{i,2}$ as above. The first division is $$\label{s_4_evendeg} I_{2,1}I_{1,1}^{-1}= \left[ \frac{s_2's_2''}{s_1's_1''}, \frac{s_2'}{s_1'}\rho, \frac{s_2''}{s_1''}(v_2 + w_2\rho + \omega) \right].$$
Proceeding as in the proof of Proposition \[t:division\_s\_1\], the next division yields $$\left[\frac{s_2}{s_2's_2''d}, u_2 + \rho, v_2 - u_2w_2 + \omega \right],$$ which justifies the latter congruence for $U$. We decompose the ideal on the right in to get a factor that matches the remaining divisor: $$\label{s_4_split} \left[ \frac{s_2's_2''d}{s_1}, S'\rho, S''(v_2 + w_2\rho + \omega) \right] \left[ \frac{s_1}{s_1's_1''d}, s_3'\rho, s_3''(v_2 + w_2\rho + \omega) \right],$$ where $$s_3' = \gcd\left( \frac{s_1}{s_1's_1''d}, \frac{s_2'}{s_1'} \right) \mbox{ and } s_3'' = \gcd\left( \frac{s_1}{s_1's_1''d}, \frac{s_2''}{s_1''} \right).$$ Note that $S'S'' = s_2's_2''d/s_1$ and $s_3's_3'' = s_1/s_1's_1''d$. Apply Lemma \[l:splitting\_s\_4\] to the right most ideal of and the remainder of the divisor to get $$\left[ \frac{s_1}{s_1's_1''d}, s_3'\rho, s_3''(v_2 + w_2\rho + \omega) \right] \left[ \frac{s_1}{s_1's_1''d}, u_1 + \rho, v_1 -w_1u_1 + \omega \right]^{-1}$$ $$= \left[ \frac{s_1}{s_1's_1''d}, \rho, v_3+ \omega \right],$$ where $v_3$ is given in Lemma \[t:inverse\_s\_3\]. This ideal gives the other congruence for $U$ and the division is complete at this step. The choice for $S$ is justified by looking at the first term in the three ideals that remain: $$S = \left(\frac{s_1}{s_1's_1''d} \right)\left(\frac{s_2's_2''d}{s_1} \right)\left(\frac{s_2}{s_2's_2''d} \right)= \frac{s_2}{s_1's_1''d}.$$ The choices for $S'$ and $S''$ are justified in . Since $s_2''(v_2 + w_2\rho + \omega) \in J$, it just remains to modify this element so that it is canonical and this justifies the choice for $V$ and $W$. The argument here is the same as in Proposition \[t:division\_s\_1\] except we have to account for the coefficient of $\omega$.
We close this section with a proposition on dividing a nonprimitive ideal by a primitive ideal. Consider an ideal of the form $\langle d \rangle I_2$, where $I_2$ is primitive, and a primitive ideal $I_1$. To compute $\langle d \rangle I_2 I_1^{-1}$, we begin by removing as much of $I_1$ from $\langle d \rangle$ as is possible. The remaining factor of $I_1$ is then removed from $I_2$. The primitive parts of the two divisions are $I_d$ and $I_m$, and their product is not necessarily primitive. While this might seem problematic, the propositions on multiplication can be used calculate the product. A proposition on multiplication will invoke this proposition, but it is invoked under the assumption that $I_1$ completely divides $\langle d \rangle$ and that there is no corresponding factor $I_2$.
[**(Nonprimitive division for Type I and II primes)**]{}\[t:nonprim\_div\_s\_1\]\
Let $I_2 = d[s_2, s_2'(u_2 + \rho), v_2 + w_2\rho + \omega]$ and $I_1 = [ s_1, s_1'(u_1+\rho), v_1 + w_1\rho + \omega]$ be such that $\langle d \rangle I_2 \subseteq I_1$. Then $IJ= I_2I_1^{-1} = (D_3)I_dI_m$, where $$I_d = [s_2, s_2'(u_2 + \rho), v_2 + w_2\rho + \omega]\left[ \frac{s_1}{D_1D_2}, \frac{s_1'}{D_1}(u_1 + \rho), v_1 + w_1\rho + \omega \right]^{-1}$$ is calculated by Proposition \[t:division\_s\_1\], $$I_m = \overline{[D_1D_2, D_1(u_1 + \rho), v_1 + w_1\rho + \omega]}$$ is calculated by Proposition \[t:inverse\_s\_1\], $$D_1 = \gcd(s_1',d), \quad D_2 = \gcd\left(\frac{s_1}{s_1'}, \frac{d}{D_1}\right), \mbox{ and } D_3 = \frac{d}{D_1D_2}.$$
We note that $\overline{I_m} \subseteq \langle d \rangle$ and $\overline{I_m} [ s_1/D_1D_2, s_1'/D_1(u_1 + \rho), v_1 + w_1\rho + \omega ] = I_1$. Therefore $\langle d \rangle \overline{I_m}^{-1} = I_m$. After this division, the factors that remain in $I_1$ are $[ s_1/D_1D_2, s_1'/D_1(u_1 + \rho), v_1 + w_1\rho + \omega ]$ and this is contained in $I_2$.
The next two propositions are stated without proof. The proofs follow a similar argument as the proof above and rely, like this proof, nearly entirely on the previously proved propositions.
[**(Nonprimitive division for Type III primes)**]{}\[t:nonprim\_div\_s\_3\]\
Let $I_2 =d [s_2, \rho, s_2'' \omega]$ and $I_1 = [ s_1, \rho, s_1'' \omega]$ be such that $\langle d \rangle I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = (D_3)I_dI_m$, where $$I_d = [s_2, \rho, s_2''\omega]\left[ \frac{s_1}{D_1D_2}, \rho, \frac{s_2''}{D_1} \omega \right]^{-1}$$ is calculated by Proposition \[t:division\_s\_3\], $$I_m = \overline{[D_1D_2, \rho, D_1 \omega]}$$ is calculated by Proposition \[t:inverse\_s\_3\], $$D_1 = \gcd(s_1'',d), \quad D_2 = \gcd\left(\frac{s_1}{s_1''}, \frac{d}{D_1}\right), \mbox{ and } D_3 = \frac{d}{D_1D_2}.$$
[**(Nonprimitive division for Type IV primes)**]{}\[nonprim\_div\_s\_4\]\
Let $I_2 = d[s_2, s_2'(u_2 + \rho), s_2''(v_2 + w_2\rho + \omega)]$ and $I_1 = [ s_1, s_1'(u_1+\rho), s_1''(v_1 + w_1\rho + \omega)]$ be such that $\langle d \rangle I_2 \subseteq I_1$. Then $J = I_2I_1^{-1} = (D_4)I_dI_m$, where $$I_d = [s_2, s_2'(u_2 + \rho), v_2 + w_2\rho + \omega]\left[ \frac{s_1}{D_1D_2D_3}, \frac{s_1'}{D_1}(u_1 + \rho), D_2(v_1 + w_1\rho + \omega) \right]^{-1}$$ is calculated by Proposition \[t:division\_s\_4\], $$I_m = \overline{[D_1D_2D_3, D_1(u_1 + \rho), D_2(v_1 + w_1\rho + \omega)]}$$ is calculated by Proposition \[t:inverse\_s\_4\], $$D_1 = \gcd(s_1',d), \ \ D_2 = \gcd(s_1'', d), \ \ D_3 = \gcd\left(\frac{s_1}{s_1's_1''}, \frac{d}{D_1D_2}\right), \mbox{ and } D_4 = \frac{d}{D_1D_2D_3}.$$
Ideal Multiplication
====================
Theoretically, ideal multiplication is the easiest operation that will be discussed since it may be achieved by simply performing linear algebra. The goal of these propositions is to eliminate much of the excess work that would be required to reduce the nine cross products arising in the multiplication of two ideals down to a basis. The extreme amount of redundancy is obvious for certain products. For example, the product of two relatively prime ideals may be computed quickly using the Chinese Remainder Theorem. Computationally, relatively prime operands are to be expected and the product may be calculated as Scheidler did in Theorem 4.4 of [@idealarithmetic].
[**(Theorem 4.4 of [@idealarithmetic])**]{}\[t:crtforideal\] Let $I_i = [ s_i, s_i'(u_i + \rho), s_i''(v_i + w_i\rho + \omega)]$ with $i = 1,2, 3$ be two ideals such that $\gcd(s_1, s_2) = 1$. Then $I_3 = I_1I_2$ is given by $$s_3 = s_1s_2, \quad s_3' = s_1's_2', \quad s_3'' = s_1''s_2'',$$ $$u_3 \equiv \left\{ \begin{array}{ll}
u_1 & \pmod{ s_1/s_1' }\\
u_2 & \pmod{ s_1/s_2'}
\end{array}
\right. , \quad
w_3 \equiv \left\{ \begin{array}{ll}
w_1 & \pmod{ s_1'}\\
w_2 & \pmod{ s_2' }
\end{array}
\right. , \quad \mbox{ and }$$ $$v_3 \equiv \left\{ \begin{array}{ll}
v_1 + u_1(w_3-w_1) & \pmod{ s_1/s_1'' }\\
v_2 + u_2(w_3 - w_2) & \pmod{ s_2/s_2'' }
\end{array}
\right. .$$
In contrast to cubic function fields of unit rank one, we can not assume that the two operands will be relatively prime. Thus, we will be forced to develop ideal multiplication systematically. The first set of propositions assumes that the product is primitive and this will be used to aid in the case where the product is not assumed to be primitive.
In proof of Proposition \[t:division\_s\_1\] the two congruences for $U$ were sufficient in guaranteeing that $U$ was determined uniquely modulo $S/S'$. This was because $s_1/(s_1'd)$ and $s_2/(s_2'd)$ were relatively prime. Had they shared a common factor, $U$ would have been determined only up to the least common multiple of $s_1/(s_1'd)$ and $s_2/(s_2'd)$. This proposition is built to handle just such a situation.
[**(Primitive Multiplication for Type I primes)**]{}\[t:prim\_mult\_s\_1\]\
Let $I_i = [s_i, s_i'(u_1 + \rho), v_i + w_i \rho + \omega]$ for $i = 1,2$ be such that $I_1I_2 = I_3$ is a primitive ideal. Then $I_3 = [S, S'(U + \rho), V + W\rho + \omega]$, where $$S = \frac{s_1s_2d_1}{d}, \quad S' = \frac{s_1's_2'd}{d_1}, \quad W = w_3 - cS', \quad V \equiv v_3 - qS'U \pmod{S},$$ $$\mbox{ and } U \equiv u_3 + k\frac{s_1s_2d_1}{s_1's_2'd^2} \pmod{S/S'}.$$ We choose $c$ to make $\deg W$ minimal and define $u_3$, $v_3$, $w_3$, $d$, $d_1$ as follows $$\begin{aligned}
d &= \gcd\left( \frac{s_1}{s_1'}, \frac{s_2}{s_2'} \right), & d_1&= \gcd(d, u_1 - u_2), \\
u_3 &\equiv u_1 \pmod{s_1d_1/s_1'sd}, & u_3 &\equiv u_2 \pmod{s_2d_1/s_2'd}, \end{aligned}$$ $$\mbox{ and k is chosen such that \quad } d_1 \ \vrule \ \frac{(u_3^3 - u_3A - FI^2)S'd_1}{S} + kA,$$ $$\begin{aligned}
w_3 &= a_1s_2w_1 + a_2s_1w_2 + a_3s_1's_2'(u_1 + u_2) + a_4s_1'(v_2+u_1w_2) \\
& \quad + a_5s_2'(v_1 + u_2w_1) + a_6(v_1w_2 + v_2w_1 - F) \\
v_3& = a_1s_2v_1 + a_2s_1'v_2 + a_3s_1's_2'(u_1u_2 + A) + a_4s_1's(u_1v_2 - FI + w_2) \\
&\quad + a_5s_2'(u_2v_1 - FI + w_1A) + a_6(v_1v_2 + w_1w_2-w_1FI - W_2FI) \end{aligned}$$ and $a_1$, $a_2$,$a_3$, $a_4$, $a_5$, and $a_6$ are given by the extended euclidian algorithm as: $$\begin{aligned}
1 & = a_1s_2 + a_2s_1 + a_3s_1's_2'I + a_4s_1'(u_1 + Iw_2) \\
& \quad+ a_5s_2'(u_2 + Iw_1) + a_6(v_1 + v_2 + w_1w_2I - E). \end{aligned}$$
Since we assume $I_3$ is primitive, it has a canonical basis of the form claimed. We begin by factoring $I_1$ and $I_2$ and deal with their product using smaller and simpler ideals. The easiest part of the product is $$[s_1', s_1'\rho, v_1 + w_1\rho + \omega ] [s_2', s_2'\rho, v_2 + w_2\rho + \omega ] = [s_1's_2', s_1's_2'\rho, V + W\rho + \omega ].$$ While we still need to find congruences for $V$ and $W$, we will return to those later and focus on the difficult part of the product: $$\label{singles_s_1} \left[ \frac{s_1}{s_1'}, u_1 + \rho, v_1 - w_1u_1 + \omega \right] \left[ \frac{s_2}{s_2'}, u_2 + \rho, v_2 - w_2u_2 + \omega \right].$$ The goal will be two split this product up into two factors. The quantity $d$ signifies common possible prime factors in this product, and $d_1$ indicates those primes that appear as squares in the product. Thus, we write the above product as $$\left[ \frac{S}{S'}, U + \rho, V + \omega \right] \left[\frac{d}{d_1}, \frac{d}{d_1}\rho, V + W\rho + \omega \right].$$ We conclude from this that $S' = s_1's_2'd/d_1$ and by equating norms that $S = s_1s_2d/d_1$. Combining the two previous statements we see that $$\left[ \frac{S}{S'}, U + \rho, V + \omega \right] = \left[ \frac{s_1d_1}{s_1'd}, u_1 + \rho, v_1 - w_1u_1 + \omega \right] \left[ \frac{s_2d_1}{s_2'd}, u_2 + \rho, v_2 - w_2u_2 + \omega \right].$$ This justifies the choice for $u_3$, and note that $u_3$ is defined uniquely modulo the least common multiple of $s_1d_1/s_1'd$ and $s_2d_1/s_2'd$. Thus we can write $U = u_3 + kS/S'd_1$ and consider $$\frac{S}{S'} \ \vrule \ N(U + \rho) \Rightarrow \frac{S}{S'} \ \vrule \ (u_3^3 - u_3A - FI^2) + kA\frac{S}{S'd_1}.$$ From the definition of $u_3$, $S/S'd_1$ divides $u_3^3 - u_3A - FI^2$ so we can conclude $$d_1 \ \vrule \ \frac{ (u_3^3 - u_3A - FI^2)S'd_1}{S} + kA$$ as claimed. This determines $U$ modulo $S/S'$ as needed. To calculate $V$ and $W$ we find any element of the form $v_3 + w_3\rho + \omega \in I_3$. Since $I_3$ is primitive and contains no index divisors, the greatest common divisor of the coefficients of $\omega$ arising from all possible products of basis elements of $I_1$ and $I_2$ must be $1$. Once this element is computed, it is a matter of subtracting multiples of the two previously calculated basis elements to ensure the third element is canonical.
The calculation of $W$ and $V$ is not as difficult as it looks. As we noted before, if $s_1$ and $s_2$ are relatively prime the above proposition is superfluous and the multiplication can be done via the Chinese Remainder Theorem. Assuming $s_1$ and $s_2$ are not relatively prime, we still expect that we will be able to write 1 as a linear combination of fewer than all six terms.
[**(Primitive Multiplication for Type II primes)**]{}\[t:prim\_mult\_s\_2\] Let $I_i = [s_i, s_i'(u_1 + \rho), v_i + w_i \rho + \omega]$ for $i = 1,2$ be such that $I_1I_2 = I_3$ is a primitive ideal. Then $I_3 = [S, S'(U + \rho), V + W\rho + \omega]$, where $$S = s_1s_2/d, \quad S' = ds_1's_2',\quad d = \gcd\left( \frac{s_1}{s_1'}, \frac{s_2}{s_2'} \right),$$ $$U \equiv f \pmod{S/S'}, \quad W \equiv I^{-1}f \pmod{S'}, \quad V \equiv f^2I^{-1} \pmod{S},$$ where $f$ is defined by $f^3 \equiv FI^2 \pmod{S}$.
We invoke Proposition \[t:wildrambasis\] to calculate $U$, $V$, and $W$.
[**(Primitive Multiplication for Type III primes)**]{}\[t:prim\_mult\_s\_3\] Let $I_i = [s_i, \rho, s_i''\omega]$ for $i = 1,2$ be such that $I_1I_2 = I_3$ is a primitive ideal. Then $$I_3 = \left[ \frac{s_1s_2}{d}, \rho, (s_1''s_2''d)\omega \right] \mbox{ where } d = \gcd\left(\frac{s_1}{s_1''}, \frac{s_2}{s_2''} \right).$$
This follows from Proposition \[sing\_wildram\].
[**(Primitive Multiplication for Type IV primes)**]{}\[t:prim\_mult\_s\_4\] Let $I_i = [s_i, s_i'(u_1 + \rho), s_i''(v_i + w_i \rho + \omega)]$ for $i = 1,2$ be such that $I_1I_2 = I_3$ is a primitive ideal. Then $I_3 = [S, S'(U + \rho), S''(V + W\rho + \omega)]$, where $$S = \frac{s_1s_2}{dd_1d_2}, \quad S' = s_1's_2'd, \quad S'' = s_1''s_2''d_1d_2.$$ To define $d_1$, $d_2$, and $d$, let $$s_{\mathfrak{q}i} = \gcd\left( \frac{s_i}{s_i's_i''}, v_i-w_iu_i \right) \mbox{ for } i = 1,2,$$ then $$d = \gcd( s_{\mathfrak{q}1}, s_{\mathfrak{q}2}), \quad d_1 = \gcd\left( s_{\mathfrak{q}2}, \frac{s_1}{s_1's_1''s_{\mathfrak{q}1}} \right), \mbox{ and } \ d_2 = \gcd\left( s_{\mathfrak{q}1}, \frac{s_2}{s_2's_2''s_{\mathfrak{q}2}} \right).$$ We defined $U = (s_1''s_2''d_1d_2)u_3 $ where $u_3$ satisfies $$u_3 \equiv u_1 \pmod{\frac{s_1}{s_1''s_1''dd_1d_2}}, \ \mbox{ and } \ u_3 \equiv u_2 \pmod{ \frac{s_2}{s_2's_2''dd_1d_2} }.$$ Finally we can choose $V$ and $W$ as $$S''W = S''w_3 - qS', \quad S''V \equiv S''v_3 - qS'U \pmod{S},$$ where $q$ is chosen so that $\deg V$ and $\deg W$ are minimal and $$\begin{aligned}
S''w_3 = & a_1s_2s_1''w_1 + a_2s_1s_2''w_2 + a_3s_1's_2'(u_1 + u_2) + a_4s_1's_2''(v_2+u_1w_2) \\
&+ a_5s_2's_1''(v_1 + u_2w_1) + a_6s_1''s_2''(v_1w_2 + v_2w_1 - F) \end{aligned}$$ $$\begin{aligned}
S''v_3 = & a_1s_2s_1''v_1 + a_2s_1s_2''v_2 + a_3s_1's_2'(u_1u_2 + A) + a_4s_1's_2''(u_1v_2 - FI + w_2) \\
&+ a_5s_2's_1''(u_2v_1 - FI + w_1A) + a_6s_1''s_2''(v_1v_2 + w_1w_2-w_1FI - W_2FI) \end{aligned}$$ where $a_i$ for $i =1, \ldots, 6$ come from the extended greatest common divisor calculation, $$\begin{aligned}
S''= & a_1s_2s_1'' + a_2s_1s_2'' + a_3s_1's_2'I + a_4s_1's_2''(u_1 + Iw_2) \\
&+ a_5s_2's_1''(u_2 + Iw_1) + a_6s_1''s_2''(v_1 + v_2 + w_1w_2I - E). \end{aligned}$$
The details of the proof are similar to those above. We will try and note only the key distinctions. We factor $I_i$ into three factors as $$I_i = J_{i,1}J_{i,2}J_{i,3} = \left[\frac{s_i}{s_i's_i''}, u_i + \rho, v_i-w_iu_i + \rho \right][s_i'', \rho, s_i''\omega][s_i', s_i'\rho, v_i + w_i\rho + \omega].$$ Since $I_3$ is primitive, all three of $\gcd(s_2'',s_1''), \gcd(s_2',s_1''),$ and $ \gcd(s_2'',s_1')$ are one. This simplifies the number of possible products to consider. We factor $J_{i,1}$ further to distinguish ramified primes (denoted with a subscript $\mathfrak{q}$) from the unramified primes: $$J_{i,1} = [ s_{\mathfrak{q}i}, \rho, \omega] \left[\frac{s_i}{s_i's_i''s_{\mathfrak{q}i} }, u_i + \rho, v_i - w_iu_i + \omega \right].$$ Now there are three possible type of products these two ideals can form. Products corresponding to a common place of ${\mathbb{F}_q}(x)$ lying below $\mathfrak{p}$ and $\mathfrak{q}$ indicate the presence of that polynomial being a factor of the coefficent of $\omega$. This justifies the choice of $d_1$ and $d_2$. There are at most single powers of $\mathfrak{q}$ in either of the two ideals that correspond to that part of the factorization. Their greatest common divisor justifies the choice of $d$. We remove these factors from their corresponding ideals in $J_{i,1}$. We can choose $u_3$ from these two divisors of $J_{i,1}$. This gives $u_3$ unique modulo $S/(S'S'')$. Since $S''$ divides $U$ this justifies the choice of $U$. Lastly, we chose $V$ and $W$ in the same manner as in the previous proposition. However, the fact that these ideals correspond to index divisors means that the greatest common divisor of the terms with $\omega$ will no longer be 1 but $S''$.
Much like Proposition \[t:prim\_mult\_s\_1\] the greatest common divisor calculation looks complicated but in general $S''$ can be found with fewer terms than the 6 given.
Now we deal with the case that the product of two ideals is not primitive. The key to these propositions is finding and removing the nonprimitive factors. The remaining product is primitive and the previous propositions may be invoked.
[**(Multiplication for Type I primes)**]{} \[t:mult\_s\_1\] For $i = 1,2$ let $I_i = [s_i, s_i'(u_1 + \rho), v_i + w_i \rho + \omega]$ be two ideals . Then $I_1I_2 = (D)I_3$ where $I_3 = I_1'I_2'J$ and $D = D_1D_2D_3$ and these quantities are as follows: $$D_1 = \gcd(s_2', s_1/s_1', u_1 + Iw_2), \quad D_2 = \gcd(s_1', s_2/s_2', u_2 + Iw_1),$$ $$D_3 = \frac{\gcd(s_1'/D_2,s_2'/D_1)}{\gcd(s_1'/D_2,s_2'/D_1, w_1-w_2)},$$ $$I_1' = \left[ \frac{s_1}{D_1D_2D_3}, \frac{s_1'}{D_2D_3}(u_1 + \rho), v_1 + w_1\rho + \omega \right],$$ $$I_2' = \left[ \frac{s_2}{D_1D_2D_3}, \frac{s_2'}{D_1D_3}(u_2 + \rho), v_2 + w_2\rho + \omega \right], \mbox{ and }$$ $$J = \langle D_3 \rangle \left( \overline{ [D_3, D_3\rho, v_1 + w_1\rho + \omega] } \ \overline{[D_3, D_3\rho, v_2 + w_2\rho + \omega] } \right)^{-1},$$ and the last calculation is done by invoking Propositions \[t:inverse\_s\_1\], \[t:nonprim\_div\_s\_1\], and \[t:prim\_mult\_s\_1\].
We factor $I_1$ and $I_2$ as in Proposition \[t:division\_s\_1\],
$$I_i = I_{i,1}I_{i,2} = [s_i', s_i'\rho, v_i + w_i\rho + \omega] \left[ \frac{s_i}{s_i'}, u_i + \rho, v_i + w_i\rho + \omega \right]$$
Of these four factors the non-primitive part of the product does not arise from $I_{1,2}I_{2,2}$. We find the non-primitive part from the product $I_{1,2}I_{2,1}$ (resp. $I_{2,2}I_{1,1}$). It suffices to consider the coefficient of $\omega$. Hence $D_1 = \gcd(s_2', s_1/s_1', u_1 + Iw_2)$ (resp. $D_2 = \gcd(s_1', s_2/s_2', u_2 + Iw_1)$). We remove $D_1$ (resp. $D_2$) from $I_{1,1}$ and $I_{2,2}$ (resp. $I_{2,1}$ and $I_{1,2}$) and rename as follows: $$I_{1,2}' = \left[\frac{s_1}{s_1'D_1}, u_1 + \rho, v_1 + w_1\rho + \omega \right] , \quad I_{1,1}'= \left[ \frac{s_1'}{D_2},\frac{s_1'}{D_2}\rho, v_1 + w_1\rho + \omega \right]$$ $$I_{2,2}' = \left[ \frac{s_2}{s_2'D_2}, u_2 + \rho, v_2 + w_2\rho + \omega \right] , \ \mbox{ and } \ I_{2,1}' = \left[ \frac{s_2'}{D_1},\frac{s_2'}{D_1}\rho, v_2 + w_2\rho + \omega \right].$$ The product $I_1I_2$ now has the form $(D_1D_2)I_{1,1}' I_{1,2}' I_{2,1}' I_{2,2,}'$ and any remaining nonprimitive factor comes from $I'_{1,1}I'_{2,1}$.
Let $$I_{1,3} = [D_3, D_3\rho, v_1 + w_1\rho + \omega ] \mbox{ and } I_{2,3} = [D_3, D_3\rho, v_2 + w_2\rho + \omega ],$$ where $D_3$ is defined above. The choice of $D_3$ is justified because $\gcd(s_1'/D_2, s_2'/D_1)$ is the possible primes that could be part of the non-primitive product. However, the previous greatest common divisor contains too many primes. For a given prime $P$ we need to be able to distinguish between $\mathfrak{pq}$ and $\mathfrak{p}^2$. If $w_1-w_2 = 0$ then the associated primes correspond to a square and that justifies the choice for the denominator in $D_3$. We justify the claim for the ideal $J$ by noting the following equalities. $$\begin{aligned}
I_{1,3}I_{2,3} &= I_{1,3}I_{2,3}\overline{I_{1,3}} \ \overline{I_{2,3}} (\overline{I_{1,3}} \ \overline{I_{2,3}})^{-1} \\
&= \langle D_3 \rangle ^2 (\overline{I_{1,3}}\ \overline{I_{2,3}})^{-1}\\
&= \langle D_3 \rangle \left( \langle D_3 \rangle / (\overline{I_{1,3}} \ \overline{I_{2,3}})\right)\\
& = \langle D_3 \rangle J\end{aligned}$$ The last ideal is the one given in the proposition statement and it is primitive. We remove the factor $I_{1,3}$ from $I_{1,2}'$ and $I_{2,3}$ from $I_{2,2}'$ to get the other two primitive ideals. The product of these three ideals is primitive and can be calculated by Proposition \[t:prim\_mult\_s\_1\].
The totally ramified primes will be much easier to deal with. For the two types, appealing to the propositions that govern their powers from Section 7 will be sufficient.
[**(Multiplication for Type II primes)**]{}\[t:mult\_s\_2\] For $i = 1,2$ let $I_i = [s_i, s_i'(u_1 + \rho), v_i + w_i \rho + \omega]$ be two ideals such that $I_1I_2 = (D)I_3$ with $I_3 = I_1'I_2'J$ and $D = D_1D_2D_3$. These quantities are given as follows: $$D_1 = \gcd \left( \frac{s_1}{s_1'}, s_2' \right), \quad D_2 = \gcd \left( \frac{s_2}{s_2'}, s_1' \right), \quad D_3 = \gcd(s_2', s_1'),$$ $$I_1' = \left[ \frac{s_1}{D_1D_2D_3}, \frac{s_1'}{D_2D_3}(u_1 + \rho), v_1 - w_1u_1 + \omega \right],$$ $$I_2' = \left[ \frac{s_2}{D_1D_2D_3}, \frac{s_2'}{D_1D_3} (u_2 + \rho), v_2 - w_2u_2 + \omega \right],$$ $$J = [ D_3, f + \rho, I^{-1}f^2 + \omega ],$$ with $f$ satisfying $f^3 \equiv FI^2 \pmod{D_3}$.
Much like the previous proposition, the key is to factor the ideals and find where the nonprimitive factors arise. Unlike the previous proposition, constructing the equivalent ideal $J$ is trivial. This is because $D_3$ is squarefree and Proposition \[t:wildrambasis\] states the form of these ramified primes.
[**(Multiplication for Type III primes)**]{}\[t:mult\_s\_3\] For $i = 1,2$ let $I_i = [s_i, \rho, s_i''\omega]$ be two ideals such then $I_1I_2 = (D)I_3$. $I_3 = I_1'I_2'J$ and $D = D_1D_2D_3$ where these quantities are given as follows: $$D_1 = \gcd\left( \frac{s_1}{s_1''}, s_2'' \right), \quad D_2 = \gcd \left( \frac{s_2}{s_2''}, s_1'' \right), \quad D_3 = \gcd(s_2'', s_1''),$$ $$I_1' = \left[ \frac{s_1}{D_1D_2D_3}, \rho , \frac{s_1''}{D_2D_3}\omega \right], \quad I_2' = \left[ \frac{s_2}{D_1D_2D_3}, \rho, \frac{s_2''}{D_1D_3} \omega \right], and$$ $$J = [ D_3, \rho, \omega ].$$
This follows in the same manner as the previous proof.
[**(Multiplication for Type IV primes)**]{}\[t:mult\_s\_4\] For $i = 1,2$ let $I_i = [s_i, s_i'(u_1 + \rho), s_i''(v_i + w_i \rho + \omega)]$ be two ideals. Then $I_1I_2 = (D)I_3$ where $$I_3 = I_1'I_2'J \mbox{ and } D = D_1D_2D_3D_4D_5D_6D_7.$$ These quantities are defined as follows: $$D_1 = \gcd(s_2'',s_1/s_1's_1'', v_1 - u_1w_1), \quad D_2 = \gcd(s_1'', s_2/s_2's_2'', v_2-w_2u_2),$$ $$D_3 = \frac{ \gcd(s_1/s_1's_1'', s_2'') } { \gcd(s_1/s_1's_1'', s_2'', v_1 - w_1u_1) }, \quad D_3 = \frac{ \gcd(s_2/s_2's_2'', s_1'') } { \gcd(s_2/s_2's_2'', s_1'', v_2 - w_2u_2) },$$ $$D_5 = \gcd( s_1'/D_4, s_2''/D_1), \quad D_6 = \gcd(s_2'/D_3, s_1''/D_2) ,$$ $$D_7 = \gcd\left( \frac{s_2''}{D_1D_5}, \frac{s_1''}{D_2D_6} \right)$$ $$I_1' = \left[ \frac{s_1}{D_2D_4D_5D_6D_7}, \frac{s_1'}{D_4D_5}\rho, \frac{s_1''}{D_2D_6D_7}(v_1 + w_1\rho + \omega) \right],$$ $$I_2' = \left[ \frac{s_2}{D_1D_3D_5D_6D_7}, \frac{s_2'}{D_3D_6}\rho, \frac{s_2''}{D_1D_5D_7}(v_2 + w_2\rho + \omega) \right], \mbox{ and }$$ $$J = [D_5D_6, \rho, \omega] [D_7, \rho, \omega + E ].$$
The proof follows in a similar manner as the previous three proofs. Again, we seek only to highlight the differences. We begin by factoring $I_1$ into three ideals as $$I_1 = I_{1,1}I_{1,2}I_{1,3} = \left[ \frac{s_1}{s_1's_1''}, u_1 + \rho, v_1 - w_1u_1 + \omega \right][s_1', s_1'\rho, v_1 + w_1\rho + \omega][s_1'', \rho, s_1''\omega],$$ and likewise with $I_2$. The quantity $D_1$ (resp. $D_2$, $D_3$, $D_4$) is the nonprimitive part from $I_{1,1}I_{2,3}$ (resp. $I_{2,1}I_{1,3}$, $I_{1,1}I_{2,2}$, $I_{2,1}I_{1,2}$). We remove these factors from the ideals and consider $I_{1,2}I_{2,3}$ (resp. $I_{2,2}I_{1,3}$). Here we are considering the case in which one ideal contains squares of the ramified prime (say, $\mathfrak{q}^2$) and the other ideal contains products of a ramified prime with its corresponding unramified prime (say, $\mathfrak{pq}$). The product of $\mathfrak{q^2}\mathfrak{pq}$ is $(P)\mathfrak{q}$. Thus we get $(D_5)[D_5, \rho, \omega]$ (resp $(D_6)[D_6, \rho, \omega]$). We remove the factor $D_5$ (resp. $D_6$) from $I_{1,2}$ and $I_{2,3}$ (resp. $I_{2,2}$ and $I_{1,3}$) and consider one last product of $I_{1,3}I_{2,3}$. This is a product where each ideal has primes of the form $\mathfrak{pq}$ and therefore the product must be of the form $(P)\mathfrak{p}$. We get $(D_7)[D_7, \rho, \omega + E]$.
We have stated the basic ideal operations necessary for arithmetic. The key now is to give a method to find a distinguished element in an ideal class. From this point forward, $\mathcal{F}/K$ will be assumed to have a totally ramified infinite place with $3\nmid \deg{FI^2}$. This latter assumption is necessary since we rely on Theorem \[t:n\_norm\]. These assumptions also ensure that the ideal class group is isomorphic to the Jacobian of the curve.
Elements of Minimal Norm
========================
The content in this section closely mirrors Section 8 of [@bauer]. Given an ideal $J = [s, s'(u + \rho),s''( v + w\rho + \omega)]$, we want to find an element in this ideal that has minimal norm. The existence of such an element (up to a constant scalar) is guaranteed by Theorem \[t:minnormexists\]. Writing the ideal as a triangular matrix and assigning a weight to each column corresponding to the norm of the associated element, the algorithm proceeds to perform elementary row operations on the matrix to find an element of minimum weight.\
$$\left[ \rule{0in}{0.3in} \right.$$
$$\begin{array}{ccc}
s & 0 & 0 \\
s'u & s' & 0 \\
s''v & s''w & s'' \\
\uparrow & \uparrow & \uparrow \\
3*\deg{FI^2} \ \ \ & 3*\deg+\deg{FI^2} \ \ \ & 3*\deg+ \deg{F^2I}
\end{array}$$
$$\left. \hspace{2.3in} \rule{0in}{.3in} \right]$$
\
By Theorem \[t:n\_norm\], the weights of distinct columns lie in distinct residue classes modulo three. If two rows have their weight coming from the same position it is possible to reduce the weight of one of the rows while still maintaining a basis for the ideal. The algorithm below just encodes the order in which to do the minimization.
[MinElement]{} \[a:n\_min1\] Minimal Element Algorithm. Let $I=[s, s'(u + \rho),s''(v + w\rho + \omega)]$. $\alpha \in I$ non-zero so that $N(\alpha)$ has minimal degree. Use the ideal to define $b_1=(b_{1,1},b_{1,2},b_{1,2})=(s,0,0)$, $b_2=(b_{2,1},b_{2,2},b_{2,2})=(s'u,s',0)$, and $b_3=(b_{3,1},b_{3,2}, b_{3,2})=(s''v,s''w,s'')$. Assign weights $w_{i,1}=3 \deg b_{i,1}$, $w_{i,2}=3 \deg b_{i_2} + \deg{FI^2}$, and, $w_{i,3}=3 \deg b_{i,3} + \deg{F^2I}$. Set $w_i=\max \{
w_{i,1},w_{i,2}, w_{i,3} \}$, and choose $a_i$ so that $w_i=w_{i,a_i}$ (i.e., $w_i=w_{i,a_i}=\deg N(b_i)$). Order the $b_i$ and their associated values so that $w_1\leq w_2\leq w_3$. $b_{2,a_2}=b_{1,a_1}c+r$ replace $b_2:=b_2-cb_1$ and recalculate $a_2,w_2$. $b_{3,a_3}=b_{1,a_1}c+r$ replace $b_3:=b_3-cb_1$ and recalculate $a_3,w_3$. $b_{3,a_2}=b_{2,a_2}c+r$ replace $b_3:=b_3-cb_2$ and recalculate $a_3,w_3$. Reorder the $b_i$’s and associated values. **Return:** $b_{1,1}+b_{1,2}\rho+b_{1,3}\omega$, the element of minimal norm.
Now that we can calculate an element of minimal norm, our goal will be to construct a canonical basis for the principal ideal generated by this element.
Canonical Basis
===============
The algorithm for finding a canonical basis for a principal ideal generated by an element of $\mathcal{O}_\mathcal{F}$ is straightforward.
[CanBasis]{}\[a:canbasis\] $a + b\rho + c\omega \in \mathcal{O}_F$ A canonical basis of the ideal $I = \langle \alpha \rangle$. Create the matrix $$\left[ \begin{array}{lll} a & b & c \cr bA-cFI & a & bI \cr -bFI & -cF & a-cE \cr
\end{array} \right].$$ Using elementary row operations transform it into a lower triangular matrix $$\left[ \begin{array}{lll} c_3 & 0 & 0 \cr c_2 & b_2 & 0 \cr c_1 & b_1
& a_1 \cr \end{array} \right] .$$ Set $d = \gcd(a_1, b_2)$, $s = c_3/d$, $s' = b_2/d$, $s'' = a_1/d$ and $u \equiv c_2/(s'd) \pmod{s/s'}$. Compute $c$ and $w$ such that $b_1/d = s'c + w$ and $deg(w) < deg(s')$. Compute $ v\equiv c_1/d - s'qu \pmod{s}$. **Return:** The ideal $d \ [s, s'(\rho + u), s''\omega + w\rho + v]$ generated by $\alpha$, given in terms of a canonical basis.
Since we used only elementary row operations, the algorithm gives a valid ${\mathbb{F}_q}[x]$-basis for the principal ideal generated by $a + b\rho + c\omega$. The latter steps in the algorithm ensure the basis is canonical.
Composition and Reduction in the ideal class group
==================================================
We have all the tools we need to do composition and reduction in the ideal class group. Given two ideals $I_1$ and$I_2$ we find a distinguished representative in the class of $I_1I_2$ as follows:
[CompRed]{} \[a:n\_compred\] Two ideals $I_1$ and $I_2$ with canonical representations. The distinguished ideal $J$ equivalent to $I_1I_2$. Calculate $I_3 = I_1I_2$. Find $\overline{I_3}$. Find $\alpha \in \overline{I_3}$ of minimal norm using Algorithm \[a:n\_min1\]. Compute $\langle \alpha \rangle = \langle d \rangle [ s, s'(u + \rho), v + w\rho + \omega]$ using Algorithm \[a:canbasis\]. Compute $J = \langle \alpha \rangle / \overline{I_3}$. **Return:** $J$.
The proof of correctness has been established in the previous sections by invoking the appropriate theorems. For almost all cubic function field in characteristic three with a totally ramified place at infinity, we have given composition and reduction in the ideal class group. There are, however, some exceptions - see Example 1 in Section 6 for a function field with a totally ramified place for which the above algorithm will fail to succeed at reduction in the ideal class group.
Example Computation
===================
We present an example to illustrate the algorithms. The field of constants is $\mathbb{F}_{3^{10}} = \mathbb{F}_3[\alpha]/(\alpha^{10} - \alpha^6 - \alpha^5 -\alpha^4 + \alpha - 1)$ and the cubic function field is $\mathbb{F}_{3^{10}}(x,y)$ where $y$ is a root of $T^3 - T + x^4 + \alpha$. Since $A = 1$, this is an Artin-Schreier extension with no finite ramification. The infinite place is totally ramified and the genus of the function field is 3. We let $$I_1 = [x, -\alpha^9 -\alpha^8 + \alpha^7 + \alpha^6 +\alpha^5 - \alpha^4 - \alpha^3 + \rho, \alpha^8 - \alpha^6 -\alpha^5 +\alpha^4 -\alpha^3 + \alpha^2 + \omega]$$ and we will find the reduced ideal in the class of $I_1^6$ following Algorithm \[a:n\_compred\].
**Step 1.** We calculating $I_1^2$ and $I_1^3$ followed by $I_1^6$ invoking Proposition \[t:prim\_mult\_s\_1\] each time. We state only the the parameters used to define $I_1^6$ which has the form $[s_2, u_2 + \rho, v_2 + \omega]$, where $$s_2 = x^6, \quad u_2 = x^4 - \alpha^9 - \alpha^8 + \alpha^7 + \alpha^6 + \alpha^5 -\alpha^4 - \alpha^3, \mbox{ and }$$ $$v_2 = (-\alpha^9 -\alpha^8 + \alpha^7 + \alpha^6 + \alpha^5 -\alpha^4 -\alpha^3)x^4 - \alpha^8 + \alpha^6 + \alpha^5 - \alpha^4 + \alpha^3 - \alpha^2 -1.$$
**Step 2.** We compute $I_3 = \overline{I_2}$. It is clear that this inverse will have the form $[s_3, s_3\rho, v_3 + w_3\rho + \omega]$. By appealing to Proposition \[t:inverse\_s\_1\], we have $s_3 = x^6$, $v_3 = -v_2$, and $w_3 = -u_2$.
**Step 3.** We apply Algorithm \[a:n\_min1\] to the above ideal. We note that the while-loop finishes in two iterations to give $a + b\rho + c\omega$ as the element of minimal norm, where $$\begin{aligned}
a & = & (\alpha^8 - \alpha^6 - \alpha^5 + \alpha^4 -\alpha^3 + \alpha^2 -1)x^2\\
b& = & (\alpha^9 + \alpha^8 - \alpha^7 - \alpha^6 - \alpha^5 + \alpha^4 + \alpha^3)x^2\\
c & = & x^2\end{aligned}$$
**Step 4.** Applying Algorithm \[a:canbasis\] to the above parameters gives $$\langle x^2 \rangle [ x^4, x^4\rho, (\alpha^8 - \alpha^6 - \alpha^5 + \alpha^4 -\alpha^3 + \alpha^2 -1) + (\alpha^9 + \alpha^8 - \alpha^7 - \alpha^6 - \alpha^5 + \alpha^4 + \alpha^3)\rho + \omega ].$$
**Step 5.** Finally, we calculate $\langle \alpha \rangle / I_3$ according to Proposition \[t:nonprim\_div\_s\_1\]. This has the form $[s_4, u_4 + \rho, v_4 + \omega]$ where $$\begin{aligned}
s_4 & = & x^2,\\
u_4& = &-\alpha^9 -\alpha^8 + \alpha^7 + \alpha^6 +\alpha^5 - \alpha^4 - \alpha^3,\\
v_4 & = & \alpha^8 - \alpha^6 -\alpha^5 +\alpha^4 -\alpha^3 + \alpha^2.\end{aligned}$$ Note that this happens to be $I_1^2$.
Conclusion
==========
This work was chiefly motivated by two sources. First we wanted to find comparable results of [@exp_cff; @aacff] in the characteristic 3 case. Given the history of the subject, it might have been sensible to attempt to compute fundamental units as in [@fun_unit]. However, a naive approach with the integral basis given in Section 3 failed to work and we are currently investigating a solution. However, the second motivation was a generalization along the lines of Bauer’s [@bauer] computation in the ideal class group assuming the function field has a totally ramified infinite place is available. The material presented here is the subject of the second author’s Ph.D. thesis.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. Perucho'
- 'M. Hanasz'
- 'J. M. Martí'
- 'H. Sol'
date: 'Received .../ Accepted ...'
title: |
Stability of hydrodynamical relativistic planar jets.\
I. Linear evolution and saturation of Kelvin-Helmholtz modes.
---
Introduction
============
Relativistic jets are found in many astrophysical objects like active galactic nuclei (see eg. Zensus 1997), microquasars (Mirabel & Rodriguez 1999) and gamma-ray bursts (Sari, Piran & Halpern 1999). The basic physical property of jets is their huge flux of kinetic energy, taken away from compact unresolved or marginally resolved central regions. The presence of jets in such objects provides a unique opportunity to study the physics of these central regions through investigations of the jet flows. However, the extraction of information about the central object is not a simple task, because jets interact in complex ways with the surrounding interstellar and intergalactic medium. On one hand the dynamical interaction of the jet matter with the ambient medium leads to the formation of shocks, turbulence, acceleration of charged particles and subsequent emission of a broad-range electromagnetic radiation, which makes jets observable. On the other hand the complex nature of the interaction which is due to the presence of flow instabilities, especially the Kelvin-Helmholtz (KH) instability, makes it difficult to distinguish those properties which are directly related to the central object from those which are due to the interaction of jets with the ambient medium. As an example one can invoke the wavelike patterns in jets, which may result from the precession of the rotation axis of the accretion disk at the jet base, or from KH instability, which finds favorable conditions at the interface of jet and external medium (Trussoni, Ferrari & Zaninetti, 1983). Very recently, the KH instability theory has been successfully used to interpret the structure of the pc-scale jet in the radio source 3C273 (Lobanov & Zensus 2001).
At kiloparsec scales, the surprisingly stable propagation of relativistic jets in some sources (e.g., Cyg A) contrasts with the deceleration and decollimation observed in other sources (e.g. 3C31). These two sources are representative examples of two kinds, which are classified into Fanaroff-Riley type I (FRI) and Fanaroff-Riley type II (FRII) radio sources (Fanaroff & Riley 1974). The morphological dichotomy of FRI and FRII sources may be related to the stability properties of relativistic jets with different kinetic powers.
This complex situation motivated us to study the interaction of relativistic jets with their ambient medium and more specifically the KH instability in detail, by applying a linear stability analysis along with numerical simulations.
The linear analysis of KH instability in relativistic jets started with the work of Turland & Scheuer (1976) and Blandford & Pringle (1976) who derived and solved a dispersion relation for a single plane boundary between the relativistic flow and the ambient medium. Next, Ferrari, Trussoni & Zaninetti (1978) and Hardee (1979) examined properties of KH instability in relativistic cylindrical jets by following the derivation of the dispersion relation in the nonrelativistic case, in the vortex sheet approximation, done by Gill (1965). They numerically solved the dispersion relation, found unstable KH modes and classified them into the fundamental (surface) and reflection (body) family of modes. The classification is related to the number of nodes, across the jet, of sound waves reflecting in between jet boundaries. An internal wave pattern is formed by the composition of oblique waves, for which the jet interior is a resonant cavity. The physical meaning of KH instability in supersonic jets has been discussed by Payne & Cohn (1985), who have shown that the presence of instability is associated with the overreflection of soundwaves (the modulus of reflection coefficient is larger than 1) on the sheared jet boundaries.
Subsequent studies include effects of magnetic field (Ferrari, Trussoni & Zaninetti 1981), the effects of the shear layer, replacing the vortex sheet in the nonrelativistic planar case (Ferrari, Massaglia & Trussoni 1981; Ray 1982; Choudhury & Lovelace 1984), and conical and cylindrical jet geometry (Birkinshaw 1984; Hardee 1984, 1986, 1987). The effects of cylindrical a shear layer have been examined by Birkinshaw (1991) in the nonrelativistic case, attempted by Urpin (2002) in the relativistic case and the presence of multiple components (jet + sheath + ambient medium) was investigated by Hanasz & Sol (1996, 1998). Other authors have investigated current-driven modes in magnetized jets (Appl & Camenzind 1992; Appl 1996) in addition to KH modes. An extension of the linear stability analysis in the relativistic case to the weakly nonlinear regime has been performed by Hanasz (1995) and led to the conclusion that Kelvin-Helmholtz instability saturates at finite amplitudes due to various nonlinear effects. An explanation of the nature of the mentioned nonlinearities has been proposed by Hanasz (1997). The most significant effect results from the relativistic character of the jet flow, namely from the fact that the velocity perturbation cannot exceed the speed of light.
A more recent approach is to perform a linear stability analysis in parallel with numerical simulations and to compare the results of both methods in the linear regime and then to follow the nonlinear evolution of the KH instability resulting from numerical simulations. Hardee & Norman (1988) and Norman & Hardee (1988) have made such a study for nonrelativistic jets in the spatial approach and Bodo et al. (1994) in the temporal approach. In the relativistic case this kind of approach was applied for the first time by Hardee et al. (1998) in the case of axisymmetric, cylindrical jets and then extended to the 3D case by Hardee et al. (2001).
Similarly to the linear stability analysis, the numerical simulations of jet evolution can be performed following both the spatial and the temporal approach, depending on the particular choice of initial and boundary conditions. In the temporal approach one considers a short slice of jet limited by periodic boundaries along the jet axis, and adds some specific perturbation, eg. an eigenmode resulting from the linear stability analysis. Due to the periodic boundary conditions the growing perturbations can only be composed of modes having a wavelength equal to the length of the computational box and/or its integer fractions (Bodo et al. 1994, 1995, 1998). Whereas the spatial approach appears more appropriate to analyze the global dynamics and morphology of the whole jet, the temporal approach is suitable for the comparison between the numerical results and analytical studies of the jet stability because, due to the fact that only part of the jet is simulated, a high effective numerical resolution is achievable with limited computer resources. Numerical simulations (Martí et al. 1997; Hardee et al. 1998, Rosen et al. 1999) demonstrate that jets with high Lorentz factors and high internal energy are influenced very weakly by the Kelvin-Helmholtz instability. Moreover, Hardee et al. (1998), Rosen et al. (1999) note that contrary to the cases with lower Lorentz factors and lower internal energies, the relativistically hot and high Lorentz factor jets do not develop modes of KH instability predicted by the linear theory. They interpret this fact as the result of a lack of appropriate perturbations generating the instability in the system. In the limit of high internal energies of the jet matter the Kelvin-Helmholtz instability is expected to develop with the highest growth rate.
In this paper we focus on the role of internal energy and the Lorentz factor for the stability of relativistic planar two-dimensional jets in the temporal approach. We chose this simple configuration in order to enhance our ability to understand the complex nonlinear evolution of the unstable KH modes in relativistic jets. For this purpose, we perform a linear analysis of the KH instability and construct perturbations analytically: the eigenmodes of the linear problem representing the most unstable first reflection mode for each set of jet parameters. These modes are then incorporated as initial states of numerical simulations. The choice of planar two-dimensional jets allows us to study symmetric and antisymmetric modes of KH instability which resemble, respectively, pinch and helical modes of cylindrical jets. Only the flutting modes of cylindrical jets, corresponding to azimuthal wavenumbers larger than 1, have no counterparts among the eigenmodes of planar two-dimensional jets. In the present paper we investigate only the symmetric modes. Our aim is to investigate physical details of the process of transition of KH instability modes from the linear to the nonlinear regime. In the next step (Paper II) we shall investigate the relations of long-term nonlinear evolution of KH instability (including the mixing phase and the properties of the evolved flow at the end of our simulations) on the initial jet parameters.
The paper is organized as follows. In Section \[sect:method\] we describe our method, i.e. perform the linear analysis of the KH instability for the relativistic planar case, then define perturbations and describe numerical algorithm and other details of numerical simulations. In Section \[sect:results\] we present and discuss results of the both linear stability analysis and the numerical simulations. In Section \[s:concl\] we summarize and conclude our paper.
Method \[sect:method\]
======================
Following the standard procedure (see eg. Gill 1965; Ferrari, Trussoni & Zaninetti 1978; Hardee 1979) we derive the dispersion relation for the Kelvin-Helmholtz modes. We focus on the simplest geometrical configuration of two-dimensional planar relativistic jets and apply the temporal stability analysis.
Linear stability analysis
-------------------------
The full set of equations describing the current problem consists of the set of relativistic equations of hydrodynamics for a perfect fluid, (e.g. Ferrari, Trussoni & Zaninetti 1978) $$\gamma^2 \!\!\left( \rho \!+\! {p\over c^2} \right)\!\! \left[\!
{\partial \vec v \over \partial t}\!+\! (\vec v \cdotp \! \vec\nabla)
\vec v \! \right]
\! + \vec\nabla p + {\vec v \over c^2} {\partial p \over \partial t}
\! = \! 0, \label{eulereq}$$ $$\gamma \!\!\left( {\partial \rho\over \partial t} \!+\! \vec v \cdotp
\vec \nabla\rho \right) \!+\! \left( \rho + {p\over c^2} \right)\!\!
\left[ {\partial \gamma \over \partial t} + \vec\nabla \cdotp (\gamma \vec v)
\right] \!\!= \! 0, \label{conteq}$$ and an (adiabatic) equation of state $$p \rho_0^{-\Gamma} = \mbox{{\rm const}}. \label{eqos}
\label{eos}$$ In the preceding equations, $c$ is the speed of light, $\rho_0$ is the particle rest mass density (i.e., $\rho_0 \equiv m n$, where $m$ is the particle rest mass and $n$ the number density in the fluid rest frame). $\rho$ stands for the relativistic density which is related to the particle rest mass density and the specific internal energy, $\varepsilon$, by $\rho = \rho_0(1+\varepsilon/c^2)$. The enthalpy is defined as $w = \rho + p/c^2$, the sound speed is given by $c_s^2 =
\Gamma p/w$, where $\Gamma$ is the adiabatic index. The relation between pressure and the specific internal energy is $p =(\Gamma - 1)
\varepsilon \rho_0$. The velocity of the fluid is represented by $\vec
v$ and $\gamma$ is the corresponding Lorentz factor.
The assumed geometry of the jet considered in the forthcoming linear stability analysis and the numerical simulations is sketched in Fig. \[f:sketch\]. First of all the considered jet is 2D-planar and symmetric with respect to the $x=0$ plane. The flow in the jet moves in the positive $z$ direction and its matter forms a contact discontinuity (a vortex sheet) with the matter of external medium at $x= -R_{\rm j}$ and $x=R_{\rm j}$. From now on all quantities representing the jet will be assigned with the ’$j$’ subscript and the quantities representing the ambient medium will be assigned with ’$a$’.
The following matching conditions are imposed on the interface between the jet and the ambient medium $$\begin{aligned}
p_a = p_j & \mbox{for} & |x|=R_j \label{pmatch} \\
h_a = h_j & \mbox{for} & |x|=R_j.\label{hmatch}\end{aligned}$$ The matching conditions express the assumption of equality of the jet and ambient pressures and the equality of transversal displacements of jet ($h_j$) and ambient ($h_a$) fluid elements adjacent to the jet boundary (at $|x| = R_j$).
In addition, the Sommerfeld radiation condition (expressing the disappearance of perturbations at the infinity) will be applied for linear perturbations.
### Equilibrium state for the linear stability analysis \[sect:equilibrium\]
We assume that the initial state is an equilibrium configuration. The initial equilibrium can be described by the following set of independent parameters: the Lorentz factor corresponding to the unperturbed longitudinal jet velocity, $v_j$, $\gamma = (1 - v_j^2/c^2)^{-1/2}$, the particle rest mass density of the jet $\rho_{0j}$ (the particle rest mass density of the ambient medium is normalized to unity: $\rho_{0a} = 1$) and the specific internal energy of the jet $\varepsilon_j$. The ambient medium is assumed to be at rest ($v_a = 0$).
The other dependent parameters describing the equilibrium state are: the internal jet Mach number $M_j = {v_j}/{c_{sj}}$ corresponding to the initial jet longitudinal velocity, the relativistic rest mass density contrast $\displaystyle{\nu =
\frac{\rho_{0j}(1+\varepsilon_j/c^2)}
{\rho_{0a}(1+\varepsilon_a/c^2)}}$ (or, equivalently, the enthalpy contrast $\displaystyle{\eta=
\frac{\rho_{0j}\left(1+(\Gamma_j-1)\varepsilon_j/c^2/
(1+\varepsilon_j/c^2)\right)}
{\rho_{0a}\left(1+(\Gamma_a-1)\varepsilon_a/c^2/
(1+\varepsilon_a/c^2)\right)}}$) and the specific internal energy of the ambient medium $\displaystyle{\varepsilon_a = \frac{(\Gamma_j - 1)
\rho_{0j}}{(\Gamma_a - 1) \rho_{0a}} \varepsilon_j}$, which is related to the specific internal energy of the jet through the pressure balance condition.
### Dispersion relation
The first step towards the dispersion relation is to reduce the equations (\[eulereq\]) - (\[hmatch\]) to the dimensionless form through the normalization of spatial coordinates to the jet radius $R_j$, velocities to the sound speed of the jet material $c_{sj}$, time to the dynamical time $R_j/c_{sj}$ and pressure to the equilibrium pressure.
The next step is to decompose each dependent quantity into the equilibrium value and the linear perturbation. After the reduction of equations to the dimensionless form and substitution of the perturbed quantities in equations (\[eulereq\])-(\[hmatch\]), we obtain the following set of dimensionless linearized equations. In the following the dimensionless quantities will be assigned the same symbols as the previous dimensional ones. The subscript ’$1$’ stands for the linear perturbation of the corresponding variable. For the jet medium we get
$$\label{linjet1}
\Gamma_j \, \gamma^2 \left(\frac{\partial{{\mbox{{\boldmath $v$}}}_{j1}}}{\partial t} +
({\mbox{{\boldmath $v$}}}_j \cdot \nabla) {\mbox{{\boldmath $v$}}}_{j1} \right) + \nabla p_{j1} +
\frac{{\mbox{{\boldmath $v$}}}_j}{c^2} \frac{\partial p_{j1}}{\partial t} = 0,$$
$$\begin{aligned}
\label{linjet2}
\frac{\partial p_{j1}}{\partial t} + {\mbox{{\boldmath $v$}}}_j \cdot \nabla p_{j1} +
\Gamma_j \nabla \cdot {\mbox{{\boldmath $v$}}}_{j1} \hspace{4cm} \\ + \gamma^2 \,
\Gamma_j \frac{{\mbox{{\boldmath $v$}}}_j}{c^2} \left(\frac{\partial
{\mbox{{\boldmath $v$}}}_{j1}}{\partial t} + ({\mbox{{\boldmath $v$}}}_j \cdot \nabla
{\mbox{{\boldmath $v$}}}_{j1})\right) = 0 \quad ,\nonumber\end{aligned}$$
where ${\mbox{{\boldmath $v$}}}_j$ is the initial (unperturbed) jet velocity in units of jet internal sound speed, $c$ is the speed of light in units of the sound speed and the normalized pressure $p_0=1$ is omitted in the equations. For the ambient medium we get
$$\label{linam1}
\frac{\Gamma_j}{\eta} \frac{\partial
{{\mbox{{\boldmath $v$}}}_{a1}}}{\partial t} + \nabla p_{a1} = 0,$$
$$\label{linam2}
\frac{\partial p_{a1}}{\partial t}
+ \Gamma_a\nabla \cdot {\mbox{{\boldmath $v$}}}_{a1} = 0.$$
The linearized matching conditions (\[pmatch\]) and (\[hmatch\]) at the jet interface read
$$\begin{aligned}
p_{a1} = p_{j1} & \mbox{for} & |x|=1 \label{pmatch1} \\
h_{a1} = h_{j1} & \mbox{for} & |x|=1 \label{hmatch1},\end{aligned}$$
where the displacements of fluid elements adjacent to the contact interface in the linear regime are related to transversal velocities by the following formulae $$v_{jx1} = \left( {\frac{{\partial}^{} {}}{{\partial} {t}^{}}} + v_j {\frac{{\partial}^{} {}}{{\partial} {z}^{}}} \right) h_{j1},
\label{hj}$$ $$v_{ax1} ={\frac{{\partial}^{} {}}{{\partial} {t}^{}}} h_{a1}.
\label{ha}$$
The following wave equations can be derived from the equations (\[linjet1\]) - (\[linam2\]), respectively for the jet
$$\begin{aligned}
\label{pj}
\gamma^2 \left({\frac{{\partial}^{} {}}{{\partial} {t}^{}}}+v_j{\frac{{\partial}^{} {}}{{\partial} {z}^{}}} \right)^2 p_{j1} -
{\frac{{\partial}^{2} {p_{j1}}}{{\partial} {x}^{2}}} \hspace{3cm}\\ - \gamma^2\left( {\frac{{\partial}^{} {}}{{\partial} {z}^{}}} +
\frac{v_j}{c^2} {\frac{{\partial}^{} {}}{{\partial} {t}^{}}}\right)^2 p_{j1} =0, \nonumber\end{aligned}$$
and for the ambient medium
$$\label{pa}
{\frac{{\partial}^{2} {p_{a1}}}{{\partial} {t}^{2}}} - \frac{\eta\Gamma_a}{\Gamma_j}\left(
{\frac{{\partial}^{2} {p_{a1}}}{{\partial} {x}^{2}}}+ {\frac{{\partial}^{2} {p_{a1}}}{{\partial} {z}^{2}}}\right) = 0.$$
It is apparent that Eqn. (\[pj\]) and (\[pa\]) describe propagation of oblique sound waves in the jet and ambient medium respectively.
The next step is to substitute perturbations of the form $$\label{eq:deltab}
\delta_{j1} = [ \delta^+_{j} F^+_{j} (x) + \delta^-_{j} F^-_{j} (x) ]
\exp i (k_{\parallel } z - \omega t) + c.c.$$ with $F^{\pm}_{j} (x) = \exp (\pm i\, k_{{j\perp}} x)$ to describe waves propagating in positive and negative $x$-directions in the beam. Here we use $k_{\parallel}$ and $k_{a,j\perp}$ for longitudinal and transverse wavenumbers. Perturbations in the external medium are of the form $$\label{eq:deltaa}
\delta_{a1} = \delta^+_a F^+_a (x) \exp\, i (k_{\parallel } z-\omega
t) +c.c.$$ where $F^+_a (x) = \exp (i\, k_{a\perp} x)$ for $|x|>1$. Assuming that ${\rm Re} (k_{a\perp}) > 0$, only outgoing waves are present in the ambient medium. After the substitution of the explicit forms of pressure perturbations to equations (\[pj\]) and (\[pa\]) we obtain the following expressions for the perpendicular components of wavevectors
$$k_{a \perp} = \left(\frac{\Gamma_j}{\eta \Gamma_a} \omega^2 -
k_\parallel^2\right)^{1/2}$$
$$k_{j \perp} = ({\omega'}^2-{k'_\parallel}^2)^{1/2}$$
which are standard relations for linear sound waves in both media (note that $\eta \Gamma_a/\Gamma_j$ is the squared sound speed of ambient medium in units of the sound speed of jet), $\omega' =
\gamma(\omega - v_j k_\parallel)$ and $k'_{\parallel} =
\gamma(k_\parallel - \frac{v_j}{c^2} \omega)$ are frequency and wavenumber of the internal sound wave in the reference frame comoving with the jet. The wavevectors ${\mbox{{\boldmath $k$}}}_j = (k_{j\perp}, k_{j\parallel})$ and ${\mbox{{\boldmath $k$}}}_a = (k_{a\perp}, k_{a\parallel})$ determine the direction of propagation of sound waves in the jet and ambient medium respectively. Vanishing of $k_{j\perp}$, for instance, would mean that the jet internal sound waves move in the axial direction. In cases $k_{j\perp} \ne 0 $ and/or $k_{a\perp} \ne 0 $ the propagation of the sound waves is oblique with respect to the jet axis.
The corresponding perturbation of the jet surface can be written as $$h_{a1}=h_{j1}= h \exp\, i (k_{\parallel } z-\omega t) +c.c. ,$$ which after the substitution to equations (\[hj\]) and (\[ha\]) reads $$v_{jx1} = -i \left(\omega - v_j k_{\parallel}\right) h_{j1} ,$$ $$v_{ax1} = -i \omega h_{j1}.$$
With the aid of the above perturbations the whole system of partial differential equations is reduced to a set of homogeneous linear algebraic equations. The dispersion relation appears as solvability condition of the mentioned set of equations, namely it arises from equating the determinant of the linear problem to zero. Within the present setup, the dispersion relation for the Kelvin-Helmholtz instability in supersonic, relativistic, two-dimensional slab jets can be written as $$\begin{aligned}
\frac{1}{\nu \Gamma_j} \frac{\omega}{\omega'} \frac{({\omega'}^2 -
{k'}^2_{\parallel})^{1/2}} {\left(\frac{\omega^2}{\nu \Gamma_a} -
k_\parallel^2\right)^{1/2}} =
\hspace{4cm}\\
\hspace{2cm} =\left\{
\begin{array}{lll}
\!\!{\rm coth}\ i ({\omega'}^2 - {k'}^2_{\parallel})^{1/2} & {\rm
for} & s = 1\\ & {\rm }& \\ \!\!{\rm th}\ i({\omega'}^2 -
{k'}^2_{\parallel})^{1/2} & {\rm for}& s=-1\\
\end{array} \nonumber
\right. \label{disprel}\end{aligned}$$ where $s = \pm 1$ is the symmetry of perturbation. In the present paper we focus on symmetric perturbations ($s=1$).
The above dispersion relation is solved numerically with the aid of the Newton-Raphson method (see, e.g., Press et al. 1992). In this way we obtain the complex frequency as a function of the parallel component of the wavenumber $k_\parallel$, since we choose the temporal stability analysis for the present investigations.
### Eigenstates of the system \[sect:eigenstates\]
Once the solutions of the dispersion relation $\omega(k)$ are found, the derivation of eigenstates of the system is straightforward. By utilizing the set of relativistic equations (\[linjet1\]) - (\[linam2\]) one can relate perturbations of gas density and velocity to the perturbation of pressure.
First, we choose the amplitude of the pressure wave in the jet, $p_j^+$, as a free constant and then relate the other constants to its value. The corresponding ’$-$’ amplitudes are computed by multiplying these values by $s$ in case of pressure and longitudinal velocity and by $-s$ in the case of transversal velocity.
The amplitude of pressure wave in the ambient medium, $p_a^+$, is found from the pressure matching condition at $x=1$:
$$p_{a}^+\exp(i\,k_{a\perp}) = p_{j}^+ \exp(i\,k_{j\perp})+s\,p_{j}^-
\exp(-i\,k_{j\perp}).$$
The remaining amplitudes of velocity perturbations are
$$v_{ax}^{+}= \frac{\eta}{\Gamma_j}
\frac{k_{a\perp}}{\omega} p_a^+,$$
$$v_{jx}^{+}= \frac{k_{j\perp}}{\Gamma_j \gamma
\omega'} p_j^{+}, \label{pj-vjperp}$$
$$v_{az}^{+} = \frac{\eta}{\Gamma_j}
\frac{k_{\parallel}}{\omega} p_a^+,$$
$$v_{jz}^{+} = \frac{k'_{\parallel}}{\Gamma_j \gamma^2
\omega'} p_j^+,$$
where $\eta$ is the ratio of enthalpies as defined earlier.
Regarding specific internal energy and rest-mass density, perturbations are calculated as follows:
$$\frac{\varepsilon_{a,j1}}{\varepsilon_{a,j}}= \frac{\Gamma_{a,j}-
1}{\Gamma_{a,j}} p_{a,j1}$$
$$\frac{\rho_{a,j1}}{\rho_{a,j}+1/c^2}=\frac{1}{\Gamma_{a,j}} p_{a,j1},$$
where $\rho_{a,j}$ refer to the relativistic density in ambient and jet media in dimensionless units, respectively, and $\rho_{a,j1}$ is the corresponding perturbation.
### Lorentz transformation of velocity perturbations
In the forthcoming discussion of results of numerical simulations we shall analyze the velocity components in the reference frame comoving with the jet. The standard Lorentz transformation formulae for the velocity components are $$\begin{aligned}
v'_z = \frac{v_z - v_j}{1- v_z v_j/c^2}, &&
v'_x = \frac{{v_x}\sqrt{1-v_j^2/c^2}}{1-v_z v_j/c^2}.
\label{eq:lorentz}\end{aligned}$$ Assuming that the Lorentz transformation is applied for linear perturbations of sufficiently small amplitude, these transformation rules can be approximated by the following formulae $$\begin{aligned}
v'_{z1} = \gamma^2 v_{z1}, &&
v'_{x1} = \gamma v_{x1}, \label{lorentz-pert}\end{aligned}$$ where $v_{x1}$, $v_{z1}$ are the velocity perturbations in the reference frame of the ambient medium and $v'_{x1}$, $v'_{z1}$ are the velocity perturbations in the jet reference frame. These transformation rules tell us that the longitudinal and transversal velocity perturbations in the jet reference frame are larger by factors of $\gamma^2$ and $\gamma$ respectively than the corresponding components in the rest (ambient medium) frame.
Numerical simulations {#ss:numsim}
---------------------
Now we proceed by performing numerical simulations of jet models to study the growth of unstable modes through the linear and non-linear phases. To this aim, we adopt as an equilibrium initial state the one described in the former section. In order to avoid the growth of random perturbations with wavelengths of the order of the cell size, we replace the transverse discontinuous profiles of equilibrium quantities by smooth profiles of the form
$$v_z(x) = \frac{v_j}{\cosh(x^m)} \label{shearlayer1}$$
$$\rho_0(x) = \rho_{0a} - \frac{\rho_{0a} - \rho_{0j}}{\cosh(x^m)}
\label{shearlayer2}$$
(see Bodo et al. 1994), where $m$ is the steepness parameter for the shear layer. Typically we used $m=40$, for which the shear layer has a width $\sim 0.1 R_j$ (see Appendix A).
On the other hand the linear solutions derived in Section \[sect:eigenstates\] correspond to an idealized case with a contact discontinuity at the jet interface. Finding of analogous solutions for the corresponding sheared boundary problem is more difficult from the numerical point of view (see Ray 1982, and Choudhury & Lovelace 1984), therefore we decided to implement corresponding smoothed eigenstates obtained for the vortex sheet limit as approximate solutions for the case of sheared boundary. The validity of this procedure is proven [*a posteriori*]{} by the convergence of theoretical growth rates and the ones determined for simulated KH modes. Since, in case of the vortex sheet approximation, all first order dynamical variables (except pressure and the transversal displacement of fluid element adjacent to the jet boundary) are discontinuous, we apply the following smoothing of the equilibrium quantities and linear perturbations at the thin layer surrounding the jet boundary
$$\delta_1=\delta_{a1}-\frac{\delta_{a1}-\delta_{j1}}{\cosh(x^m)},
\label{smooth-pert}$$
where $\delta$ stands for any perturbed variable. Such smoothed eigenstates are subsequently implemented as small-amplitude perturbations in the initial equilibrium model of the simulations.
Our first aim is to reproduce the linear evolution of unstable modes by means of hydrodynamical simulations. For this reason and in order to make the picture of evolution of the KH instability as clear and simple as possible, we perturb the equilibrium with the most unstable first reflection mode, which is smoothed at the jet interface according to formula (\[smooth-pert\]). The length of the computational grid is in each simulation equal to the wavelength of the mentioned mode. Defining the initial state in such a way, we know precisely which growth rate to expect in numerical simulations. The difference between the expected and computed growth rates is a measure of the performance of the code in describing the linear regime and gives us confidence on the numerical results from the non-linear evolution.
The simulations have been performed using a high–resolution shock capturing code to solve the equations of relativistic hydrodynamics in Cartesian planar coordinates. The code is an upgrade of that used to study large scale (Martí et al. 1997) as well as compact relativistic jets (Gómez et al. 1997). The equations of relativistic hydrodynamics are written in conservation form and the time variation of the proper rest mass, momentum and total energy within numerical cells are calculated through the fluxes across the corresponding cell interfaces. Fluxes are calculated with an approximate Riemann solver that uses the complete characteristic information contained in the Riemann problems between adjacent cells (Donat & Marquina 1996). It is based on the spectral decomposition of the Jacobian matrices of the relativistic system of equations derived in Font et al. (1994) and uses analytical expressions for the left eigenvectors (Donat et al. 1998). The spatial accuracy of the algorithm is improved up to third order by means of a conservative monotonic parabolic reconstruction of the pressure, proper rest-mass density and the spatial components of the fluid four-velocity, as in [GENESIS]{} (see Aloy et al. 1999, and references therein). Integration in time is done simultaneously in both spatial directions using a total-variation-diminishing (TVD) Runge-Kutta scheme of high order (Shu & Osher 1988). See Martí et al. (1997) for the differential equations as well as their finite–difference form, and a description of exhaustive testing of the hydrodynamical code. Besides relativistic density, momentum and energy, the code also evolves a passive scalar representing the jet mass fraction. This allows us to distinguish between ambient and jet matter helping us to characterize processes like jet/ambient mixing or momentum exchange.
The parameters of the simulations presented in this paper are listed in Table \[tab:param\]. Values of the parameters were chosen to be close to those used in some simulations by Hardee et al. (1998) and Rosen et al. (1999) and are chosen to span a wide range in thermodynamical properties as well as beam flow Lorentz factors. In all simulations, the density in the jet and ambient gases are $\rho_{0j}=0.1$, $\rho_{0a}=1$ respectively and the adiabatic exponent $\Gamma_{j,a}=4/3$.
Since the internal rest mass density is fixed, there are two free parameters characterizing the jet equilibrium: the Lorentz factor and the jet specific internal energy displayed in columns 2 and 3 of Table \[tab:param\]. Models whose names start with the same letter have the same thermodynamical properties. Beam (and ambient) specific internal energies grow from models A to D. Three different values of the beam flow Lorentz factor have been considered for models B, C and D. The other dependent parameters mentioned in Section \[sect:equilibrium\] are displayed in columns 4-10 of Table \[tab:param\]. Note that given our choice of $\rho_{0j}$, the ambient media associated to hotter models are also hotter. The next three columns show the longitudinal wavenumber together with oscillation frequency and the growth rate of the most unstable reflection mode. The following three columns display the same quantities in the jet reference frame. Next two columns show the perpendicular wavenumbers of linear sound waves in jet and ambient medium respectively. The last column shows the linear growth rate of KH instability in the jet reference frame expressed in dynamical time units, i.e. in which time is scaled to $R_j/c_{sj}$. All other quantities in the table, are expressed in units of the ambient density, $\rho_{0a}$, the speed of light, $c$, and the jet radius, $R_j$.
As it is apparent in Table \[tab:param\] growth rates tend to increase with the specific internal energy of the beam and to decrease with the beam flow Lorentz factor. Note that, in the jet reference frame, models with the same thermodynamical properties tend to have (within $\approx 10 $ %) the same growth rates. Note also that in the jet reference frame and in dynamical time units, all the models have comparable (within a factor 1.5) linear growth rates.
$
\begin{array}{|c|cc|ccccccc|ccc|ccc|ccc|}
\hline
{\rm Model} & \gamma & \varepsilon_j & \varepsilon_a & c_{sj} & c_{sa} &
p & \nu & \eta & M_j & k_{\parallel} & \omega_r & \omega_i
&{k'_\parallel} & \omega'_{r} & \omega'_i & k_{j\perp} &
k_{a\perp} & {\omega'}^{\rm dyn}_i\\
\hline
{\rm A05} & 5 & 0.08 & 0.008 & 0.18 & 0.059 & 0.0027 & 0.11 & 0.11 & 5.47 & 0.30 & 0.20 & 0.026 & 1.32 & 7.20 & 0.13 & 7.08 & 0.53 & 0.73 \\
{\rm B05} & 5 & 0.42 & 0.042 & 0.35 & 0.133 & 0.014 & 0.14 & 0.15 & 2.83 & 0.69 & 0.49 & 0.055 & 2.62 & 7.32 & 0.28 & 6.84 & 1.08 & 0.79 \\
{\rm C05} & 5 & 6.14 & 0.614 & 0.55 & 0.387 & 0.205 & 0.44 & 0.51 & 1.80 & 2.00 & 1.60 & 0.114 & 5.73 & 9.98 & 0.57 & 8.17 & 1.07 & 1.05 \\
{\rm D05} & 5 & 60.0 & 6.000 & 0.57 & 0.544 & 2.000 & 0.87 & 0.90 & 1.71 & 2.63 & 2.18 & 0.132 & 7.02 & 11.56 & 0.66 & 9.18 & 0.24 & 1.15 \\
{\rm B10} & 10 & 0.42 & 0.042 & 0.35 & 0.133 & 0.014 & 0.14 & 0.15 & 2.88 & 0.50 & 0.41 & 0.031 & 3.59 & 10.28 & 0.31 & 9.64 & 0.94 & 0.90 \\
{\rm C10} & 10 & 6.14 & 0.614 & 0.55 & 0.387 & 0.205 & 0.44 & 0.51 & 1.83 & 1.91 & 1.72 & 0.055 & 9.77 & 17.67 & 0.55 & 14.72 & 1.49 & 1.01 \\
{\rm D10} & 10 & 60.0 & 6.000 & 0.57 & 0.544 & 2.000 & 0.87 & 0.90 & 1.73 & 2.00 & 1.81 & 0.063 & 9.67 & 16.58 & 0.63 & 13.47 & 0.20 & 1.10 \\
{\rm B20} & 20 & 0.42 & 0.042 & 0.35 & 0.133 & 0.014 & 0.14 & 0.15 & 2.89 & 0.46 & 0.39 & 0.014 & 6.51 & 18.76 & 0.28 & 17.60 & 0.90 & 0.81 \\
{\rm C20} & 20 & 6.14 & 0.614 & 0.55 & 0.387 & 0.205 & 0.44 & 0.51 & 1.83 & 1.44 & 1.37 & 0.027 & 13.89 & 25.38 & 0.54 & 21.24 & 1.28 & 0.99 \\
{\rm D20} & 20 & 60.0 & 6.000 & 0.57 & 0.544 & 2.000 & 0.87 & 0.90 & 1.74 & 2.00 & 1.91 & 0.029 & 18.11 & 31.43 & 0.58 & 25.68 & 0.31 & 1.01 \\
\hline
\end{array}
$
The initial numerical setup consists of a steady two-dimensional slab jet model (see Fig. \[f:sketch\]). As stated above, a thin shear layer between the ambient medium and the jet is used instead of the vortex sheet. Due to symmetry properties, only half of the jet ($x>0$) has to be computed. Reflecting boundary conditions are imposed on the symmetry plane of the flow, whereas periodical conditions are settled on both upstream and downstream boundaries. The numerical grid covers a physical domain of one wavelength along the jet (3 to 20 $R_j$; see Table \[tab:param\]) and 100 $R_j$ across (200 $R_j$ in the case of models D). The size of the transversal grid is chosen to prevent loses of mass, momentum and energy through the boundaries. 400 numerical zones per beam radii are used in the transverse direction across the first 3 $R_j$. From this point up to the end of the grid, 100 (200, in case of models D) extra numerical zones growing geometrically have been added. The width growth factor between contiguous zones is approximately 1.08 for models A, B and C and 1.04, for models D. Along the jet, a resolution of 16 zones per beam radius has been used. The agreement of the linear stability analysis with the results of numerical simulations depends on the grid resolution (and the width of the initial shear layer). The applied resolution of $400
\times 16$ grid zones per beam radius is chosen on the base of several tests, which are presented in detail in Appendix A.
The steady model is then perturbed according to the selected mode, with an absolute value of the pressure perturbation amplitude inside the beam of $p_j^{\pm}=10^{-5}$. This means that those models with the smallest pressure, like model A, have relative perturbations in pressure three orders of magnitude larger than those with the highest pressures, D. However this difference seems not to affect the linear and postlinear evolution.
Results \[sect:results\]
========================
Following the behaviour of simulated models we find that the evolution of the perturbations can be divided into a linear phase, a saturation phase and mixing phase. This section is devoted to describing the properties of the linear and saturation phases in our models focusing on the influence of the basic parameters. The remaining fully nonlinear evolution will be discussed in Paper II. Our description shares many points with that of Bodo et al. (1994) for the case of classical jets.
In order to illustrate the growth of perturbations and determine the duration of the linear and saturation phases in our simulations, we plot in Fig. \[fig:amplitudes\] the amplitudes of the perturbations of the longitudinal and transversal velocities, inside the jet and in the jet reference frame, together with the pressure oscillation amplitude. We plot also the growth of the imposed eigenmodes resulting from the linear stability analysis. Both the velocity perturbations are transformed from the ambient rest frame to the unperturbed jet rest frame using the Lorentz transformation rules for velocity components, given by Eqn. (\[eq:lorentz\]).
We define the characteristic times in the following way. During the linear phase the ratios of the exponentially growing amplitudes of pressure, longitudinal velocity and transversal velocity remain constant by definition. We define the end of the linear phase ($t_{\rm lin}$) as the moment at which one of quantities starts to depart from the initial exponential growth. Within the set of our models the first quantity to break the linear behaviour is the longitudinal velocity perturbation.
Later on the transversal velocity saturates, i.e. stops growing at the saturation time ($t_{\rm sat}$). We call the period between $t_{\rm
lin}$ and $t_{\rm sat}$ the saturation phase. We find also that the pressure perturbation amplitude reaches a maximum. This moment is denoted by $t_{\rm peak}$. In Paper II we will see that this peak announces the entering of the fully nonlinear regime. The choice of $t_{\rm lin}$, $t_{\rm sat}$, $t_{\rm peak}$ has been illustrated in Fig. \[fig:amplitudes\] (top panel).
Table \[tab:phases\] collects times of the linear and saturation phases in the different models (columns 2-4). The last column shows the saturation time in dynamical units and in the jet reference frame. The change of reference frame eliminates the effects coming from the jet flow Lorentz factor that stretches out the rhythm of evolution in the ambient rest frame. Dynamical time units are adapted to the characteristic time of evolution of each model. Hence this change of units and reference frame allows us to compare the relevant scales of evolution of all the models directly.
$
\begin{array}{|c|ccc|c|}
\hline
{\rm Model} & t_{\rm lin} & t_{\rm sat} & t_{\rm peak}
& {t'}_{\rm sat}^{\rm dyn}\\
\hline
{\rm A05} & 180 & 380 & 380 & 13.69 \\
{\rm B05} & 125 & 200 & 205 & 13.99 \\
{\rm C05} & 100 & 125 & 130 & 13.74 \\
{\rm D05} & 105 & 120 & 130 & 13.71 \\
\hline
{\rm B10} & 235 & 380 & 385 & 13.29 \\
{\rm C10} & 210 & 245 & 250 & 13.46 \\
{\rm D10} & 180 & 225 & 225 & 12.86 \\
\hline
{\rm B20} & 450 & 760 & 780 & 13.29 \\
{\rm C20} & 270 & 645 & 775 & 17.72 \\
{\rm D20} & 350 & 480 & 500 & 13.71 \\
\hline
\end{array}
$
Linear phase {#ss:linear}
------------
In the linear phase the ratios of oscillation amplitudes of different dynamical variables (density, pressure and velocity components) are constant. We have used this property to identify the end of the linear phase with the time $t_{\rm lin}$ at which one of the variables deviates from linear growth in a systematic way. Note that our definition is different from that of Bodo et al. (1994) who associate the end of the linear phase with the formation of the first shocks inside the jet.
We note that in our simulations, during the linear phase, the growth of perturbations of each dynamical variable follows the predictions of the linear stability analysis with relatively good accuracy. On average the growth rates measured for all our numerical experiments are about 20 % smaller than corresponding growth rate resulting from the linear stability analysis. This small discrepancy may be partially a result of the application of the shearing layer in simulations and the vortex sheet approximation in the linear stability analysis, and also to a lack of transversal resolution in the numerical simulations (see Appendix A). In all cases the amplitude of the longitudinal velocity oscillation is the first quantity to stop growing. The reason for the limitation of the velocity oscillations is obvious: the oscillations of velocity components (corresponding to sound waves propagating in the jet interior) cannot exceed the speed of light. This limitation (specific to relativistic dynamics) is easily noticeable in the jet reference frame, but it is obscured by the Lorentz factor (in the first and second power) in the reference frame of the ambient medium, as it follows from the formulae (\[lorentz-pert\]).
\
As seen in Fig. \[fig:amplitudes\], model C20 evolves distinctly from the other models: at $t = 270$, the linear phase ends before the saturation of longitudinal speed. See the middle panels of Fig. \[fig:fcd20s\] (especially the pressure panel) which could be responsible for the peculiar evolution of model C20. Note also that model C20 is the only one that has a theoretical perturbation growth rate smaller than the numerical one (probably associated with the excited short wavelength mode). At the end of the linear phase, the values of quantities like density, pressure and flow Lorentz factor are still very close to the corresponding background values (the perturbation in pressure is between 10 % and 50% of the background pressure in all the models). Figure \[fig:prepert\] shows the distribution of pressure at the end of the linear phase for two representative models B05 and D05.
Saturation phase
----------------
As it has been mentioned, the end of the linear phase coincides with the limitation of the longitudinal oscillation velocity. At time $t_{\rm lin}$ the transversal velocity component is smaller than the speed of light, by an order of magnitude approximately, so the transversal velocity perturbation has still room to grow. We defined $t_{\rm sat}$ as the time corresponding to the saturation of the transversal velocity and saturation phase as the period between $t_{\rm lin}$ and $t_{\rm sat}$.
We find that the longitudinal velocity perturbation amplitude reaches almost the speed of light (more than $0.9 c$) while the transversal perturbation amplitude stops its growth at the level of approximately $0.5c$ for all presented simulations. This can be explained in the following way. We remind that the eigenmodes are built of oblique sound waves overreflected at the jet/ambient medium interface. This means that the amplitude of the reflected wave is larger than the amplitude of the incident one. Sound waves are longitudinal waves, so the gas oscillation velocity can be split into the longitudinal and transversal components separately for the incident and reflected waves. Let us consider locally the incident wave and the reflected wave in the jet medium, in a fixed point close to the jet boundary. Then the longitudinal velocity components of the incident and reflected waves sum in phase, while the transversal ones sum in counterphase. This means that while the amplitude of the total longitudinal velocity oscillation component grows and approaches the speed of light, the oscillation amplitude of the total transversal component is a difference of the two velocities, each smaller than the speed of light. Therefore the oscillation amplitude of the total transversal component has to take values significantly smaller than the longitudinal one. The duration of the saturation phase depends on the Lorentz factor and the specific internal energy with a tendency to increase with the former and to decrease with the latter (exception made of models C10 or D10 and C20). It ranges between a few tens of (absolute) time units to a few hundreds.
We note that the saturation times expressed in dynamical time units and in the jet reference frame (see Table \[tab:phases\]) are almost equal for all models. This similarity can be explained by the following argument. First, the amplitude of pressure perturbations $p_1^\pm=10^{-5}$ is the same for all simulations. Second, the amplitudes of pressure and transversal velocity perturbations are related by formula (\[pj-vjperp\]). The perturbation of perpendicular velocity is then transformed to the jet reference frame following (\[lorentz-pert\]). Following the results of numerical experiments, the transversal velocity grows with the linear growth rate until the transversal velocity perturbation reaches the upper limit, i.e. $|v_{j\perp}^{+}|
\exp (\omega_i t_{\rm sat}) \simeq 0.5c $. If we express $t_{\rm sat}
$ in dynamical time units then in the jet reference frame
$${t'}_{\rm sat}^{\rm dyn} \simeq \frac{\gamma}{c_{sj} \omega_i}
\log \left(\frac{c \Gamma_j}{2p_1^+}
\left|\frac{\omega'}{k_{j\perp}}\right|\right) \label{tsat}.$$
Since the term $\omega'/k_{j\perp}$ under the logarithm is close to the jet sound speed and is varying only slightly in between models, while the other factors under the logarithm are constants, the saturation time is proportional to the inverse growth rate. From the linear stability analysis, the latter one measured in the jet reference frame and expressed in dynamical time units is almost equal for all the models (see last column in Table \[tab:param\]). Substitution of numbers into the formula (\[tsat\]) provides $t'_{\rm sat}$ in the range $10 \div 15$, which is consistent with the value 13.5 in Table \[tab:phases\]. However, the remarkable convergence of saturation times for all the models is difficult to explain in view of the additional randomness in the evolution of perturbations, which is apparent in Fig. \[fig:amplitudes\].
During the saturation phase, the jet inflates (Bodo et al. 1994 called this phase the expansion phase) and deforms due to transversal oscillations. On the other hand, the saturation time $t_{\rm sat}$ coincides (within a few time units) with the appearance of an absolute maximum in the pressure distribution (at $t_{\rm peak}$) at the jet boundary, and with the start of the mixing phase (to be studied in Paper II). Figure \[fig:fab05s\]-\[fig:fcd20s\] display snapshots of several quantities close to the end of the saturation phase before the distortion of the jet boundary.
The structure of perturbations at the end of the saturation phase is quite similar in all models. The longitudinal wavelength of perturbations is given by the linear stability analysis and it is constant because of the fixed length of the computational domain. Along with the longitudinal wavelength, the opening angle of oblique waves, far from the jet symmetry plane, is given by the linear analysis. The opening angle of oblique waves is enlarged in the vicinity of the jet interface. Closer to the jet symmetry plane the perturbations are already in the nonlinear phase, which can be recognized by the apparent presence of oblique shock fronts in the jet itself and the ambient medium. These shocks form as a natural consequence of the nonlinear steepening of oblique sound waves as their amplitude becomes large. Practically all the four quantities displayed for a given model (i.e. pressure, rest-mass density, Lorentz factor and internal energy) show the same structure, so there is no need to discuss them separately. Therefore we treat the pressure distribution as a representative quantity. The following properties of the flow patterns can be noticed in Figs. \[fig:fab05s\]-\[fig:fcd20s\]:
1. The interface between jet and ambient medium forms a regular sinusoidal pattern. The amplitudes are all comparable and have values of the order of $0.5 - 1.0 R_j$. In all cases the departure from a regular sinusoidal shape is rather small.
2. The structure of oblique shocks crosses the jet interface in such a way that the deformed interface is almost parallel to the shock front. One should notice however that the oblique shock moves in the flow direction in the external medium and in the counter-flow direction in the jet medium, as can be deduced from the distribution of hot post-shock and cold pre-shock material. These two oppositely moving shock waves fit together because of the jet background flow.
3. The highest gas pressure is always located on the jet symmetry plane (the brightest spots on pressure plots), where the crossing shocks form the familiar x-pattern, but the pressure enhancement on the jet interface is almost as strong as on the jet symmetry plane.
4. The shocks are much stronger (i.e., the jumps between pre- and post-shock pressure larger) in the cold cases. In the case of hot models the jumps are smoother. This is a straightforward consequence of the larger ratios of the perturbated velocity to the sound speed in the cold cases.
5. There is no significant variation of the structure of nonlinear perturbations for different values of the Lorentz factor.
6. Model C20 displays a structure of internal waves, which is absent in other cases. This fine structure is probably connected to the distinct evolution of this model in the postlinear phase (see Sect. \[ss:linear\] and Fig. \[fig:amplitudes\]).
Therefore the differences between models just before the end of the saturation phase can be considered as minor.
Summary \[s:concl\]
===================
This is the first of a series of papers devoted to the effects of relativistic dynamics and thermodynamics in the development of KH instabilities in planar relativistic jets, accross both linear and fully nonlinear regimes. To this aim, we have performed a linear stability analysis and high-resolution numerical simulations for the most unstable first reflection modes in the temporal approach, for three different values of the jet Lorentz factor $\gamma$ (5, 10 and 20) and a few different values of specific internal energy of the jet matter (from $0.08$ to $60.0 c^2$). In this paper we concentrate in the early stages of evolution of the KH instability, namely the linear and saturation phases. The present paper is also intended to set the theoretical and numerical background for the whole series of papers.
Our simulations describe the linear regime of evolution of the excited eigenmodes of the different models with high accuracy. The growth rates of the perturbed modes in the vortex sheet approximation were determined with an average relative error of 20%.
In all the examined cases the longitudinal velocity perturbation is the first quantity that departs from linear growth when it reaches a value close to the speed of light in the jet reference frame. The reason for this saturation, specific to relativistic dynamics, is not so obvious in the reference frame of the external medium where it saturates at a smaller value ($\approx c/\gamma$, where $\gamma$ is the bulk flow Lorentz factor in the jet).
The saturation phase extends from the end of the linear phase up to the saturation of the transversal velocity perturbation (at approximately $0.5c$ in the jet reference frame). The saturation times for the different numerical models have been explained from elementary considerations, i.e. from properties of linear modes provided by the linear stability analysis and from the limitation of the transversal perturbation velocity.
The limitation of the components of the velocity perturbation at the end of the linear and saturation phases allows us to conclude that the relativistic nature of the flow appears to be responsible for the departure of the system from linear evolution. This behaviour is consistent with the one deduced by Hanasz (1995, 1997) with the aid of analytical methods.
At the saturation time the perturbation structure is close to the structure of the initial perturbation (the one corresponding to the most unstable first reflection mode), except that the oblique sound waves forming the perturbation became steep due to their large amplitude. It is interesting to note that the oblique shocks are stronger (i.e., the pressure jumps are larger) in the colder cases.
Our simulations, performed for the most unstable first reflection modes, confirm the general trends resulting from the linear stability analysis: the faster (larger Lorentz factor) and colder jets have smaller growth rates. As we mentioned in the Introduction, Hardee et al. (1998) and Rosen et al. (1999) note an exception which occurs for the hottest jets. These jets appear to be the most stable in their simulations (see also the simulations in Martí et al. 1997). They suggest that this behaviour is caused by the lack of appropriate perturbations to couple to the unstable modes. This could be partially true as fast, hot jets do not generate overpressured coocons that let the jet run directly into the nonlinear regime. However, from the point of view of our results, the high stability of hot jets may have been caused by the lack of radial resolution that leads to a damping in the perturbation growth rates. We note that the agreement between the linear stability analysis and numerical simulations of KH instability in the linear range has been achieved for a very high radial resolution of 400 zones/$R_j$, which appears to be especially relevant for hot jets. Finally, one should keep in mind that the simulations performed in the aforementioned papers only covered about one hundred time units, well inside the linear regime of the corresponding models for small perturbations.
The high accuracy of our simulations in describing the early stages of evolution of the KH instability (as derived from the agreement between the computed and expected linear growth rates and the consistency of the saturation times) establishes a solid basis to study the fully nonlinear regime, to be done in the second paper of this series. In this paper (Paper II) we will show that the similarities found in the evolution of all the models accross the linear and saturation phases is lost and very different nonlinear evolution is found depending on the initial jet parameters.
This work was supported in part by the Spanish Dirección General de Enseñanza Superior under grant AYA-2001-3490-C02 and by the Polish Committee for Scientific Research (KBN) under grant PB 404/P03/2001/20. M.H. acknowledges financial support from the visitor program of the Universidad de Valencia. M.P. has benefited from a predoctoral fellowship of the Universidad de Valencia ([*V Segles*]{} program). The authors want also to acknowledge financial support from the Spanish-French [*Picasso*]{} program and the Polish-French [*Joumelage*]{} program in the period 1997-98 when the research on relativistic jet stability was initiated.
Influence of the numerical resolution in the description of the linear regime of KH modes of relativistic flows
===============================================================================================================
Growth of the instability depends critically on the numerical viscosity of the algorithm. Hence our first aim was to look for suitable numerical resolutions by comparing numerical and analytical results for the linear regime. We performed a number of simulations based on model A05 changing longitudinal and radial resolution, and also the exponent $m$ for the shear layer steepness (see Eqn. \[shearlayer1\] and \[shearlayer2\]). Figure \[fig:linear2\] shows results for several of those simulations.
A shear layer was included in order to avoid non-steadiness in initial conditions given by discontinuous separation between the beam and the external medium. Therefore, if for some transversal resolution we introduced a shear layer which was too thin, we found a similar non-steady behavior. However, in order to reproduce the linear regime, we needed a steep enough shear layer, due to the fact that theory was developed for a discontinuous separation between both media. We can see in figure \[fig:linear2\] that $m=10$ models are considerably damped with respect to the theoretical growth, independent of the transversal resolution. Hence resolution perpendicular to the flow appeared to be essential, requiring very high resolutions (400 zones/$R_j$) and thin shear layers ($m=40$) with 40 to 45 zones, i.e., an equilibrium between steepness and number of cells. Lower transversal resolutions and/or thicker shear layers led to non-satisfactory results, with a slow or damped growth. A very low longitudinal resolution results in damping of growth rate, too, as can be seen from comparison of $m=20$ models, but $r_z=16$ seems to give a reasonable growth rate compared to theory. This (small) resolution of 16 zones/$R_j$ along the jet was taken as a compromise between accuracy and computational efficiency.
Aloy, M.A., Ibáñez, J.M$^{\underline{\rm a}}$, Martí, J.M$^{\underline{\rm a}}$, and Müller, E. 1999, Astrophys. J. Suppl. Ser., 122, 151
Appl, S. 1996, A&A, 314, 995
Appl, S., Camenzind, M. 1992, A&A, 256, 354
Blandford, R.D, Pringle, J.E. 1976, MNRAS, 176, 443
Birkinshaw, M. 1984, MNRAS, 208, 887
Birkinshaw, M. 1991, MNRAS, 252, 505
Bodo, G., Massaglia, S., Ferrari, A., and Trussoni, E. 1994, A&A, 283, 655
Bodo, G., Massaglia, S., Rossi, P., Rosner, R., Malagoli, A., Ferrari, A. 1995, A&A, 303, 281
Bodo, G., Rossi, P., Massaglia, S., Ferrari, A., Malagoli, A., Rosner, R. 1998, A&A, 333, 1117
Choudhury, S.R, Lovelace, R.V.E. 1984, ApJ, 283, 331
Donat, R., Font, J.A., Ibáñez, J.M$^{\underline{\rm a}}$, Marquina, A. 1998, J. Comput. Phys., 146, 58
Donat, R., and Marquina, A. 1996, J. Comput. Phys., 125, 42
Fanaroff, B.L., Riley, J.M. 1974, MNRAS, 167, 31
Ferrari, A., Massaglia, S., Trussoni, E. 1981, MNRAS, 198, 1065
Ferrari, A., Trussoni, E., and Zaninetti, L. 1978, A&A, 64, 43
Ferrari, A., Trussoni, E., Zaninetti, L. 1981 MNRAS, 196, 1051
Font, J.A., Ibáñez, J.M$^{\underline{\rm a}}$, Marquina, A., and Martí, J.M$^{\underline{\rm a}}$ 1994, A&A, 282, 304
Gill, A.E. 1965, Phys. Fluids, 8, 1428
Gomez, J.L., Martí, J.M$^{\underline{\rm a}}$, Marscher, A.P., Ibáñez, J.M$^{\underline{\rm a}}$, Marcaide, J.M. 1995, ApJ, 449, L19
Gómez, J.L., Martí, J.M$^{\underline{\rm a}}$, Marscher, A.P., Ibáñez, J.M$^{\underline{\rm a}}$, and Alberdi, A. 1997, ApJ, 482, 33
Hanasz, M., 1995, PhD Thesis, Nicholas Copernicus University, Toru[ń]{}.
Hanasz, M., Sol, H. 1996, A&A, 315, 355
Hanasz, M., Sol, H. 1998, A&A, 339, 629
Hanasz, M., 1997, in [*Relativistic jets in AGNs*]{}, Ostrowski, M., Sikora, M., Madejski, G., Begelman, M., eds, Krak[ó]{}w, p. 85 (astro-ph 9711275)
Hardee, P.E. 1979, ApJ, 234, 47
Hardee, P.E. 1984, ApJ, 287, 523
Hardee, P.E. 1986, ApJ, 303, 111
Hardee, P.E. 1987, ApJ, 318, 78
Hardee, P.E. 2000, ApJ, 533, 176
Hardee, P.E. Norman, M.L. 1988, ApJ, 334, 70
Hardee, P.E., Rosen, A., Hughes, P.A., Duncan, G.C. 1998, ApJ, 500, 559
Hardee, P.E., Hughes, P.A., Rosen, A., Gomez, E.A. 2001, ApJ, 555, 744
Lobanov, A.P., Zensus, J.A. 2001, Science, 294, 128.
Martí, J.M$^{\underline{\rm a}}$, Müller, E., Font, J.A., Ibáñez, J.M$^{\underline{\rm a}}$, and Marquina, A. 1997, ApJ, 479, 151
Mirabel, I.F., Rodriguez, L.F. 1999, Ann. Rev. A&A, 37, 409
Norman, M.L., Hardee, P.E. 1988, 334, 80
Payne, D.G, Cohn, H. 1985, ApJ, 191, 655
Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T., 1992, [*Numerical recipes*]{}, Cambridge University Press.
Ray, T.P. 1982, MNRAS, 198, 617
Rosen, A., Hughes, P.A., Duncan, G.C., Hardee, P.E. 1999, ApJ, 516, 729
Roy Choudhury, S., and Lovelace, R.V.E. 1984, ApJ, 283, 331
Sari, R., Piran, T., Halpern, J.P. 1999, ApJ 519, L17
Shu, C.W., and Osher, S.J. 1988, J. Comput. Phys., 77, 439
Trussoni, E. Ferrari, A., Zaninetti, L. 1983, in [ *Astrophysical jets. Proceedings of the International Workshop, Turin, Italy*]{}, Dordrecht, Reidel, p. 281
Turland, B.D., and Scheuer, P.A.G. 1976, MNRAS, 176, 421
Urpin, V. 2002, A&A, 385, 14
Zensus, J.A. 1997, ARA&A, 35, 607
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This short review aims to clarify upon the origins of so–called Eddington–Barbier relationships, which relate the emergent specific intensity and the flux to the photospheric source function at specific optical depths. Here we discuss the assumptions behind the original derivation of Barbier (1943). We also point to the fact that Milne had already formulated these two relations in 1921.'
author:
- Frédéric Paletou
title: 'On Milne–Barbier–Unsöld relationships'
---
Introduction
============
The theory of radiation transfer is fundamental in astrophysics. Besides the *in situ* exploration of various bodies in the solar system, cosmic rays astrophysics and the recent spectacular advent of gravitational wave detections, remote sensing of radiation remains the primary means through which we advance our knowledge of celestial bodies.
Analytical solutions were successively derived, mostly during the first half of the XX*th* century, and beyond the advent of numerical computing, into the 1960’s. After pioneering contributions of Schuster (1905) and Schwarzschild (1906), important results were further established, in particular for the case of stellar photospheres in radiative equilibrium. They relate, for instance, to the temperature distribution in a so-called “gray atmosphere”, and to the associated limb–darkening law of radiation. One may also mention the $\sqrt{\varepsilon}$ surface value for simplified radiation transfer *out of* local thermodynamical equilibrium (see e.g., Hubeny 1987; Lambert et al. 2016).
Fundamental elements of radiative transfer can be found in the texts of Rutten (2003) and Hubeny & Mihalas (2014).
Eddington–Barbier relationships
===============================
The so-called Eddington–Barbier relationships constitute fundamental analytic results, systematically presented in most textbooks and lectures about radiative transfer in astrophysics. In most cases, they are introduced and derived assuming that the source function is just a *linear* function of the optical depth:
$$S(\tau)= a + b \tau \, ,$$
where $a$ and $b$ are arbitrary coefficients.
It is then easy to derive the *emergent* specific intensity, from a plane-parallel semi-infinite atmosphere, according to:
$$I(\mu)=
\int_{0}^{\infty} { S(\tau) e^{-(\tau / \mu)}
\left( d\tau/\mu \right) } \, .
\label{eq:emergent}$$
Here $\mu$ is the usual cosine of the angle of the ray to the vertical direction. Given equation (1), the specific intensity is merely:
$$I(\mu)= a + b \mu \, ,$$
that is, the source function at optical depth $\tau=\mu$ i.e.,
$$I(\mu)= S(\tau=\mu) \, .$$
In other words, this means also that, for a given line of sight the emergent intensity equals the source function at a depth found after crossing an optical depth unity *along the line of sight*.
A second relationship can also be derived for the *emergent flux* defined as:
$${\cal{F}} = 2 \pi
\int_{0}^{1} { I(\mu) \mu d\mu } \, .
\label{eq:flux}$$
The quantity ${\cal{F}}$ is relevant to spatially *unresolved* objects, like most stars (besides the Sun), and it is easy to show that the emergent flux is then characterized by the source function at optical depth $\tau=2/3$.
Many commonly-read textbooks such as Athay (1972), Mihalas (1978), the very popular $e-$book of Rutten (2003), and even the recent Hubeny & Mihalas (2014) omit however to cite any original publication establishing first these two classical relationships.
Original derivation
===================
The origin of the derivation of these relationships can be found in an article of French astronomer *Daniel Barbier* published in 1943, although he did not address explicitly the case of the specific intensity there. The original derivation of Barbier starts with the following Taylor series expansion for the source function:
$$S(\tau)=S(\tau_{*}) + (\tau-\tau_{*})S^{\prime} (\tau_{*})
+ {1 \over 2} (\tau-\tau_{*})^2 S^{\prime\prime} (\tau_{*})
\label{eq:taylor}$$
that we truncate here at 2nd order. In this expression $S^{\prime}$ and $S^{\prime\prime}$ are respectively the first and the second derivatives vs. $\tau$ of the source function. This expansion is introduced into Eq.(\[eq:emergent\]), and straightfoward integrations give the following expression for the emergent specific intensity:
$$\begin{aligned}
I(\mu) & = & S(\tau_{*}) + (\mu-\tau_{*})S^{\prime} (\tau_{*}) \nonumber \\
& + & (\mu^2 - \mu \tau_{*} + \frac{1}{2} \tau_{*}^2) S^{\prime\prime}
(\tau_{*}) \, .
\label{eq:Barbier} \end{aligned}$$
In his original article of 1943, D. Barbier does *not* give an expression for the emergent specific intensity, but does give the emergent “total flux”, where intensity is integrated over $\mu$ (equation 5). However, we can adopt his argument, and *choose $\tau_{*}$ which makes the term in $S^{\prime}$ vanish, and which minimizes that in $S^{\prime\prime}$*. It is therefore obvious that:
$$\tau_{*}=\mu \, .$$
In such a case, the emergent specific intensity is:
$$I(\mu)=S(\tau=\mu) +
\frac{1}{2} \mu^2 S^{\prime\prime} (\tau=\mu) \, ,
\label{eq:BarbierSolI}$$
which is indeed identical to $S(\tau=\mu)$, if one assumes that the source function is no more than *linear* in the optical depth, so that $S^{\prime\prime}(\tau)=0$.
Discussion
==========
Barbier (1943) cites Eddington quite precisely, pointing at *p. 330* of his famous textbook *The internal constitution of stars* (1926). In this chapter, *The Outside of a Star*, Eddington states that the *effective temperature* of the angle-dependent atmospheric radiation should be the temperature of the layer where $\tau \approx \mu$. This may have inspired Barbier to adopt a Taylor series expansion method.
Soon after Barbier’s contribution, Unsöld (1948; 1949, in English) makes explicit these classical relationships, *both* for the specific intensity and for the emergent flux. Unsöld (1955) gives direct credit to Barbier in his famous textbook, *Physik der Sternatmosphären*. However, he wrongly cites Barbier (1944), instead of Barbier (1943)! About Barbier’s method based on a Taylor series expansion of the source function $S$ around a certain optical depth $\tau_{*}$ which has to be determined, he writes, originally in German, that: *“a method of approximation, proposed by A.S. Eddington and better argumented by D. Barbier is still very useful and interesting”*. Unsöld coins it also the *“$x=\cos \vartheta$–Methode”* of Eddington and Barbier, where $x$ is used for optical thickness $\tau$, and $\mu=\cos \vartheta$.
Kourganoff’s (1952) textbook gives a proper citation and description of Barbier’s original contribution in his §18.2: *“After giving Barbier’s demonstration, which is known as the $\tau_{*}$-method”*, which is however immediately followed by: *“we shall explain why it seems to us to be unsatisfactory”...* Modern texts reflect Kourganoff’s clear statement that: *“Now Barbier’s demonstration (or a direct calculation) shows that all of the Eddington–Barbier relations are rigorously true if the source function is a linear function of $\tau$”.* And despite a critical discussion in the remaining of §18, Kourganoff concludes with: *“The Eddington–Barbier relations, apart from the applications which have already been made by Barbier himself and by Unsöld, are extremely useful when one wants to represent the connexion between the source function and certain observable quantities like $I(0,\mu)$ and ${\cal{F}}(0)$”.*
With time, citations to Barbier (1943) vanish, although they appear in textbooks such as those of Zirin (§6.10., 1988) and Castor (§5.7., 2004). This is, surprinsingly, *not* the case for the famous and comprehensive Mihalas (1978) textbook. However, in its §2-2 an exercise (2-5) is directly inspired by the method used by Barbier (1943). It may be unfortunate though, that the revised and expanded textbook by Hubeny & Mihalas (2014) does not elaborate on the original derivation of one of the most famous result of analytical radiation transfer.
A lost contribution of Milne?
=============================
During the course of our investigations on the original contributions leading to the so-called Eddington–Barbier relations, we also went back to an article of V.V. Ivanov (1991) in the proceedings of the Trieste conference *Stellar Atmospheres: beyond classical models*.
At the end of section *History of ART*, where ART stands for “Analytical Radiative Transfer”, Ivanov writes this somewhat intriguing statement: *“The standard Eddington approximate form of the temperature distribution in a grey atmosphere,*
$$T^4 = (3/4) T_{\rm eff}^4 (\tau + 2/3) \, ,
\label{eq:greyatmos}$$
*belongs not to Eddington, but to Milne. In 1917, he introduced the approximation known today as the Eddington approximation.”*
However, it seems that Milne did not publish any astrophysical result before 1921, according to Tayler (1996) for instance. This led us to look back with some care to the early contributions of Milne in the domain of radiative transfer and stellar atmospheres. And his first article of 1921, *Radiative equilibrium in the outer layers of a star: the Temperature Distribution and the Law of Darkening* contains in Eqs. (36) and (37) both Eddington–Barbier relations, for the specific intensity and for the flux, as shown in Fig. (1).
![In this excerpt of Milne’s (1921) article, $B(\tau)$ should also read as the source function and Eqs. (36) and (37) are just the Eddington–Barbier relationships respectively on the specific intensity and on the flux, published more than twenty years before Barbier (1943).[]{data-label="Fig1"}](Milne1921_EB.png){width="8.25"}
In this article, the derivation is formally distinct from the one used by Barbier (1943). Milne is inspired by the method already used by Schwarzschild and Jeans. He uses, in particular, simplifications for lower and upper boundaries incident radiation which are assumed to be distinct but independent of direction. Then he gives an *integral form* of the “temperature distribution” for a semi–infinite atmosphere, with no incoming radiation on its surface. First, he shows that a linear source function is a possible solution at *large* optical depths, and even at this stage the first part of his Eq.(24) already contains the Eddington–Barbier relationship for the specific intensity, when (inappropriately though) setting... $\tau=0$. He goes however further and implements a method of successive approximations, which leads to a second approximation for the source function of the form:
$$\begin{aligned}
S(\tau) & = & a+2b\tau + \frac{1}{2}e^{-\tau}(b-a-b\tau) \nonumber \\
& + & \frac{1}{2}(a\tau+b\tau^2) \int_{\tau}^{\infty} {{{e^{-y}} \over {y} } dy}
\, .
\label{eq:Milne} \end{aligned}$$
This expression has indeed the same form as the first approximation, linear in $\tau$, for large optical depths but now, the first exponential integral $E_1(\tau)$ appears, as in the Schwarzschild–Milne equations (see e.g., Hubeny & Mihalas 2014). It is also more accurate in the sense that it exhibits the often forgotten *singularity* in $dS/d\tau$ at $\tau=0$ (see e.g., Chevallier et al. 2003 and references therein).
Both the so-called Eddington–Barbier relationships appear once Milne, *“for completeness”*, goes back to his first approximation, when the source function is just linear in the optical depth. At this stage, Milne does *not* comment any further on his Eqs. (36) and (37). This may also explain why these intermediate results of Milne were neither used, nor cited further. Perhaps they may also have been eclipsed by other aspects of Milne’s important contributions in the early 1920’s?
For instance, following the remark of Ivanov (1991), in his first 1921 article Milne establishes also a limb–darkening law such that:
$$\frac{I(0,\mu)}{I(0,1)} = \frac{3}{5} (\mu + 2/3) \, ,
\label{eq:limdrk}$$
improving on Schwarzschild and Jeans. It is also fully consistent with the assumption known as the “Eddington approximation”, which leads to the (constant) $2/3$ appearing in Eq.(\[eq:greyatmos\]).
Conclusion
==========
One may argue that Unsöld (1948; 1949, in English) makes the classical relations explicit, for the first time, *both* for the specific intensity and for the emergent flux, following Barbier (1943). However they were *both* expressed in Milne (1921), but cited neither by Barbier, nor by Unsöld (or anyone else, to the best of our knowledge) more than twenty years after.
After our investigations, we would therefore propose to the astrophysical community to shift, at last, from the “Eddington–Barbier” usual designation to the *fairer* “Milne–Barbier–Unsöld” relationships.
Finally, we also report that despite its legacy, Barbier (1943) has been cited only four times so far, according to the ADS service! This is also the case for Unsöld (1948; and it is even worse for his 1949 article, although written in English)...
This review was initiated after a question of one of *Université de Toulouse* astrophysics master student, Mrs Alexandra Le Reste, during a lecture of Sept. 2017. We also wishes to thank Prof. R.J. “Rob” Rutten for fruitful and enjoyable discussions, as well as Drs Céline Reylé (*Observatoire de Besançon*, France) and Torsten Böhm (CNRS, IRAP, Toulouse, France) for their kind assistance, and finally Dr. Ivan Hubeny, for his encouragement to make available this note in English. This research has made use of NASA’s Astrophysics Data System, a fantastic tool to search and read again the historical astrophysical literature.
[99]{} Athay, R.G. 1972, Radiation Transport in Spectral Lines, D. Reidel, Dordrecht
Barbier, D. 1943, *Annales d’Astrophysique*, 6, 113 (text in French)
Barbier, D. 1944, *Annales d’Astrophysique*, 7, 115 (text in French)
Castor, J.I. 2004, Radiation Hydrodynamics, Cambridge University Press
Chevallier, L., Paletou, F., Rutily, B. 2003, A&A, 411, 221
Eddington, A.S. 1926, The Internal Constitution of Stars, Cambridge University Press
Hubeny, I. 1987, A&A 185, 332
Hubeny, I & Mihalas, D. 2014, Theory of Stellar Atmospheres, Princeton University Press
Ivanov, V.V. 1991, In: Stellar Atmosphere: Beyond Classical Models, Dordrecht, D. Reidel Publishing Co.
Kourganoff, V. 1952, Basic Methods in Transfer Problems, Clarendon Press, Oxford
Lambert, J., Paletou, F., Josselin, E., Glorian, J.-M., 2016, Eur. J. Phys, 37, 015603
Mihalas, D. 1978, Stellar Atmospheres (2nd edition), San Franciso, Freeman
Milne, E.A. 1921, Monthly Notices of the Royal Astronomical Society, 81, 361
Rutten, R.J. 2003, Radiative Transfer in Stellar Atmospheres, Lecture Notes, Utrecht University (see also: http://www.staff.science.uu.nl/$\sim$rutte101/\
Radiative\_Transfer.html)
Schuster, A. 1905, Astrophys. J., 21, 1
Schwarzschild, K. 1906, *Nachrichten von der Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse*, 41 (text in German; see also Menzel D.H., Milne E.A. 1966, Selected papers on the transfer of radiation, New-York, Dover)
Tayler, R.J. 1996, Quarterly Journal of the Royal Astronomical Society, 37, 355
Unsöld, A. 1948, *Annalen der Physik*, 438, 124
Unsöld, A. 1949, *Memorie della Società Astronomica Italiana*, 20, 49 (in English)
Unsöld, A. 1955, *Physik der Sternatmosphären*, Springer (in German)
Zirin, H. 1988, Astrophysics of the Sun, Cambridge University Press
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a random walk on one-dimensional inhomogeneous graphs built from Cantor fractals. Our study is motivated by recent experiments that demonstrated superdiffusion of light in complex disordered materials, thereby termed Lévy glasses. We introduce a geometric parameter $\alpha$ which plays a role analogous to the exponent characterizing the step length distribution in random systems. We study the large-time behavior of both local and average observables; for the latter case, we distinguish two different types of averages, respectively over the set of all initial sites and over the scattering sites only. The “single long jump approximation" is applied to analytically determine the different asymptotic behaviours as a function of $\alpha$ and to understand their origin. We also discuss the possibility that the root of the mean square displacement and the characteristic length of the walker distribution may grow according to different power laws; this anomalous behaviour is typical of processes characterized by Lévy statistics and here, in particular, it is shown to influence average quantities.'
author:
- |
Alessandro Vezzani$^{\rm a,b}$$^{\ast}$, Raffaella Burioni$^{\rm b,c}$, Luca Caniparoli$^{\rm d}$ and Stefano Lepri$^{\rm e}$ [^1]\
$^{\rm a}$[[Centro S3, CNR-Istituto di Nanoscienze, Via Campi 213A, 41125 Modena Italy]{} $^{\rm b}$[[Dipartimento di Fisica, Università degli Studi di Parma, viale G.P.Usberti 7/A, 43100 Parma, Italy]{}]{}; $^{\rm c}$[[INFN, Gruppo Collegato di Parma, viale G.P. Usberti 7/A, 43100 Parma, Italy]{}]{}; $^{\rm d}$[[International School for Advanced Studies SISSA, via Beirut 2/4, 34151, Trieste, Italy]{}]{}; $^{\rm e}$[[Istituto dei Sistemi Complessi, Consiglio Nazionale delle Ricerche, via Madonna del Piano 10, I-50019 Sesto Fiorentino, Italy]{}]{}]{}\
title: Local and average behavior in inhomogeneous superdiffusive media
---
Introduction
============
The laws of Brownian motion crucially relies on the hypothesis that the steps for the diffusing particle are small (with finite variance) and uncorrelated. Whenever these assumptions are violated, the standard diffusion picture breaks down and anomalous phenomena emerge [@bouchaud; @klages].
In particular, transport processes where the step length distribution has a diverging variance have been theoretically studied in detail. Among those, one of the most interesting is the so-called Lévy walks [@Blumen1989; @Klafter1990], in which particles perform independent steps $l$ at constant velocity, with a distribution $\lambda(l)$ following an algebraic tail $\sim l^{-(1+\alpha)}$. Such a distribution is said to be heavy-tailed and it is known to have a diverging variance for $\alpha< 2$. Since transport is thereby dominated by very long steps, the mean square displacement increases faster than linearly with time, hence the name superdiffusion.
Among the many possible experimental applications, our work is motivated by the recent realization of materials termed *Lévy glasses*, where light rays propagate through an assembly of transparent spheres embedded in a scattering medium [@Barthelemy08; @Bertolotti10]. If the diameter of the spheres is power-law distributed, light can indeed perform anomalous diffusion. Owing to their ease of fabrication and tunability, such a novel material offers an unprecedented opportunity to study anomalous transport processes in a systematic and controllable way.
An important feature of the experimental samples is that the walk is correlated: light that has just crossed a large glass microsphere has a higher probability of being backscattered at the following step and thus to perform a jump of roughly the same length. While the case of uncorrelated jumps is well understood [@ann], the correlation effects, that are expected to deeply influence the diffusion properties [@Fogedby94], are still to be characterized. To this aim, quenched Lévy processes have been studied on one dimensional systems [@klafter; @beenakker]. More recently, different aspects regarding the scaling properties of random-walk distributions, the relations between the dynamical exponents and the different average procedures have been discussed in a common framework [@noi2].
In order to get a deeper insight on the effect of step-length correlations, a class of deterministic, one-dimensional models called *Cantor graphs* has been introduced [@noi]. Random walks on these structures perform correlated long jumps induced by the underlying fractal topology. As the latter is generated by deterministic rules, diffusion properties can be studied in a simpler way than in the random case. Here we extended to this deterministic topologies some of the results proved in [@noi2] for random structures. In particular, we introduce a geometric parameter $\alpha$ which plays the same role as the exponent characterizing the step length distribution $\lambda(l)$ in random systems. Three kinds of statistical averaging are introduced: (i) a local one, namely the average of all trajectories starting from a given initial site (whose asymptotic behaviour being expected to be independent of the site choice); (ii) an averaging over *all* possible initial sites of the graph and (iii) averages where *only* scattering sites are considered as initial conditions. In the random case, the differences between such averaging procedures have been evidenced in [@klafter; @noi2]. On determinisitc structures, average procedures (i) and (ii) have been discussed in [@noi]. Here we complete the picture studying the effect of averaging over scattering sites. We evidence that the behavior of the mean square displacement is similar to random case; while the probability density displays a more complex structure being given by a non-trivial time-dependent superposition of step functions. We remark that, in experiments [@Barthelemy08; @Bertolotti10], light enters the sample with a scattering event and averages of type (iii) are the most physically sensible quantities to compare with. Interestingly, in Lévy processes the root of the mean square displacement and the characteristic length of the distributions may grow according to different power laws; here, in particular, this strongly anomalous behavior characterizes average quantities.
The paper is organized as follows. In the next section we introduce the directed Cantor graphs and we define a simple random walk on these structures. We then discuss the relevant physical quantites and the average procedures. Section 4 is devoted to discuss the scaling hypothesis and the single long jump approximation, which allows to evaluate the tails of the density distribution. Finally, in Section 5 we discuss our results evidencing differences among averaging procedures. In particular for the case (i) and (ii) we review the results presented in [@noi] presenting new numerical data, while for the case of average over scattering sites, which has not been discussed so far for the deterministic structures, we provide a sketch of the derivation of the dynamical exponents within the single long jump approximation and we compare the result with numerical simulations. In general the new simulations evidence that the probability density presents different asymptotic behaviors depending on the average procedures, clearly supporting both the scaling hypothesis and the single long jump approach. Moreover the asymptotic behavior of the mean square displacement have been tested for a wider range of $\alpha$’s.
Random walks on Cantor graphs
=============================
In paper [@noi] we have introduced a class of graphs, denoted as directed Cantor graphs, as a one-dimensional, deterministic counterpart of the geometric structure of the Lévy glass materials mentioned above [@Barthelemy08]. Indeed, random walks on the Cantor graphs display a superdiffusive Lévy like motion analogous to the the one of light in such inhomogeneous glassy material.
The class of graphs we will consider is defined by two parameters, denoted as $n_r$ and $n_u$, and describing the growth of the fractal from generation $G-1$ to generation $G$. In particular the fractal $\cal G$ of generation $G$ is built connecting $n_r$ fractals of generation $G-1$ by $n_r-1$ unidirectional bubbles of length $L_G=n_u^{G-1}$, as shown in Figures \[cantor\_walk\_var\] and \[cantor\_walk\_nat\].
A simple random walk [@rw] is naturally defined on these structures: undirected links connect sites in both directions while directed links have to be crossed only in the prescribed way. A site is called *bidirectional* if the walker placed on that site can move in both directions (and the two possible moves are performed with probability $1/2$). A site is instead called *unidirectional* if the walker is allowed to move only in one direction (in this case the move is performed with probability one, i.e. ballistic motion). All the links have unitary length and are crossed at constant velocity $v$. The number of bidirectional sites $N_b$ and unidirectional sites $N_u$ present at generation $G$ is given by: $$N_{b}=2n_r^G,\qquad
N_{u} =
\frac{(n_u-1)n_r^G-\ (n_r-1)n_u^G}{n_r-n_u}+1
\label{nu}$$ where we use the convention of counting only once the couple of sites of the bubbles at the same distance from the origin. The total number of sites at generation $G$ is hence $N_G=N_b+N_u$. In the following, an important parameter for the description of the structure will be the ratio $\alpha=\log(n_r)/\log(n_u)$. For $\alpha<1$, the graphs are called [*slim*]{}, as the fraction of bidirectional sites vanishes in the thermodinamic limit, i.e. $\lim_{G \to \infty}N_b/N_G=0$. On the contrary, for $\alpha>1$ graphs are called [*fat*]{}, since a finite fraction of bidirectional sites is present, and $\lim_{G \to \infty}N_b/N_G>0$.
![The generation $G=3$ of the graph with $n_u=3,4,5$, in the $n_r=2$ case.[]{data-label="cantor_walk_var"}](plot_graph_n_u345_n_s3.eps){width="50.00000%"}
![Three examples of Cantor graphs of generation $G=2$ with $n_r=2,3,4$ and $n_u=3,4,5$.[]{data-label="cantor_walk_nat"}](plot_graph_n_s357.eps){width="50.00000%"}
Physical quantities and averages
================================
Let $P_i(r,t)$ be the probability of arriving at distance $r$ starting from $i$ in $t$ steps. In general, $P_i(r,t)$ depends on the starting site $i$. However for large enough space scales (i.e. $r$ much larger than the distance between $i$ and an origin $i=1$) the asymptotic properties of $P_i(r,t)$ are expected to be site independent and hence to describe a property of the whole graph, i.e. $P_i(r,t)\sim P_1(r,t)$. In particular, the asymptotic behaviour of the mean square displacement $$\langle r_i^2(t)\rangle = \int_{0}^{vt} P_i(r,t) r^2 dr
\label{r2loc}$$ should be independent of $i$. The integration cutoff in (\[r2loc\]) is provided by the fact that the walker covers at most a distance $vt$ in a time $t$ ($v=1$ in the following). Another important quantity whose asymptotic behaviour depends only on the large scale topology of the structure is the resistivity $R_i(r)$ i.e the number of bidirectional sites whose distance from $i$ is smaller than $r$.
On inhomogeneous structures, average and local properties are in general different [@burioni], and in structures characterized by long tails, different averaging procedures can yield different results as well [@klafter; @noi2]. In particular, on Cantor graphs one can distinguish between the average over all starting sites, i.e.: $$\bar P(r,t)= \lim_{G \to \infty}\frac{ \sum_{i\in{\cal G}} P_i(r,t)}{N_G}
\label{avgtot}$$ and the averages on processes beginning with a scattering event: $$\tilde P(r,t)= \lim_{G \to \infty}\frac{ \sum_{i\in{\cal G}_b} P_i(r,t)}{N_b}
\label{avgsct}$$ where ${\cal G}_b$ is the set of the bidirectional sites belonging to graphs of generation $G$. The same averaging procedures can be introduced also for different quantities. One can consider the average resistivity $\bar{R}(r)$ and $\tilde{R}(r)$ and, as in (\[r2loc\]), one can also define the average mean square distances $\langle \bar{r}^2(t)\rangle$ and $\langle
\tilde{r}^2(t)\rangle$.
Since the resistivity can be evaluated by simply counting the number of sites in a given generation, one obtain the following asymptotic behaviours: $$R_1(r)\sim \tilde{R}(r)
\sim\left\{\begin{array}{lc} r^\alpha, & \mathrm{if}\ \alpha<1 \\
\\
r & \mathrm{if}\ \alpha \geq 1
\end{array}\right.
\label{Rloc}$$
$$\bar{R}(r)\sim
\left\{\begin{array}{lc} 0 & \mathrm{if}\ \alpha<1 \\
\\
r & \mathrm{if}\ \alpha \geq 1.
\end{array}\right.
\label{Rav}$$
These results represent the deterministic analog of the expressions for the resistivity obtained in a random sample in [@beenakker].
The scaling hypothesis and the single long jump approximation
=============================================================
The most general scaling hypothesis for the probabilities $P_1(r,t)$ is: $$P_1(r,t)=\ell_1^{-1}(t)f_1(r/\ell_1(t))+g_1(r,t)
\label{sal}$$ with a convergence in probability $$\lim_{t\to\infty}\int_0^{t} |P_1(r,t)-\ell^{-1}_1(t)f_1(r/\ell_1(t))|dr=0
\label{sal2}$$ The leading contribution to $P_1(r,t)$ is hence $\ell_1^{-1}(t)f_1(r/\ell_1(t))$ which is significantly different from zero only for $r \lesssim \ell_1(t)$. The subleading term $g_1(r,t)$, with $\lim_{t \to \infty}\int |g_1(r,t)| dr=0$ describes the behavior at larger distances, i.e. $\ell_1(t)\ll r < t$. Notice that, if $g_1(r,t)$ does not vanish rapidly enough, it can nevertheless provide important contributions to $\langle r_1^2(t)\rangle$. The same scaling ansatz should be valid also for the average probabilities $\bar{P}(r,t)$ and $\tilde{P}(r,t)$ by introducing suitable averaged scaling length and scaling functions $\bar{f}(r/\bar{\ell})$, $\tilde{f}(r/\tilde{\ell})$, and suitable averaged long distance corrections $\bar{g}(r,t)$, $\tilde{g}(r,t)$.
In [@noi; @cates] it has been proved that the growth of the characteristic length can be directly related to the growth of the resistance. We define the exponent describing the growth of the correlation length as: $$\ell_1(t)\sim t^{d_s/2}.
\label{sal3}$$ so that in analogy with standard definition of random walks [@Orbach], we get $P_1(0,t)\sim t^{-d_s/2}$. Then using the scaling relations proved in [@noi; @cates] we obtain $$R_1(r)\sim r^{2/d_s-1}.
\label{scalR}$$ Analogous relations hold for the average quantities, in terms of the average scaling lengths. Introducing the known results for the resistivity (\[Rloc\],\[Rav\]), one obtains the following behaviors for the scaling lenghts: $$\ell_1(t)\sim \tilde{\ell}(t)
\sim \left\{\begin{array}{lc} t^{\frac 1{1+\alpha}}, & \mathrm{if}\ \alpha<1 \\
\\
t^{\frac 1 2} & \mathrm{if}\ \alpha \geq 1
\end{array}\right.
\label{elloc}$$
$$\bar{\ell}(t)\sim
\left\{\begin{array}{lc} t & \mathrm{if}\ \alpha<1 \\
\\
t^{\frac 1 2} & \mathrm{if}\ \alpha \geq 1
\end{array}\right.
\label{elav}$$
Let us now discuss the behaviour of the mean square displacements. When only lengths of order $ r \lesssim \ell(t)$ provide significant contributions to the integral (\[r2loc\]), the standard relation $\langle r^2 (t) \rangle \sim
\ell^2(t)$ holds and the asymptotic behaviour coincides with those given by (\[elloc\],\[elav\]). However it is known that, in presence of long tailed distributions, anomalies with respect to this behaviour can be present. In particular it has been evidenced in [@klafter; @noi2] that a key role is played by long jumps, leading the walker to a distance $r\gg
\ell(t)$. In the random case, consideration of a single long jump actually accounts for the asymptotic behaviour [@noi2]. Indeed, these processes can give rise to two different types of corrections to $P(r,t)$. First, they can produce a zero-measure function $g(r,t)$, providing a significant contribution to $\langle r^2 (t) \rangle$; second, the scaling function $f(x)$ can feature a long tail, breaking the proportionality between $\langle r^2 (t) \rangle$ and $ \ell^2(t)$. Here we will evidence that these anomalies, originating from the single long jump, are also present in the deterministic graphs and they are deeply influenced by the averaging procedures.
Results
=======
Local behaviour
---------------
![Dynamical scaling of the probability for initial site $i=1$ on the graph obtained with $n_r=2$ and $n_u=3$. Note the fast decay of the scaling function. The stepwise structure of the scaling function is due to the fractality of the graph.[]{data-label="dynscal"}](scalP23local.eps){width="50.00000%"}
Let us first consider the local properties. In this situation we can focus on processes starting from the origin $i=1$ of the graph. Indeed, if $r$ is much larger than the distance between $i$ and $1$, we expect that $P_i(r,t)$ should behave as $P_1(r,t)$ and hence the asymptotic behaviour should be the same for any starting point. When starting from the origin, in a time $t$ the walker typically covers a distance $\ell_1(t)$ and, in the deterministic graphs, within such distance there are no bubbles of length larger than $\ell_1(t)$. Therefore, long jumps do not occur and $\langle r^2_1(t)\rangle\sim \ell_1^2(t)$. Therefore [@noi] $$\langle r^2_1(t)\rangle
\sim \left\{\begin{array}{lc} t^{\frac 2{1+\alpha}}, & \mathrm{if}\ \alpha<1 \\
\\
t & \mathrm{if}\ \alpha \geq 1
\end{array}\right.
\label{msqloc}$$ Figure \[dynscal\] reports, an example the probability density $P_1(r,t)$ obtained by a Montecarlo simulation. The data provide a clear evidence that the scaling function presents a fast decay confirming our hypothesis of no long jump. As explained in [@noi] the fractal structure give rise to log-periodic oscillations which can be discarded considering peculiar sequence of times (in this case $t=6^k$) , However, such oscillations do not change the general framework of the scaling hypothesis. The growth of $\langle r_1^2(t) \rangle$ is plotted in Figure \[x2\], the continuous lines represent the expected behaviors (\[msqloc\]).
![Growth of the mean square displacements for initial site $i=1$. Power-law fits give exponents in very good agreement with the theoretical values, in equation (\[msqloc\]). []{data-label="x2"}](r2local.eps){width="50.00000%"}
Averages over all sites
-----------------------
When averaging over all the sites, it has been evidenced [@noi] that for $\alpha<1$ the motion is always ballistic, while for $\alpha>1$ the situation is much more complex, since the walker can perfom a single jump much larger than $\bar{\ell}(t)$. Tipically, such a long jump occurs at the first step because, with a random choice of the starting point, the probability of belonging to a large bubble is much larger at $t=0$ than during the rest of the evolution. In this situation, one can estimate the zero measure correction to $\bar{P}(r,t)$ obtaining $\bar{g}(r,t)\sim t^{-\alpha+1} \delta(r-t)$, i.e. a peak associated with the ballistic motion of the particle, weighted by a factor $ t^{-\alpha+1} $ representing the probability of belonging to a bubble larger than $t$ at the initial time. The behavior of $P(r,t)$ is illustrated in the simulations of figure \[pscaling\_73\_medio\] evidencing the presence of a scaling regime for $r \lesssim \bar{\ell}(t)$ and of balistic peaks at large $r$ whose height evolves as $t^{-\alpha+1}$ (dased line). Even if $\bar{g}(r,t)$ provides a subleading contribution to $\bar{P}(r,t)$, the integral in (\[r2loc\]) is dominated by $\bar{g}(r,t)$ for $1<\alpha<2$ and by $\bar{f}(r/\bar{\ell}(t))$ only when $\alpha>2$. Therefore the behaviour of the mean square displacement is $$\langle \bar{r}^2(t)\rangle
\sim \left\{\begin{array}{lc} t^{3-\alpha}, & \mathrm{if}\ 1<\alpha<2 \\
\\
t & \mathrm{if}\ \alpha \geq 2
\end{array}\right.
\label{r2avg}$$ while the motion is purely ballistic for $\alpha<1$. Figure \[x2\_73\_92\] shows that the predicted exponents (\[r2avg\]) are well verified.
![Dynamical scaling of $\bar{P}(r,t)$ in the average case. The data refer to the case $n_r=7$ $n_u=4$. Ballistic peaks at $r= t$ scale as $t^{1-\alpha}$ (dashed line), while for $r \lesssim \bar{\ell}(t)$ the scaling hypothesis is well satisfied.[]{data-label="pscaling_73_medio"}](scalP74all.eps){width="50.00000%"}
![Growth of the mean square displacements when average over scattering over the whole sample is considered. We compare the numerical results with the theoretical preditions (\[r2avg\]) in both the regimes $1<\alpha<2$ and $\alpha>2$.[]{data-label="x2_73_92"}](r2meanall.eps){width="50.00000%"}
Averages over bidirectional sites
---------------------------------
In [@noi2] it has been shown that, for random systems, averages over bidirectional scattering sites can provide different results with respect to averages over all starting points and a scaling approach has been discussed, based on the single long jump approximation. Here we introduce an analogous argument for the deterministic graph.
![The dynamical scaling of $\tilde{P}(r,t)$ in the case of averages over bidirectional sites. Here we plot the case $n_r=4$ and $n_u=3$. For $r \lesssim \tilde{\ell}(t)$ the scaling hypothesis (\[sal\]) is very well verified with $\tilde{\ell}(t)$ growing as indicate in (\[elloc\]). For larger $r$ the behaviour is much more complicated since the tails are composed by a superpositions of step functions.[]{data-label="scalPmeansc"}](scalP43sc.eps){width="50.00000%"}
![Growth of the mean square displacement when averaged over bidirectional sites is consider. The results are compared with the theoretical predictions (\[msqav\]) given by the continuous lines. Different values of $n_r$ and $n_u$ are considered, showing that the predictions (\[msqav\]) are verified in the different regimes of $\alpha$’s.[]{data-label="r2meansc"}](r2meansc.eps){width="50.00000%"}
In the case of average over bidirectional sites, the single long jump does not occur necessarily at the first step. In particular for the deterministic graph the probability of performing a jump of length $L_k=n_u^k\gg \tilde{\ell}(t)$ in a time $t$ is $N(t) n_r^{-k}$ where $N(t)$ is the number of bidirectional sites visited by the walker in a time $t$ and $n_r^{-k}$ is the probability that a bidirectional site belongs to a bubble of length $L_k$. Discarding this long jump, the distance crossed by the walker in a time $t$ is of order $\tilde{\ell}(t)$, and therefore, according to the behaviour of the resistivity described by equations \[Rloc\], $N(t)\sim\tilde{\ell}(t)^\alpha$ for $\alpha<1$ and $N(t)\sim\tilde{\ell}(t)$ for $\alpha \geq 1$. The main difference with respect to the random case is that now the only possible lengths of the long jumps are given by the sizes $L_k$ of the bubbles in the fractal. Hence, for $\tilde{\ell}(t)<r<t$, $\tilde{P}(r,t)$ is a complex step function where both integers $n_u$ and $n_r$ plays a non trivial role. However, the contribution to the mean square displacement can be evaluated as follows: $$N(t)\left(\sum_{\tilde{\ell}(t)<L_k<t} \frac{n_u^{2k}}{n_r^{k}}+t^2 \sum_{L_k>t} \frac{1}{n_r^{k}}\right)
\label{msqav1}$$ where the first sum is related to the bubbles of length $L_k$ ($\tilde{\ell}(t)<L_k<t$) providing a contribution to the mean square displacements of order $L_k^2=n_u^{2k}$, while the second sum comes from the bubbles of length larger than $t$, providing a contribution $t^2$. Expanding equation \[msqav1\] for large times one obtains the asymptotic behaviours $$N(t)\left(\sum_{\tilde{\ell}(t)<L_k<t} \frac{n_u^{2k}}{n_r^k}+ t^2 \sum_{L_k>t} \frac{1}{n_r^k} \right)\sim t^{\frac{2+2 \alpha+\alpha^2}{1+\alpha}}
\label{msqav2}$$ for $\alpha<1$ and $$N(t)\left(\sum_{\tilde{\ell}(t)<L_k<t} \frac{n_u^{2k}}{n_r^k}+t^2 \sum_{L_k>t} \frac{1}{n_r^k} \right)\sim t^{5/2-\alpha}
\label{msqav3}$$ for $\alpha>1$. The first expression is always dominant with respect $\bar{\ell}(t)$, while the second expression becomes subleading for $\alpha> 3/2$. The overall behaviour of the mean square displacement is summarized as follows: $$\langle \tilde{r}^2 (t) \rangle \sim
\begin{cases}
t^{\frac{2+2 \alpha-\alpha^2}{1+\alpha}} & \mathrm{if}\ 0<\alpha<1 \\
t^{\frac{5}{2}- \alpha} & \mathrm{if}\ 1 \leq \alpha \leq 3/2 \\
t & \mathrm{if}\ 3/2 < \alpha
\label{msqav}
\end{cases}$$
Equations (\[msqav\]) extend the results of the random case to the deterministic topology described by the directed Cantor graphs. Clearly the complex shape of the scaling function (\[msqav1\]) determines the presence of logperiodic oscillations superimposed to (\[msqav\]), which is a a typical behaviour of fractal structures [@periodic]. Figure (\[scalPmeansc\]) evidences that the dymanical scaling is well verified for $r \lesssim \tilde{\ell}(t)$, while at larger distances $\tilde{P}(r,t)$ is characterized by a superposition of slowly decaying step functions as predicted by (\[msqav1\]). Figure (\[r2meansc\]) evidences by means of Montecarlo simulations that Equations (\[msqav\]) are well verified for large times in the whole range of $\alpha$’s.
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge useful discussion with P. Barthelemy, J. Bertolotti, R. Livi, D.S. Wiersma, K. Vynck. This work is partially supported by the MIUR project PRIN 2008 *Non linearity and disorder in classical and quantum processes*.
[10]{}
J.P. Bouchaud and A. Georges, Phys. Rep. [195]{} (1990) p. 127; D. ben-Avraham and S. Havlin, *Diffusion and Reactions in Fractals and Disordered Systems*, Cambridge University Press, 2004
R.Klages, G. Radons and I.M. Sokolov (Eds.) *Anomalous Transport: Foundations and Applications* (Wiley, VCH Berlin), 2008.
A. Blumen, G. Zumofen, and J. Klafter, Phys. Rev. A **40**, 3964 (1989).
J. Klafter, A. Blumen, G. Zumofen, and M. F. Shlesinger, Physica A, **168** 637 (1990).
P. Barthelemy, J. Bertolotti and D.S. Wiersma, Nature [453]{} 495 (2008).
J. Bertolotti, K. Vynck, L. Pattelli, P. Barthelemy, S. Lepri, D.S. Wiersma Adv. Funct. Mat. 20 Issue 6 , 965 - 968 (2010)
T. Geisel, J. Nierwetberg and A. Zacherl Phys. Rev. Lett. [54]{} 616 (1985), M. F. Shlesinger, G. M. Zaslavski and J. Klafter Nature, [363]{} 31 (1993), G. Zumofen and J. Klafter Phys. Rev. E [47]{}, 851 (1993).
H.C. Fogedby, Phys. Rev. Lett. [73]{} 2517 (1994), R. Kutner and P. Maass, J. Phys. A: Math. Gen. [31]{}, 2603 (1998), M. Schulz, Phys. Lett. A , [298]{}, 105 (2002).
E. Barkai, V. Fleurov, J. Klafter, Phys. Rev. E [61]{} 1164 (2000).
C.W.J. Beenakker, C.W. Groth, A.R. Akhmerov, Phys. Rev. B [79]{}, 024204 (2009).
R. Burioni, L. Caniparoli, S. and A. Vezzani Phys. Rev. E 81, 060101 (2010)
R. Burioni, L. Caniparoli, S. Lepri and A .Vezzani Phys. Rev. E [81]{}, 011127 (2010).
E. W. Montroll and G. H. Weiss, J. Math. Phys. [6]{} (1965) p. 167.
R. Burioni and D. Cassi, J. Phys. A [38]{}, R45-R78 (2005).
M.E. Cates, J. Physique [46]{}, (1985) p.1059. S. Alexander and R. Orbach, J. Physique Lett. [43]{} (1982) p. L62. P.J. Grabner and W.Woess, Stoc. Proc. Applic. 69, (1997) p. 127.
[^1]: $^\ast$Corresponding author. Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We use polarised neutron diffraction to study the induced magnetization density of near optimally doped Ba(Fe$_{0.935}$Co$_{0.065}$)$_2$As$_2$ ($T_C$=24 K) as a function of magnetic field ($1 \! < \! \mu_0 H \! < \! 9$ T) and temperature ($2 \! < \! T \! < \! 300$ K). The $T$-dependence of the induced moment in the superconducting state is consistent with the Yosida function, characteristic of spin-singlet pairing. The induced moment is proportional to applied field for $\mu_0 H \leq \mbox{9 T} \approx \mu_0 H_{c2}/6$. In addition to the Yosida spin-susceptibility, our results reveal a large zero-field contribution $M (H \rightarrow 0,T \rightarrow 0)/H \approx 2/3 \chi_{\mathrm{normal}}$ which does not scale with the field or number of vortices and is most likely due to the van Vleck susceptibility. Magnetic structure factors derived from the polarization dependence of 15 Bragg reflections were used to make a maximum entropy reconstruction of the induced magnetization distribution in real space. The magnetization is confined to the Fe atoms and the measured density distribution is in good agreement with LAPW band structure calculations which suggest that the relevant bands near the Fermi energy are of the $d_{xz/yz}$ and $d_{xy}$ type.'
author:
- 'C. Lester'
- 'Jiun-Haw Chu'
- 'J. G. Analytis'
- 'A. Stunault'
- 'I. R. Fisher'
- 'S.M. Hayden'
bibliography:
- 'BaFeCoAs\_D3.bib'
title: 'A polarized neutron diffraction study of the field-induced magnetization in the normal and superconducting states of Ba(Fe$_{1-x}$Co$_{x}$)$_2$As$_2$ ($x$=0.65)'
---
Introduction
============
The discovery of superconductivity in iron-based materials such as LaO$_{1-x}$F$_x$FeAs [@Kamihara2008a] has ignited intense interest in this field. The common features of the iron-based superconductors appear to be that they are semi-metals which have electron and hole Fermi surface pockets, separated by a $(\pi,\pi)$ wavevector [@Guo_note]. Experiments have demonstrated the existence of strong spin excitations with this same wavevector for superconducting compositions. It is widely believed that iron-based superconductivity is mediated by these spin excitations resulting in superconducting states such as $s_{\pm}$, where the sign of the gap changes between different Fermi surface sheets.
Of particular interest in the iron-based superconductors is the structure of the superconducting gap and the nature of the low-energy electronic states. Penetration depth [@Fletcher2009a] and thermal conductivity studies [@Reid2010a] of a number of materials (e.g. LaFePO, KFe$_2$As$_2$ and BaFe$_2$(As$_{1-x}$P$_x$)$_2$) show evidence for low-energy quasiparticle excitations which could be due to nodes in the superconducting order parameter.
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ is a good system to study since it is possible to grow large single crystals with homogeneous doping. It has been studied widely by various probes including angle-resolved photoemission spectroscopy (ARPES) [@Terashima2009a], scanning tunnel microscopy (STM) [@Yin2009a], penetration depth [@Gordon2010a; @Luan2011a], $\mu$SR [@Williams2009a], heat capacity [@Gang2010a; @Gofryk2011a] and thermal conductivity [@Reid2010a]. Even in this single system, different gap characters have been proposed as a function of doping, including fully and nodally gapped structures [@Terashima2009a; @Yin2009a; @Gordon2010a; @Williams2009a; @Gang2010a; @Gofryk2011a; @Reid2010a].
Here we use half-polarized neutron diffraction to measure the susceptibility and induced magnetization in the normal and superconducting states of near optically doped Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x=0.065$). Our measurements shed light on the electronic structure and the nature of the low energy electronic states in both phases. By measuring the flipping ratios of a number of Bragg peaks, we are able to extract the spatial Fourier components of the induced magnetization density $\mathbf{M(r)}$. In a metal, this provides information about the electronic states near the Fermi energy. We compare our results with a band structure calculation. In addition to measuring the $\mathbf{M(r)}$, we also made a detailed study of the temperature and magnetic field dependence of the induced magnetization by measuring a single Bragg peak in detail.
The paper is organised as follows. In Sec. \[Sec:Background\] we introduce the polarized beam method used in our experiment and report our unpolarized structural refinement of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x=0.065$). In Sec. \[Sec:MagDens\] we report our determination of the real space magnetization density $\mathbf{M(r)}$ induced by an applied magnetic field. We also present the results of a LAPW calculation of the magnetization density distribution. In Sec. \[Sec:TBdep\], we report measurements of the induced magnetization *in the superconducting state* as a function of magnetic field and temperature. We discuss the significance of our observations with respect to the superconductivity in Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x=0.065$) and other experimental results. This is followed by a summary of our conclusions.
Background {#Sec:Background}
==========
Polarized Neutron Diffraction Studies of the Induced Magnetization
-------------------------------------------------------------------
Polarized neutron scattering experiments can directly measure the real-space magnetization density $\mathbf{M}(\mathbf{r})$ in the unit cell, induced by a large magnetic field $\mu_0 H$. Due to the periodic crystal structure, the applied magnetic field induces a magnetization density with spatial Fourier components $\mathbf{M} (\mathbf{G})$, where $\mathbf{G}$ are the reciprocal lattice vectors, such that
$$\label{Eq:M_r} \bf{M(r)} = \frac{1}{\nu_{0}}
\sum_{\bf{G}}{M(G)}\rm{exp(-}\it{i}\bf{G}\cdot\bf{r})$$
and $\nu_{0}$ is the volume of the unit cell. The Fourier components of the magnetization density are given by $$\label{Eq:M_G} \mathbf{M(G)} = \int_{\mbox{unit cell}} \mathbf{M(r)} \exp(i \mathbf{G
\cdot r})\rm{d}\mathbf{r}.$$
Neutrons interact with matter through the strong nuclear force and electromagnetic interaction. For neutrons with initial and final spin polarisations $\mathbf{\sigma}_i$ and $\mathbf{\sigma}_f$, the total scattering cross section is
$$\begin{aligned}
\label{Eq:xsection}
\left( \frac{d \sigma}{d \Omega} \right)_{\sigma_{i} \rightarrow \sigma_{f}}
& \propto &
\left| \langle \sigma_{i} |
\frac{\gamma r_{0}}{2 \mu_{B}}
\bbox{\sigma} \cdot \mathbf{\hat{G}} \times
\left\{ \mathbf{M(G)} \times
\mathbf{\hat{G}} \right\} \right. \nonumber \\
& + &
\left. F_{N}(\mathbf{G})
| \sigma_{f} \rangle \right|^{2},\end{aligned}$$
where $\gamma r_0=5.36\times 10^{-15}$ m and $F_N(\mathbf{G})$ is the nuclear structure factor. The sign of the first (magnetic) term in Eq. \[Eq:xsection\] can be changed by reversing the incident neutron polarization. Thus we are able to isolate the interference term between the nuclear and magnetic scattering. In this experiment we measure the flipping ratio $R$, defined as the ratio of the cross-sections with neutrons parallel and anti-parallel to the applied magnetic field. Because the induced moment is small, the experiment is carried out in the limit $(\gamma r_{0}/2 \mu_{B}) M(\mathbf{G})/F_N(\mathbf{G})\ll 1$. In this limit, the flipping ratio derived from Eq. \[Eq:xsection\] is, $$\label{Eq:Flipping}
R = \frac{|F_{N}(\mathbf{G})-
(\gamma r_{0} /2 \mu_{B}) {M}(\mathbf{G})|^{2}}
{|F_{N}(\mathbf{G})+
(\gamma r_{0}/2 \mu_{B}) {M}(\mathbf{G})|^{2}}
\approx
1 - \frac{2 \gamma r_{0}}{\mu_{B}}
\frac{{M}(\mathbf{G})}
{F_{N}(\mathbf{G})}.$$ As the nuclear structure factors $F_{N}(\mathbf{G})$ are known from the crystal structure, Eq. \[Eq:Flipping\] directly gives $M({\bf G})$.
Experimental Details
--------------------
The Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x=0.065$) single crystal used in this study was prepared by a self-flux method [@Chu2009a], had approximate dimensions 6 $\times$ 1.5 $\times$ 0.2 mm$^3$ and a mass of $\sim$1.8 mg. Similar samples were used in our previous studies [@Lester2009a; @Lester2010a]. Resistivity and magnetization measurements on crystals from the same batch identified the superconducting transition temperature $T_c$(onset)=24 K and showed no evidence of magnetic order down to 2 K. The bulk susceptibility in a 5 T field measured using a SQUID magnetometer was $\chi_{ab}=1.22 \times 10^{-3}$ $\mu_B$ T$^{-1}$ f.u.$^{-1}$. We note that the expected upper critical field applied in the $ab$ plane for this composition is $H_{c2,ab}=$55 T [@Kano2009a].
Neutron scattering experiments were performed at the Institut Laue-Langevin, Grenoble, France. An initial unpolarized structural refinement was performed using the 4-circle D9 spectrometer with $\lambda=0.837(1)$ Å. The results are shown in Table \[table:D9\].
[lcccc]{} Atom & & $z$& B (Å$^2$)\
Ba & 2a & $(0\ 0\ 0)$ & & 0.023(16)\
Fe/Co & 4d & $(\frac{1}{2}\ 0\ \frac{1}{4})$ & & 0.110(10)\
As & 4e & $(0\ 0\ z)$ & $0.35352$ & 0.083(11)\
\
\
Polarized beam measurements of the flipping ratio were made on the D3 spectrometer. The sample was mounted on a thin aluminium post with the $[1\bar{1}0]$ direction vertical and parallel to the applied field. Data was collected with an incident wavelength $\lambda=0.825$ Å and a 0.5 mm Er filter (to reduce higher order contamination in the incident beam). Flipping ratios for equivalent reflections were collected, averaged and corrected for the finite beam polarization and extinction effects. The sample was cooled through $T_c$ at each field measured when collecting data in the superconducting state.
Induced Magnetization Distribution {#Sec:MagDens}
==================================
Results
-------
----- ----- ----- -------- ---------------- -------- -----------------
$h$ $k$ $l$
2 2 0 0.3611 0.83 $\pm$ 0.3 36.16 3.1 $\pm$ 0.9
0 0 2 0.0780 4.98 $\pm$ 0.3 -16.88 -10.0 $\pm$ 0.5
1 1 2 0.1967 3.75 $\pm$ 0.5 19.93 7.5 $\pm$ 0.8
2 2 2 0.3694 1.90 $\pm$ 0.6 -16.54 -3.1 $\pm$ 1.0
0 0 4 0.1560 6.49 $\pm$ 0.5 12.19 8.0 $\pm$ 0.6
1 1 4 0.2386 3.08 $\pm$ 0.5 -24.46 -7.6 $\pm$ 1.1
2 2 4 0.3933 0.8 $\pm$ 1.3 12.04 0.9 $\pm$ 1.5
1 1 6 0.2956 1.61 $\pm$ 0.2 32.68 5.5 $\pm$ 0.9
0 0 8 0.3120 1.58 $\pm$ 0.4 29.32 4.9 $\pm$ 0.8
2 2 8 0.4772 1.18 $\pm$ 0.7 28.90 1.8 $\pm$ 1.5
0 0 10 0.3900 1.12 $\pm$ 0.5 -25.57 -2.8 $\pm$ 1.2
1 1 10 0.4298 4.52 $\pm$ 1.2 10.42 4.4 $\pm$ 1.2
2 2 10 0.5315 3.60 $\pm$ 2.9 -25.11 -1.1 $\pm$ 2.2
0 0 12 0.4681 0.72 $\pm$ 0.6 23.42 2.9 $\pm$ 1.2
1 1 12 0.5017 3.29 $\pm$ 4.5 -12.16 -4.6 $\pm$ 1.9
----- ----- ----- -------- ---------------- -------- -----------------
: For each Bragg reflection $(hkl)$, the table shows: $\sin\theta/\lambda$, the measured flipping ratio $R$, structure factor $F_N(\mathbf{G})$ calculated from the structure in Table \[table:D9\], and the determined magnetic structure factor $M(\mathbf{G})$ [@SF_sign]. Data collected at $T=30$ K and $\mu_0 H=9$ T. \[table:D3\]
We measured the magnetization induced by a 9 T magnetic field applied along the $[1\bar{1}0]$ direction in the normal state at $T$=30 K. Table \[table:D3\] shows the measured flipping ratios under these conditions and the extracted Fourier components of the magnetization density. Fig. \[Fig:FormFactor\] shows $|M(\mathbf{G})|$ plotted against $\sin \theta /\lambda = |\mathbf{G}|/4 \pi$. If Fe were the only magnetic atom, then the sign of $M(\mathbf{G})$ is determined by the geometric structure factor for the Fe atoms and $|M(\mathbf{G})|$ is the effective form factor of the Fe atom. The solid line in Fig. \[Fig:FormFactor\] is the standard calculated isotropic atomic form factor for Fe$^{2+}$ [@Brown1992]. Deviations from an isotropic form factor are expected at larger $\theta$ or $|\mathbf{G}|$. Our results are in broad agreement with a recent study of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ $(x=0.066)$ [@Prokes2011a]. The main differences between the present data and the data presented in Ref. are: (i) the present study has higher statistical accuracy (ii) the value of $M(\mathbf{G})$ for $\mathbf{G}$=(002) is larger relative to the other $\mathbf{G}$ points in the present study. Our extinction model shows that the (002) peak has the largest extinction correction factor of 1.26.
![The magnetic structure factors $|M(\mathbf{G})|$ measured for a 9 T field applied along the $[1\bar{1}0]$ direction. The solid line is the Fe$^{2+}$ form factor [@Brown1992] scaled to the measured susceptibility. []{data-label="Fig:FormFactor"}](FormFactor.eps){width="0.95\linewidth"}
We used the maximum entropy method (MaxEnt) [@Skilling1989a; @Papoular1990a; @Gull1989a; @Brown2010a] to make a model-free reconstruction of the magnetization density in real space. Flipping ratios for reflections of the type $(hhl)$ were collected, this allowed the reconstruction of the magnetization density projected down the $[1\bar{1}0]$ direction onto the $(110)$ plane as illustrated in the left panel of Fig. \[Fig:UnitCell\]. The result of the reconstruction is shown in the right panel of Fig. \[Fig:UnitCell\]. As expected, the magnetization density is localized mostly on the Fe atoms. The magnetization “cloud” appears to be slightly extended along the $\langle 110 \rangle$ directions. Our results are in broad agreement with Ref. [@Prokes2011a], however we observe no significant magnetization density on the Ba sites. A recent study of the paramagnetic state of the parent antiferromagnet BaFe$_2$As$_2$ [@Brown2010a] is also broadly consistent with our results. The main difference in the BaFe$_2$As$_2$ case is that the magnetization extends more towards the As atoms.
![(left) The conventional tetragonal unit cell of Ba(Fe,Co)$_2$As$_2$. We measure the magnetization density integrated perpendicular to the $(110)$-type plane shown. The magnetization density shown is the result of the VCA calculation shown in Fig. \[Fig:MagDensity\]. (right) Projected magnetization distribution reconstructed from the experimental data in Table. \[table:D3\]. The reconstructed magnetization map shows $\frac{1}{4}$ of the area of the plane shown in the left panel.[]{data-label="Fig:UnitCell"}](UnitCell.eps){width="0.95\linewidth"}
Electronic Structure Calculations
---------------------------------
Induced form factor measurements have been widely used to determine the nature of the electrons responsible for paramagnetism in solids. In metals, the induced magnetization arises from a redistribution of electrons between up and down states near the Fermi energy. Thus, we probe the nature of the electronic wavefunctions for states near the Fermi energy. In order to understand our results further, we have carried out electronic structure calculations using the WEIN2k package [@Blaha2001a].
The linear augmented plane-wave (LAPW) method [@Singh2006a] was used to obtain the electronic structure and spin density of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x=0.065$). We used a full-potential LAPW method with the generalized gradient approximation (GGA). In the case of doped compositions, we used the virtual crystal approximation (VCA) [@Nordheim1931a]. We used the lattice parameters and atom position for the Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ $(x=0.065)$ structure shown in Table. \[table:D3\] for both compositions. The muffin-tin radii were chosen to be 2.5, 2.37 and 2.11 atomic units (a.u.) for Ba, Fe and As respectively, with the quantity $R_{\mathrm{MT}}K_{\mathrm{max}}$ set to 7, where $R_{\mathrm{MT}}$ is the smallest muffin-tin radius and $K_{\mathrm{max}}$ is the plane wave cutoff. For integrations we used 726 $k$-points in the irreducible Brillouin zone. In the LAPW method, the charge (spin) density is represented by a plane wave expansion in the interstitial region and as a combination of radial functions times spherical harmonics inside the muffin-tin spheres.
{width="0.95\linewidth"}
Fig. \[Fig:BandStructure\] shows the bandstructure calculated in the tetragonal phase for Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ for $x$=0 and $x$=0.065. Our calculations generally agree with others in the literature [@Singh2008a; @Thirupathaiah2010a; @Andersen2011a; @Colonna2011a], in particular, they show that the states near the Fermi energy are predominately of $\displaystyle{d_{xz,yz}}$ and ${\displaystyle d_{xy}}$ character.
![Magnetization density of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ calculated using the LAPW method and VCA approximation for $x=0$ (left panel) and $x=0.065$ (right panel). Calculations were carried out with a small fixed ferromagnetic moment to mimic the effect of an applied field. \[Fig:MagDensity\]](MagDensity_VCA.eps){width="0.95\linewidth"}
Spin polarized calculations were carried out in the tetragonal state with the ferromagnetic magnetization constrained to be 0.01 $\mu_{B}$ per unit cell to mimic the effect of an applied magnetic field. The results were then scaled to the measured magnetization at $\mu_0 H=$9 T for comparison purposes. The results are shown in Fig. \[Fig:MagDensity\]. The $x=0.065$ calculation shows reasonable agreement with our maximum entropy reconstruction in that the magnetization density is extended along the $\langle 110 \rangle$ direction. However, the reconstruction from our experimental data does not show the two maxima along $(\xi,\xi,1/2)$ predicted in the calculation. This is presumably because our data have insufficient Fourier components to resolve these features. We also carried out a calculation for BaFe$_2$As$_2$ ($x=0$) in the paramagnetic tetragonal state with the same structural parameters (but no Co potential) to demonstrate the sensitivity of the magnetization distribution to the bandstructure. Notice that the change in electronic structure between $x=0$ and $x=0.065$ causes a rotation of the calculated pattern in Fig. \[Fig:MagDensity\].
Temperature and Field Dependence of the Magnetization {#Sec:TBdep}
=====================================================
The measurement of the field and temperature dependence of bulk magnetization $M(H,T)$ in the mixed state of a superconductor provides information about the nature of the superconductive pairing. Thermal conductivity $\kappa(H,T,\theta)$ [@Reid2010a] and specific heat measurements $C=\gamma(H) T$ [@Moler1997a] in the mixed state have been extremely useful in probing the low energy quasi particles and identified the gap structure of a number of superconductors. In particular, the field dependence of the electronic contributions to $\kappa$ and $C$ in the $T\rightarrow0$ limit is sensitive to the symmetry of the superconducting gap function [@Shakeripour2009a] (see Fig. \[Fig:MCK\_comparison\]).
![Schematic field dependence of the low temperature magnetization $M$, specific heat $C$, and thermal conductivity $\kappa$ for an $s$-wave superconductor. The behavior for $C$ and $\kappa$ are based on V$_3$Si [@Boaknin2003a].[]{data-label="Fig:MCK_comparison"}](MCK_comparison.eps){width="0.85\linewidth"}
Complementary information is contained in $M(H,T)$. However, studies of the bulk magnetization in the mixed state of superconductors are not possible by conventional means, e.g. SQUID magnetometery, because of the presence of a large diamagnetic contribution. Polarized neutron diffraction and NMR Knight shift measurements [@Ning2008a; @Oh2011a] can be used to make magnetization measurements in the mixed state. The polarized neutron diffraction technique used here is unique because it directly measures the total magnetization including spin and orbital contributions. This technique was first used by Shull and Wedgwood in 1963 to study V$_3$Si [@Shull1966a] and has subsequently been applied to such varied superconductors as UPt$_3$ [@Stassis1986a], YBa$_2$Cu$_3$O$_{7-\delta}$ [@Boucherle1993a] and Sr$_2$RuO$_4$ [@Duffy2000a].
Results
-------
![The temperature dependence of the susceptibility and induced moment of V$_3$Si [@Shull1966a] (top panel) and Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ $x$=0.065 (bottom panel) measured using polarised neutron scattering. The solid lines are the Yosida behaviour expected for a singlet order parameter.[]{data-label="Fig:TDep"}](TDep.eps){width="0.85\linewidth"}
![(Top panel) Field dependence of the induced magnetization $M(\mathbf{G})$ of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ $x$=0.065 measured at $T=2$ K. $R$ is the corresponding flipping ratio. (Bottom panel) $M(\mathbf{G})/H$. The dashed line corresponds to the approximation $M_{\mathrm{spin}} \propto H^2$. \[Fig:HDep\]](HDep.eps){width="0.85\linewidth"}
We used the (002) Bragg peak to make our measurement since this requires little correction to give the $\mathbf{G}=0$ magnetization, $M(\mathbf{G}=(002))/M(\mathbf{G}=0)=0.979$ for an isotropic Fe$^{2+}$ ion. Fig. \[Fig:TDep\] shows the temperature dependence the induced magnetization of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ ($x$=0.065) for $\mu_0 H=9$ T. The temperature dependence shows the characteristic Yosida dependence expected for singlet pairing [@Yosida1958a], this is consistent with measurements of the NMR Knight shift in the same compound [@Ning2008a; @Oh2011a]. There is a large susceptibility in the $T\rightarrow0$ limit. A large residual contribution is also observed in V$_3$Si [@Shull1966a] and this has been attributed to the van Vleck contribution [@Clogston1962a]. The presence of large van Vleck contribution has also been inferred [@Grafe2008a; @Graser2009a] from NMR knight shift measurements on other iron-based superconductors. Fig. \[Fig:HDep\](top panel) shows the field dependence of the induced magnetization. $M(\mathbf{G}) \propto H$ over the field range of the present experiment ($\mu_0 H < 9$ T). When we plot the susceptibility $M(\mathbf{G})/H$ \[see Fig. \[Fig:HDep\](bottom panel)\], we find that the value as $T \rightarrow 0$ and $H \rightarrow 0$ is about $\frac{2}{3}$ of the normal state value measured at $T_c$ in the present experiment.
Interpretation
--------------
We measured the static response at finite wavevector $\mathbf{G}$ to an (approximately) uniform magnetic field, that is, $\chi(\mathbf{G}) \equiv \chi(\mathbf{G},0)=M({\mathbf{G}})/H(\mathbf{G}=0)$ [@M_sign; @SF_sign]. The atomic nature of solids means that the induced magnetization $M(\mathbf{r})$ is spatially modulated on an atomic scale (see Fig. \[Fig:UnitCell\]). Neutrons diffract from this modulation. In the mixed state of a superconductor, there is an additional diamagnetic magnetization (which gives rise to the vortex lattice), which is not detected in the present experiment. The signal from the vortex lattice is only present at small wavevectors (scattering angles) and can be studied by neutrons using small angle scattering techniques. In the following discussion we do not include the superconducting diamagnetic response.
The magnetic susceptibility in $d$-band metals has several components: atomic diamagnetic, van Vleck (“orbital” or “interband”) and spin. The atomic diamagnetic contribution in BaFe$_2$As$_2$ has been estimated [@Brown2010a] to be small and is neglected here. Only the spin contribution is expected to be suppressed by singlet Cooper pairing, thus we write the spatially averaged induced magnetization in the mixed state as: $$M= \chi_{\mathrm{orb}} H + M_{\mathrm{spin}}(T,H),$$ where $\chi_{\mathrm{orb}}$ is the orbital susceptibility and $M_{\mathrm{spin}}$ is the spin magnetization. In an $s$-wave superconductor the density of states due to the introduction of vortices is $\propto N_F H/H_{c2}$ [@Volovik1993a; @Caroli1964a]. Thus the spin magnetization should vary as $M_{\mathrm{spin}}(T \rightarrow 0,H) \propto H^2$. The temperature dependence is given by the Yosida function [@Yosida1958a], $M_{\mathrm{spin}}(T,H \rightarrow 0) \propto Y(T)$. The present experiments (Fig. \[Fig:HDep\]) show that the differential susceptibility $dM/dH$ has a large finite value in the $H\rightarrow0$ limit and $dM/dH \approx$ constant for $\mu_0 H < 9$ T. This is consistent with the finite $\chi(T\rightarrow 0,H \rightarrow 0)$ response being due to a van Vleck contribution. It should be noted that there is also NMR evidence for a residual susceptibility in BaFe$_{2}$(As$_{0.67}$P$_{0.33}$)$_2$ [@Nakai2010a] and BaFe(Fe$_{0.93}$Co$_{0.07}$)$_{2}$As$_{2}$ [@Oh2011a]. Specific heat measurements also suggest that there can be a sizeable residual quasiparticle density of states in BaFe(Fe$_{1-x}$Co$_{x}$)$_{2}$As$_{2}$ [@Gang2010a; @Gofryk2011a; @Residual_gam]. We cannot rule out the possibility that this is related to the residual susceptibility observed by neutron scattering.
There has been considerable debate about the nature of the superconducting gap in Ba(Fe$_{1-x}$Co$_{x}$)$_{2}$As$_{2}$. In principle, a detailed measurement of the $T$-dependence of the induced moment could be used to distinguish between different models for the gap. Unfortunately, the statistical noise in the present data is relatively high. Thus we make only a basic comparison with a singlet $s$-wave state. Within the statistical error of our data, the temperature dependent component of the induced moment $M_{\mathrm{spin}}$ is well described by a Yosida temperature dependence (see Fig. \[Fig:TDep\]) with $\Delta=1.78 k_B T=41$ K$=3.5$ meV.
Conclusions
===========
We have used a polarised neutron diffraction technique to measure the induced magnetization density of near optimally doped Ba(Fe$_{0.935}$Co$_{0.065}$)$_2$As$_2$ ($T_C$=24 K) as a function of magnetic field and temperature. The induced magnetization is confined to the Fe atoms with an oblate distribution spread out in the $a-b$ plane. The distribution is in reasonable agreement with a full potential LAPW band structure calculation which suggests that the relevant bands near the Fermi energy are of the $d_{xz/yz}$ and $d_{xy}$ type.
The $T$-dependence of the induced moment in the superconducting state is consistent with the Yosida function characteristic of spin-singlet pairing, and the induced moment is proportional to applied field for $\mu_0 H \leq \mbox{9 T} \approx \mu_0 H_{c2}/6$. We observe a large residual susceptibility $M (H \rightarrow 0,T \rightarrow 0)/H \approx 2/3 \chi_{\mathrm{normal}}$. This is most easily interpreted as being due to the van Vleck contribution present in other $d$-band systems, but may also signal a residual quasiparticle density of states.
Acknowledgements
================
We thank P. J. Brown, A. Carrington, P. J. Hirschfeld and I. I. Mazin for helpful discussions. Work at Stanford was supported by the Department of Energy, Office of Basic Energy Sciences under contract DE-AC02-76SF00515.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A new method to constrain the distance of blazars with unknown redshift using combined observations in the GeV and TeV regimes will be presented. The underlying assumption is that the Very High Energy (VHE) spectrum corrected for the absorption of TeV photons by the Extragalactic Background Light (EBL) via photon-photon interaction should still be softer than the extrapolation of the gamma-ray spectrum observed by Fermi/LAT. Starting from the observed spectral data at VHE, the EBL-corrected spectra are derived as a function of the redshift z and fitted with power laws. Comparing the redshift dependent VHE slopes with the power law fits to the LAT data an upper limit to the source redshift can be derived.
The method is applied to all TeV blazars detected by LAT with known distance and an empirical law describing the relation between the upper limits and the true redshifts is derived. This law can be used to estimate the distance of unknown redshift blazars: as an example, the distance of PKS 1424+240 is inferred.
author:
- |
E.Prandini[^1], G. Bonnoli, L. Maraschi, M. Mariotti\
F. Tavecchio
title: 'Blazars distance indications from [*Fermi*]{} and TeV data'
---
Introduction
============
Source Name $z$\[real\] [*Fermi*]{}/LAT slope TeV slope $z^*$
---------------- --------------- ----------------------- --------------- ----------------- -- -- --
Mkn 421 0.030 1.78 $\pm$ 0.03 2.3 $\pm$ 0.1 0.08 $\pm$ 0.02
Mkn 501 0.034 1.73 $\pm$ 0.06 2.3 $\pm$ 0.1 0.10 $\pm$ 0.02
1ES 2344$+$514 0.044 1.76 $\pm$ 0.27 2.9 $\pm$ 0.1 0.20 $\pm$ 0.06
Mkn 180 0.045 1.91 $\pm$ 0.18 3.3 $\pm$ 0.7 0.20 $\pm$ 0.12
1ES 1959$+$650 0.047 1.99 $\pm$ 0.09 2.6 $\pm$ 0.2 0.09 $\pm$ 0.04
BL Lacertae 0.069 2.43 $\pm$ 0.10 3.6 $\pm$ 0.5 0.23 $\pm$ 0.12
PKS 2005$-$489 0.071 1.91 $\pm$ 0.09 3.2 $\pm$ 0.2 0.19 $\pm$ 0.04
W Comae 0.102 2.02 $\pm$ 0.06 3.7 $\pm$ 0.2 0.23 $\pm$ 0.05
PKS 2155$-$304 0.116 1.87 $\pm$ 0.03 3.4 $\pm$ 0.1 0.22 $\pm$ 0.01
1ES 0806$+$524 0.138 2.04 $\pm$ 0.14 3.6 $\pm$ 1.0 0.23 $\pm$ 0.15
1ES 1218$+$304 0.182 1.63 $\pm$ 0.12 3.1 $\pm$ 0.3 0.21 $\pm$ 0.08
1ES 1011$+$496 0.212 1.82 $\pm$ 0.05 4.0 $\pm$ 0.5 0.49 $\pm$ 0.12
S5 0716$+$714 0.310$^{a,b}$ 2.16 $\pm$ 0.04 3.4 $\pm$ 0.5 0.21 $\pm$ 0.09
PG 1553+113 0.400$^c$ 1.69 $\pm$ 0.04 4.1 $\pm$ 0.2 0.57 $\pm$ 0.05
3C66A 0.444$^a$ 1.93 $\pm$ 0.04 4.1 $\pm$ 0.4 0.34 $\pm$ 0.05
3C279 0.536 2.34 $\pm$ 0.03 4.1 $\pm$ 0.7 0.75 $\pm$ 0.72
: TeV blazars used in this study. The sources used in this study are listed in the first column, their redshift (second column), their [*Fermi*]{}/LAT slope (third column), the VHE slope of the observed differential energy spectrum fit (fourth column) and the value $z^*$ (last column). $^a$Uncertain; $^b$from [@nilsson08]; $^c$from [@danforth10]. Detailed references can be found in [@prandini10].[]{data-label="table_values"}
The extragalactic TeV sky catalogue ($E>100$ GeV), counts nowadays 45 objects[^2]. Many of these sources have been recently detected also at GeV energies by the [*Fermi*]{} satellite [@abdo09], allowing for the first time a quasi-continuous coverage of the spectral shape of extragalactic VHE emitters over more than 4 decades of energy. The large majority of extragalactic TeV emitting objects are blazars, radio-loud active galactic nuclei with a relativistic jet closely oriented toward the Earth, as described in [@urry]. Here, we discuss a method, recently published in [@prandini10], to derive an upper limit on the redshift of a blazar, based on the comparison between the spectral index at GeV energies as measured by LAT (unaffected by the cosmological absorption up to redshifts far beyond those of interest here) and the TeV spectrum corrected for the absorption. Starting from the derived limits, we find a simple law relating these values to real redshift, which can be used to guess the distance of unknown redshift blazars. We assume a cosmological scenario with $h=0.72$, $\Omega_M=0.3$ and $\Omega_\Lambda=0.7$.
Results
=======
The photon flux emitted by a blazar in both GeV and TeV regimes can be well approximated by power laws, of the form $dN/dE = f_0 (E/E_0)^{-\Gamma}$, where $\Gamma$ is the power-law index. At VHE, the photons of the spectrum interact with the extragalactic background light (EBL), via electron-positron pair creation. Quantitatively, the effect is an exponential attenuation of the flux by a factor $\tau (E,z)$, where $\tau$ is the optical depth, a function of both photon energy and source redshift. Thus, the observed differential energy spectrum from a blazar, $F_{\rm obs}$, is related to the emitted one, $F_{\rm em}$, according to $F_{\rm obs}(E)=e^{-\tau(E)} F_{\rm em}(E)$.
In order to estimate a safe upper limit to the source distance, we can reasonably assume that the intrinsic spectrum at TeV energies cannot be harder than that in the adjacent GeV band. Indeed, from the brightest objects studied at both GeV and TeV energies it appears that the SED is continuous, with a broad peak not requiring additional spectral components [@aharonian09]. Hence, a natural assumption is to require that the slope measured in the GeV energy range is a limit value for the power-law index of the de-absorbed TeV spectrum.
For the study, we consider the blazar sample listed in table \[table\_values\] and containing all the extragalactic TeV emitters located at redshift larger than $z=0.01$, detected by LAT after taking 5.5 months of data [@abdo09]. In order to estimate the redshift $z^*$ for which the TeV spectral slope equals to the GeV one, the measured spectral points of each source have been corrected for the corresponding absorption factor [@franceschini08], starting from redshift $z=0.01$, and the resulting spectrum fitted with a power law. The procedure, applied in fine steps of redshift, is iterated until the slope of the de-absorbed spectrum equals to the one measured by LAT. The corresponding redshift, $z^*$, reported in table \[table\_values\], is the limit value on the source distance.
Among the 16 sources considered in this study, 14 blazars have well-known redshift and are used to test the method, while the remaining two blazars (3C 66A and S5 0716+714) have uncertain redshift, and are considered separately. The errors on $z^*$ are estimated taking into account both errors on the TeV and LAT slopes. Fig. \[correlationplot\] shows the comparison between the known redshift, $x$-axis, and $z^*$. All the $z^*$ lie above the bisector (dashed line) meaning that their values are larger than those of the the real redshift $z[true]$. This is expected since we are not considering the presence of the intrinsic break in the blazar spectra, and [*confirms that the method can be used to set safe upper limits on blazars distance.*]{} The only exceptions are the two sources with uncertain distance, S 0716+714 and 3C 66A (open circles). This could be either due to some intrinsic properties of the sources or to a wrong estimate of their distances. In the latter case, our method would constrain, at two sigma level, the redshift of S5 0716+714 below $0.39$ and that of 3C 66A below $0.44$.
![$z^*$ versus true redshift derived with the procedure described in the text. The open points are the two uncertain redshift sources, namely 3C 66A and S5 0716+714, not used in the fit calculation (continuous line). The dashed line is the bisector.[]{data-label="correlationplot"}](prandini_fig01.eps){width="4.in"}
In [@stecker10], a linear expression for the steepening of the observed TeV slope due to EBL absorption is derived. Since in our procedure $z^*$ is related to this steepening, it is natural to assume that also $z^*$ and $z$\[true\] are related by a linear function, of the form $z^*=A+Bz$\[true\]. The meaning of the coefficients is rather transparent: basically $A$ is a measure of the intrinsic spectral break of the sources, while, following [@stecker10], $B$ is a measure (increasing values for decreasing EBL level) of the optical depth of the EBL model used. We interpolate with this linear function the data with well-known distance of figure \[correlationplot\]. The linear fit (continuous line) has a probability of $62\%$. Once derived this empirical relation, one can use it to [*determine the redshift*]{} of sources with uncertain distance. For S5 0716+714 the reconstructed redshift is $z[rec]$ = 0.11 $\pm$ 0.05, while that of 3C 66A is $z[rec]$ = 0.21 $\pm$ 0.05. The error quoted is estimated in [@prandini10].
The redshift of PKS 1424+240
============================
As a final example of application, we use our procedure on PKS 1424+240, a blazar of unknown redshift recently observed in the VHE regime by Veritas [@acciari10]. The slope spectrum measured by [*Fermi*]{}/LAT between $0.2$ and $300$ GeV is $1.85\pm0.05$. The corresponding $z^*$ redshift at which the de-absorbed TeV spectrum slope becomes equal to it, is $0.382\pm0.105$, figure \[1424\_plot\], using the EBL model [@franceschini08]. This result is in agreement with the value of $0.5\pm0.1$, reported in [@acciari10], calculated applying the same procedure but only simultaneous [*Fermi*]{} data. Our estimate on the most probable distance for PKS 1424+240, obtained by inverting the $z^*$ formula, is $z$\[rec\]$=0.24\pm0.05$, where, as before, the error quoted is estimated in [@prandini10].
![Measured (open points) and deabsorbed (filled points) spectrum of PKS 1424+240 at redshift $z$ = 0.382.[]{data-label="1424_plot"}](prandini_fig02.eps){width="3.5in"}
GB, LM and FT acknowledge a 2007 Prin-MIUR grant for financial support.
. submitted to ApJ
[^1]: e-mail: [email protected]
[^2]: For an updated list see: http://www.mppmu.mpg.de/$\sim$rwagner/sources/
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Deutsches Elektronen-Synchrotron DESY\
22603 Hamburg, Germany\
E-mail:
- |
Markus Diehl\
Deutsches Elektronen-Synchrotron DESY\
22603 Hamburg, Germany\
E-mail:
- |
Tomas Kasemets\
Nikhef and Department of Physics and Astronomy, VU University Amsterdam\
De Boelelaan 1081, NL-1081 HV Amsterdam, the Netherlands\
E-mail:
title: Double parton scattering for perturbative transverse momenta
---
Introduction {#ss:introduction}
============
In the collision of two hadrons, double parton scattering (DPS) describes interactions in the form of two hard processes, each initiated by a separate set of partons. DPS was already considered long ago [@Landshoff:1975eb; @Landshoff:1978fq] and its understanding is relevant for studying physics at particle colliders such as the LHC [@Bartalini:2011jp; @Proceedings:2016tff]. In these proceedings we will focus on DPS with color singlet final states. An example of such a process is e.g. double Drell-Yan (DY), for which a factorization formula was first written down in the Refs. [@Paver:1982yp; @Mekhfi:1983az].
In DPS processes, correlators are described by double parton distribution functions (DPDFs) and double transverse momentum dependent distributions (DTMDs). In configuration space the correlator is described by a combination of the parameters $z_1$, $z_2$ and $y$, see Fig. \[f:labeling\] for an illustration. The transverse distance $\bm{y}$ is a measure for the separation between the two hard processes [@Paver:1982yp; @Mekhfi:1983az; @Diehl:2011yj]. In momentum space the situation is described by a combination of the momenta $k_{1}$, $k_2$ and $r$. For the discussion in these proceedings, we will consider the short-distance expansion, where the two hard processes have a large spatial separation and where both transverse momenta $\bm{k}_1$ and $\bm{k}_2$ are perturbative. As such, $\bm{z}_1$ and $\bm{z}_2$ are small compared with a nonperturbative scale $1/\Lambda$ and $\bm{y}$ is of the order $1/\Lambda$. Although we use these approximations, many results we present are valid beyond it.
![The momenta involved in double parton scattering illustrated for a double parton correlator. The momentum space configuration involves the momenta $k_{i}$ and $r$, whereas $\textcolor{blue}{z_{i}}$ and $\textcolor{blue}{y}$ (blue color online) are the configuration space variables.[]{data-label="f:labeling"}](labeling_2){width="50.00000%"}
In these proceedings, we will give details of factorizing the DPS soft factor into $z_{1}$, $z_{1}$ and $y$-dependent parts. An evolution kernel $K$, familiar from the context of single parton scattering (SPS) [@Collins:1981uk; @Collins:1981uw; @Collins:1984kg; @Collins:2011zzd], is related to this soft factor. Furthermore, one can choose the two hard processes to have separate renormalization scales $\mu_{1}$ and $\mu_{2}$. There is a scale $\zeta$ on top that plays the role of a rapidity regularization scale. We will solve their corresponding evolution equations. Finally, we provide the matching equations of DTMDs onto DPDFs for processes with a colorless final state, which is the main goal of these proceedings. For more details we refer to a forthcoming paper [@Buffing:2016].
Soft factors and color {#ss:soft_factors}
======================
In DPS not only the two hard interactions have to be treated, also Wilson lines have to be taken into account. First of all there are Wilson lines needed for ensuring gauge invariance of the correlators. Such Wilson lines come from diagrams with gluons coupling to partons involved in the hard scatterings. Furthermore, the presence of a soft factor is required, coupling the two correlators to each other. Factorization proofs ensure that the different contributions to the cross section factorize, see e.g. Ref. [@Diehl:2015bca] and references therein. The soft factor ensures a cancellation of the rapidity divergence and is required to define the subtracted DTMDs from the unsubtracted ones. We note that in the situation where all transverse distances are short (this is different from our situation), the soft factor for DPS has been calculated at the two loop level [@Vladimirov:2016qkd].
We focus on a single pair of two Wilson lines first, which is illustrated in Fig. \[f:SPSsoft\]. Each double line in the Feynman graph represents a Wilson line given by $$\begin{aligned}
W_{ij}(\bm{z},v) = \mathcal{P}\exp\left[
-igt^{a}\int_{-\infty}^{0}d\lambda\, v\,A^{\alpha}(z + \lambda v)\right]_{ij}^{z^{+} = z^{-} = 0}\end{aligned}$$ and similarly for the adjoint representation. For the Wilson lines, the first argument indicates the position of the gauge field, whereas $v$ is a vector that is associated with the rapidity.
![Illustration of interactions in a single pair of two Wilson lines in the soft function. As explained in the main text, short-distance interactions (red online) are located in the boxed section in the center of this figure.[]{data-label="f:SPSsoft"}](soft-fact-match){width="25.01290%"}
Consider a short-distance expansion of the Wilson line operator structure of the soft function that contains Wilson lines of the form $W(\bm{\xi}\pm\tfrac{1}{2}\bm{z},v)$ around $\bm{z}=0$. Short-distance interactions are then located closest to the point $\xi^{+} = \xi^{-} = 0$ of the Wilson lines, since this will result in a minimization of the number of Eikonal propagators with a large momentum. This is illustrated in Fig. \[f:SPSsoft\] for a single pair of two Wilson lines, where these kind of contributions are located in the boxed region in the center. It is necessary to keep open indices for the Wilson lines in this region, since this obect will appear in the matching procedure. A second pair of two Wilson lines should be included for DPS with the same type of short-distance interactions as the pair of two Wilson lines in Fig. \[f:SPSsoft\]. Although these short-distance interactions for the Wilson lines then are pairwise at the same point, the structure still contains nonperturbative interactions between Wilson lines separated by $\bm{y}$. In Fig. \[f:SPSsoft\] this corresponds to the gluons outside the box. Schematically, the structure of the soft function reads $$\begin{aligned}
S(\bm{z}_1, \bm{z}_2, \bm{y}) & = C_{s}(\bm{z}_1)\, C_{s}(\bm{z}_2)\,S(\bm{y}), \label{e:soft-fact-match_A}\end{aligned}$$ where $S(\bm{y})$ has half as many indices as $S(\bm{z}_1, \bm{z}_2, \bm{y})$, since it includes the open indices in the middle of the soft function. The coefficients $C_{s}$ are the contributions to the soft factor that contain all the $\bm{z}_{1}$ or $\bm{z}_{2}$ contributions respectively.
![(a) The soft function for double parton scattering decomposed in individual Wilson lines. Color projectors act on the indices in the middle. We have shown for $\bm{z}_{1} = \bm{z}_{2} = \bm{0}$ that it is equivalent to contract the indices in the middle and let the color projectors act on the indices at the ends of the Wilson lines, schematically giving us the structure in (b). Note that the picture includes gluon exchanges between the Wilson lines.[]{data-label="f:softstructure"}](soft-fact-indices_2 "fig:"){width="42.00000%"} ![(a) The soft function for double parton scattering decomposed in individual Wilson lines. Color projectors act on the indices in the middle. We have shown for $\bm{z}_{1} = \bm{z}_{2} = \bm{0}$ that it is equivalent to contract the indices in the middle and let the color projectors act on the indices at the ends of the Wilson lines, schematically giving us the structure in (b). Note that the picture includes gluon exchanges between the Wilson lines.[]{data-label="f:softstructure"}](soft-fact-indices_contracted_2 "fig:"){width="42.00000%"}\
(a) (b)
The most general soft factor for DPS is illustrated in Fig. \[f:softstructure\](a). Regarding the soft function in this figure, the indices $i$ at the top and the indices $l$ at the bottom represent the start and end of the Wilson lines at the two correlators, whereas the indices $j$ and $k$ in the middle involve the coupling of these Wilson lines. The indices are not contracted yet because we wish to do matching later on. Hard scattering in the DPS couples four parton lines, for which we choose to insert color projectors $P_{R}$ in order to simplify the color structure. The index $R$ labels the different color configuration that are possible (singlet, octet, etc.) [@Kasemets:2014yna], since the color structure for DPS has more possible configurations than for the SPS case. Two color projectors coupling the various indices in the middle of Fig. \[f:softstructure\](a) are required, since there are two processes involved. Examples are the color singlet and octet quark projectors, given as $$\begin{aligned}
P_{1}^{j_{1}j_{1}^{\prime}\,k_{1}k_{1}^{\prime}} = \frac{1}{N_c}\delta_{j_{1}j_{1}^{\prime}}\delta_{k_{1}k_{1}^{\prime}}, \hspace{15mm}
P_{8}^{j_{1}j_{1}^{\prime}\,k_{1}k_{1}^{\prime}} = 2 t_{j_{1}j_{1}^{\prime}}^{a}t_{k_{1}k_{1}^{\prime}}^{a}. \label{e:colorprojectors} \end{aligned}$$ For gluons more color projectors exist and mixed quark-gluon projectors also have to be considered [@Diehl:2011yj; @Kasemets:2014yna], since in double parton scattering one of the partons could be a quark and the other one a gluon. We generalize the notation for the projectors introduced in Eq. \[e:colorprojectors\] as $P_{R}^{j_{i}j_{i}^{\prime}\,k_{i}k_{i}^{\prime}}$. In SPS the situation is simple and we would have $P_{1}^{jj^{\prime}}$ only.
Coming back to simplifying the soft function, we would in particular like to simplify the nonperturbative contribution $S(\bm y)$. Transforming the soft function in Fig. \[f:softstructure\](a) to that in Fig. \[f:softstructure\](b) is a convenient and simple way of ensuring this. For the nonperturbative sector, where it is desirable to have as few functions as possible, a simplification is especially helpful, since a procedure transforming the soft function in Fig. \[f:softstructure\](a) to that in Fig. \[f:softstructure\](b) reduces the number of open indices in the soft factor significantly. This transformation can be achieved by using color projectors for $S(\bm y)$ for $\bm{z}_{1} = \bm{z}_{2} = \bm{0}$.
In order to prove that the Wilson lines can be contracted in a way that would allow the above sought simplification, we have to prove that we can commute color projectors acting on the Wilson lines in the soft factor through the Wilson lines. Then, the projectors would no longer be acting on the indices in the middle, but on indices at the ends of the Wilson lines and we could contract the Wilson lines, which would reduce the number of open indices. The identity we have to prove is $$\begin{aligned}
W^{}_{ij}\, P_R^{jj',k'k}\, W^\dagger_{j'i'} & = W^{}_{jk}\, P_R^{ii',j'j}\, W^\dagger_{k'j'}, \label{e:comm-quark}\end{aligned}$$ which is illustrated graphically in Fig. \[f:softmatch\]. We have proven this identity using the color Fierz identity $$\begin{aligned}
2t_{ii^{\prime}}^{a}2t_{jj^{\prime}}^{a} & = \delta_{ij^{\prime}}\delta_{i^{\prime}j} - \frac{1}{N_c} \delta_{ii^{\prime}}\delta_{jj^{\prime}}.\end{aligned}$$ Note that this relation only holds for the collinear situation, where the Wilson lines are at the same transverse position. For our purposes this is fine, since we will use it to study matching. The factorization also works for adjoint Wilson lines, which we need as soon as gluons are involved, implying $$\begin{aligned}
W^{}_{ab}\, P_R^{bb',cc'}\, W^\dagger_{b'a'} & = W^{}_{bc}\, P_R^{aa',bb'}\, W^\dagger_{c'b'}. \label{e:comm-glu}\end{aligned}$$
![Illustration of moving the color projector through the Wilson line structure. As explained before and illustrated in a previous figure, this allows for contracting the Wilson lines in the center region for $\bm{z}_{1} = \bm{z}_{2} = \bm{0}$. As before, note that the picture includes gluon exchanges between the Wilson lines.[]{data-label="f:softmatch"}](soft-match_2){width="42.86190%"}
Using the color projector identity in Fig. \[f:softmatch\], we can relate the full soft factor in Fig. \[f:softstructure\](a) to the soft function in Fig. \[f:softstructure\](b), where the indices in the middle are contracted. Rather than making the color projections at the color indices of the fields at $\xi^{+} = \xi^{-} = 0$, the same projection can be made for the indices of the fields at infinity. The fact that this relation holds implies that the collinear soft matrix for DPS is diagonal in the color representations of the left and right moving Wilson lines. The color structure of Eq. \[e:soft-fact-match\_A\] then reads $$\begin{aligned}
{}^{RR'}S_{a_1 a_2}(\bm{z}_1, \bm{z}_2, \bm{y}) & = {}^{R\,}C_{s, a_1}(\bm{z}_1)\, {}^{R\,}C_{s, a_2}(\bm{z}_2)\,{}^{RR}S(\bm{y})\, \delta_{RR'}^{} \,. \label{e:soft-fact-match_B}\end{aligned}$$ Using projector identities one can also show that the soft factor ${}^{RR^{\prime}}S(\bm{y})$ is color diagonal. It follows from the proof of the above equation that the soft factor ${}^{RR'}S_{a_1 a_2}(\bm{z}_1, \bm{z}_1, \bm{y})$ is diagonal in the color representations $R$ and $R^{\prime}$ in the limit $|\bm{z}_1|,\,\,|\bm{z}_2|\,\,\ll\,\,|\bm{y}|$.
Now we come back to DPDs. They evolve with the evolution kernel ${}^{RR^{\prime}}K$, which is closely related to the soft factor ${}^{RR^{\prime}}S$ in a similar way as for SPS, but more complicated. The multiplicative structure of Eq. \[e:soft-fact-match\_B\] results in $$\begin{aligned}
{}^{RR'}K_{a_1 a_2}(\bm{z}_i,\bm{y}; \mu_i) & = \delta_{RR'}^{}\, \bigl[{}^{R}K_{a_1}(\bm{z}_1;\mu_1) + {}^{R}K_{a_2}(\bm{z}_2;\mu_2) + {}^{R}{J(\bm{y}; \mu_i)}\bigr], \label{e:CS-gen-match}\end{aligned}$$ see a forthcoming paper for the details [@Buffing:2016]. In Eq. \[e:CS-gen-match\] and all following equations, $\mu_{i}$ implies a dependence on both $\mu_1$ and $\mu_2$ and similarly for the other parameters where this notation is used. It should further be stressed that in contrast to SPS both the soft function ${}^{RR'}S$ and the evolution kernel ${}^{RR'}K$ are matrix valued expressions. The fact that the evolution kernel in Eq. \[e:CS-gen-match\] is a sum of three separate contributions simplifies dealing with evolution.
Evolution equations {#ss:evolution}
===================
A description of DTMDs involves rapidity and scale parameters, for which evolution equations have to be derived and solved. Earlier descriptions of evolution for DTMDs using a different framework can be found in literature, see e.g. the Refs. [@Diehl:2011yj; @Diehl:2011tt]. We give results for the situation $|\bm{z}_1|,\,\,|\bm{z}_2|\,\,\ll\,\,|\bm{y}|$, but as was the topic of the talk of M. Diehl at this conference, many of our results have a wider applicability than the small distance expansion only. In these proceedings, on the other hand, we are looking at the small distance expansion, where the two hard processes are separated from each other. As such, we have two different renormalization scales, $\mu_{1}$ and $\mu_{2}$, for which we need two separate evolution equations, namely $$\begin{aligned}
\frac{\partial}{\partial \log\mu_1}\,{}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta) & = \gamma_{F, a_1}(\mu_1, x_1\zeta/x_2)\,{}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta) \label{e:RG-TMD}\end{aligned}$$ and a similar equation for the $\mu_2$ evolution. Furthermore, the DTMD carries a color representation index $R$. Here, the $\gamma_{F,a_i}$ are anomalous dimensions of the DTMDs, equal to the same objects in the TMD evolution [@Collins:2011zzd; @Aybat:2011zv]. Also, $\gamma_{F,a}$ depends only on whether one is dealing with (anti)quarks, involving fundamental Wilson lines, or with gluons, involving adjoint Wilson lines. In the $\mu$-evolution equations, the anomalous dimensions $\gamma_{F}$ depend on the rapidity regularization scale $\zeta$. A closer analysis shows that $\zeta$ has to be rescaled with $x_{1}/x_{2}$ or $x_{2}/x_{1}$. This rescaling is required, since we have chosen to use a single $\zeta$ scale per DTMD. It then follows from Eq. \[e:RG-TMD\] that the $\mu$ evolution of DTMDs is given by $$\begin{aligned}
{}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta) = & {}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_{0i},\zeta) \nonumber \\[0.2em]
& \qquad \times\exp\biggl[ \int_{\mu_{01}}^{\mu_1} \frac{d\mu}{\mu}\,\gamma_{F,a_1}(\mu, x_1\zeta/x_2) + \int_{\mu_{02}}^{\mu_2} \frac{d\mu}{\mu}\,\gamma_{F,a_2}(\mu, x_2\zeta/x_1) \biggr] \label{e:RG-TMD-sol}\end{aligned}$$ from the starting scales $\mu_{01}$ and $\mu_{02}$ for $\mu_{1}$ and $\mu_{2}$. The rapidity evolution of DTMDs is given by $$\begin{aligned}
\frac{\partial}{\partial\log \zeta}
{}^{R}{F_{a_1 a_2}}(x_i,\bm{z}_i,\bm{y}, \mu_i,\zeta) & = \frac{1}{2} \sum_{R'}{}^{RR'}{K_{a_1 a_2}(\bm{z}_i,\bm{y}; \mu_i)}\,{}^{R'}{F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y}; \mu_i,\zeta)} \, , \label{e:CS-TMD}\end{aligned}$$ where we can split the evolution kernel $K$ in the short-distance limit into the three separate contributions ${}^{R}K_{a_1}(\bm{z}_1;\mu_1)$, ${}^{R}K_{a_2}(\bm{z}_2;\mu_2)$ and ${}^{R}{J(\bm{y}; \mu_i)}$ as in Eq. \[e:CS-gen-match\].
The anomalous dimensions $\gamma_{F}$ in the $\mu$ evolution equations contain a $\zeta$ dependence and the evolution kernel $K$ in the $\zeta$ evolution equation is $\mu$-dependent. Before we can write down the full DTMD evolution equation, it has to be understood how they evolve. The evolution kernel ${}^{RR^{\prime}}K$ can be written as a sum of terms as in Eq. \[e:CS-gen-match\]. Taking the derivative with respect to $\mu_1$ of the separate contributions gives $$\begin{aligned}
\frac{\partial}{\partial \log\mu_1}\, {}^{R} K_a(\bm{z}; \mu_1) & = - {}^{R}\gamma_{K,a}(\mu_1), \label{e:CS-coll-RG_K} \\
\frac{\partial}{\partial \log\mu_1}\, {}^{R}J(\bm{y}; \mu_i) &= - {}^{R\,}\gamma_J(\mu_1) \label{e:CS-coll-RG_J}\end{aligned}$$ and similar equations for the derivative with respect to $\mu_2$. The anomalous dimensions in the Eqs. \[e:CS-coll-RG\_K\] and \[e:CS-coll-RG\_J\] satisfy $$\begin{aligned}
\gamma_{K,a}(\mu) &= {}^{R}\gamma_{K,a}(\mu) + {}^{R}\gamma_J (\mu) . \label{e:AD-sum}\end{aligned}$$ Furthermore, the rapidity dependence of the anomalous dimension $\gamma_{F,a_1}$ that came from the $\mu$-scale equation is given by $$\begin{aligned}
\frac{\partial}{\partial \log\zeta}\, \gamma_{F,a}(\mu, \zeta) &= - \frac{1}{2}\mskip 1.5mu \gamma_{K,a}(\mu) \, . \label{e:cusp}\end{aligned}$$ Note that the $\mu$ dependence of the anomalous dimensions is understood to be through the coupling $\alpha_{s}(\mu)$ only. Combining the above information regarding the evolution with respect to $\mu_1$, $\mu_2$ and $\zeta$, the solution of the evolution equations for DTMDs is then given by $$\begin{aligned}
{}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta) & = {}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_{0i},\zeta_0) \nonumber \\
& \quad \times\exp\,\biggl\{ \int_{\mu_{01}}^{\mu_1} \frac{d\mu}{\mu}\,\biggl[\gamma_{F,a_1}(\mu, \mu^2) - \gamma_{K,a_1}(\mu) \log\frac{\sqrt{x_1\zeta/x_2}}{\mu} \biggr] \nonumber \\
& \qquad\qquad + \int_{\mu_{02}}^{\mu_2} \frac{d\mu}{\mu}\,\biggl[ \gamma_{F,a_2}(\mu, \mu^2) - \gamma_{K,a_2}(\mu) \log\frac{\sqrt{x_2\zeta/x_1}}{\mu} \biggr] \nonumber \\
& \qquad\qquad + \Bigl[ {}^{R}K_{a_1}(\bm{z}_1,\mu_{01}) + {}^{R}K_{a_2}(\bm{z}_2,\mu_{02}) + {}^{R}J(\bm{y},\mu_{0i}) \Bigr]\log\frac{\sqrt{\zeta}}{\sqrt{\zeta_0}} \biggr\} \label{e:evsolving}\end{aligned}$$ for the starting scales $\mu_{01}$, $\mu_{02}$ and $\zeta_{0}$.
Matching {#ss:matching}
========
The matching equation for DTMD/DPDF matching is given by [@Buffing:2016] $$\begin{aligned}
& {}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta) & = \sum_{b_1 b_2}
{}^{R\,}C\!_{a_1 b_1}(x_1',\bm{z}_1;\mu_{1},\mu_{1}^2) \underset{x_1}{\otimes}
{}^{R\,}C\!_{a_2 b_2}(x_2',\bm{z}_2;\mu_{2},\mu_{2}^2) \underset{x_2}{\otimes}
{}^{R}F_{b_1 b_2}(x_i',\bm{y};\mu_{i},\zeta), \label{e:matching}\end{aligned}$$ where the convolution between two functions $A$ and $B$ is given by $$\begin{aligned}
A(x') \underset{x}{\otimes} B(x') &= \int_{x}^1 \frac{dx'}{x'}\, A(x')\,B\biggl(\frac{x}{x'}\biggr). \label{e:conv-def}\end{aligned}$$ The summation over $b_{1}$ and $b_{2}$ is over parton species and polarization as in Ref. [@Diehl:2011yj]. The two coefficient functions are both TMD/PDF matching coefficient functions, a statement that can be seen more easily when writing down the formalism for DTMD/DPDF matching at the level of operators [@Buffing:2016].
In Eq. \[e:evsolving\] the evolution of DTMDs from a set of starting scales is given. Combining this result with the matching equation in Eq. \[e:matching\], we can write the matching equation for the DTMD in the short-distance limit, where the rapidity dependence of the coefficient functions will be split off in separate terms containing ${}^{R}K\!_{a_1}(\bm{z}_1,\mu_{01})$ and ${}^{R}K\!_{a_2}(\bm{z}_2,\mu_{02})$. It can be shown that this transforms the matching equation into $$\begin{aligned}
& {}^{R}F_{a_1 a_2}(x_i,\bm{z}_i,\bm{y};\mu_i,\zeta)
\nonumber \\
&\quad = \sum_{b_1 b_2}
{}^{R\,}C\!_{a_1 b_1}(x_1',\bm{z}_1;\mu_{01},\mu_{01}^2) \underset{x_1}{\otimes}
{}^{R\,}C\!_{a_2 b_2}(x_2',\bm{z}_2;\mu_{02},\mu_{02}^2) \underset{x_2}{\otimes}
{}^{R}F_{b_1 b_2}(x_i',\bm{y};\mu_{0i},\zeta_0)
\nonumber \\
& \qquad \times \exp\, \biggl\{
\int_{\mu_{01}}^{\mu_1} \frac{d\mu}{\mu}\,
\biggl[ \gamma_{F,a_1}(\mu, \mu^2) - \gamma_{K,a_1}(\mu) \log\frac{\sqrt{x_1\zeta/x_2}}{\mu} \biggr]
+ {}^{R}K\!_{a_1}(\bm{z}_1,\mu_{01}) \log\frac{\sqrt{x_1\zeta/x_2}}{\mu_{01}}
\nonumber \\
& \qquad\qquad\;\, + \int_{\mu_{02}}^{\mu_2} \frac{d\mu}{\mu}\,
\biggl[ \gamma_{F,a_2}(\mu, \mu^2) - \gamma_{K,a_2}(\mu) \log\frac{\sqrt{x_2\zeta/x_1}}{\mu} \biggr]
+ {}^{R}K\!_{a_2}(\bm{z}_2,\mu_{02}) \log\frac{\sqrt{x_2\zeta/x_1}}{\mu_{02}}
\nonumber \\
& \qquad\qquad\;\, + {}^{R}J\!(\bm{y},\mu_{0i}) \log\frac{\sqrt{\zeta}}{\sqrt{\zeta_0}} \biggr\}. \label{e:small-z-evolved}\end{aligned}$$ For color singlet configurations the above combined evolution and matching equation in essence consists of doubling the single TMD/PDF formalism. For other color configurations there is an additional Sudakov suppression coming from ${}^{R}J\!(\bm{y},\mu_{0i})$, which is zero for color singlet configurations. Note that the evolution kernel contributions ${}^{R}K\!_{a_1}(\bm{z}_1,\mu_{01})$ and ${}^{R}K\!_{a_2}(\bm{z}_2,\mu_{02})$ have a color dependence.
At the level of the cross section for DPS in proton-proton collisions, two DTMDs have to be involved. The matching equation for a cross section contribution involving two such objects is given by $$\begin{aligned}
W_{\text{large $\bm{y}$}} &= \sum_{c_1 c_2 d_1 d_2} \sum_{R} \exp\, \biggl\{
\int_{\mu_{01}}^{\mu_1} \frac{d\mu}{\mu}\,
\biggl[\gamma_{F,a_1}(\mu, \mu^2)
- \gamma_{K,a_1}(\mu) \log\frac{{Q_1^2}}{\mu^2} \biggr]
+ {}^{R}K\!_{a_1}(\bm{z}_1,\mu_{01})
\log\frac{{Q_1^2}}{\mu_{01}^2} \nonumber \\
& \qquad\qquad + \int_{\mu_{02}}^{\mu_2} \frac{d\mu}{\mu}\,
\biggl[ \gamma_{F,a_2}(\mu, \mu^2)
- \gamma_{K,a_2}(\mu) \log\frac{{Q_2^2}}{\mu^2} \biggr]
+ {}^{R}K\!_{a_2}(\bm{z}_2,\mu_{02}) \log\frac{{Q_2^2}}{\mu_{02}^2} \biggr\} \nonumber \\[0.3em]
& \quad \times
{}^{R\,}C\!_{b_1 d_1}(\bar{x}_1',\bm{z}_1;\mu_{01},\mu_{01}^2) \underset{\bar{x}_1}{\otimes}
{}^{R\,}C\!_{b_2 d_2}(\bar{x}_2',\bm{z}_2;\mu_{02},\mu_{02}^2) \,\underset{\bar{x}_2}{\otimes} \nonumber \\[0.3em]
& \quad \times
{}^{R\,}C\!_{a_1 c_1}(x_1',\bm{z}_1;\mu_{01},\mu_{01}^2) \underset{x_1}{\otimes}
{}^{R\,}C\!_{a_2 c_2}(x_2',\bm{z}_2;\mu_{02},\mu_{02}^2)\,\underset{x_2}{\otimes} \nonumber \\
& \quad \times \bigl[ \Phi(\nu\bm{y}) \bigr]^2\,
\exp\,\biggl[{}^{R}J\!(\bm{y},\mu_{0i})\log\frac{\sqrt{Q_1^2 \mskip 1.5mu Q_2^2}}{\zeta_0} \,\biggr]
{}^{R}F_{d_1 d_2}(\bar{x}_i,\bm{y};\mu_{0i},\zeta_0)\,
{}^{R}F_{c_1 c_2}(x_i,\bm{y};\mu_{0i},\zeta_0). \label{e:W-large-y}\end{aligned}$$ In this equation, the energies $Q_{1}$ and $Q_2$ of the two hard partonic interactions are related to the scale parameters $\zeta$ and $\overline{\zeta}$ of the two DTMDs through the relation $\zeta\overline{\zeta} = Q_{1}^{2}Q_{2}^{2}$. Coming from Eq. \[e:small-z-evolved\], each DTMD contributes two coefficient functions, giving a grand total of four for the cross section contribution. In addition, there is a $\bm{y}$ contribution in the form of $\bigl[ \Phi(\nu\bm{y}) \bigr]^2$. This function regulates the ultraviolet region and ensures that the integral converges at small distances [@Diehl:2016khr]. It should be noted that the Eqs. \[e:small-z-evolved\] and \[e:W-large-y\] are valid for both quarks and gluons.
Discussions and conclusions {#ss:conclusions}
===========================
In our work we use the short-distance expansion, valid if the two partons initiating the two different hard processes in DPS have perturbative transverse momenta $\bm{k}_1$ and $\bm{k}_2$. We also consider the large $\bm{y}$ situation, such that the hard processes are spatially well separated from each other. In this limit the DPS soft function can be factorized in three separate contributions, namely $\bm{z}_1$, $\bm{z}_2$ and $\bm{y}$-dependent ones. From this, it follows that the evolution kernel has three separate terms. Using this important result we have given the evolution equations for DPDs and solved them. We furthermore presented the matching equations for DTMDs as well as for the cross section contribution for the production of colorless final states.
An important result is that the matching equation for the DTMDs, Eq. \[e:small-z-evolved\], has two coefficient functions, with each of them equal to a single TMD/PDF coefficient function. The reason for this is the use of the short-distance expansion and considering the large $\bm{y}$ situation. As such, for DPS the coefficient functions can be recycled from the coefficient functions for the TMD/PDF matching. In a derivation at the level of operators this is apparent from the start. In our forthcoming paper [@Buffing:2016] we will give the matching coefficients for all polarization modes, complementing results for TMDs in SPS [@Collins:2011zzd; @Aybat:2011zv; @Bacchetta:2013pqa; @Echevarria:2015uaa].
TK is supported by the European Community under the “Ideas” program QWORK (contract 320389).
[99]{} P.V. Landshoff, J.C. Polkinghorne and D.M. Scott, *Phys. Rev.* [**D 12**]{}, 3738 (1975). P.V. Landshoff and J.C. Polkinghorne, *Phys. Rev.* [**D 18**]{}, 3344 (1978). P. Bartalini e.a., \[[arXiv:1111.0469 \[hep-ph\]]{}\]. H. Jung, D. Treleani, M. Strikman and N. van Buuren, *Proceedings, 7th International Workshop on Multiple Partonic Interactions at the LHC (MPI@LHC 2015)*, DESY-PROC-2016-01 (2016). N. Paver and D. Treleani, *Nuovo Cim* [**A 70**]{}, 215 (1982). M. Mekhfi, *Phys. Rev.* [**D 32**]{}, 2371 (1985). M. Diehl, D. Ostermeier and A. Sch[ä]{}fer, *JHEP* [**03**]{}, 089 (2012) \[[arXiv:1111.0910 \[hep-ph\]]{}\]. J.C. Collins and D.E. Soper, *Nucl. Phys.* [**B 193**]{}, 381 (1981), erratum: [**B 213**]{}, 545 (1983). J.C. Collins and D.E. Soper, *Nucl. Phys.* [**B 194**]{}, 445 (1982). J.C. Collins, D.E. Soper and G.F. Sterman, *Nucl. Phys.* [**B 250**]{}, 199 (1985). J. Collins, *Foundations of perturbative QCD*, Cambridge University Press (2013). M.G.A. Buffing, M. Diehl and T. Kasemets, preprint number NIKHEF-2016-028, in preparation. M. Diehl, J.R. Gaunt, D. Ostermeier, P. Pl[ö]{}[ß]{}l and A. Sch[ä]{}fer, *JHEP* [**01**]{}, 076 (2016) \[[arXiv:1510.08696 \[hep-ph\]]{}\]. A. Vladimirov, \[[arXiv:1608.04920 \[hep-ph\]]{}\]. T. Kasemets and P.J. Mulders, *Phys. Rev.* [**D 91**]{}, 014015 (2015) \[[arXiv:1411.0726 \[hep-ph\]]{}\]. M. Diehl and A. Sch[ä]{}fer, *Phys. Lett.* [**B 698**]{}, 389 (2011) \[[arXiv:1102.3081 \[hep-ph\]]{}\]. S.M. Aybat and T.C. Rogers, *Phys. Rev.* [**D 83**]{}, 114042 (2011) \[[arXiv:1101.5057 \[hep-ph\]]{}\]. M. Diehl and J.R. Gaunt, *Proceedings, 7th International Workshop on Multiple Partonic Interactions at the LHC (MPI@LHC 2015)*, 121 \[[arXiv:1603.05468 \[hep-ph\]]{}\]. A. Bacchetta and A. Prokudin, *Nucl. Phys.* [**B 875**]{}, 536 (2013) \[[arXiv:1303.2129 \[hep-ph\]]{}\]. M.G. Echevarria, T. Kasemets, P.J. Mulders and C. Pisano, *JHEP* [**07**]{}, 158 (2015) \[[arXiv:1502.05354 \[hep-ph\]]{}\].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Daniel B. Smith\
Jian Liu
bibliography:
- 'actin-references.bib'
title: Supplement
---
Equilibration
=============
The simulations did appear to be sampling from an equilibrium distribution. For the simulations where the force-velocity relationship was loosely flat, the velocity converged to an equilibrium quite rapidly, as can be seen in Figure S1. The curves shown are the average of 10 simulations with the highlighted region representing the estimated $\pm$ standard deviation.
A)
{width="0.9\linewidth"}
B)
{width="0.9\linewidth"}
However, where the force-velocity curve was sharp, the velocity converged only at non-physiological time-scales, if at all. This can be seen in Figure S2.
A)
{width="0.9\linewidth"}
B)
{width="0.9\linewidth"}
Stall Force
===========
We estimated the stall force per filament by defining the system to be stalled when an increase in force of 170 pN led to a decrease in velocity of less than 1%. The stall forces per filament we observed ranged between 0.9 pN and 2.4 pN for $1\leq\kappa\leq5$/s/filament with a mean of 1.3 pN.
Velocity Reduction to Small Forces
==================================
Our model qualitatively reproduces the large velocity reduction in response to small forces observed in experiment [@Prass2006]. The graph below shows the equilibrium velocity $\displaystyle \nicefrac{v}{v_{free}}$ in response to a 170 pN force. We believe that this is due to the fact that the leading edge must be sufficiently slowed down for trailing filaments to catch up to the leading edge. Thus, the force-independent velocities must be significantly slower than $v_{free}$.
{width="0.6\linewidth"}
Spatial Restriction of Branching
================================
There is experimental evidence that new filaments only branch in a small zone bordering the membrane. To test how well our results would hold up under that type of condition, we ran another set of simulations restricting where filaments could branch. For Figure S4, we kept all of the conditions identical to simulations in the main text except for that filaments only branched in a zone of $N\delta$ away from the leading edge.
{width="0.6\linewidth"}
We also tried allowing filaments to branch directly at the leading edge. This is not necessarily physical as filament tips appear to nucleate new branching sites [@Pantaloni2000]. The results we found are not similar to current experiments as can be seen in Figure S5.
{width="0.6\linewidth"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Various experiments have found a boundary slip in hydrophobic microchannel flows, but a consistent understanding of the results is still lacking. While Molecular Dynamics (MD) simulations cannot reach the low shear rates and large system sizes of the experiments, it is often impossible to resolve the needed details with macroscopic approaches. We model the interaction between hydrophobic channel walls and a fluid by means of a multi-phase lattice Boltzmann model. Our mesoscopic approach overcomes the limitations of MD simulations and can reach the small flow velocities of known experiments. We reproduce results from experiments at small Knudsen numbers and other simulations, namely an increase of slip with increasing liquid-solid interactions, the slip being independent of the flow velocity, and a decreasing slip with increasing bulk pressure. Within our model we develop a semi-analytic approximation of the dependence of the slip on the pressure.'
author:
- 'Jens Harting, Christian Kunert, and Hans J. Herrmann'
title: Lattice Boltzmann simulations of apparent slip in hydrophobic microchannels
---
During the last century it was widely assumed that the velocity of a Newtonian liquid at a surface is always identical to the velocity of the surface. However, in recent years well controlled experiments have shown a violation of the no-slip boundary condition in sub-micron sized geometries. Since then, experimental [@bib:lauga-brenner-stone] and theoretical works [@bib:theory], as well as computer simulations [@bib:lbsims; @bib:barrat-bocquet-99; @bib:koplik-banavar-willemsen-89; @koplik-banavar-98; @bib:cieplak-koplik-banavar-01; @bib:thompson-robbins-1990; @bib:troian97] have tried to improve our understanding of boundary slip. The complex behavior of a fluid close to a solid interface involves the interplay of many physical and chemical properties. These include the wettability of the solid, shear rate, pressure, surface charge, surface roughness, as well as impurities and dissolved gas. Since all those quantities have to be determined very precisely, it is not surprising that our understanding of the phenomenon is still unsatisfactory. Due to the large number of different parameters, a significant dispersion of the results can be observed for ostensibly similar systems [@bib:lauga-brenner-stone], e.g. observed slip lengths vary between nanometres [@bib:churaev-sobolev-somov-84] and micrometers [@bib:tretheway-meinhart-0204] and while some authors find a dependence of the slip on the flow velocity [@bib:varvel], others do not [@cheng-giordano-02; @bib:tretheway-meinhart-0204]. Most computer simulations apply Molecular Dynamics (MD) and report increasing slip with decreasing liquid density [@bib:koplik-banavar-willemsen-89; @bib:thompson-robbins-1990] or liquid-solid interactions [@bib:cieplak-koplik-banavar-01; @bib:nagayama-cheng-2004], while slip decreases with increasing pressure [@bib:barrat-bocquet-99]. These simulations are usually limited to some tens of thousands of particles, lengths scales of nanometres and timescales of nanoseconds. Also, shear rates are orders of magnitude higher than in any experiment [@bib:lauga-brenner-stone]. We overcome these limitations using the lattice Boltzmann (LB) algorithm – a powerful method for simulating fluid dynamics [@bib:succi]. Rather than tracking individual atoms and molecules, the dynamics of the single-particle distribution function $\eta$ of mesoscopic fluid packets is described. In contrast to MD simulations, this method is less computationally demanding and allows to simulate experimentally accessible length and time scales. Our ansatz differs from other LB approaches where slip is introduced by generalizing no-slip bounce back boundary conditions to allow specular reflections with a given probability [@bib:lbsims] or where the viscosity is modified due to local density variations [@bib:nie-doolen-chen]. In both cases, parameters determining the properties at the boundaries are not easily mappable to experimentally available values. Our approach is based on Shan and Chen’s multi-phase LB model [@bib:shan-chen]. Here, interactions between different species are modelled by mesoscopic forces between the phases. This naturally opens the way to introduce similar interactions between each fluid species and the channel walls, where the strength of the interaction is determined by the fluid densities, free coupling constants, and a wall interaction parameter which is treated in a similar manner as a local fluid density. The model allows the simulation of multi-phase flows along hydrophobic boundaries and is introduced in the following. However, in order to study the influence of hydrophobicity on the boundary slip and to demonstrate the basic properties of the model, we focus on single phase flow in this paper. Results of multi-phase simulations will be presented in a future work. A multi-phase LB system can be represented by a set of equations [@bib:mplb] $$\label{LBeqs}
\begin{array}{cc}
\eta_i^{\alpha}({\bf x}+{\bf c}_i, t+1) - \eta_i^{\alpha}({\bf x},t) =
\Omega_i^{\alpha}, & i= 0,1,\dots,b\mbox{ ,}
\end{array}$$ where $\eta_i^{\alpha}({\bf x},t)$ is the single-particle distribution function, indicating the amount of species $\alpha$ with velocity ${\bf c}_i$, at site ${\bf x}$ on a D-dimensional lattice of coordination number $b$ (D3Q19 in our implementation), at time-step $t$. For the collision operator $\Omega_i^{\alpha}$ we choose the Bhatnagar-Gross-Krook (BGK) form$$\label{Omega}
\Omega_i^{\alpha} =
% -\frac{1}{\tau^{\alpha}}(\eta_i^{\alpha}({\bf x},t) - \eta_i^{\alpha
%\, eq}({\bf x},t))\mbox{ ,}
-\frac{1}{\tau^{\alpha}}(\eta_i^{\alpha}({\bf x},t) - \eta_i^{\alpha
\, eq}({\bf u}^{\alpha}({\bf x},t),\eta^{\alpha}({\bf x},t)))\mbox{ ,}$$ where $\tau^{\alpha}$ is the mean collision time for component $\alpha$ and determines the fluid viscosity. The system relaxes to an equilibrium distribution $\eta_i^{\alpha\,eq}$ which can be derived imposing restriction on the microscopic processes, such as explicit mass and momentum conservation for each species [@bib:lbneq]. $\eta^{\alpha}({\bf x},t)\equiv\sum_i \eta_i^{\alpha}({\bf x},t)$ is the fluid density and ${\bf u}^{\alpha}({\bf x},t)$ is the macroscopic velocity of the fluid, defined as $\eta^{\alpha}({\bf x},t){\bf u}^{\alpha}({\bf x},t)
\equiv \sum_i \eta_i^{\alpha}({\bf x},t){\bf c}_i$. Interactions between different fluid species are introduced as a mean field body force between nearest neighbors [@bib:shan-chen]: $$\label{Eq:force}
{\bf F}^{\alpha \bar{\alpha}}({\bf x},t) \equiv -\psi^{\alpha}({\bf x},t)\sum_{\bf
\bar{\alpha}}g_{\alpha \bar{\alpha}}\sum_{\bf
x^{\prime}}\psi^{\bar{\alpha}}({\bf x^{\prime}},t)({\bf x^{\prime}}-{\bf
x})\mbox{ ,}$$ where $\psi^{\alpha}({\bf x},t)=(1 - e^{-\eta^{\alpha}({\bf
x},t)/\eta_0})$ is the so-called effective mass with $\eta_0$ being a reference density that is set to 1 in our case [@bib:shan-chen]. $g_{\bar{\alpha}\alpha}$ is a force coupling constant, whose magnitude controls the strength of the interaction between component $\alpha$ and $\bar{\alpha}$. The dynamical effect of the force is realized in the BGK collision operator in Eq. (\[Omega\]) by adding to the velocity ${\bf u}$ in the equilibrium distribution an increment $\delta{\bf u}^{\alpha} = {\tau^{\alpha}{\bf
F}^{\alpha \bar{\alpha}}}/{\eta^{\alpha}}$. For the interaction of the fluid components with the channel walls we apply mid-grid bounce back boundary conditions [@bib:succi] and assign interaction properties to the wall which are similar to those of an additional fluid species. I.e., we specify constant values for the force coupling constant $g_{\bar{\alpha}\alpha}=g_{wall,\alpha}$ and the density $\eta^{\bar{\alpha}}=\eta^{wall}$ at wall boundary nodes of the lattice. This results in a purely local force as given in Eq. \[Eq:force\] between the flow and the boundaries. Even though one could argue that a single parameter to tune the fluid-wall interaction would be sufficient, we keep our approach as close as possible to the original idea of Shan and Chen in order to benefit from the experience obtained from other works using the original model. Furthermore, the additional parameter allows more flexibility when simulating not only a single fluid, but a multi-phase system. The fluid-wall interaction can be linked to a contact angle between fluid droplets and solid walls as it is often used to quantitatively describe hydrophobic interactions [@bib:contact-angle]. Recently, Benzi et al. have shown how to compute the contact angle within the Shan-Chen model [@bib:benzi-06]. The same authors also developed an approach to model apparent slip which is related to ours, but instead of using only local fluid-solid interactions, they add an exponential decay of the interaction with distance from the wall [@benzi-biferale-05]. We simulate pressure driven flow between two infinite planes (Poiseuille flow), where pressure driven boundary conditions are implemented in a similar way as in most experiments: a fixed pressure is set at the channel inlet and an open boundary at the outlet. The outlet is realized by interpolating the particle distribution function at the end of the channel as given by $ \eta_i^{\alpha}({\bf x},t)=2\eta_i^{\alpha}({\bf
x-1},t)-\eta_i^{\alpha}({\bf x-2},t)$ leading to a linear pressure gradient. Already in 1823, Navier proposed a boundary condition where the fluid velocity at a surface is proportional to the shear rate at the surface, i.e. $v_z(x_0)=\beta\partial v_z(x)/\partial x$ at $x=x_0$ [@bib:Navier]. Following his hypothesis the velocity in flow direction ($v_z$) at position $x$ between the planes is given by $$v_z(x)=\frac{1}{2 \mu}\frac{\partial P}{\partial z}
\left[ h^2-x^2-2h\beta \right],
\label{eq:plattenprofil}$$ where $2h$ is the distance between the planes, and $\mu$ the viscosity. In contrast to a no-slip formulation, the last term in Eq. \[eq:plattenprofil\] linearly depends on the slip length $\beta$. Since $\beta$ is typically of the order of nanometers or micrometers, it can be neglected in macroscopic experiments. In order to obtain $\beta$ from our data, we measure the pressure gradient ${\partial P}/{\partial z}$ at the center of the channel and the velocity profile between the two planes at a fixed position $z$. $\beta$ is then obtained by a least square fit with Eq. \[eq:plattenprofil\].
Our simulation parameters are as follows: the lattice size is kept constant with the channel length ($z$ direction) being 256 sites, the distance between the plates $2h$ being 60 sites ($x$ direction). We approximate infinite planes by using a 16 sites wide channel with periodic boundaries in $y$ direction. In order to assure a fully equilibrated system we simulate for at least 40000 time steps before measuring and assured our results being independent of the discretization level by comparing to simulations of 28 and 124 sites wide channels. Each data point in the figures below corresponds to about six hours simulation time on eight IBM Power 4 1.7GHz CPUs. All units in this paper are in lattice units with the lattice constant $c$ and timestep $\Delta t$ set to 1 if not stated otherwise.
The dependence of the slip length $\beta$ on the interaction parameter $g_{wall,\alpha}$ is studied for $\eta^{wall}$=1.0 and 5.0. The bulk pressure $P=\rho c_s^2$, where $\rho$ is the fluid density and $c_s=1/\sqrt{3}$ the speed of sound, is kept at $P$=0.11, while the flow velocity is set to $V$=0.033. As shown in Fig. \[fig:slipvsgbr\]a we vary $g_{wall,\alpha}$ from 0.06 to 0.22 and find a steady increase of $\beta$ for increasing $g_{wall,\alpha}$. As expected, the curve for $\eta^{wall}$=5.0 is growing substantially faster than for $\eta^{wall}$=1.0. The maximum available $\beta$ are at about 5.2 for $g_{wall,\alpha}$=0.26 and $\eta^{wall}$=1.0. At these strong fluid-wall interactions, the force as given in Eq. \[Eq:force\] becomes very large and results in a large area of low fluid density close to the wall. Increasing the interaction even further results in numerical instabilities due to too steep density gradients. In order to study the dependence of the slip on other parameters, the coupling constant $g_{wall,\alpha}$ is kept constant at 0.08 from now on. Fig.\[fig:slipvsgbr\]b depicts the dependence of $\beta$ on $\eta^{wall}$ for different bulk pressures $P$=0.033, 0.1, and 0.3 and fixed flow velocity $V=3.5\cdot 10^{-3}$ in the system. While all three graphs grow constantly with increasing $\eta^{wall}$, the one for $P$=0.033 grows the fastest demonstrating that absolute values for $\beta$ are higher for lower pressure.
We have measured the magnitude of the boundary slip over a very wide range of flow velocities $V$ from $1\cdot 10^{-4}$ to $3\cdot 10^{-2}$ for wall interactions $\eta^{wall}$=0.0, 0.5, 1.0, and 2.0. $V$ is measured at the center of the channel and given on a logarithmic scale in Fig. \[fig:betavsv\]. For $\eta^{wall}$=0.0 we do not find any boundary slip confirming that our method properly reproduces no slip behavior in the interaction free case. With increasing wall interactions, we achieve an increase of the magnitude of $\beta$ to up to $\simeq$1.1 for $\eta^{wall}$=2.0. We are not able to find any velocity dependence of $\beta$, but find constant slip for fixed fluid-wall interactions, which is consistent with many experiments [@cheng-giordano-02; @cheikh-koper-03]. The fluctuations of the data for very low flow velocities are due to numerical uncertainties of the fit at very low curvature of the parabolic velocity profile. For $V>0.01$ we find a slight deviation of $\beta$ from the constant measurements. This is due to a small variation of the bulk pressure from $P$=0.097 for $V=1\cdot 10^{-4}$ to $P$=0.106 for $V=0.03$ that cannot easily be avoided for technical reasons. We have checked for a few data points that $\beta$ stays constant if $P$ can be kept at exactly fixed values, too. The slip length being independent of the flow velocity is consistent with many experiments and computer simulations, like the MD simulations of Cottin-Bizonne et al. [@bib:cottin-bizone] and the experiments of Cheng et al. [@cheng-giordano-02] and Baudry et al. [@baudry-charlaix-01]. We speculate that an increase of $\beta$ with increasing flow velocity as measured by some experiments [@bib:varvel] is due to surface roughness of the channel boundaries or other nonlinear effects. Since our model is not able to treat roughness on an atomic scale, we do not expect to conform with those results. MD simulations which find a non-constant value for $\beta$ operate at very high shear rates which are orders of magnitude higher than what can be obtained by our approach [@bib:troian97].
Computing the exact slip in dependence of the interaction parameters from first principle analytically is a very hard or even impossible task since our interaction as given in Eq. \[Eq:force\] modifies the equilibrium distribution in the BGK operator. Therefore, we present a semi-analytic approximation which utilizes the common two-layer model. Here, it is assumed that a thin fluid layer with thickness $\delta$ and different viscosity as the bulk fluid exists near the channel walls. As calculated by various authors [@bib:theory], within this model the slip length can be computed as $\beta=(\mu_{bulk}/\mu_1-1)\delta$, where $\mu_{bulk}$ is the viscosity of the bulk fluid, and $\mu_1$ the viscosity close to the wall. Since the dynamic viscosity is given by the kinematic viscosity times the fluid density, $\mu=\rho\nu=\rho(2\tau^\alpha -1)/6$ [@bib:succi], we write $\beta=(\rho_{bulk}/\rho_1-1)\delta$. $\rho_{bulk}$ can be measured in the channel’s center and $\rho_1$ at the first lattice site next to the wall. Fig. \[fig:rhowall\] shows the dependence of $\rho_1$ on $g_{wall,\alpha}$ for $\eta^{wall}$=1.0, 5.0, $P$=0.11, and $V$=0.033. Since $\rho_1$ cannot easily be computed analytically, we postulate an interaction term that depends on the bulk density and the fluid-wall interaction as well as a free fit parameter $k$, $${\cal I}=k{\bf F}^{wall,\alpha}({\bf x},t)/\rho_{bulk}({\bf x},t)$$ and fit $\rho_1$ with an exponential function $\rho_1=\rho_{bulk}({\bf x},t)\exp(-{\cal I})$. With only a single value for $k$ we are able to utilize this equation to fit $\rho_1$ for all our simulation parameters. $k$ is found to be 8.35 for our data. The lines in Fig. \[fig:rhowall\] illustrate the good quality of our approximation. A similar approach is applied to model the thickness of the layer at the wall which strongly depends on the fluid-wall interaction and bulk density. Here, we set $\delta=\exp({\cal
I})$. As a result, $\beta$ can be estimated by $\beta=(\exp({\cal I})-1)\exp({\cal I})$.
The semi-analytic approximation is used to fit the dependence of the slip length $\beta$ on the bulk pressure $P$. Fig. \[fig:betavsp\] shows the simulation data (symbols) and the approximation (lines) for wall interactions $\eta^{wall}$=0.5, 1.0, and 2.0. The bulk pressure is varied from 0.03 to 0.33. We find a decrease of $\beta$ with increasing pressure $P$. An increase of $\eta^{wall}$ leads to an increasing slope of the curves and to higher absolute values for $\beta$. Furthermore, we find a decrease of the slip with increasing bulk pressure. These results qualitatively agree with MD simulations [@bib:barrat-bocquet-99; @bib:cieplak-koplik-banavar-01]. Even with a single value for the fit parameter $k$, the semi-analytic description of $\beta$ agrees very well for low fluid-wall interactions. For strong interactions ($\eta^{wall}$=2.0), the fit qualitatively reproduces the behavior of the slip length. Higher order terms in the exponential ansatz for $\delta$ are needed for a better agreement.
To demonstrate that our approach is able to achieve experimentally available length and time scales, we scale our simulations to the experimental setup of Tretheway and Meinhart [@bib:tretheway-meinhart-0204]. They use a 30$\mu$m high and 300$\mu$m wide microchannel with typical flow velocities of $V=10^{-2}$mm/s. For water, they measure a slip length of 0.92$\mu$m. The Reynolds number $Re=2hV/\nu$ in their experiment is $\simeq$0.3. To reproduce the observed slip, we set $g_{wall,\alpha}$=0.16 and $\eta^{wall}=1.0$ (see Fig. \[fig:slipvsgbr\]a). We are able to cover a wide range of flow velocities, i.e. for the setup given above, velocities can range from as low as $1\times 10^{-4}$ to as high as 0.05 corresponding to $Re$ between 0.038 and 19. The Knudsen number is given by $Kn=\nu/(c_s 2h)$ which corresponds to $4.8\times 10^{-3}$ for the simulations presented here. At these low $Kn$, the hydrodynamic approach is well valid. However, it has been shown that the LB method can be applied for $Kn$ much larger than 1 if one uses modified boundary conditions [@toschi-succi-05]. Our mesoscopic force is expected to be able to properly describe fluid-wall interactions in such systems as well.
In conclusion, we presented a new approach to investigate boundary slip in hydrophobic microchannels by means of a multi-phase LB model. In contrast to MD simulations, our model is able to reach the length and time scales of typical experiments and is applicable for a wide range of realistic flow velocities. We qualitatively reproduced the dependence of slip on the hydrophobicity of the channel walls and found constant slip for varying flow velocities. The decrease of the slip with increasing pressure can be approximated by a semi-analytic approach. Our results are consistent with MD simulations [@bib:cottin-bizone; @bib:barrat-bocquet-99; @bib:cieplak-koplik-banavar-01] and experiments [@bib:tretheway-meinhart-0204].
We would like to thank G. Giupponi, M. Hecht, N. González-Segredo, and V.S.J. Craig for fruitful discussions and acknowledge the Neumann Institute for Computing for providing access to their IBM p690 system.
[40]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , ** (, ), chap. .
, ****, (). , ****, ().
, ****, (). , , , , in ** ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, (). , ****, ().
, ****, ().
, , , ****, (). , , , ****, ().
, ****, ().
, ****, ().
, ** (, ).
, , , ****, ().
, ****, (). , ****, ().
, , , ****, (). , ****, ().
, , , , ****, (). , , , ****, (). , ****, (). , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, , , , , , ****, (). , , , , , ****, ().
, ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Motion of test particles along rotating curved trajectories is considered. The problem is studied both in the laboratory and the rotating frames of reference. It is assumed that the system rotates with the constant angular velocity $\omega = const$. The solutions are found and analyzed for the case when the form of the trajectory is given by an Archimedes spiral. It is found that particles can reach infinity while they move along these trajectories and the physical interpretation of their behaviour is given. The analogy of this idealized study with the motion of particles along the curved rotating magnetic field lines in the pulsar magnetosphere is pointed out. We discuss further physical development (the conserved total energy case, when $\omega \ne
const$) and astrophysical applications (the acceleration of particles in active galactic nuclei) of this theory.
address:
- 'Dipartimento di Fisica Generale, Universitá degli Studi di Torino, Via Pietro Giuria 1, I-10125 Torino; and Abdus Salam International Centre for Theoretical Physics, I-34014 Trieste, Italy'
- 'Department of Physics, Tbilisi State University, Chavchavadze ave. 2, Tbilisi 380028, Georgia'
- 'Centre for Plasma Astrophysics, Abastumani Astrophysical Observatory, Kazbegi str. $2â$, Tbilisi 380060, Georgia'
author:
- Andria Rogava
- George Dalakishvili
- Zaza Osmanov
title: Centrifugally driven relativistic dynamics on curved trajectories
---
Introduction
============
Rotation and relativity are those two features of motion, which do [*not*]{} easily match with each other. Still in astrophysics, with its abundance of extremely strong electromagnetic and gravitational fields, there are situations where motion [*is*]{} both rotational and relativistic. Most prominent examples include swirling astrophysical jets in active galactic nuclei (AGNs) and quasars, innermost regions of black hole accretion disks, accretion columns in X-ray pulsars and plasma outflows in radio pulsar magnetospheres. In these [*kinematically complex astrophysical flows*]{}, where rotation is interlaced with the relativistic motion of particles, the coexistence of these two features of the motion leads to observationally puzzling phenomena with sophisticated and ill-understood physical background. The interest to these flows is not new, but the upgrade of highly idealized models to more realistic, astrophysically relevant levels is still related with major theoretical and computational difficulties.
Some important and basic theoretical issues, related with the relativistic rotation, are not uniquely defined and often evoke controversial interpretations. One of the most notable examples is the [*“centrifugal force reversal"*]{} effect, originally found in \[1\], and later \[2-4\] studied in detail. It was argued that under certain conditions the centrifugal force [*attracts*]{} towards the rotation axis both for Schwarzschild and Kerr black holes. In Ernst spacetime, which represents the gravitational field of a mass embedded in a magnetic field, the centrifugal force acting on a particle in circular orbit was reported \[5\] to reverse its sign even twice! In the simplest case of the Schwarzschild spacetime, strictly and essentially speaking, it was found that below the radius of the spatially circular photon orbit an [*increase*]{} of the angular velocity of a test particle causes more [*attraction*]{} rather than additional centrifugal [*repulsion*]{}. This effect was interpreted by Abramowicz \[4\] in terms of the centrifugal force reversal - it was stated that in such cases the centrifugal force [*attracts*]{} towards the axis of rotation!
This interpretation was criticized by de Felice \[6,7\] (see also \[8\]), who argued that the discovered effect could be attributed to the strength of the gravitational field and be explained in a way which preserves the repulsive character of the centrifugal force. The spirit of this approach — to save the intuitively appealing nature of the centrifugal force as of [*“something which pushes things away"*]{} \[7\] — is theoretically valid and practically convenient. After all, in general relativity, there is no implicit way to define the centrifugal force: in any case one needs to introduce some sort of “3+1" spacetime splitting and dub as the “centrifugal force" some Newtonian-like expression, which looks like it \[7\]. Moreover, de Felice found several interesting examples of the ambiguity of the global concept of ’outwards´ and pointed out at the deep interrelation of this problem with the definition of the centrifugal force in relativity. Abramowicz studied further the problem of the local and the global meaning of ’inwards´ and ’outwards´ \[9\] and showed that the centrifugal force always repels outwards in the [*local*]{} sense, while it may attract inwards, towards the centre of the circular motion, in the [*global*]{} sense! The theoretical scheme for the operationally unambiguous definition of the inward direction was suggested by de Felice and Usseglio-Tomasset \[10\] and later this approach was used for the geometrical definition of the generalized centrifugal force \[11\].
Therefore the effect, discovered in \[1\], is indubitably a genuine relativistic effect, although its [*interpretation*]{} in terms of the “reversal" of the centrifugal force is not implicit and is largely the matter of definition.
Same is true for another rotational effect, disclosed by Machabeli and Rogava \[12\], on the basis of the relatively simple and idealized special- relativistic [*gedanken experiment*]{}: motion of a bead within a rigidly and uniformly rotating massless linear pipe. It was shown that even if the starting velocity of the bead is nonrelativistic, after an initial phase of usual centrifugal acceleration, while the bead acquires high enough relativistic velocity, it starts to [*decelerate*]{} and after reaching the light cylinder changes the character of its motion from centrifugal to centripetal. It was found that when the initial velocity $v \ge \sqrt{2}/2$ the motion of the bead is decelerative all the way from the pivot to the light cylinder.
Certainly no real pipe may stay absolutely rigid, especially nearby the light cylinder. Besides, in order to maintain the uniform rotation of such a device, one needs an infinite amount of energy. Therefore the setup considered in \[12\] was [*highly idealized*]{}. The constant rotation rate assumption was replaced in \[13\] with a more realistic one: the total energy of the system “rotator+pipe+bead" was assumed to be constant. It was found that the moving bead acquires energy from the slowing down rotator, but under favorable conditions the bead deceleration still happens.
The results of \[12\] were interpreted by its authors in terms of the centrifugal force reversal. De Felice disputed this interpretation \[14\] and argued that, also in this case, like in above-mentioned general-relativistic examples, the generalized definition of the centrifugal force may guarantee the absolutely repulsive character of the force. He pointed out that an inertial observer will never [*see*]{} the bead reaching the light cylinder, because all light signals from the bead are infinitely redshifted. It was also shown that the vanishing of the radial velocity of the bead at the light cylinder can be interpreted in terms of the corresponding vanishing of the bead’s proper time.
Despite the controversy of interpretations it is generally believed that rotational relativistic effects could operate in different astrophysical situations and might, hopefully, lead to detectable observational appearances. Recently, Heyl \[15\] suggested that the observed QPO frequency shifts in bursters are caused by a geometrical effect of the strong gravity, similar to the Abramowicz-Lasota centrifugal force “reversal"[^1]. As regards Machabeli-Rogava [*gedanken*]{} experiment, it implies that radially constrained, relativistic and rotationally (centrifugally) driven motion shows inevitable radial deceleration near the light cylinder. Evidently this effect might occur in a number of astrophysical situations, where motion is constrained, rotational and strongly (special) relativistic.
One of the most important class of astrophysical flows, where this effect could show up, is [*centrifugally driven outflows*]{}. In the context of pulsar emission theory they were first considered in the late 1960s by Gold \[17,18\] (for recent studies see e.g. Ref.19 and 20). For accreting black holes, both of galactic and extragalactic origin, Blandford and Payne \[21\] first noted that centrifugally driven outflows from accretion disks could be responsible for the launch of jets, if the poloidal field direction is inclined at an angle less than $60^\circ$ to the radial direction[^2].
Recently Gangadhara and Lesch \[23\] suggested that centrifugal acceleration, taking place as a consequence of the bead-on-the-wire motion similar to the Machabeli-Rogava \[12\] [*gedanken*]{} experiment, may account for the acceleration of particles to very high energies by the centrifugal forces while they move along rotating magnetic filed lines of the rotating AGN magnetosphere. They claimed that the highly nonthermal, X-ray and $\gamma$-ray emission in AGNs arises via the Comptonization (inverse-Compton scattering) of ultraviolet photons by centrifugally accelerated electrons. The same processes was critically re-examined by Rieger and Mannheim \[24\] and it was found that the rotational energy gain of charged particles is efficient but substantially limited not only by the Comptonization but also by the effects of the relativistic Coriolis forces. The specific nature of the propagation of electromagnetic radiation in the rotating frame of reference \[14,25\] is another aspect of this problem, which still needs to be taken into account.
The whole philosophy of the ‘pipe-bead’ (ot ‘rotator-pipe-bead’) [*gedanken*]{} experiments was to mimic the common situation in relativistic and rotating astrophysical flows, where the plasma particles are doomed to move along the field lines of governing magnetic fields. While we consider relatively small length scales, the shape of the field lines can be approximated as being straight. However, on larger length scales the curvature of the field lines turns out to be important for the physics of the plasma streams, which are guided by them. The natural question arises: how the motion of the bead changes when the pipe is curved? In other words, how the dynamics of particles, prescribed to move along the fixed trajectories, change when the shape of their involuntary tracks of motion is [*not*]{} straight!? Obviously, this is not only a mere theoretical curiosity, but the issue which might have a tangible practical importance. In astrophysical situations the role of the “pipes" is played by the magnetic field lines, and the latter are [*always*]{} curved. Therefore, it is clear that the study of the motion of test particles along prescribed curved rotating trajectories is a necessary and important step for the ultimate building of a physically meaningful model of centrifugally driven relativistic particle dynamics for rotating magnetospheres of pulsars and AGNs.
It is the purpose of this paper to address the above stated issue. In particular, in the next section, we develop special-relativistic theory of the motion of centrifugally driven particles on fixed nonstraight trajectories. The formalism is developed both for the laboratory frame (LF) and for the frame of reference rotating with the system (rotating frame, or RF). Equations of motion are derived and solved numerically. The detailed study is given only for the case when the angular velocity of the rotation is constant. However, we also outline the formalism for the astrophysically more realistic case of the conservative ‘rotator-pipe-bead’ system with perceptible exchange of energy between the bead and the rotator, leading to the variability of the angular velocity of the whole system. In the final section of the paper we discuss the results, consider the directions and aims of the future study, suggest and discuss those astrophysical situations, where the obtained results could be useful for the clarification of puzzling observational appearances of related astronomical objects.
Main consideration
==================
The ideal two-dimensional system, which we are going to consider, consists of three basic parts: the device of the mass $M$ and the moment of inertia ${\cal I}$, rotating with the angular velocity $\omega(t)$, hereafter referred as the [*rotator*]{}; the massless but absolutely rigid [*pipe*]{} steadily attached to the rotator; and the small [*bead*]{} of the mass $m$ and the radius equal to the internal cross-section radius of the pipe. The bead is put inside the pipe and can slide along the pipe without a friction. Evidently, instead of the pipe-bead dichotomy, since we are considering the two-dimensional layout, one may think about the ’wire-on-bead´ analogy, which is sometimes used \[24\].
Contrary to the \[12\], where a straight pipe case was studied, now we let the pipe to be an arbitrarily flat curve, mathematically defined by: $$\varphi\equiv
\varphi(r), \eqno(1a)$$ with $[d\varphi/dr \equiv{\varphi}'(r)]$: $$\Phi \equiv r{\varphi}'(r). \eqno(1b)$$
The dynamics of the system may be studied basing on two alternative assumptions:
1. It makes the task simpler to suppose that the kinetic energy of the rotator $E_M$ is huge and always $E_M \gg E_m$; i.e., despite the exchange of the energy with the moving bead, $E_M$ stays practically constant. Hence, the angular velocity of the whole “rotator-pipe-bead" system (henceforth referred as the RPB system) stays constant: $$\omega = const. \eqno(2a)$$
In this case the first part of the triple RPB system (the rotator) continuously supplies the bead with energy and helps to keep the angular velocity of rotation constant. Therefore, the problem reduces to the study of the pipe-bead double system (the PB system) with the [*constant*]{} rotation rate. In the case of the [*straight*]{} pipe ($\varphi=\varphi_0$) the problem has exact analytic solution, found and analyzed in \[12\].
2. It is more realistic to assume that the rotator energy $E_M$ is finite, so that the whole RPB system is conservative $E_{tot}\equiv E_M+E_m=const$. There is a perceptible energy exchange between the rotator and the bead: both the energy $E_M$ and the angular momentum $L_M$ of the rotator [*are*]{} variable and, consequently, the angular velocity of the rotation can [*not*]{} stay constant: $$\omega \ne const. \eqno(2b)$$ The problem with the straight pipe and variable rotation rate $\omega(t)$ has no analytic solution. It was studied numerically in \[13\].
With either (2a) or (2b) assumptions the pipe is always assumed to be the passive part of the system. In order to mimic a magnetic field line it is assumed to be massless, having no share in the energy and/or momentum balance of the whole system. Still the role of the pipe — as the dynamic link between the rotator and the bead — is significant: it provides the prescribed “guiding" of the bead motion in the rotating frame of reference and makes the trajectory of the bead known in advance.
In this paper the dynamics of the [*gedanken*]{} system is studied in detail only under the first, easier, assumption of the constant angular velocity. The rout to the solution of the problem under the second assumption is also given, but its full study needs separate consideration and will be published elsewhere.
There are two natural frames of reference, in which the dynamics of this system could be studied. The first, inertial one, is the laboratory frame (LF), where the observer measures the angular velocity of the rotator (and the pipe) to be $\omega(t)$, while the angular velocity of the bead is equal to: $$\Omega(t)=\omega(t)+{\varphi}'(r)v(t), \eqno(3)$$ and the dynamics of the moving bead is governed by the pipe reaction force acting on it. Note that $v(t) \equiv dr/dt$ is the radial velocity of the bead relative to the LF.
The second frame, rigidly attached to the rotator and rotating with it (hereafter referred as the rotating frame, or the RF), is non-inertial, but quite convenient for the inspection of the motion of the bead along the curved pipe. This original approach implies embodying of the form of the pipe into the metric of the rotating frame. It was used in \[12\] for the straight pipe case and proved to be quite efficient for the case when $\omega(t)$ is assumed to be a constant. On the contrary, as we shall see later, the LF treatment appears to be handier when the second (2b) approximation ($E_{tot}=const$ and $L_{tot}=const$, rather than $\omega(t)=const$) is chosen. That is why it is important to consider the problem both in the LF and in the RF.
Uniformly rotating PB system
----------------------------
First, let us consider the problem in the laboratory frame of reference (LF) and ascertain that it admits full (numerical) solution of the associated initial value problem. Second, let us consider the same problem in the rotating frame of reference (RF). We shall see that when the (2a) assumption of the constancy of the rotation rate is used the latter approach is mathematically easier and provides fuller information about the dynamics of the system.
### LF treatment
The most straightforward way to approach the problem is to consider it in the laboratory frame of reference, in which the spacetime is Minkowskian: $$ds^2 =- dT^2 + dX^2 + dY^2=- dT^2 + dr^2 + r^2d\phi^2. \eqno(4)$$
We use geometrical units, in which $G=c=1$. Note that azimuthal angle $\phi$, as measured in the LF, is related with the azimuthal angle $\varphi$, measured in the RF, via the obvious expression: $\phi = \varphi + \omega t$. The pipe reaction force ${\bf F}$ is the dynamic factor constraining the bead to move along the pipe. It is easy to see (from the 4-velocity normalization $g_{\alpha\beta}U^{\alpha}U^{\beta}=-1$) that the Lorentz factor of the moving bead is: $$\gamma(t)=[1-r^2\Omega^2-v^2]^{-1/2}. \eqno(5)$$
The angle between the radius-vector of a point of the pipe and the tangent to the same point is given by the relation: $$\alpha = \arctan \Phi, \eqno(6)$$ and the components of the reaction force, acting in the radial and azimuthal directions, are $$F_r = -|F|\sin{\alpha} = -{{\Phi}\over{\sqrt{1+\Phi^2}}}|{\bf F}|,
\eqno(7a)$$ $$F_{\phi} = |F|\cos{\alpha} = {{1}\over{\sqrt{1+\Phi^2}}}|{\bf F}|,
\eqno(7b)$$ respectively.
Defining the physical components of the bead relativistic momentum $[m(t) \equiv m_0\gamma(t)]$: $$P_r \equiv mv, \eqno(8a)$$ $$P_{\phi} \equiv mr{\Omega}, \eqno(8b)$$ we can write the two components of the equation of motion in the following way: $$\dot{P}_r - {\Omega}P_{\phi}=F_r, \eqno(9a)$$ $$\dot{P}_{\phi} + {\Omega}P_r=F_{\phi}. \eqno(9b)$$
Combining these equations we can, first, derive the equation: $$\dot{P}_r+{\Phi}\dot{P}_{\phi}+{\Omega}
({\Phi}P_r-P_{\phi})=0. \eqno(10a)$$ It is easy to calculate that: $$\dot{\Omega}={\varphi}'\dot{v}+{\varphi}''v^2, \eqno(10b)$$ $$\dot{m}=m{\gamma}^2[(\Omega+r\varphi''v)r{\Omega}v+(v+r^2\varphi'
\Omega)\dot{v}]; \eqno(10c)$$ and using these relations together with (8) we can easily derive the explicit equation for the radial acceleration of the bead: $$\ddot{r}={{r\omega\Omega-{\gamma}^2rv(\varphi'+\omega v)
(\Omega+r\varphi''v)}\over{\gamma^2\Delta^2}}, \eqno(11)$$ where $$\Delta \equiv [1-{\omega}^2r^2+\Phi^2]^{1/2}. \eqno(12)$$
The Eq. (11) being of the form $\ddot{r}=G(\dot{r},r)$ admits full numerical solution, as the standard initial value problem, providing the initial position of the bead, $r_0$, its initial velocity, $v_0$, and the shape of the pipe, $\varphi(r)$, are specified.
Defining the spatial vector of the 2-velocity ${\bf v} \equiv
(v, r \Omega)$, we can calculate the absolute value of the reaction force $|{\bf F}|$ using the equation \[26\]: $$\dot{m}={\bf F}\cdot{\bf v}, \eqno(13a)$$ which, in our case, leads to: $$\sqrt{g_{rr}}\dot{m}=r \omega |F|. \eqno(13b)$$
It is also easy to verify that the following quantity: $$\Psi \equiv m(t) - \omega r P_{\varphi} = m_0\gamma(1-r^2\omega
\Omega)=const(t), \eqno(14)$$ is the [*constant*]{} in time. This allows to find the solutions of the problem as functions of the specific value of this constant. In the next subsection we will see what is the physical meaning of this parameter - it turns out to be proportional to the proper energy of the moving bead in the RF.
One important class of a possible shape of the curved trajectory is [*Archimedes spiral*]{}, given by the formula: $$\varphi(r)=ar,~~~~a=const. \eqno(15)$$ In this case, since $\varphi''=0$, from (11) it is easy to see that $\ddot{r} \sim
\Omega$, while (14) implies that $|{\bf F}| \sim \Omega$ as well. Therefore, in the case of the Archimedes spiral trajectory, we can predict that the asymptotic behavior of the functions $\Omega(t)$, $v(t)$, $\dot{v}(t)$, and $|{\bf F}(t)|$ will be similar.
### RF treatment
We see the LF treatment allows to solve the problem and to obtain the complete information about the dynamics of the bead motion along the fixed nonstraight (curved) trajectories. However it is quite instructive and much more convenient to consider the same problem in the frame of reference, rotating with the pipe-bead system (rotating frame - RF). In order to do this we, first, need to switch from (4) to the frame, rotating with the angular velocity $\omega$. Employing the transformation of variables: $$T=t, \eqno(16a)
µ$$ $$X=rcos \phi = rcos(\varphi + \omega t), \eqno(16b)$$ $$Y=rsin \phi = rsin(\varphi + \omega t), \eqno(16c)$$ we arrive to the metric: $$ds^2 \equiv = -(1- \omega^2r^2)dt^2 +
2\omega^2dtd\varphi + r^2d\varphi^2 + dr^2. \eqno(17)$$
For the straight pipe ($\varphi=\varphi_0$) case (17) reduces to the metric $ds^2= -(1- \omega^2r^2)dt^2+ dr^2$, which was basic metric for the \[12\] study. Now, for a curved pipe, defined by the equation (1), (17) reduces to the following form: $$ds^2 = -(1- \omega^2r^2)dt^2 + 2\omega r\Phi dtdr +
(1 + \Phi^2)dr^2. \eqno(18)$$
For the resulting metric tensor $$\|g_{{\alpha}{\beta}}\|
\equiv
{\left(
\matrix{-(1-\omega^2r^2), & \omega r\Phi \cr
\omega r\Phi, & 1+\Phi^2 \cr}
\right)}, \eqno(19)$$ we can easily find out that $$\Delta \equiv [-\det(g_{{\alpha}{\beta}})]^{1/2}=(1-{\omega}^2r^2+
{\Phi}^2)^{1/2}, \eqno(20)$$ and, apparently it is the same function $\Delta$, defined previously by (13).
For this relatively simple, but nondiagonal, two-dimensional spacetime we can develop the $``1+1"$ formalism. Doing so we follow as a blueprint the well-known “3+1´´ formalism, widely used in the physics of black holes \[27-29\]. Namely, we introduce definitions of the [*lapse function*]{}: $$\alpha \equiv{{\Delta}\over{\sqrt{g_{rr}}}}=\sqrt{{{1-\omega^2r^2+\Phi^2}
\over{1+\Phi^2}}}, \eqno(21)$$ and the one-dimensional vector $\vec{\beta}$ with its only component: $$\beta^r\equiv {{g_{tr}}\over{g_{rr}}}={{{\omega}r{\Phi}}\over
{1+\Phi^2}}. \eqno(22)$$
Within this formalism (18) can be presented in the following way: $$ds^2=-\alpha^2dt^2+g_{rr}(dr+\beta^r dt)^2. \eqno(23)$$
Note that for the metric tensor (19) $t$ is the cyclic coordinate and, moreover, in the RF the motion of the bead inside the pipe is [*geodesic*]{} - there are no external forces acting on it. Hence the [*proper energy*]{} of the bead, $E$, must be a conserved quantity. Employing the definition of the four velocity $U^{\alpha}{\equiv}dx^{\alpha}/d{\tau}$ we can write: $$E \equiv- U_{t}= -U^t{\left[g_{tt} + g_{tr}v\right]} =
{\it const}. \eqno (24)$$
On the other hand, the basic four-velocity normalization condition $g_{{\alpha}{\beta}}U^{\alpha}U^{\beta}=-1$ requires $$U^t={\left[-g_{tt}-2g_{tr}v-g_{rr}v^2\right]}^{-1/2}, \eqno(25a)$$ this equation, written explicitly, has the following form: $$U^t={\left[1-\omega^2r^2-2\omega r \Phi v-(1+\Phi^2)v^2\right]}^
{-1/2}. \eqno(25b)$$ Recalling the expression (3) for the angular velocity of the bead $\Omega(t)$, measured in the LF, and the definition (5) of the Lorentz factor $\gamma(t)$ in the same frame of reference we can easily see that: $$U^t=[1-r^2\Omega^2-v^2]^{-1/2}=\gamma(t). \eqno(25c)$$
It is important to note that the conserved proper energy of the bead, $E$, defined by (24) may be written in terms of the $\Omega(t)$ and $\gamma(t)$ functions simply as: $$E=\gamma(t)[1-r^2(t)\omega\Omega(t)]=const. \eqno(26)$$ Taking time derivative of this relation and rearranging the terms we will finally arrive to exactly the same Eq. (11) for the radial acceleration of the bead $\ddot{r}$ as in the LF treatment. Note also that ${\Psi}=m_0 E$.
But the convenience of the RF treatment goes much further. From (24) and (25a) we can derive the explicit quadratic equation for the velocity: $$(g_{tr}^2+E^2g_{rr})v^2+2g_{tr}(g_{tt}+E^2)v+g_{tt}(g_{tt}+E^2)=0,
\eqno(27)$$ with the obvious solution: $$v=\dot{r}={{\sqrt{g_{tt}+E^2}}\over{(g_{tr}^2+E^2g_{rr})}}
{\left[-g_{tr}\sqrt{g_{tt}+E^2} \pm E \Delta \right]}. \eqno(28)$$
The $``1+1"$ formalism helps to write equivalents of the same equations in a more elegant form. Namely, if we define the radial velocity $$V^r \equiv {{1}\over{\alpha}}{\left(v+\beta^r\right)},
\eqno(29)$$ and corresponding Lorentz factor: $$\tilde{\gamma}
\equiv (1-V^2)^{-1/2}, \eqno(30)$$ then, instead of (25c), we will simply have: $$U^t = \tilde{\gamma}/\alpha, \eqno(31)$$ while from the (24) we obtain: $$E = \tilde{\gamma}[\alpha - (\vec{\beta}\cdot{\vec V})]. \eqno(32)$$ Instead of (27) we will have \[$V^2 \equiv
g_{rr}V^rV^r=V_rV^r$, $\beta^2 \equiv g_{rr}\beta^r\beta^r=\beta_r
\beta^r$\]: $$(\beta^2+E^2)V^2-2\alpha({\vec{\beta}}{\cdot}{\vec V})+(\alpha^2-
E^2)=0, \eqno(33)$$ with the solution: $$V^r={{1}\over{\beta^2+E^2}}{\left[\alpha\beta^r \pm E
\sqrt{{{E^2+\beta^2-\alpha^2}\over{g_{rr}}}}~\right]}. \eqno(34)$$
Note that the RF Lorentz factor defined by (30) and the LF Lorentz factor, specified by (5) do [*not*]{} equal each other $$\tilde{\gamma}(t) \ne \gamma(t), \eqno(35)$$ which is the manifestation of the obvious fact that Lorentz factor is not an invariant physical quantity. One can see that for the $\tilde{\gamma}(t)$ the following quadratic equation holds: $$(\alpha^2-\beta^2){\tilde{\gamma}}^2-2\alpha E \tilde{\gamma}
+(E^2+\beta^2)=0, \eqno(36)$$ with the following solution: $$\tilde{\gamma}(t)={{1}\over{\alpha^2-\beta^2}}{\left[\alpha E
\pm|\beta|\sqrt{\beta^2+E^2-\alpha^2}~\right]}. \eqno(37)$$
Therefore, above developed theory allows us to look for the solution of the initial value problem for the bead moving along an arbitrarily curved pipe. The thorough consideration of different particular cases is beyond the scope of the present paper. Instead, we will give representative solutions for one of the simplest kinds of a spiral – Archimedes spiral – defined by (15).
The scheme for the complete inspection of the problem for any given initial value problem is the following: First, one specifies the initial location of the bead ($r_0$) and its initial radial velocity ($v_0$). The values of the $\varphi'(r_0)$ and $\Phi(r_0)$ are fixed as soon as we specify the form of the pipe $\varphi(r)$. The initial values for the $\Omega(0)$ and $\gamma(0)$ are given by (3) and (5): $$\Omega_0=\omega+\varphi'(r_0)v_0, \eqno(38)$$ $$\gamma_0=[1-r_0^2\Omega_0^2-v_0^2]^{-1/2}, \eqno(39)$$ while the value of the bead proper energy, according to (26), is given as: $$E=\gamma_0[1-r_0^2\omega\Omega_0]. \eqno(40)$$
Working with the Eq.(28), as the first order ordinary differential equation for the radial position $r(t)$ of the bead at any moment of time, we can subsequently calculate all other physical variables. On the Fig.1 the set of solutions is given for the case of the Archimedes spiral with $a=-5$, rotating with the angular velocity $\omega=2$ for the bead, which initially was situated right over the pivot of the rotator ($r_0=0$) and had initial radial velocity $v_0=0.1$.
These plots tell us that in the limit of large distance from the rotator the value of the radial velocity tends to the asymptotic value: $$\liminf{v(t)}=v_\infty \equiv -\omega/a, \eqno(41)$$ which, in this case, is equal to $v_\infty = 0.4$.
The angular velocity of the bead in the LF tends to zero, as well as the absolute value of the pipe reaction force, implying that at infinity the bead asymptotically reaches the limit of the force-free motion. This limit is understandable also analytically, because from (3) and (28) we can see that: $$v\to\frac{\omega}{|a|}+\frac{E\sqrt{a^2-\omega^2}}{\omega a^2
r^2}, \eqno(42)$$ $$\Omega\to \frac{E\sqrt{a^2-\omega^2}}{\omega
a r^2}. \eqno(43)$$
From these expressions it is clear that this regime is accessible iff the condition $|a|>\omega$ holds! Otherwise, the particle is not able to reach the infinity.
Since the shape of the function $r(t)$ is almost linear it is instructive to make plots for the functions $v(r)$ for different values of the initial radial velocity $v_0$, but with all other parameters of the initial value problem being the same. On the Fig.2 we plotted these functions for eight different values of the initial radial velocity. We see that when $v_0=v_\infty$ the movement of the particle is force-free (geodesic) and uniform during the whole course of the motion. Physically it means that for this particular value of the $v_0$ the shape of the pipe follows the geodesic trajectory of the bead, in the RF, for the metric (17) on the rotating 2D disk, so the bead moves freely, without interacting with the walls of the pipe. When $v_0<v_\infty$, the particle moves with positive acceleration and asymptotically reaches the force-free regime in the infinity. While, when $v_0>v_\infty$ the character of the motion is decelerative, but the force-free limit is reached, again, when the bead heads to infinity.
One more example of the latter behavior, similar to the case shown on the Fig. 1, but plotted for the initial velocity $v_0=0.5>v_\infty=0.4$ is given on the Fig.3. Here we see that, unlike the case given on the Fig.1, the acceleration of the bead is negative all the time and it reaches zero “from below", taking less and less negative values. While the angular velocity of the bead relative to the LF $\Omega(t)$ is also negative from the beginning but its absolute value decreases and reaches the zero as the particle tends to the infinity.
Conserved energy case
---------------------
When the bead accelerates, it continuously takes energy from the rotator. So, if one needs to keep the rotation rate constant, one needs to supply the system with energy from outside. This is, certainly, less realistic setup than the assumption that the “rotator-pipe-bead" system is conservative, viz. its total energy $E_{tot}$ is constant. In this case, however, the bead acceleration can not be permanent, because asymptotically it extracts all energy from the rotator and reaches the regime: $\omega(t)\to 0$, $E_M \to 0$ and $E_m \to E_{tot}$. Clearly, in this situation, it is more convenient to study the dynamics in the laboratory frame of reference (LF), in which the rotator and the pipe are rotating rigidly with the time-dependent angular velocity $\omega(t)$. As regards the bead, since the shape of the pipe is curved, its angular velocity relative to the LF is given by (3). Since the pipe is considered to be massless and absolutely rigid, it does not contribute any amount of energy and/or angular momentum to the total energy $E_{tot}$ and angular momentum $L_{tot}$ of the system. The rotator for simplicity is assumed to be a sphere of the radius $R$ and the mass $M$ having the inertia moment $${\cal I} = {2 \over 5}MR^2, \eqno(44)$$ the energy $$E_M={{\cal I} \over 2}\omega^2(t),\eqno(45)$$ and the angular momentum $$L_M={\cal I} \omega(t). \eqno(46)$$
Note that (44-46) are nonrelativistic expressions. If initially, at $t=t_0$, $$\omega_0R \ll 1, \eqno(47)$$ then it will remain nonrelativistic during the whole course of the motion, because it is assumed that the bead constantly extracts energy from the rotator, while the latter slows down so that the angular velocity $\omega(t)$ is a monotonically decreasing function of time. The (47) condition seems to be valid for the known fastest rotators in the Nature — pulsars. For the Crab pulsar, for instance, $R=1.2\times10^6cm$, $\omega_0=190.4Hz$ and consequently $\omega_0R/c\simeq7.6\times10^{-3}$. Even for the fastest millisecond pulsars $\omega_0R/c\le0.25$. This justifies the usage of nonrelativistic (44-46) expressions in our analysis.
The remaining part of the threefold system — the bead — is assumed to be of the rest mass $m_0$. Its angular velocity and radial velocity relative to the LF, at any given moment of time, are $\Omega(t)$ and $v(t) \equiv \dot{r}$, respectively. Even when the initial radial velocity of the bead is nonrelativistic ($v_0 \ll
1$), it is still necessary to write relativistic expressions for its energy and angular momentum, because the bead gains energy, accelerates and sooner or later its motion becomes relativistic. Therefore, its energy and angular momentum must be written as: $$E_m=m(t)=m_0\gamma(t), \eqno(48)$$ $$L_m=m_0\gamma(t)r^2(t)\Omega(t)=m(t)r^2(t)\Omega(t), \eqno(49)$$ where the LF Lorentz factor $\gamma(t)$ is defined by (5).
The system “rotator-pipe-bead" is conservative, there is no energy inflow from outside. In this sense it principally differs from the one considered in the previous section, where either the external energy source was necessary to keep the rotation rate $\omega$ constant or the rotator was assumed to possess an infinite amount of energy. Now, since the system is conservative, its dynamics are governed by the conservation laws of its total energy $E_{tot} \equiv E_M+E_m$ and total angular momentum $L_{tot}
\equiv L_M+L_m$: $${{{\cal I} \over 2}}~\omega^2(t)+m_0\gamma(t)=E_{tot}, \eqno(50)$$ $${\cal I} \omega(t)+m_0\gamma(t)r^2\Omega(t)=L_{tot}. \eqno(51)$$
And the solution of the problem reduces to the solution of these equations, linked with (9), for two unknown functions of time $r(t)$ and $\omega(t)$ for an arbitrary initial value problem: initial location of the bead $r_0=r(0)$ and the initial value of the rotation rate $\omega_0=\omega(0)$ of the whole system. The detailed study of this problem is beyond the scope of this paper and will be given in a separate publication.
Conclusion
==========
The purpose of the present paper was to study the dynamics of relativistic rotating particles with prescribed, curved trajectories of motion in the rotating frame of reference. The work is a natural generalization of the [*gedanken*]{} “pipe-bead" experiment considered Machabeli and Rogava \[12\]. In that paper the authors considered the case of the [*straight*]{} rotating pipe and they found out that when the velocity of the bead, driven by the rotation of the whole device and sliding along/within the pipe, is high enough the character of the motion changes from the accelerated to the decelerated one. In particular, it was found that when the bead starts moving from the pivot ($r=0$) of the rotating pipe with initial velocity $v_0 >
\sqrt{2}/2$, the motion is decelerative from the very beginning.
In this paper we consider the motion of rotationally driven particles along flat trajectories of arbitrarily curved shape. The practical motivation for this approach and its importance are related with the following two facts:
1. The ‘pipe-bead’ (or the ‘bead-on-the-wire’) [*gedanken*]{} experiment is considered as a model for the study of dynamics of centrifugally driven relativistic particles in rotating magnetospheres, in various classes of astrophysical objects, like pulsars \[17-18,12,20\] and AGNs \[21-24\]. The role of “pipes" is played by the magnetic field lines.
2. The shape of magnetic field lines is always curved. It implies that for the large-scale, global dynamics of charged particles — driven by centrifugal forces and moving along curved field lines of rotating magnetospheres — it is important to know what qualitative changes occur when the form of the field lines is not linear but curved.
In this paper we studied this problem, on the level of the idealized [*gedanken*]{} experiment, both in the laboratory (LF) and in the rotating (RF) frames of reference. For the simple example of the Archimedes spiral we found that the dynamics of such particles may involve both accelerative and decelerative modes of motion.
One important difference from the linear pipe case \[12\] is that for the case of a curved pipe the motion of the bead is not any more radially bounded: there exist regimes of motion when the bead may reach infinity. This result has simple physical explanation. For the case of the linear pipe, rotating with the constant angular velocity $\omega_0$, the natural limit of the radial motion was given by the light cylinder radius, defined as $R_L
\equiv \omega_0^{-1}$. Now, in the case of the curved pipe, even when it rotates with the same constant rate, the bead can slide in the azimuthal direction, following the curvature of the pipe and having a variable angular velocity ${\Omega}(t)$. It means that now the role of the [*effective light cylinder*]{} is played by $R_L(t)=
{\Omega}(t)^{-1}$, and, hence, all those radial distances become accessible, where $r(t)<R_L(t)$. Therefore, if both $r(t)$ and $R_L(t)$ are monotonically increasing functions, but the former stays always smaller than the latter (evidently it was the case in above considered examples for the Archimedes spiral) then the bead can reach infinity.
Moreover, we found that there are special solutions, which are force-free during the whole course of the motion. These are simply geodesics in the two-dimensional rotating metric (17). In the LF the motion of the bead in this case is radial, because its angular velocity ${\Omega}(t)$ stays zero all the time and, correspondingly, the light cylinder is at the infinity from the very beginning. The form of the trajectory in the RF in this case, $\varphi(r)=-(v_0/\omega_0)r$, is simply that trace, which a free bead could leave on the surface of the rotating disk during the course of its geodesic motion. Intuitively it is evident that if the pipe has this particular form the bead slides within it freely, without interaction with the walls of the pipe.
We considered only one, simple, subclass of spiral trajectories as the representative example of the solutions, but the developed theory may be used for the study of the dynamics of particles moving along arbitrarily shaped flat trajectories. It means that this approach may find wide applications to different astrophysical situations where rotation impels plasma particles to move along curved magnetic field lines.
We also gave basic equations and outlined the scheme for the solution of the more general version of the same problem, where the angular velocity of the rotating system is not assumed to be constant. Instead, it is assumed that the system rotator-pipe-bead is [*conservative*]{} and the rotator is allowed to exchange perceptible portions of energy with the bead.
The formalism developed and the results found in this paper suggest some important directions of the further research, which could be physically and astrophysically relevant.
First, it seems worthy to consider the case of the charged bead, which naturally would radiate while performing its nonuniform motion along the curved pipe. The radiative energy losses and the change of the angular momentum of the bead due to radiation could affect the dynamics of the bead and the whole system, considered, again, to be conservative. In the astrophysical context it could be interesting to see how the radiation of the bead would appear for the distant observer.
Second, it is quite natural to try to extend the analysis for the 3-D fixed trajectories of motion, considering a family of axisymmetric trajectory lines and imitating the structure of the pulsar dipole magnetic field. This could bring us at least one step closer to the understanding of the radiative processes in pulsar magnetospheres.
Third, sooner or later, we should address the fluid (plasma) problem and try to see how a continuous stream of fluid particles would behave, moving along rotating curved trajectories. This will comprise one more step closer to the reality of the pulsar environment and could help to relate with each other total energy losses of a pulsar, estimated through its slowing rate, and its radiation losses and energy taken away by centrifugally driven plasma, forming eventually the pulsar wind.
These efforts would, hopefully, bring us to the construction of the unified theory of the pulsar magnetosphere, where the inertial aspects of the particle dynamics would be taken into due account. One could then try to test the theory with the existing empirical (observational) data about the energy deposited by pulsars into their winds and the energy they lose via their radio emission. This way we could have a clue as of how important inertial processes (often unfairly neglected) are in the dynamics of pulsars.
Similar problems can be addressed also in the context of the centrifugal acceleration of particles in the jets in AGNs \[21-24\].
Yet another promising field of application is the accretion of plasma on strongly magnetized neutron stars, which is believed to lead to the appearance of [*X-ray pulsars*]{}. The final stage of the accretion is dominated by the dipole magnetic field of the accreting neutron star. It is normally assumed that the motion of accreted plasma particles is guided to the magnetic poles of the star by the rotating family of polar magnetic field lines. So, in a certain sense, this problem is of the same kinematic nature as the one related with radio pulsars, except that this time plasma moves from the region outside of the light cylinder towards the star. Certainly here, again, radiative effects and collective plasma effects are essential, so the simple ’one-particle´ treatment can give only very approximate picture of the involved physical processes. But taking into account the plasma fluid effects and the role of the radiation on the dynamics of infalling plasma streams, one could try to show how important the rotational (inertial) processes are for the dynamics of the flows infalling on strongly magnetized neutron stars and what is the influence of these processes on the observational appearance of related X-ray sources.
Acknowledgements
================
The authors are grateful to George Machabeli and Swadesh Mahajan for valuable discussions. A.D. and Z.O. are grateful to the Abdus Salam International Centre for Theoretical Physics for the hospitality and support during their visits to the Centre as a Regular Associate and a Young Collaborator, respectively.
F. de Felice, Rendiconti di Matematica, Serie VII [**10**]{}, 59 (1990)
F. de Felice, Mon. Not. R. Astr. Soc. [**252**]{}, 197 (1991)
F. de Felice and S. Usseglio-Tomasset, Class. Quantum Grav. [**8**]{}, 1817 (1991)
F. de Felice and S. Usseglio-Tomasset, Class. Quantum Grav. [**10**]{}, 353 (1993)
D. Bini, F. de Felice and R. T. Jantzen, Class. Quantum Grav. [**16**]{}, 2105 (1999)
G. Z. Machabeli, and A. D. Rogava, Phys.Rev. A [**50**]{}, 98 (1994).
G. Z. Machabeli, I.S. Nanobashvili and A. D. Rogava, Izv. Vuz. Radiofizika, [**39**]{}, 39 (1996)
F. de Felice, Phys.Rev. A [**52**]{} (1995)
J. S. Heyl, Ap.J., [**542**]{}, L45 (2000)
M. A. Abramowicz, W. Kluźniak and J. P. Lasota, Astron. Astrophys. [**374**]{}, L16 (2001).
T. Gold, Nature [**218**]{}, 731 (1968).
T. Gold, Nature [**221**]{}, 25 (1969).
R. T. Gangadhara, Astron. Astrophys. [**314**]{}, 853 (1996).
I. Contopoulos D. Kazanas and C. Fendt, Ap.J. [**511**]{}, 351 (1999).
R. D. Blandford and D. G. Payne, Mon. Not. R. Astr. Soc. [**199**]{}, 883, (1982).
X. Cao, Mon. Not. R. Astr. Soc. [**291**]{}, 145, (1997).
R. T. Gangadhara and H. Lesch, Astron. Astrophys. [**323**]{}, L45 (1997).
F. M. Rieger and K. Mannheim, Astron. Astrophys. [**353**]{}, 473 (2000).
Z. N. Osmanov, G. Machabeli and A.D. Rogava, Phys. Rev. A [**66**]{}, 042103 (2002).
W. Rindler [*Special Relativity*]{} (Oliver and Boyd, Edinburgh and London 1960).
K. S. Thorne, & D. A. MacDonald, Mon. Not. R. Astr. Soc. [**198**]{}, 339, (1982).
D. A. MacDonald & K. S. Thorne, Mon. Not. R. Astr. Soc. [**198**]{}, 345, (1982).
K. S. Thorne, R. H. Price, & D. A. MacDonald, eds. [*Black Holes: The Membrane Paradigm*]{} (Yale University Press, New Haven 1986).
[^1]: However later \[16\] it was found that the Heyl’s calculations contained the sign error and the real effect could hardly account for the observed frequency shifts in the type I X-ray bursts.
[^2]: For a rapidly rotating Kerr black hole the critical angle can be as large as $90^\circ$ \[22\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report a trigonometric parallax measurement with the Very Long Baseline Array for the water maser in the distant high-mass star-forming region [G75.30+1.32]{}. This source has a heliocentric distance of $9.25^{+0.45}_{-0.40}$ kpc, which places it in the Outer arm in the first Galactic quadrant. It lies 200 pc above the Galactic plane and is associated with a substantial enhancement at the border of a large molecular cloud. At a Galactocentric radius of 10.7 kpc, [G75.30+1.32]{} is in a region of the Galaxy where the disk is significantly warped toward the North Galactic Pole. While the star-forming region has an instantaneous Galactic orbit that is nearly circular, it displays a significant motion of 18 [km s$^{-1}$]{} toward the Galactic plane. The present results, when combined with two previous maser studies in the Outer arm, yield a pitch angle of about $12\degr$ for a large section of the arm extending from the first quadrant to the third.'
author:
- 'A. Sanna, M. J. Reid, T. M. Dame, K. M. Menten, A. Brunthaler, L. Moscadelli, X. W. Zheng and Y. Xu'
title: 'TRIGONOMETRIC PARALLAXES OF MASSIVE STAR-FORMING REGIONS. IX. THE OUTER ARM IN THE FIRST QUADRANT'
---
Introduction
============
The section of the Galaxy beyond the solar circle in the first quadrant is arguably the most difficult in which to study Galactic structure. Heliocentric distances are large (e.g., @Mead1988 [@Digel1990]), Population [i]{} objects scarce, the rotation curve poorly constrained, and the disk warped and flared. Still, two well-defined spiral arms are easily recognized in large-scale 21 cm and CO surveys of the region (e.g., @Dame2001; see also our Figure \[fig4\] below). The nearer one, the Perseus arm, straddles the solar circle in the first quadrant, apparently crossing it near a longitude of $\sim 50\degr$; the other is 2–3 kpc farther from the Galactic center and is generally called the Outer arm (although other names such as “Cygnus-Outer”, “Norma-Cygnus”, and “Perseus+I” have been applied; @Vallee2008). We note in passing that an even more distant arm was recently identified in both 21 cm and CO emissions in the first quadrant [@Dame2011]; at about 15 kpc from the Galactic center this may be the distant end of the Scutum-Centaurus arm.
In general, the Outer arm contains much less molecular gas than arms at smaller Galactic radii (e.g., Perseus). Even so, some star-forming regions are detected in the arm at distances of $\sim5$ kpc toward the anticenter direction, and this section of the arm has been located by trigonometric parallaxes [@Honma2007; @Hachisuka2009], through kinematic distance estimates (e.g., @Russeil2007), and via photometric observations of open clusters (e.g., @Pandey2006). An accurate knowledge of the gas distribution and large-scale motions in the outer regions of the Galaxy may provide strong constraints on the rotation curve of the Milky Way. In this respect, trigonometric parallax measurements of high-mass star-forming regions (HMSFRs) yield both an accurate distance to the source (within 10% accuracy) and its full-space proper motion around the Galaxy (within a few [km s$^{-1}$]{}). So far, these observations have provided evidence for a nearly flat rotation curve up to about 13 kpc from the Galactic center in both the third [@Honma2007] and second Galactic quadrants [@Hachisuka2009; @Reid2009b].
In this paper, we further constrain the outer-Galaxy rotation curve and the structure of the Outer arm in the first quadrant using masers associated with the HMSFR (IRAS 201444+3726). The longitude and large negative local standard of rest (LSR) velocity of this region, $-57$ [km s$^{-1}$]{}, place it within the locus of the Outer arm as traced by the gas. We conducted multi-epoch, astrometric, Very Long Baseline Array (VLBA) observations of its 22 GHz water masers ($>50$ Jy; @Brand1994) in order to measure the trigonometric parallax and 3-dimensional velocity vectors of this star-forming region.
Observations and Data Analysis
==============================
We conducted VLBA[^1] observations to study the $6_{16}-5_{23}$ H$_2$O maser emission (rest frequency 22.235079 GHz) toward the HMSFR . In order to measure the trigonometric (annual) parallax and Galactic proper motion of this source, we used phase-referencing observations by fast switching between the maser target and two extragalactic continuum sources, J2015+3710 and J2018+3812. The former one was found in the VCS2 survey and has an absolute position known to better than $\pm1$ mas [@Fomalont2003], whereas the latter belongs to the VERA 22 GHz Fringe Search Survey (@Petrov2007; see Figure \[calib\]). Two strong fringe-finders (3C345 and 3C454.3) were observed for bandpass, single-band delay, and instrumental phase-offset calibration. We also employed four blocks of geodetic-like observations, in order to remove total atmospheric delays for each antenna [@Reid2009a]. Detailed source information is summarized in Table \[tab1\].
At first, we carried out a preparatory survey with the VLA in BnA-configuration under program BM272 on 2007 October 5, in order to determine the maser position with sub-arcsecond accuracy. The VLBA observations were scheduled under program BM272H at four epochs: 2008 November 10, 2009 May 6, 13, and November 13. These dates were optimized to sample the yearly peaks of the right-ascension parallax sinusoid and at the same time remove all correlation among the parallax and proper motion parameters (e.g., @Sato2010). We employed four adjacent intermediate frequency (IF) bands, each 8 MHz wide, in dual circular polarization; each band was correlated to produce 256 spectral channels. The bandwidth was wide enough to detect even weak continuum sources (a few mJy beam$^{-1}$) and the channel width of 0.42 [km s$^{-1}$]{} was narrow enough to resolve well the masers line’s spectral components. The third IF band was centered on an LSR velocity (V$_{\rm LSR}$) of –56.0 [km s$^{-1}$]{}, covering the range of previously detected water maser emission [@Brand1994], whereas the remaining three IFs were spaced by 108 [km s$^{-1}$]{} each one from the other. The data were processed with the VLBA correlation facility in Socorro (New Mexico) using an averaging time of about 0.9 s, which limited the instantaneous field of view of the interferometer to about $2''$ (i.e. without significant amplitude losses). Data were reduced with the NRAO Astronomical Image Processing System (AIPS) following the procedure described in @Reid2009a. A total-power spectrum of the 22.2 GHz masers toward [G75.30+1.32]{} is shown in Figure \[spectrum\].
Results
=======
Maser emission
--------------
We mapped a range of LSR velocities within $\pm 30$ [km s$^{-1}$]{} about the mean velocity of the emission ($\approx 56$ [km s$^{-1}$]{}) and a field of view of $\pm 1''$ about the peak position of the reference feature No. 2 (see Table \[tab3\]). Since [G75.30+1.32]{} has not been well studied in the past, we imaged a large field-of-view and checked for differences between the total-power spectrum (Figure \[spectrum\]) and the flux density from the detected maser features (Table \[tab3\]). With the interferometer, we recovered more than $90\%$ of the emission detected with single-dish measurements. In addition to the emission reported in Table \[tab3\], we found a strong, maser line at about $-49$ [km s$^{-1}$]{}in the spectra of the second and third epochs. This emission came from about $6''$ northward of the reference feature and probably represents another forming star in the region. However, analysis of this source is beyond the scope of this paper.
We identified 38 distinct water maser features, distributed within an area of about $ 0\farcs6 \times 0\farcs9 $. We use the term *feature* to refer to multiple “spots” spatially overlapping in contiguous velocity channels and identify features with individual masing cloudlets (e.g., @Sanna2010a). We used the AIPS task JMFIT to fit an elliptical Gaussian brightness distribution to the brightness structure of each maser emission center at each epoch. The measured properties of individual features are presented in Table \[tab3\]. Their intensities range from about 10 to 0.1 Jy beam$^{-1}$ and span LSR velocities from $-48$ [km s$^{-1}$]{} for the most redshifted feature (No. 17) to $-66$ [km s$^{-1}$]{} for the most blueshifted one (No. 18). One-third of the maser features persisted over the 1 year time-baseline of our observations. For features lasting at least 6 months, we modeled their relative position shifts with time by a linear fit to get the internal proper motions of the maser distribution. At a distance of 9.25 kpc, the magnitude of the relative motions with respect to feature No. 2 ranges from 4 [km s$^{-1}$]{} for feature No. 6 to 84 [km s$^{-1}$]{} for feature No. 34, with an average accuracy of about 2 [km s$^{-1}$]{}.
Distance and Galactic proper motion {#distance_results}
-----------------------------------
We measured the parallax and proper motion of the masers relative to the two QSOs by differencing their positions over time (see @Reid2009a for a detailed discussion). For the purposes of a maser reference position, we used the spot at $-54.3$ [km s$^{-1}$]{} associated with feature No. 5 (in Table \[tab3\], it is reported the peak LSR velocity for each feature). This spot was isolated, nearly point-like, and fairly strong at all epochs ($\rm SNR \gg 100$). Table \[tab2\] and Figure \[parallax\] show the results of the parallax and proper motion fitting for both background sources.
Since systematic errors usually dominate over random noise, we estimated “a posteriori” the quality of the measurements from the fit residuals (see @Reid2009a). While both calibrators had a small angular distance from the maser source (see Table \[tab1\]), J2018+3812 had a deconvolved size (e.g., $\rm \approx 0.5~mas \times 0.2~mas$ at $-5\degr$, from the first epoch data) more than two times that of J2015+3710. That affected the parallax signature in the north direction for the fainter calibrator J2018+3812. Thus, to optimally combine the measurements of the two background sources, we determined their relative weights as follow. First, we fitted data for each background source separately to give preliminary estimates of parallax and proper motion. Since we expect no detectable proper motion for the extragalactic sources, we re-fitted the data from the two sources independently, holding the proper motions fixed at the average values from the first fit. This procedure was iterated, adjusting the individual error floors to yield values of chi-squared per degree of freedom near unity. The error-floors determined in this manner were $\pm 0.01$ mas in both the E–W and N–S directions for J2015+3710 and $\pm 0.03$ mas in the E–W direction and $\pm 0.08$ mas in the N–S direction for J2018+3812. These individual error floors were used in a final “combined” parallax reported in Table \[tab2\]. Note that, since the targets of these observations were high-declination sources, we avoided low-elevation data. That explains the accurate N–S offset measurements for the point-like calibrator J2015+3710, compared to low-declination sources (e.g., @Reid2009a). The parallax of [G75.30+1.32]{} from the combined fit is $0.108 \pm 0.005$ mas, corresponding to a distance of $9.25^{+0.45}_{-0.40}$ kpc from the Sun. For the IAU value of the distance to the Galactic center, R$_0 = 8.5$ kpc, our measured distance translates to a Galactocentric radius of 10.8 kpc.
To determine a secular proper motion for the HMSFR (as opposed to a single maser spot), we need to correct the combined fit values in Table \[tab2\] for the proper motion of the reference spot with respect to the forming star. We estimated this contribution from the average proper motion of all the maser features with respect to the reference spot of the parallax. The average internal velocity components are $-0.28 \pm 0.09 $ mas yr$^{-1}$ toward the east and $-0.64 \pm 0.17 $ mas yr$^{-1}$ toward the north, where we report the standard error of the mean. Thus, the total motion of the whole source is estimated to be $-2.37 \pm 0.09 $ mas yr$^{-1}$ and $-4.48 \pm 0.17 $ mas yr$^{-1}$ in the east and north directions, respectively. At our measured distance these values correspond to $-104 \pm 4$ [km s$^{-1}$]{} and $-196 \pm 7$ [km s$^{-1}$]{} eastward and northward, respectively. Completing the kinematic information, we have assumed an LSR velocity of $-57.0 \pm 2.2$ [km s$^{-1}$]{} for the HMSFR [G75.30+1.32]{}, obtained from the CS(2$-$1) line survey by @Bronfman1996, in agreement with the median velocity of the maser features ($-57.2$ [km s$^{-1}$]{}).
We can transform the 3-dimensional velocity from the equatorial heliocentric reference frame, in which they were measured, to a reference frame rotating with the Galaxy, to estimate the peculiar motion of the HMSFR (i.e. the deviation from a circular rotation). Adopting a flat rotation curve for the Milky Way with $\Theta = 239$ [km s$^{-1}$]{} [@Brunthaler2011], a current “best-estimate” for the distance to the Galactic center of R$_0 = 8.3$ kpc [@Reid2009b; @Brunthaler2011], and the revised *Hipparcos* measurements of the solar motion [@Schoenrich2010], we find peculiar velocity components for [G75.30+1.32]{} of $(U_s, V_s, W_s)=( 11.3 \pm 5.3, 1.1 \pm 8.8, -17.8 \pm 5.4)$ [km s$^{-1}$]{}, where U$_s$, V$_s$ and W$_s$ are directed toward the Galactic center, in the direction of Galactic rotation and toward the North Galactic Pole, respectively. In Table \[outer\], we summarize these results together with the peculiar proper motions of sources previously measured along the Outer arm.
Discussion
==========
The source [G75.30+1.32]{} is located in the first Galactic quadrant at a Galactocentric radius of about 11 kpc, far beyond the Galactocentric radius of the Perseus arm in this direction ($\sim 8$ kpc; lower right panel in Figure \[fig4\]). The Outer arm between Galactic longitudes of $20\degr-90\degr$ at zero latitude can be clearly traced in 21 cm emission with $\rm V_{LSR} \leqslant -20$ [km s$^{-1}$]{} (upper panel of Figure \[fig4\]). The longitude-velocity diagram in Figure \[fig4\] shows that [G75.30+1.32]{} (marked by a cross) is definitely within the locus of the Outer Arm, although it lies about 10 [km s$^{-1}$]{} higher in velocity than the central velocity of the arm at that longitude ($\sim -67$ [km s$^{-1}$]{}). We note that, with a Galactic latitude of $b=1.32\degr$ and a distance of 9.25 kpc, [G75.30+1.32]{} is located $\sim 213$ pc above the Galactic plane. An analysis of both the CO survey from @Dame2001 and the survey from @Kalberla2005 toward the Galactic latitude of [G75.30+1.32]{} shows that the star-forming region sits on the side of a fairly bright CO cloud and lies close to a substantial enhancement (lower left panel in Figure \[fig4\]). A further comparison of the Galactic coordinates and distance of [G75.30+1.32]{} with a vertical cross-section of the Galactic plane points out the coincidence with a prominent warp of the Milky Way, particularly pronounced in the neutral hydrogen gas (e.g., @Nakanishi2003).
The Galactic warp is a displacement of the outer disk from the midplane, generally upward (toward the North Galactic Pole) between the first and second quadrants and downward at diametrically opposite directions in the third and fourth. This midplane displacement can be also detected in faint CO line emission and peaks close to the Galactic location of [G75.30+1.32]{}, at a Galactocentric azimuth $\theta=124\degr$, toward the North Galactic pole (e.g., @Nakanishi2006, their Figure 8). The Galactocentric azimuth is defined as the azimuth angle around the rotation axis of the Galaxy, taken so that $\theta=180\degr$ points to the Sun and $\theta=90\degr$ is parallel to $\ell=90\degr$. The difference of $\rm V_{LSR}$ at $b =0\degr$ and $b =1.32\degr$ may be related to the dynamics of the warp together with the peculiar motion of [G75.30+1.32]{}perpendicular to the plane ($-18$ [km s$^{-1}$]{}) and directed toward the Galactic plane. We note that this last value differs by about six times the mean value for HMSFRs located within the Perseus arm ($\sim -3$ [km s$^{-1}$]{}; @Reid2009b).
Two ideas have been forwarded to explain the Galactic warp (see @Sellwood2010 and reference therein): 1) as a consequence of the ongoing accretion of matter by the Galactic halo at much larger radii than the edge of the disk and 2) as a result of tidal interactions due to the proximity of the Magellanic Clouds ($\ell \sim 280\degr$) and the Sagittarius dwarf galaxy (e.g., Figure 1 in @Bailin2003). In the accretion scenario, gravitational torques between the inner halo and the outer disk produce a misalignment of the outer parts of the Galaxy with respect to the plane of the inner disk. This happens because of a redistribution of angular momentum through the system as long as freshly accreted matter streams from the outer to the inner parts of the Galaxy (e.g., @Binney1992 [@Jiang1999]). In this framework, the present result ($W_s$) may be interpreted as an inward motion through the Galactic warp toward the inner Galactic disk.
Using R$_0 = 8.3$ kpc, $\Theta_0 = 239$ [km s$^{-1}$]{}, and a flat rotation curve beyond the solar circle, the *revised* kinematic distance [@Reid2009b] at the Galactic coordinates of the source with an $\rm V_{LSR}=-57$ [km s$^{-1}$]{} is 9.9 kpc, with no kinematic distance ambiguity as the source is in the outer Galaxy. The small discrepancy ($\sim1\sigma$) between the kinematic distance of [G75.30+1.32]{} and its trigonometric distance of 9.25 kpc can be accounted for by the small peculiar motion of [G75.30+1.32]{} ($\sim7$ [km s$^{-1}$]{}) toward to the Sun direction (Table \[outer\]). On the one hand, this small difference hints at the validity of a nearly flat rotation curve up to the outer regions of the first Galactic quadrant ($d\Theta/dR < 3 \rm~ km~s^{-1}~kpc^{-1} $; e.g., @Brand1993 [@Reid2009b]). On the other hand, the small values of the peculiar motion of [G75.30+1.32]{} along the Galactic plane (Table \[outer\]) may be interpreted in the framework of a density-wave-induced spiral perturbation of the gravitational potential in a rotating galaxy [@Lin1969]. These perturbations give rise to systematic, residual velocities within the spiral arms of the order of 10 [km s$^{-1}$]{}[@Lin1969]: inside the corotation radius, at the inner edge of an arm these streaming motions are directed toward the galactic center and counter to galactic rotation; at the outer edge, residual velocities toward the galactic center decrease to zero and point in the direction of the galactic rotation. Outside the corotation radius residual velocities should be reversed (e.g., Figure 2 in @Mel'Nik1999; @Mel'Nik2001). By assuming a spiral pattern speed of $\rm 20~km~s^{-1}~kpc^{-1} $ (e.g., @Bissantz2003) and a Galactic rotation of about 240 [km s$^{-1}$]{}, the corotation radius would be located at about 12 kpc. [G75.30+1.32]{} shows a residual velocity toward the Galactic center consistent with the star-forming region inside the corotation radius (Table \[outer\]). Furthermore, the streaming motions of star-forming regions measured so far along the Outer arm (Table \[outer\]) show positive values toward the Galactic center and suggest a corotation radius above a Galactocentric distance of 13 kpc. On the contrary, @Russeil2007 located the corotation radius between the Perseus arm and the Outer arm in order to explain the opposite sign of residual velocities of their regions. Note that, this estimate accounted only for residual radial velocities ($\rm V_{LSR}$). As soon as the BeSSeL survey[^2] and the VERA project[^3] increase the sample of 3-dimensional velocity measurements along the Outer arm, we should be able to assess these issues with stronger statistical support.
Finally, we can estimate the pitch angle of the Outer arm between Galactic longitudes from $75\degr$ to $196\degr$. We mention as a caveat that spiral arms in other galaxies can have “kinks” and variations of pitch angle over large portions of an arm (e.g., @Seigar1998). For the simplest case of a logarithmic spiral arm (e.g., @Reid2009b), we fitted the parallax data of [G75.30+1.32]{} (first quadrant) together with the two previous (trigonometric) measurements of star-forming regions along the Outer arm (Table \[outer\]). This approach formally yields a pitch angle of $12.1\degr \pm 4.2 \degr$ which covers about $120\degr$ of longitude between the first and third Galactic quadrants. In Figure \[pitch\], we display this result with the logarithm of the Galactocentric radius as a function of the Galactocentric longitude ($\beta$, defined as 0 toward the Sun). While perfect-logarithmic, spiral arm sections should appear as straight lines in such a plot, the measurement for the star-forming region WB 89–437 deviates from the fit by about three times the parallax uncertainty ($\approx 600$ pc). This behavior may be explained when considering the spread of star-forming regions within an arm with a width of several hundreds of pc. The mean value derived here is in good agreement with a weighted average of published data for Galactic spiral arms, which gives a mean pitch angle of $12.8\degr$ [@Vallee2008], and agrees also with a more recent estimate of $13.5\degr$ for the Outer arm based on modeling of FIR cooling lines from the Galactic interstellar medium [@Steiman-Cameron2010]. Furthermore, it is consistent with the first direct estimate of the Perseus arm pitch angle ($16.5\degr \pm 3.1 \degr$) made with the trigonometric distances of its HMSFRs [@Reid2009b].
This work was partially funded by the ERC Advanced Investigator Grant GLOSTAR (247078). Y.X. was supported by the Chinese NSF through grants NSF 11073054, NSF 10733030, and NSF 10621303.
[*Facilities:*]{} .
Bailin, J. 2003, , 583, L79
Benjamin, R. A., et al. 2005, , 630, L149
Binney, J. 1992, , 30, 51
Bissantz, N., Englmaier, P., & Gerhard, O. 2003, , 340, 949
Brand, J., & Blitz, L. 1993, , 275, 67
Brand, J., et al. 1994, , 103, 541
Bronfman, L., Nyman, L.-A., & May, J. 1996, , 115, 81
Brunthaler, A., et al. 2011, arXiv:1102.5350
Dame, T. M., Hartmann, D., & Thaddeus, P. 2001, , 547, 792
Dame, T. M., & Thaddeus, P. 2011, , 734, L24
Dehnen, W., & Binney, J. J. 1998, , 298, 387
Digel, S., Thaddeus, P., & Bally, J. 1990, , 357, L29
Fomalont, E. B., Petrov, L., MacMillan, D. S., Gordon, D., & Ma, C. 2003, , 126, 2562
Hachisuka, K., Brunthaler, A., Menten, K. M., Reid, M. J., Hagiwara, Y., & Mochizuki, N. 2009, , 696, 1981
Honma, M., et al. 2007, , 59, 889
Hou, L. G., Han, J. L., & Shi, W. B. 2009, , 499, 473
Kalberla, P. M. W., Burton, W. B., Hartmann, D., Arnal, E. M., Bajaja, E., Morras, R., P$\ddot{o}$ppel, W. G. L. 2005, , 440, 775
Jiang, I.-G., & Binney, J. 1999, , 303, L7
Leung, H. O., & Thaddeus, P. 1992, , 81, 267
Levine, E. S., Blitz, L., & Heiles, C. 2006, Science, 312, 1773
Lin, C. C., Yuan, C., & Shu, F. H. 1969, , 155, 721
Mead, K. N., & Kutner, M. L. 1988, , 330, 399
Mel’Nik, A. M., Dambis, A. K., & Rastorguev, A. S. 1999, Astronomy Letters, 25, 518
Mel’Nik, A. M., Dambis, A. K., & Rastorguev, A. S. 2001, Astronomy Letters, 27, 521
Nakanishi, H., & Sofue, Y. 2003, , 55, 191
Nakanishi, H., & Sofue, Y. 2006, , 58, 847
Pandey, A. K., Sharma, S., & Ogura, K. 2006, , 373, 255
Petrov, L., Hirota, T., Honma, M., Shibata, K. M., Jike, T., & Kobayashi, H. 2007, , 133, 2487
Reid, M. J., Menten, K. M., Brunthaler, A., Zheng, X. W., Moscadelli, L., & Xu, Y. 2009a, , 693, 397
Reid, M. J., et al. 2009, , 700, 137
Russeil, D., Adami, C., & Georgelin, Y. M. 2007, , 470, 161
Sanna, A., Moscadelli, L., Cesaroni, R., Tarchi, A., Furuya, R. S., & Goddi, C. 2010, , 517, A71
Sato, M., Reid, M. J., Brunthaler, A., & Menten, K. M. 2010, , 720, 1055
Sch[ö]{}nrich, R., Binney, J., & Dehnen, W. 2010, , 403, 1829
Sellwood, J. A. 2010, arXiv:1006.4855
Seigar, M. S., & James, P. A. 1998, , 299, 685
Steiman-Cameron, T. Y., Wolfire, M., & Hollenbach, D. 2010, , 722, 1460
Taylor, J. H., & Cordes, J. M. 1993, , 411, 674
Vall[é]{}e, J. P. 2008, , 135, 1301
[llllrcccc]{}
G75.30+1.32 & 20 16 16.0117 & 37 35 45.807 & & & $0.87\times0.43 \ at -15.9\degr$ & 4.499 & 0.007 & $-56.8$\
J2015+3710 & 20 15 28.7297 & 37 10 59.514 & 0.4 & $-139$ & $0.88\times0.45 \ at -15.3\degr$ & 2.301 & 0.010 &\
J2018+3812 & 20 18 42.8500 & 38 12 41.700 & 0.8 & +58 & $0.89\times0.45 \ at -13.2\degr$ & 0.079 & 0.002 &\
[lcccrrrr]{}
1 & ,3,4 & –53.89 & 9.84 & $491.02 \pm 0.04 $ & $ -32.78 \pm 0.03 $ & $-0.11 \pm 0.11 $ & $ 0.54 \pm 0.04 $\
2 & ,2,3,4 & –56.84 & 8.89 & $ 0 \pm 0.01 $ & $ 0 \pm 0.01 $ & $ 0 $ & $ 0 $\
3 & ,2,3,4 & –56.42 & 7.87 & $ 1.37 \pm 0.03 $ & $ -0.91 \pm 0.03 $ & $-0.31 \pm 0.05 $ & $ 0.38 \pm 0.06 $\
4 & 2,3, & –56.00 & 6.66 & $ 4.22 \pm 0.04 $ & $ -2.27 \pm 0.03 $ & $ 0.51 \pm 0.08 $ & $-0.41 \pm 0.04 $\
5 & ,2,3,4 & –54.73 & 4.92 & $297.69 \pm 0.03 $ & $ -32.02 \pm 0.01 $ & $ 0.59 \pm 0.03 $ & $ 0.70 \pm 0.01 $\
6 & 1,2,3, & –55.57 & 4.80 & $ 3.51 \pm 0.02 $ & $ -1.95 \pm 0.01 $ & $ 0.05 \pm 0.09 $ & $ -0.07 \pm 0.05$\
7 & 2,3, & –63.58 & 3.08 & $304.16 \pm 0.05 $ & $ -38.64 \pm 0.03 $ & $ 0.62 \pm 0.06 $ & $ 0.28 \pm 0.03 $\
8 & 2,3, & –58.52 & 2.54 & $490.89 \pm 0.07 $ & $ -32.84 \pm 0.03 $ & $ 0.26 \pm 0.09 $ & $ 0.31 \pm 0.04 $\
9 & 2,3, & –61.05 & 2.09 & $530.13 \pm 0.09 $ & $ 56.01 \pm 0.14 $ & $ 1.13 \pm 0.18 $ & $ 0.70 \pm 0.23 $\
10 & ,2,3,4 & –55.57 & 1.50 & $ 2.71 \pm 0.02 $ & $ -1.50 \pm 0.02 $ & $ 0.00 \pm 0.04 $ & $ -0.38 \pm 0.02$\
11 & 2,,4 & –57.26 & 1.16 & $540.14 \pm 0.04 $ & $ 83.24 \pm 0.03 $ & $ 0.80 \pm 0.06 $ & $ 0.42 \pm 0.04 $\
12 & 2, & –55.15 & 1.02 & $532.29 \pm 0.04 $ & $ 90.97 \pm 0.03 $ & &\
13 & ,2,3,4 & –55.57 & 0.82 & $298.30 \pm 0.06 $ & $ -32.21 \pm 0.03 $ & $ 0.86 \pm 0.06 $ & $ 0.63 \pm 0.06 $\
14 & 1,2,,4 & –56.00 & 0.70 & $-74.65 \pm 0.02 $ & $ 224.93 \pm 0.04 $ & $ 0.57 \pm 0.04 $ & $ -1.39 \pm 0.07$\
15 & 1,2,3, & –59.79 & 0.66 & $240.94 \pm 0.04 $ & $-103.11 \pm 0.02 $ & $ 0.39 \pm 0.04 $ & $ 0.56 \pm 0.03 $\
16 & 2,3, & –56.00 & 0.58 & $ 5.57 \pm 0.05 $ & $ -2.94 \pm 0.04 $ & $-0.03 \pm 0.09 $ & $ -0.43 \pm 0.07$\
17 & 1 & –48.41 & 0.55 & $363.93 \pm 0.01 $ & $-110.72 \pm 0.01 $ & &\
18 & 1,2,3, & –66.11 & 0.54 & $304.20 \pm 0.02 $ & $ -38.95 \pm 0.04 $ & $ 0.47 \pm 0.02 $ & $ 0.37 \pm 0.03 $\
19 & 4 & –55.15 & 0.50 & $-73.77 \pm 0.04 $ & $ 224.36 \pm 0.03 $ & &\
20 & ,2,3,4 & –59.79 & 0.48 & $-75.77 \pm 0.01 $ & $ 223.34 \pm 0.01 $ & $ 0.69 \pm 0.03 $ & $ -1.21 \pm 0.04$\
21 & 4 & –58.52 & 0.37 & $535.22 \pm 0.02 $ & $ 90.11 \pm 0.02 $ & &\
22 & ,3,4 & –52.63 & 0.32 & $313.84 \pm 0.05 $ & $-635.91 \pm 0.04 $ & $ 0.15 \pm 0.12 $ & $ -0.38 \pm 0.09$\
23 & 1 & –54.73 & 0.32 & $-74.32 \pm 0.01 $ & $ 226.09 \pm 0.03 $ & &\
24 & 2,3, & –57.68 & 0.28 & $-74.40 \pm 0.06 $ & $ 212.61 \pm 0.06 $ & $ 0.70 \pm 0.08 $ & $ -1.50 \pm 0.10$\
25 & ,2,3,4 & –63.58 & 0.24 & $302.19 \pm 0.01 $ & $ -33.75 \pm 0.02 $ & $ 0.18 \pm 0.03 $ & $ 0.17 \pm 0.05 $\
26 & 4 & –55.57 & 0.22 & $489.33 \pm 0.03 $ & $ -34.53 \pm 0.03 $ & &\
27 & 4 & –62.32 & 0.22 & $305.18 \pm 0.03 $ & $ -40.37 \pm 0.04 $ & &\
28 & ,3 & –54.31 & 0.20 & $301.96 \pm 0.04 $ & $ -34.03 \pm 0.04 $ & &\
29 & 2,3, & –58.10 & 0.18 & $240.70 \pm 0.05 $ & $-102.68 \pm 0.04 $ & $-0.35 \pm 0.12 $ & $ 0.73 \pm 0.08 $\
30 & 1 & –50.52 & 0.16 & $240.23 \pm 0.01 $ & $-103.19 \pm 0.02 $ & &\
31 & 1 & –54.73 & 0.16 & $298.91 \pm 0.03 $ & $ -33.57 \pm 0.04 $ & &\
32 & ,2,3 & –56.84 & 0.13 & $ 8.90 \pm 0.02 $ & $ -4.92 \pm 0.03 $ & $ 0.26 \pm 0.09 $ & $ -0.60 \pm 0.07$\
33 & 3 & –53.89 & 0.12 & $ 8.49 \pm 0.04 $ & $ -5.05 \pm 0.04 $ & &\
34 & 1,2,3, & –59.79 & 0.11 & $-83.93 \pm 0.03 $ & $ 242.82 \pm 0.05 $ & $-0.57 \pm 0.05 $ & $ 1.82 \pm 0.08 $\
35 & 1 & –58.94 & 0.10 & $-76.87 \pm 0.01 $ & $ 231.80 \pm 0.02 $ & &\
36 & 2, & –58.94 & 0.09 & $-75.07 \pm 0.06 $ & $ 213.21 \pm 0.04 $ & &\
37 & 2, & –61.47 & 0.07 & $-73.28 \pm 0.05 $ & $ 211.22 \pm 0.05 $ & &\
38 & 4 & –50.52 & 0.07 & $-83.85 \pm 0.06 $ & $ 246.97 \pm 0.07 $ & &\
[lllll]{}
$-54.3$ & J2015+3710 & $0.108 \pm 0.005$ & $-2.088 \pm 0.016 $ & $-3.847 \pm 0.019 $\
$-54.3$ & J2018+3812 & $0.096 \pm 0.008$ & $-2.153 \pm 0.023 $ & $-3.751 \pm 0.134 $\
& & & &\
$-54.3$ & Combined & $ 0.108 \pm 0.005$ & $-2.093 \pm 0.012 $ & $-3.845 \pm 0.014 $\
& & & &\
\
& & & &\
& & $ 0.108 \pm 0.005$& $-2.37 \pm 0.09 $ & $-4.48 \pm 0.17 $\
[lrrcrrrc]{}
G75.30+1.32 & 75.30 & +1.32 & 10.7 & $ 11.3 \pm 5.3$ & $1.1 \pm 8.8 $ & $-17.8 \pm 5.4 $ &\
WB 89–437 & 135.28 & +2.80 & 13.2 & $ 14.8 \pm 4.8$ & $2.4 \pm 8.3 $ & $0.9 \pm 9.8 $ & 1\
S 269 & 196.45 & $-1.68$ & 13.4 & $ 5.0 \pm 3.5$ & $5.8 \pm 2.8 $ & $-4.4 \pm 1.5 $ & 2\
& & & & & &\
[^1]: The VLBA is operated by the National Radio Astronomy Observatory (NRAO). The NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
[^2]: see the web page at the following URL, http://www.mpifr-bonn.mpg.de/staff/abrunthaler/BeSSeL/index.shtml
[^3]: see the web page at the following URL, http://veraserver.mtk.nao.ac.jp/index.html
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We research proximinality of $\mu$-sequentially compact sets and $\mu$-compact sets in measurable function spaces. Next we show a correspondence between the Kadec-Klee property for convergence in measure and $\mu$-compactness of the sets in Banach function spaces. Also the property $S$ is investigated in Fréchet spaces and employed to provide the Kadec-Klee property for local convergence in measure. We discuss complete criteria for continuity of metric projection in Fréchet spaces with respect to the Hausdorff distance. Finally, we present the necessary and sufficient condition for continuous metric selection onto a one-dimensional subspace in sequence Lorentz spaces $d(w,1)$.'
author:
- 'Maciej Ciesielski$^*$ and Grzegorz Lewicki'
title: Best approximation properties in spaces of measurable functions
---
[^1]
[ . ]{}
[ Fréchet spaces, Banach function spaces, Lorentz spaces, Kadec-Klee property, metric selection, extreme points, $\mu$-compactness.]{}
Introduction
============
The geometrical structure of Banach spaces has been investigated tremendously and applied to approximation theory by many authors [@CheHeHudz; @CiesKamPluc; @HKL; @Hu-Ku; @kurc]. Since the deep motivation of study of the geometry of Banach spaces has been developed, during the decades selected global properties corresponding to a metric have been evaluated [@Cal; @CKKP; @hk1; @Kam_extrem; @Kam; @KMGam]. Namely, the monotonicity and rotundity properties of Banach spaces (for example rotundity, uniform rotundity, strict monotonicity and uniform monotonicity) are crucial key in investigation of existence and uniqueness at the best approximation problems [@CheHeHudz; @Hu-Ku; @kurc]. If the global structure of Banach space disappoints, then the natural question appears of researching the applicable local structure of Banach space in approximation problems. Namely, local approach of rotundity, monotonicity and Kadec Klee property for global and local convergence in measure, resp. (for example extreme points, points of lower and upper monotonicity, $H_g$ and $H_l$ points) with application to the best approximation problems has been evolved recently by [@CiesKamPluc; @CieKolPan; @CieKolPlu]. In view of the previous results, a natural expectation that the metric structure of Fréchet spaces plays an analogous rule as Banach structure in approximation theory is researched in this paper. The next point of our interest in this paper is an existence of continuous metric selection onto a one-dimensional subspace of the sequence Lorentz spaces. It is worth to noticing that the complete condition under which there exists a continuous metric selection in $C_0(T)$ and $L^1$ was researched in [@Deutsch]. In the spirit of these results we investigate criteria for continuity of the metric selection onto a one-dimensional subspace of the Lorentz spaces $d(w,1)$. For more information concerning various concepts of continuity of the metric projection operator the reader is referred to [@Brown1] - [@Brown5] and [@Wulbert]. The last point of our consideration in this paper is devoted to the property $S$ that was established unexplicitly for the first time in [@KPS] and applied to provide the Kadec-Klee property for convergence in measure in Lorentz spaces $\Lambda_\phi$.
In section 2 we recall the necessary terminology.
The section 3 is devoted to investigation of proximinality of $\mu$-sequentially compact sets and $\mu$-compact sets in measurable function spaces. We present examples and properties of $\mu$-sequentially compact and $\mu$-compact sets. We positively answer the essential question under which criteria in a Fréchet space $X$, equipped with a $F$-norm, $\mu$-sequentially compactness yields proximinality of closed subset of $X$. We also discuss some special proximinal sets and Chebyshev sets in Fréchet spaces. Finally, in this section we apply Kedec-Klee property for global convergence in measure to establish proximinality of $\mu$-compact sets in Banach function spaces. We also present a class of Lorentz spaces $\Lambda_\phi$ and $\Gamma_{p,w}$ which possess a property that any nonempty $\mu$-compact subset in these spaces is proximinal.
In the next section 4 we characterize a continuity of metric projection operator and property $(S)$ in Banach function spaces and Fréchet spaces. Namely, we show examples of Lorentz spaces $\Lambda_\phi$ and $\Gamma_{p,w}$ and also some specific Fréchet spaces which satisfy property $S$. We also prove that the direct sum of Fréchet spaces equipped with $F$-norm satisfying property $S$ is a Fréchet space equipped with $F$-norm and has Kadec-Klee property for local convergence in measure. Finally, we establish the necessary and sufficient criteria for continuity of the metric projection with respect to the Hausdorff distance in Fréchet function space with the Kadec-Klee property for local convergence in measure. It is worth noticing that in our considerations we do not restrict ourselves to the case of Banach spaces but we also consider Fréchet spaces not necessarily locally convex. Observe that in abstract approximation theory there are very few papers concerning Fréchet spaces. Mainly the case of Banach spaces is considered (see for example taken in a random way volume of Journal of Approximation Theory or Constructive Approximation).
The last section 5 is devoted to the characterization of the metric selection for the metric projection onto a one-dimensional subspace in sequence Lorentz spaces $d_{(w,1)},$ which surprisingly is a highly non-trivial problem (compare with [@Deutsch], Th. 6.3 and Cor. 6.6). We present full criteria for continuity of the metric selection onto a one-dimensional subspace in sequence Lorentz space $d_{(w,1)}$. We also discuss some examples of the metric projection onto a one-dimensional subspace in sequence Lorentz spaces which does not admit a continuous metric selection and also admits a continuous metric selection.
Preliminaries
=============
Let $\mathbb{R}$, $\mathbb{R}^+$ and $\mathbb{N}$ be the sets of reals, nonnegative reals and positive integers, respectively. Denote as usual by $S_X$ (resp. $B_X$) the unit sphere (resp. the closed unit ball) of a Banach space $(X,\left\Vert \cdot \right\Vert _{X}).$ A point $x\in{B_X}$ is called an [*extreme point*]{} of $B_X$ if for any $y,z\in{B_X}$ such that $y+z=2x$ we have $y=z$. A Banach space $X$ is said to be [*strictly convex*]{} if any element $x\in{S_X}$ is an extreme point of $B_X$. Define by $(T,\Sigma,\mu)$ a measure space and by $L^{0}(T)$ the set of all (equivalence classes of) extended real valued $\mu$ measurable functions on $T$. For simplicity we use the short notation $L^0=L^0([0,\alpha))$ with the Lebesgue measure $m$ on $[0,\alpha)$, where $\alpha =1$ or $\alpha =\infty$. A Banach lattice $(E,\Vert \cdot \Vert _{E})$ is called a *Banach function space* (or a *Köthe space*) if it is a sublattice of $L^{0}$ satisfying the following conditions
- If $x\in L^0$, $y\in E$ and $|x|\leq|y|$ m-a.e., then $x\in E$ and $\|x\|_E\leq\|y\|_E$.
- There exists a strictly positive $x\in E$.
By $E^{+}$ we denote the positive cone of $E$, i.e. $E^{+}={\{x \in E:x \ge 0,m\textnormal{-a.e.}\}}$. We denote $A^{c}=[0,\alpha)\backslash A$ for any measurable set $A$. A space $E$ has the *Fatou property* if for any $\left( x_{n}\right)\subset{}E^+$, $\sup_{n\in \mathbb{N}}\Vert x_{n}\Vert
_{E}<\infty$ and $x_{n}\uparrow x\in L^{0}$, then $x\in E$ and $\Vert x_{n}\Vert _{E}\uparrow\Vert x\Vert_{E}$. A space $E$ has the semi-*Fatou property* if conditions $0\leqslant
x_{n}\uparrow x\in E$ with $x_{n}\in E$ imply $\Vert x_{n}\Vert _{E}\uparrow
\Vert x\Vert _{E}$. For any function $x\in L^{0}$ we define its *distribution function* by $$d_{x}(\lambda) =m\left\{ s\in [ 0,\alpha) :\left\vert x\left(s\right) \right\vert >\lambda \right\},\qquad\lambda \geq 0,$$ and its *decreasing rearrangement* by $$x^{\ast }\left( t\right) =\inf \left\{ \lambda >0:d_{x}\left( \lambda
\right) \leq t\right\}, \text{ \ \ } t\geq 0.$$ Given $x\in L^{0}$ we denote the *maximal function* of $x^{\ast }$ by $$x^{\ast \ast }(t)=\frac{1}{t}\int_{0}^{t}x^{\ast }(s)ds.$$Two functions $x,y\in{L^0}$ are said to be *equimeasurable* (shortly $x\sim y$) if $d_x=d_y$. A Banach function space $(E,\Vert \cdot \Vert_{E}) $ is called *rearrangement invariant* (r.i. for short) or *symmetric* if whenever $x\in L^{0}$ and $y\in E$ with $x \sim y,$ then $x\in E$ and $\Vert x\Vert
_{E}=\Vert y\Vert _{E}$. For more properties of $d_{x}$, $x^{\ast }$ and $x^{\ast \ast }$ see [@BS; @KPS]. Let $0<p<\infty $ and $w\in L^{0}$ be a nonnegative weight function, the Lorentz space $\Gamma _{p,w}$ is a subspace of $L^{0}$ such that $$\Vert x\Vert _{\Gamma _{p,w}}:=\left( \int_{0}^{\alpha }x^{\ast \ast
p}(t)w(t)dt\right) ^{1/p}<\infty .$$ Additionally, we assume that $w$ is from class $D_{p}$, i.e. $$W(s):=\int_{0}^{s}w(t)dt<\infty \mathnormal{\ \ \ }\text{\textnormal{and}}\mathnormal{\ \ }W_{p}(s):=s^{p}\int_{s}^{\alpha }t^{-p}w(t)dt<\infty$$for all $0<s\leq 1 $ if $\alpha =1 $ and for all $0<s<\infty $ otherwise. These two conditions guarantees that Lorentz space $\Gamma _{p,w}$ is nontrivial. It is well known that $\left( \Gamma _{p,w},\Vert \cdot \Vert
_{\Gamma _{p,w}}\right) $ is a r.i. quasi-Banach function space with the Fatou property. It was proved in [@KMGam] that in the case when $\alpha =\infty $ the space $\Gamma _{p,w}$ has order continuous norm if and only if $\int_{0}^{\infty }w\left( t\right) dt=\infty .$ The spaces $\Gamma _{p,w}$ were introduced by A.P. Calderón in [@Cal] in a similar way as the classical Lorentz spaces $\Lambda_{p,w}$ that is a subspace of $L^0$ with $$\left\Vert x\right\Vert _{\Lambda_{p,w}}=\left( \int_{0}^{\alpha}(x^{\ast }(t))^{p}w(t)dt\right) ^{1/p}<\infty,$$where $p\geq 1$ and the weight function $w$ is nonnegative and nonincreasing (see [@Loren]). The space $\Gamma _{p,w}$ is an interpolation space between $L^{1}$ and $L^{\infty }$ yielded by the Lions-Peetre $K$-method [@BS; @KPS]. Clearly, $\Gamma _{p,w}\subset \Lambda _{p,w}.$ The opposite inclusion $\Lambda _{p,w}\subset \Gamma _{p,w}$ is satisfied if and only if $w\in B_{p}$ (see [@KMGam]). It is worth mentioning that the spaces $\Gamma _{p,w}$ and $\Lambda _{p,w}$ are also connected by Sawyer’s result (Theorem 1 in [@Sawy]; see also [@Step]), which states that the Köthe dual of $\Lambda _{p,w}$, for $1<p<\infty $ and $\int_{0}^{\infty }w(t)dt=\infty $, coincides with the space $\Gamma _{p^{\prime },\widetilde{w}}$, where $1/p+1/p^{\prime }=1$ and $\widetilde{w}(t)=\left( t/\int_{0}^{t}w(s)ds\right) ^{p^{\prime }}w(t)$.
Let $(X,\tau)$ be a topological vector space, where $X$ is a vector space and $\tau$ is topology. A topological vector space $(X,\tau)$ is said to be a [*Fréchet space*]{} if its topology $\tau$ is induced by a translation invariant metric $d$, i.e. $d(x+z,y+z)=d(x,y)$ for any $x,y,z\in{X}$, and also $(X,d)$ is complete. Unless we say otherwise we consider a Fréchet space $X$ with a topology $\tau$ induced by a $F$-norm, i.e. a mapping ${\ensuremath{\left\Vert\cdot\right\Vert_{}^{}}}:X\rightarrow\mathbb{R}^+$ satisfying the triangle inequality and the following conditions; $(i)$ ${\ensuremath{\left\Vertx\right\Vert_{}^{}}}=0\Leftrightarrow{x=0}$, and $(ii)$ ${\ensuremath{\left\Vertx\right\Vert_{}^{}}}={\ensuremath{\left\Vert-x\right\Vert_{}^{}}}$.
Let $ X \subset L^0(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|$. A point $x\in{X}$ is said to be an $H_g$ *point* in $X$ if for any $(x_n)\subset{X}$ such that $x_n\rightarrow{x}$ globally in measure and ${\ensuremath{\left\Vertx_n\right\Vert_{}^{}}}\rightarrow{\ensuremath{\left\Vertx\right\Vert_{}^{}}}$, we have $\left\Vert x_n-x\right\Vert\rightarrow{0}$. A point $x\in{X}$ is said to be an $H_l$ *point* in $X$ if for any $(x_n)\subset{X}$ such that $x_n\rightarrow{x}$ locally in measure and ${\ensuremath{\left\Vertx_n\right\Vert_{}^{}}}\rightarrow{\ensuremath{\left\Vertx\right\Vert_{}^{}}}$, we have $\left\Vert x_n-x\right\Vert\rightarrow{0}$. We say that the space $X$ has the *Kadec-Klee property for global convergence in measure* (*Kadec-Klee property for local convergence in measure*) if each $x\in{X}$ is an $H_g$ point (an $H_l$ point) in $X$.
Let $ X $ be a Fréchet space with a $F$-norm ${\ensuremath{\left\Vert\cdot\right\Vert_{}^{}}}$ and let $ Y \subset X$ be a nonempty subset. For $x \in X$ define $$P_Y(x) = \{ y \in Y: \| x - y\| = dist(x,Y)\}.$$ Any $y \in P_Y(x)$ is called a best approximant in $Y$ to $x$ and the mapping $ x \rightarrow P_Y(x)$ is called [*the metric projection.*]{} A nonempty set $ Y \subset X$ is called [*proximinal*]{} if $ P_Y(x) \neq \emptyset$ for any $ x \in X.$ A nonempty set $Y$ is said to be a [*Chebyshev set*]{} if it is proximinal and $ P_Y(x)$ is a singleton for any $ x \in X.$ A continuous mapping $S : X \rightarrow Y$ is called a [*continuous metric selection*]{} if $Sx \in P_Y(x)$ for any $ x \in X.$
Proximinality in spaces of measurable functions
===============================================
We start with the necessary notion.
\[mucompact\] Let $(T, \Sigma, \mu)$ be a measure space and let for $ t \in T$ $$Z_t = \mathbb{R} \cup \{ - \infty, \infty \}.$$ Assume that $ Z = \Pi_{t\in T} Z_t$ is equipped with the Tychonoff topology, which will be denoted by $ \tau.$ A set $ C \subset Z$ is called $\mu$-sequentially compact if for any sequence $ \{ c_n\} \subset C$ there exists a subsesquence $ \{ c_{n_k}\}$ and $ c \in Z$ such that $c_{n_k}(t) \rightarrow c(t)$ $ \mu$-a.e..
\[counting\] If $ T = \mathbb{N}, \Sigma = 2^{\mathbb{N}}$ and $ \mu$ is the counting measure, then by diagonal argument any set $C \subset Z$ is $ \mu$-sequentially compact.
Now we present some examples of $ \mu$-sequentially compact sets. First we investigate $\mu$-sequentially compactness of a set of all increasing functions on $\mathbb{R}$.
\[increasing\] Let $ T = \mathbb{R},$ $ \mu$ be a nonatomic measure on $T$ and let $\Sigma$ denote the $\sigma$-algebra of Borel subsets of $T.$ Let $$C = \{ f: \mathbb{R} \rightarrow \mathbb{R}: f \hbox{ is increasing} \}.$$ Then $C$ is $\mu$-sequentially compact.
Fix a sequnece $ \{ c_n \} \subset C.$ Let $ A_n = cl^{\tau}\{c_k: k \geq n\}.$ Since $ Z$ is $\tau$-compact, $$A = \bigcap_{n =1}^{\infty} A_n \neq \emptyset.$$ Fix $ c \in A.$ First assume that $ |c(t)| < + \infty $ for any $ t \in \mathbb{R}.$ Since $ c_n$ are increasing it is easy to see that $c$ is also increasing. Since $ c$ is increasing, it has at most countable number points of discontinuity. Let us denote this set by $ D_o.$ Since $ \mu$ is a nonatomic, Borel maesure,$\mu(D_o) = 0.$ Now fix a countable set $ E= \{t_1,...\} \subset F= \mathbb{R} \setminus D_o$ dense in $F.$ Now we find a subsequence $ \{n_k\} $ such that $ c_{n_k}(t_i) \rightarrow c(t_i)$ for any $i \in \mathbb{N}.$ To do that, set $$U_k =\{ g \in A_k: |g(t_i) - c(t_i)| < \frac{1}{k} \hbox{ for } i=1,...,k \}.$$ By definition of $c,$ $\tau$ and $ A $ we can select a stricly increasing sequence $ \{ n_k \}$ such that $ c_{n_k} \in \{c_l: l > n_{k-1}\}$ for any $ k \in \mathbb{N}.$ Clearly, $ c_{n_k}(t_i) \rightarrow c(t_i)$ for any $i\in \mathbb{N}.$ Now we show that $c_{n_k}(t) \rightarrow c(t) \in \mathbb{R}$ for any $ t \in F.$ Assume to the contrary that there exists $ t \in F$ such that $c_{n_k}(t)$ does not converge to $c(t).$ Hence without loss of generality, passing to a subsequence if necessary, we can assume that there exists $ d>0$ such that $ |c_{n_k}(t) - c(t)| >d$ for any $ k \in \mathbb{N}.$ First assume that for infinite number of $k$ $c_{n_k}(t) - c(t) >d.$ Then for any $ s> t$ $$c_{n_k}(s) - c(t) \geq c_{n_k}(t) - c(t) > d.$$ Since $ c$ is continuous at $t$ select $ t_i \in E$ such that $ t < t_i$ and $ c(t_i) - c(t) < \frac{1}{3d}.$ Since for $ k \geq k_o $ $ |c_{n_k}(t_i) -c(t_i)| < \frac{1}{3d},$ we have for $ k \geq k_o,$ $$0< d <c_{n_k}(t) - c(t) \leq c_{n_k}(t_i) - c(t)$$ $$\leq |c_{n_k}(t_i) - c(t_i)| + |c(t_i) - c(t)| < \frac{2}{3d} < d,$$ which is a contradiction. If for infinite number of $k$ $c(t)-c_{n_k}(t) >d,$ reasoning in the same way, we get a contradiction. Now assume that there exists $ s_o \in F$ such that $ |c(s_o)| = +\infty.$ Without loss of generality we can assume that $ c(s_o) = +\infty.$ Let $$s_1 = \inf \{ t \in \mathbb{R}: c(t) = +\infty \}$$ $(s_1 = -\infty $ if for any $s\in\mathbb{R}$ $c(s)= +\infty ).$ Since $ c_{n_k}$ is increasing for any $k\in\mathbb{N}$, $c(s) = + \infty$ for any $ s > s_1.$ If $ s_1 \in \mathbb{R}, $ and $c(u) = -\infty $ for some $ u \leq s_1,$ then put $$u_1 = \sup \{ t \in \mathbb{R}: c(t) = -\infty \}.$$ If $ u_1 = s_1, $ then $c_{n_k}(t) \rightarrow +\infty $ for any $ t > s_1$ and $c_{n_k}(t) \rightarrow -\infty $ for any $ t < s_1.$ If $ u_1 < s_1$ then for any $ t \in (u_1,s_1) $ $ |c(t)| < +\infty.$ Reasoning as in the first part of the proof we get our claim.
\[convex\] Let $ T = \mathbb{R},$ $ \mu$ be a nonatomic measure on $T$ and let $ \Sigma$ denote the $\sigma$-algebra of Borel subsets of $T.$ Let $$C = \{ f: \mathbb{R} \rightarrow \mathbb{R}: f \hbox{ is convex} \}.$$ Then $C$ is $\mu$-sequentially compact.
Fix a sequence $ \{c_n \} \subset C.$ If there exists a subsequence $ \{n_k\} $ such that $ c_{n_k}$ is increasing for any $k\in\mathbb{N}$ or $c_{n_k}$ is decreasing for any $k\in\mathbb{N}$ then applying Theorem \[increasing\] we get our claim. In the opposite case, select for any $ n\in\mathbb{N}$, $ a_n \leq b_n$ such that $ c_n$ attains its global minimum at $ t \in [a_n,b_n].$ Let $$a = \inf \{ a_n : n \in \mathbb{N}\}.$$ Without loss of generality, passing to a convergent subsequence, if necessary we can assume that $ a_n \rightarrow a.$ If $ a = -\infty ,$ then for any $ l \in \mathbb{N}$ there exists $ n_o(l) \geq l$ such that $ c_n|_{(-l, +\infty)}$ are increasing for $n \geq n_o(l).$ By Theorem \[increasing\] and the diagonal argument we can select a subsequence $\{n_k\} $ such that $c_{n_k}(t) \rightarrow c(t)$ $ \mu$-a.e.. If $ a > -\infty$ then for any $ l \in \mathbb{N}$ there exists $ n_o(l) \geq l$ such that $ c_n|_{(a+\frac{1}{l}, +\infty)}$ are increasing for $n \geq n_o(l)$ and $ c_n|_{(-\infty, a-\frac{1}{l})}$ are decreasing. Reasoning as in the proof of Theorem \[increasing\] and applying the diagonal argument we can select a subsequence $\{n_k\} $ such that $c_{n_k}(t) \rightarrow c(t)$ $ \mu$-a.e.. The proof is complete.
Now we prove important for further applications
\[partition\] Let $ T = \mathbb{R}$ and let $ \mu$ a nonatomic measure on $T$ and let $ \Sigma$ the $\sigma$-algebra of Borel subsets of $T.$ Fix $ \{ t_n\}_{n \in \mathbb{Z}},$ $ t_n < t_{n+1},$ a partition of $ T.$ Assume that $ C \subset Z$ is such that for any $i \in \mathbb{N}$ $$C|_{(t_i, t_{i+1})}= \{ c|_{(t_i, t_{i+1})}:c \in C\}$$ is $\mu$-sequentially compact. Then $C$ is $\mu$-sequentially compact.
By Def. \[mucompact\], applying the diagonal argument we select a a subsequence $ \{ n_k \} $ and $ A_i \subset (t_i, t_{i+1}),$ $ \mu (A_i) =0,$ such that $ c_{n_k}(t) \rightarrow c(t)$ for any $t \in \mathbb{R} \setminus \bigcup_{i\in \mathbb{Z}} A_i.$ Since $ \mu \left( \bigcup_{i\in \mathbb{Z}} A_i \right) =0,$ we get our claim.
It is clear that in Theorem \[increasing\] and Theorem \[convex\] the set of real numbers $T$ can be replaced by any open interval $(a,b)$ where $ a, b \in \mathbb{R} \cup \{ +\infty, -\infty \}.$
Now we present other possible constructions of $\mu$-sequentially compact sets.
\[constructions\] Let $(T, \Sigma, \mu)$ be a measure space.
- If $C$ is $\mu$-sequentially compact then any nonempty set $ D \subset C$ is $\mu$-sequentially compact.
- If $ C_i$ for $ i=1,..n$ are $\mu$-sequentially compact, then $C = \bigcup_{j=1}^{n}C_i$ is $\mu$-sequentially compact.
- Let $C_1$ and $C_2$ be $\mu$-sequentially compact subsets such that $C_i\subset\Pi_{t\in T} Z_t$ for $i=1,2$, where $Z_t = [f(t), +\infty)$ and $f(t) >0$ for $t \in T.$ Then $C_1C_2 = \{ c_1c_2 : c_1 \in C_1, c_2 \in C_2 \}$ and $ aC_1 + bC_2$ with $ a, b\geq 0$ are $\mu$-sequentially compact sets.
- If $C_1$ and $C_2$ are $\mu$-sequentially compact, then $ C_1\vee{C_2} = \{c_1\vee{}c_2: c_1 \in C_1, c_2 \in C_2 \}$ and $ C_1\wedge{C_2} = \{ c_1\wedge{}c_2: c_1 \in C_1, c_2 \in C_2 \}$ are $\mu$-sequentially compact sets, where $(c_1\vee{}c_2)(t) = \max(c_1(t),c_2(t))$ and $(c_1\wedge{}c_2)(t) = \min(c_1(t),c_2(t)).$
- Let $ f : T \rightarrow \mathbb{R}_+$ be a function. If $C$ is $\mu$-sequentially compact, then $ P_f(C)$ is also $\mu$-sequentially compact, where for $ c \in C$ and $t \in T,$ $ P_f(c)(t) = c(t)$ if $ |c(t)| \leq f(t)$ and $ P_f(c)(t)= sgn(c(t))f(t)$ in the opposite case.
The proofs of conditions from $(a)$ to $(d)$ follows immediately from definition of the limit, Def. \[mucompact\] and the fact that countable union of sets of measure zero has also measure zero. To prove (e) fix a sequence $\{d_n\} \subset P_f(C).$ Let $ d_n = P_f(c_n), $ where $c_n \in C$ for any $n \in \mathbb{N}.$ Since $ C$ is $\mu$-sequentially compact, we can select a subsequence $ \{ c_{n_k}\}$ such that $ c_{n_k} \rightarrow c(t) \in \mathbb{R} \cup \{ -\infty, +\infty \}$ $ \mu$-a.e. If $ |c(t)| > f(t),$ then for $ k \geq k_o(t)$ $ |c_{n_k}(t)| > f(t), $ and consequently $$d_{n_k}(t) =P_f(c_{n_k})(t) = f(t) \rightarrow f(t).$$ If $ |c(t)| = f(t),$ then passing to a convergent subsequence, if necessary, we can assume that $ d_{n_k}(t) \rightarrow f(t)$ or $ d_{n_k}(t) \rightarrow -f(t).$ Finally, if $ |c(t)| < f(t),$ then $ d_{n_k}(t) = P_f(c_{n_k}) = c_{n_k}(t)$ for $ k \geq k_o(t)$ and consequently $ d_{n_k}(t) \rightarrow c(t).$ The proof is complete.
\[F1\] Let $(T, \Sigma, \mu)$ be a measure space such that $\mu$ is $ \sigma$-finite. Let $ T = \bigcup_{n=1}^{\infty}T_n, $ where $ T_n \in \Sigma,$ $ T_n \subset T_{n+1}$ and $ \mu(T_n) < \infty.$ Put $$L_o(T) = \{ f:T \rightarrow \mathbb{R} : f \hbox{ is } \mu-measurable \}.$$ Let $ (X , \| \cdot \|)$ be a Fréchet space such that $ X \subset L_o(T).$ Assume that for any $\{ x_n\} \subset X$ and $x \in X$ if $ \| x_n-x\| \rightarrow 0,$ then $x_n \rightarrow x$ locally in measure. If $ C \subset X$ is $\mu$-sequentially compact then $cl(C)$ is also $\mu$-sequentially compact.
Fix $ \{d_n\} \subset cl(C).$ Choose for any $n \in \mathbb{N}$ $ c_n \in C$ such that $ \| c_n -d_n \| < \frac{1}{n}.$ By our assumption $ d_n - c_n \rightarrow 0$ locally in measure. Hence for any $ i \in \mathbb{N} $ there exists a subsequence $ \{n_k\}$ (depending on $i$) such that $ c_{n_k} - d_{n_k} \rightarrow 0 $ $ \mu$-a.e. on $ T_i.$ Applying the diagonal argument we can assume that $ c_{n_k} - d_{n_k} \rightarrow 0$ $ \mu$-a.e. on $T.$ Since $C$ is $\mu$-sequentially compact, passing to a convergent subsequence if necessary, we can assume that $ c_{n_k}(t) \rightarrow c(t) \in \mathbb{R} \cup \{ -\infty, +\infty \}$ $ \mu$-a.e.. Consequently $ d_{n_k}(t) \rightarrow c(t)$ $ \mu$-a.e., as required.
Now we present the main result concerning proximanality of $ \mu$-sequentially compact sets.
\[proxym\] Let $(T, \Sigma, \mu)$ be a measure space. Let $ X \subset L_o(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|$ satisfying the Fatou property. Let $C \subset X$ be a $\mu$-sequentially compact closed subset of $ X.$ Assume that for any $ \{ c_n\} \subset C,$ if there exists $ M > 0$ such that for any $n\in\mathbb{N}$, $ \| c_n\| \leq M$ and $ c_n(t) \rightarrow c(t) \in \mathbb{R}$ $\mu$-a.e., then $ c \in C.$ Let $ f \in X \setminus C.$ Assume that there exists $ \{ c_n \} \subset C$ such that $ \| f -c_n\| \rightarrow dist(f,C)$ and $$\label{restriction}
\mu(\{ t \in T: \liminf_{n\rightarrow\infty}|c_n(t)|=+\infty \}) =0.$$ Then $$P_C(f) = \{ c \in C: \| f-c\| = dist(f,C)\} \neq \emptyset.$$
Fix $ f \in X \setminus C$ and $ \{ c_n\} \subset C$ such that $ \| f -c_n\| \rightarrow dist(f,C)$ satisfying (\[restriction\]). Since $ C$ is $\mu$-sequentially compact, by our assumptions, there exists a subsequence $\{n_k\}$ and $ c \in C$ such that $ c_{n_k}(t) \rightarrow c(t)$ $\mu$-a.e.. Hence $ |(f-c_{n_k})(t)| \rightarrow |(f-c)(t)|$ $\mu$-a.e.. Put for $ k \in \mathbb{N}$ $ g_k(t) = \inf \{ |(f-c_{n_l})(t)|, l \geq k\}.$ Notice that $g_k \uparrow |f-c|$ $\mu$-a.e. and $ \sup\{ \|g_k\|: k \in \mathbb{N} \} < \infty. $ By the Fatou property, $$\| f-c \| = \lim_k \|g_k\|.$$ Since $ X$ is a Fréchet function space, for each $ k$ $ \| g_k \| \leq \| f-c_{n_k}\|,$ which shows that $ \| f-c\| = dist(f,C).$ Hence $P_C(f) \neq \emptyset,$ as required.
Now we present some applications of Theorem \[proxym\].
Let $(T, \Sigma, \mu)$ be a $\sigma$-finite measure space. Let $ X \subset L_o(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|$ satisfying the Fatou property. Let $ C \subset X$ be such that for any $ M >0$, $C \cap B(0,M)= C\cap\{ x \in X: \|x\| \leq M \}$ is compact in the topology of local convergence in measure. Then $C$ is proximinal in $X.$
Fix $ f \in X \setminus C$ and $ \{ c_n \} \subset C$ such that $ \| f - c_n \| \rightarrow dist(f,C).$ Let $T = \bigcup_{n=1}^{\infty} T_n$ such that $ \mu(T_n) < \infty .$ By the triangle inequality, $$\|c_n \| \leq dist(f,C) + \|f\| + K\quad\textnormal{with }K>0.$$ By our assumption for any $i \in \mathbb{N}$ there exists a subsequence $ \{ n_k \} $ and $ c^i \in C|_{T_i}$ such that $ c_{n_k} \rightarrow c^i$ in measure on $ T_i.$ Passing to a convergent subsequence, if necessary, we can asume that $ c_{n_k} \rightarrow c^i$ in $ \mu$-a.e. on $ T_i$. Applying diagonal argument, we can choose a subsequence $ \{ n_k \} $ such that $ c_{n_k} \rightarrow c\in C$ $\mu$-a.e. on $T.$ By Theorem \[proxym\] $ P_C(f) \neq \emptyset,$ as required.
First we show that (\[restriction\]) is not very restrictive.
\[proxym1\] Let $(T, \Sigma, \mu)$ be a $\sigma$-finite measure space. Let $ X \subset L_o(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|$ satisfying the Fatou property. Assume that for any measurable $ A \subset T$ if $ \mu(A) >0$ then $$\lim_{k\rightarrow \infty } \| k\chi(A)\| = + \infty,$$ where $ \chi(A)$ denotes the characteristic function of $A.$ If $ \{c_n\} \subset X$ is such that $ \| c_n\| < M $ for some $ M \in \mathbb{R}$ and any $n\in\mathbb{N}$ then $ \{c_n\} $ holds (\[restriction\]).
Fix $ \{c_n\} \subset X$ is such that $ \| c_n\| < M $ for some $ M \in \mathbb{R}$ and any $n\in\mathbb{N}.$ Let $A= \{ t \in T: \liminf_{n\rightarrow\infty}|c_n(t)|=+\infty \}.$ Assume to the contrary that $ \mu(A) >0.$ Without loss of generality we can assume that $ \mu(A) < \infty $ and $c_{n}(t) \rightarrow + \infty $ for any $ t \in A.$ Put for $ n \in \mathbb{N},$ $ g_n(t) = \inf_{m\geq n} c_m(t).$ It is clear that $ g_n(t) \uparrow +\infty $ for any $t \in A.$ Fix $ k \in \mathbb{N}.$ Define for any $ n\in\mathbb{N},$ $$A_{k,n} = \{ t \in A: g_n(t) > k\}.$$ Observe that $ A_{n,k} \subset A_{n+1,k} $ and $ \bigcup_{n=1}^{\infty} A_{n,k} = A.$ Hence $ \lim_n \mu(A_{n,k}) = \mu(A).$ Notice that $$\| c_n\| \geq \| g_n \chi(A_{n,k}) \| \geq \| k \chi(A_{n,k}) \|.$$ By the Fatou property $$\lim_n \| k \chi(A_{n,k}) \| = \|k \chi(A)\|$$ and consequently $$\liminf_n \| c_n \| \geq \|k \chi(A)\|$$ for any $ k \in\mathbb{N} .$ Since $k$ was arbitrary, $\liminf_n \| c_n \| = +\infty,$ which leads to a contradiction.
A Fréchet space $(X, \| \cdot \|)$ is called $ s$-convex for some $ s \in (0,1]$ if $ \|tx\| = |t|^s\|x\|$ for any $ x \in X$ and $ t \in \mathbb{R}.$ It is clear that any $s$-convex Fréchet function space (in particular any Banach function space) satisfies the assumptions of Lemma \[proxym1\]. Also any Orlicz space generated by an increasing function $ \phi$ satisfying $ \lim_{t \rightarrow +\infty} \phi(t) = + \infty$ satisfies the assumptions of Lemma \[proxym1\].
Now, applying Theorem \[proxym\], Theorem \[increasing\] and Theorem \[convex\], we can show
\[increasing1\] Let $ T = \mathbb{R},$ $ T = \mathbb{R}_+$ or $ T = (a,b).$ Let $(T, \Sigma, \mu)$ be a $\sigma$-finite, nonatomic measure space. Let $ X \subset L_o(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|$ satisfying the Fatou property. Assume that for any measurable $ A \subset T$ if $ \mu(A) >0$ then $$\lim_{k\rightarrow \infty } \| k\chi(A)\| = + \infty,$$ where $ \chi(A)$ denotes the characteristic function of $A.$ LeT $ C $ denote the set of all increasing (decreasing, resp.) functions on $T,$ or let $ C $ denote the set of all convex (concave, resp.) functions on $T.$ If $ C \subset X$ then $C$ is proximinal in $X$ (in particular $C$ is closed in $X).$
Assume that $C$ is the set of all increasing functions on $T.$ Fix $ f \in X$ and $ \{ c_n \} \subset C$ be such that $ \| f - c_n\| \rightarrow 0.$ By Theorem \[increasing\], there exists a subsequence $ \{ n_k\}$ such that $ c_{n_k}(t) \rightarrow c(t) \in [-\infty, +\infty]$ $ \mu$-a.e.. Since the sequence $ \| c_n\|$ is bounded, by our assumptions $c(t) \in \mathbb{R}$ $\mu$-a.e.. It is easy to see that $c$ is also an increasing function. By Theorem \[proxym\] we get our result. If $C$ is the set of all decreasing functions on $T$ applying the above reasoning to $-C$ we get our claim. If $C$ is the set of all convex (concave, resp.) functions on $T$ applying Theorem \[convex\] instead of Theorem \[increasing\] and reasoning as above, we get our conclusion.
\[var1\] Applying Theorem \[increasing1\] and Theorem \[constructions\] we can find other examples of proximinal subsets of Fréchet function spaces $X$ contained in $ L_o(T).$
\[pieces\] Let $ T = (a,b)$ and let $ \{t_n\} \subset (a,b),$ $t_n < t_{n+1}$ be a partition of $(a,b).$ Let for $ n \in \mathbb{N},$ $(T_n=(t_n,t_{n+1}), \Sigma_n, \mu_n)$ be a $\sigma$-finite, nonatomic measure space. Let $$L_{o,n} = \{f: T_n \rightarrow \mathbb{R}, f \hbox{ is }\mu_n\hbox{-measurable}\}.$$ Let $ X_n \subset L_{o,n}$ be a Fréchet space equipped with an F-norm $ \| \cdot \|_n.$ Define $$X = \{ f:(a,b)\rightarrow \mathbb{R}: f|_{T_n} \in X_n\}$$ and for $ f \in X$ $$\| f\| = \sum_{n=1}^{\infty} \frac{\|f|_{T_n}\|_n}{2^n(1+\|f|_{T_n}\|_n)}.$$ Let for $ n \in \mathbb{N}$ $ C_n \subset X_n$ be a proximinal set in $ X_n.$ Put $$C = \{ f\in{X}: f|_{T_n} \in C_n\}.$$ Then $ C$ is proximinal in $X.$
First observe that for any $ f \in X$ and $ c \in C,$ $$\| f - c\| = \sum_{n=1}^{\infty} \frac{\|(f-c)|_{T_n}\|_n}{2^n(1+\|(f-c)|_{T_n}\|_n)}$$ $$\geq \sum_{n=1}^{\infty} \frac{dist_n(f|_{T_n},C_n)}{2^n(1+dist_n(f|_{T_n},C_n)))},$$ where for any $g \in X_n$ $dist_n(g,C_n)$ denotes the distance of $g$ to $C_n$ with respect to $ \| \cdot \|_n.$ For $ n \in \mathbb{N}$ fix $ c_n \in P_{C_n}(f|_{T_n}).$ Define $ c \in C$ by $c(t) = c_n(t) $ for $ t \in T_n$. Since $ T_n \cap T_m = \emptyset $ for $ n\neq m,$ it is easy to see $c$ is well defined. Notice that, $$\| f - c\| =\sum_{n=1}^{\infty} \frac{dist_n(f|_{T_n},C_n)}{2^n(1+dist_n(f|_{T_n},C_n)))},$$ which shows that $ c \in P_C(f).$ The proof is complete.
\[var2\] If for any $ n\in \mathbb{N}$ $ C_n $ denote the set of all increasing (decreasing, resp.) functions on $T_n,$ or $ C_n $ denote the set of all convex (concave, resp.) functions on $T_n$ and $X_n \subset L_{o,n}$ is a Fréchet function space satysfying the assumptions of Theorem \[increasing1\] (see also Rem. \[var1\]), then by Theorem \[increasing\] we can apply Theorem \[pieces\] to $ \{ X_n\} $ and $ \{C_n\}.$
Now we apply Theorem \[pieces\] and Remark \[var2\] to construct Chebyshev subsets in Fréchet function spaces.
\[pieces1\] Let $X_n,$ $C_n,$ $\| \cdot \|_n$ for any $n\in\mathbb{N}$ and $C$, $X$ be as in Theorem \[pieces\]. Assume that for any $ n \in \mathbb{N}$ the set $C_n$ is a Chebyshev subset of $X_n$ with respect to $ \| \cdot \|_n$. Then $C$ is a Chebyshev subset of $X$ with respect to $ \| \cdot \|.$
By Theorem \[pieces\] for any $ f \in X$ $ P_C(f) \neq \emptyset . $ Moreover for any $ c \in C,$ $$\| f - c\| = \sum_{n=1}^{\infty} \frac{dist_n(f|_{T_n},C_n)}{2^n(1+dist_n(f|_{T_n},C_n)))}.$$ This implies that for any $ n \in \mathbb{N}$ $ c|_{T_n} \in P_{C_n}(f|_{T_n}).$ Since $C_n$ is a Chebyshev set in $ X_n,$ $ c|_{T_n}$ is uniquely determined for any $ n \in \mathbb{N}.$ Hence $ c \in P_C(f)$ is uniquely determined, which shows that $C$ is a Chebyshev subset of $X.$
Now we present an example of a Chebyshev set in a Fréchet function space.
\[interesting\] Let $ T = (a,b)$ and let $ \{t_n\} \subset (a,b),$ $t_n < t_{n+1}$ be a partition of $(a,b).$ Let for $ n \in \mathbb{N},$ $(T_n=(t_n,t_{n+1}), \Sigma_n, \mu_n)$ be a $\sigma$-finite, nonatomic measure space. Let $$L_{o,n} = \{f: T_n \rightarrow \mathbb{R}, f \hbox{ is }\mu_n\hbox{-measurable}\}.$$ Let $ X_n \subset L_{o,n}$ be a strictly convex, reflexive Banach space equipped with a norm $ \| \cdot \|_n.$ Fore example $ X_n = L^{p_n}(T_n),$ where $ 1 < p_n < \infty.)$ Define $$X = \{ f:(a,b) \rightarrow \mathbb{R}: f|_{T_n} \in X_n\}$$ and for $ f \in X$ $$\| f\| = \sum_{n=1}^{\infty} \frac{\|f|_{T_n}\|_n}{2^n(1+\|f|_{T_n}\|_n)}.$$ Let for $ n \in \mathbb{N},$ $ C_n \subset X_n$ be a closed, convex subset of $ X_n.$ Put $$C = \{ f:(a,b) \rightarrow \mathbb{R}: f|_{T_n} \in C_n\}.$$ Since for any $n\in\mathbb{N},$ $ X_n$ is reflexive and $C_n$ is closed and convex it follows that $C_n$ is a Chebyshev subset of $X_n$ with respect to $\|\cdot\|_n$. Hence $C_n$ is a Chebyshev set with respect to $F$-norm $\frac{\|\cdot\|_n}{1+\|\cdot\|_n}$. By Theorem \[pieces1\] $C$ is a Chebyshev subset of $X.$ In particular, $ C_n $ can be the set of all increasing (decreasing, resp.) functions on $T_n$ or $ C_n $ can be the set of all convex (concave, resp.) functions on $T_n.$ It is clear that $C$ is a Chebyshev set.
Now we research $\mu$-compactness in Banach function spaces.
A nonempty subset $C\subset{Z}$ is said to be $\mu$-compact if and only if for any sequence $(c_n)\subset{C}$ there exist $(n_k)\subset\mathbb{N}$ and $c\in{C}$ such that $c_{n_k}$ converges to $c$ in measure $\mu.$
Let $C$ be a nonempty subset of $L^*=\{x^*:x\in{L^0}\}$, where $x^*$ is a decreasing rearrangement of $x\in{L^0}$. Then, by Theorem \[increasing\] and Theorem \[constructions\] we get immediately $C$ is $\mu$-sequentially compact. Additionally, if we assume that $c(\infty)=0$ for any $c\in{C}$, then $C$ is $\mu$-compact. Indeed, since $C$ is $\mu$-sequentially compact for any sequence $(c_n)\subset{C}$ there exist $(n_k)\subset\mathbb{N}$ and $c\in{C}$ such that $c_{n_k}$ converges to $c$ $m$-a.e. Hence, $c_{n_k}$ converges to $c$ locally in measure and by assumption that each element $a\in C$, $a=a^*$ and $a(\infty)=0$ it follows that $c_{n_k}$ converges to $c$ in measure.
Now we present properties of $\mu$-compact sets. The proof of the below result follows immediately by definition of $\mu$-compactness and similarly as in case of Theorem \[constructions\].
\[constructions2\] Let $(T, \Sigma, \mu)$ be a measure space.
- If $C$ is $\mu$-compact then any nonempty set $ D \subset C$ is $\mu$-compact.
- If $ C_i$ for $ i=1,..n$ are $\mu$-compact, then $C = \bigcup_{j=1}^{n}C_i$ is $\mu$-compact.
- Let $C_1$ and $C_2$ be $\mu$-compact subsets such that $C_i\subset\Pi_{t\in T} Z_t$ for $i=1,2$, where $Z_t = [f(t), +\infty)$ and $f(t) >0$ for $t \in T.$ Then $ aC_1 + bC_2$ with $ a, b\geq 0$ are $\mu$-compact sets. Additionally, if $\mu(T)<\infty$ then $C_1C_2 = \{ c_1c_2 : c_1 \in C_1, c_2 \in C_2 \}$ is $\mu$-compact.
- If $C_1$ and $C_2$ are $\mu$-compact, then $ C_1\vee{C_2} = \{c_1\vee{}c_2: c_1 \in C_1, c_2 \in C_2 \}$ and $ C_1\wedge{C_2} = \{ c_1\wedge{}c_2: c_1 \in C_1, c_2 \in C_2 \}$ are $\mu$-compact sets, where $(c_1\vee{}c_2)(t) = \max(c_1(t),c_2(t))$ and $(c_1\wedge{}c_2)(t) = \min(c_1(t),c_2(t)).$
- Let $ f : T \rightarrow \mathbb{R}_+$ be a function. If $C$ is $\mu$-compact, then $ P_f(C)$ is also $\mu$-compact, where for $ c \in C$ and $t \in T,$ $P_f(c)(t) = sgn(c(t))\{|c(t)|\wedge{f(t)}\}$.
\[thm:app-com\] Let $E$ be a symmetric Banach function space with semi-Fatou property and $H_g$ property and let $C\subset{E}$ be nonempty $\mu$-compact subset of $E$. Then for every $x\in{E}\setminus{C}$ and for any minimizing sequence $(c_n)\subset{C}$ of $\operatorname{dist}(x,C)$ there exist a subsequence $(c_{n_k})\subset(c_n)$ and $c\in{P}_{C}(x)$ such that $${\ensuremath{\left\Vertc_{n_k}-c\right\Vert_{}^{}}}\rightarrow{0}.$$
Assume that $(c_n)\subset{C}$ is a minimizing sequence of $\operatorname{dist}(x,C)>0$, i.e. $$\label{equ:thm:H_g:in:approx}
\lim_{n\rightarrow\infty}{\ensuremath{\left\Vertx-c_n\right\Vert_{}^{}}}=\inf_{d\in{C}}{\ensuremath{\left\Vertx-d\right\Vert_{}^{}}}>0.$$ Since $C$ is $\mu$-compact there exist $(c_{n_k})\subset(c_n)$ and $c\in{C}$ such that $c_{n_k}$ converges to $c$ globally in measure. Therefore, $x-c_{n_k}$ converges to $x-c$ globally in measure. Consequently, by assumption that $E$ has $H_g$ property and by Lemma 3.8 [@CieKolPlu] it follows that $${\ensuremath{\left\Vertx-c\right\Vert_{}^{}}}\leq\liminf_{k\rightarrow\infty}{\ensuremath{\left\Vertx-c_{n_k}\right\Vert_{}^{}}}.$$ Hence, by condition we obtain $$\begin{aligned}
\operatorname{dist}(x,C)\leq{\ensuremath{\left\Vertx-c\right\Vert_{}^{}}}\leq\liminf_{k\rightarrow\infty}{\ensuremath{\left\Vertx-c_{n_k}\right\Vert_{}^{}}}\leq\lim_{k\rightarrow\infty}{\ensuremath{\left\Vertx-c_{n_k}\right\Vert_{}^{}}}=\operatorname{dist}(x,C).
\end{aligned}$$ Finally, according to assumption that $E$ has $H_g$ property we get $${\ensuremath{\left\Vertc_{n_k}-c\right\Vert_{}^{}}}\rightarrow{0}.$$
Let $\phi$ be a concave increasing function such that $\phi(0^+)=0$ and let $C\subset{\Lambda_\phi}$ be a nonempty $\mu$-compact set and $x\in\Lambda_\phi\setminus{C}$. Then for any minimizing sequence $(c_n)\subset{C}$ of $\operatorname{dist}(C,x)$ there exist $(n_k)\subset\mathbb{N}$ and $c\in{P_C(x)}$ such that $${\ensuremath{\left\Vertc_{n_k}-c\right\Vert_{\Lambda_{\phi}}^{}}}\rightarrow{0}.$$
By Corollary 1.3 [@ChDSS] it follows that $\lambda_\phi$ has the Kadec Klee property for global convergence in measure. Since $\Lambda_{\phi}$ satisfies Fatou property [@Loren] we have $\Lambda_\phi$ has also semi-Fatou property and by Theorem \[thm:app-com\] we complete the proof.
Let $0<p<\infty$ and $w\geq{0}$ be weight function on $I$. Let $C\subset{\Lambda_\phi}$ be a nonempty $\mu$-compact set and $x\in\Gamma_{p,w}\setminus{C}$. Then for any minimizing sequence $(c_n)\subset{C}$ of $\operatorname{dist}(C,x)$ there exist $(n_k)\subset\mathbb{N}$ and $c\in{P_C(x)}$ such that $${\ensuremath{\left\Vertc_{n_k}-c\right\Vert_{\Gamma_{p,w}}^{}}}\rightarrow{0}.$$
Immediately, by Theorem 4.1 [@CieKolPlu] we obtain that $\Gamma_{p,w}$ has the Kadec Klee property for global convergence in measure. Furthermore, it is well known that $\Gamma_{p,w}$ has Fatou propert and so it has also semi-Fatou property. Hence, by Theorem \[thm:app-com\] it follows that for any minimizing sequence $(c_n)\subset{C}$ there exist $(n_k)\subset\mathbb{N}$ and $c\in{P_C(x)}$ such that ${\ensuremath{\left\Vertc_{n_k}-c\right\Vert_{\Gamma_{p,w}}^{}}}\rightarrow{0}$.
Property (S) and continuity of metric projection operator
=========================================================
First we introduce the property $(S).$
\[Kadec\] Let $(T, \Sigma, \mu)$ be a measure space. Let $ X \subset L_o(T)$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|.$ It is said that $X$ has property [**(S)**]{} if there exists $ 0<p < \infty $ such that for any $ \{ c_n \} \subset X,$ $ c \in X$ and $f \in X,$ $$\| f-c_n\|^p = \|f-c\|^p + \|c_n-c\|^p + o(1)$$ provided $ c_n \rightarrow c$ locally in measure.
Now we present an some Banach function spaces with property $(S)$.
Let $0<\lambda<\infty$ and $\phi$ be concave increasing function on $[0,\alpha)$ with $\phi(0^+)=0$. Replacing $x$ by $f-c$ and $y_n$ by $c_n-c$ in Proposition 1.2 [@KPS] we get the Lorentz space $\Lambda_\phi$ with the property $(S)$ equipped with the norm given by $${\ensuremath{\left\Vertx\right\Vert_{\Lambda_\phi}^{}}}=\int_{0}^{\alpha}{x^*(t)\phi'(t)}dt.$$
Let $p\geq{}1$ and $w$ be a nonegative weight function on $[0,\alpha)$ with $0<\alpha<\infty$. Taking $x=f-c$ and $y_n=c_n-c$ in the proof of Theorem 4.1 [@CieKolPlu] we can easily observe that the Lorentz space $\Gamma_{p,w}$ holds the property $(S)$.
The next examples show some special Fréchet function spaces which satisfy the property $(S)$.
\[one\] Let $(T, \Sigma, \mu)$ be a finite measure space. For $ f \in L_o(T) $ define $$\| f\| = \int_T \frac{|f(t)|}{1+|f(t)|} d\mu(t).$$ It is clear that for any $ \{ c_n \} \subset X$ and $ c \in X$, $ \| c_n - c\| \rightarrow 0$ if and only if $c_n \rightarrow c$ locally in measure. Hence in this case property (S) is satisfied.
\[two\] Let $(T, \Sigma, \mu)$ be a $\sigma$-finite measure space. Assume that $ T = \bigcup_{n=1}^\infty T_n, $ where $ T_n \subset T_{n+1}$ and $ \mu(T_n) < \infty. $ Define for $f \in L_o(T)$ $$\|f\| = \sum_{n=1}^{\infty} \frac{1}{2^n} \int_{T_n} \frac{|f(t)|}{1+|f(t)|} d\mu(t).$$ It is clear that for any $ \{ c_n \} \subset X$ and $ c \in X$, $ \| c_n - c\| \rightarrow 0$ if and only if $c_n \rightarrow c$ locally in measure. Hence in this case property (S) is satisfied.
The next theorem shows a large class of Fréchet function spaces satisfying the Kadec-Klee property for local convergence in measure.
\[FS1\] Let for $ n \in \mathbb{N}$ $(T_n, \Sigma_n, \mu_n)$ be a sequence of measure spaces such that $ \mu(T_n) < \infty $ and $ T_n \cap T_m = \emptyset $ if $ m \neq n.$ Let $ T = \bigcup_{n=1}^{\infty} T_n$ be equipped with the measure $$\mu(S) = \sum_{n=1}^{\infty} \mu_n(S\cap T_n).$$ Let $$L_{o,n} = \{f: T_n \rightarrow \mathbb{R}, f \hbox{ is }\mu_n\hbox{-measurable}\}.$$ Let $ X_n \subset L_{o,n}$ be a Fréchet function space equipped with an F-norm $ \| \cdot \|_n$ with the Fatou property. Define $$X = \bigoplus_{n=1}^{\infty} X_n \subset L_o(T)$$ and for $ f= (f_1,f_2,...) \in X$ $$\| f\| = \sum_{n=1}^{\infty} \frac{\|f_n\|_n}{2^n(1+\|f_n \|_n)}.$$ Assume that for any $n \in \mathbb{N},$ $ \| \cdot \|_n$ satisfies property $(S)$ with $p_n \in (1, +\infty).$ Then $(X, \| \cdot \|)$ satisfies the Kadec-Klee property for local convergence in measure.
Fix $ c =(c_1,c_2, ...) \in X$ $f= (f_1,f_2,...) \in X$ and $ \{ c_n\} \subset X$ such that $ c_n \rightarrow c$ locally in measure. Assume to the contrary that $ \| f-c_n \| \rightarrow \| f-c\| $ and $ \| c_n -c\|\nrightarrow{0}$. Passing to a subsequence and relabelling if necessary, we can assume $$\label{d}
\| c_n -c\| > d>0$$ for any $ n \in \mathbb{N}.$ Since for any $j \in \mathbb{N},$ $ \mu(T_j) < \infty,$ again passing to a convergent subsequence, if necessary, and applying diagonal argument we can assume that $ c_n(t) \rightarrow c(t)$ $\mu$-a.e. Hence, $ |f(t) -c_n(t)| \rightarrow |f(t)-c(t)|$ $\mu$-a.e. By the Fatou property, for any $j \in \mathbb{N}$ we get $$\| f_j - (c)_j\|_j \leq \liminf_n \| f_j - (c_n)_j\|_j.$$ Since $ \| f-c_n\| \rightarrow \|f-c\|,$ by the above inequality, for any $j \in \mathbb{N},$ $$\lim_n \| f_j - (c_n)_j\|_j =\| f_j - c_j\|_j.$$ By property (S) applied to $ \| \cdot \|_j,$ $ (c_n)_j , c_j$ and $ f_j$ we get $$\| (c_n)_j - c_j\|_j \rightarrow_n 0.$$ Consequently $ \| c_n - c\| \rightarrow_n 0,$ which is a contradiction with (\[d\]).
Now we apply the Kadec-Klee property for local convergence in measure to the problem of continuity of the metric projection operator.
\[projop\] Let $ X$ be a Fréchet space and let $ C\subset X$ be a proximinal subset of $X.$ Let for $ x \in X, $ $$P_C(x) = \{ c \in C: \| x-c\| = dist(x,C)\}.$$ The metric projection operator is a mapping $P$ from $X$ to $ 2^C$ defined by $ Pf = P_C(f).$
\[cont1\] Let $\mu$ be a $\sigma$-finite measure and let $(X,\| \cdot \|) \subset L_o(T) = L_o(T, \Sigma, \mu)$ be a Fréchet function space with the Fatou property and $X\in(H_l).$ Let $C \subset X$ be a proximinal subset of $X.$ Assume for any $g \in X$ and $ c_n \subset C$ if $ \| g-c_n\| \rightarrow dist(g,C)$ then there exists $c \in C$ and a subsequence $ \{ c_{n_k}\}$ such that $ c_{n_k} \rightarrow c$ locally in measure. Let $ \{ f_n\} \subset X$, $f \in X$ and $ \| f_n-f\| \rightarrow 0.$ Then $$\lim_n (\sup\{ dist(c, P_C(f)): c \in P_C(f_n)\}) =0.$$
Assume to the contrary that there exist $ \{ f_n\} \subset X$ and $f \in X$ such that $ \| f_n-f\| \rightarrow 0$ and $$\limsup_n (\sup\{ dist(c, P_C(f)): c \in P_C(f_n)\}) >0.$$ Passing to a subsequence if necessary, we can assume that $$\sup\{ dist(c, P_C(f)): c \in P_C(f_n)\} > d >0$$ for any $n \in \mathbb{N}.$ For any $ n \in \mathbb{N}$ select $ c_n \in P_C(f_n)$ such that $$dist(c_n, P_C(f)) > \frac{d}{2}.$$ Notice that $$dist(f,C) \leq \| f-c_n\| \leq (dist(f_n,C) + \|f_n-f\|) \rightarrow_n dist(f,C).$$ Hence, since $ \| f_n-f\| \rightarrow 0,$ $ \| f-c_n\| \rightarrow dist(f,C).$ By our assumptions there exists a subsequence $ \{ n_k\}$ and $ c \in C,$ such that $ c_{n_k} \rightarrow c$ locally in measure. Since $ \mu$ is $\sigma$-finite, applying the diagonal argument and passing to a convergent subsequence, if necessary we can assume that $ c_{n_k}(t) \rightarrow c(t)$ $\mu$-a.e.. By the Fatou property, $$\begin{aligned}
\|f-c\| &\leq \liminf_{k\rightarrow\infty} \| f -c_{n_k}\| \leq \liminf_{k\rightarrow\infty} \| f-f_{n_k}\| + \lim_{k\rightarrow\infty} dist(f_{n_k},C)\\
&= dist(f,C) \leq \|f-c\|.\end{aligned}$$ Consequently, $ c \in P_C(f)$ and $ \|f-c_{n_k} \| \rightarrow \| f-c\|.$ By assumption that $X\in(H_l)$, it follows that $ \| c_{n_k} -c\| \rightarrow 0.$ Hence, $$\frac{d}{2} < dist(c_{n_k}, P_C(f)) \leq \| c_{n_k} - c\| \rightarrow 0,$$ which is a contradiction.
\[cont2\] Let $X$, $ \{f_n\}\subset X,$ $f \in X$ and $C \subset X$ be such as in Theorem \[cont1\]. Assume that $ P_C(f)$ is a singleton. Then $$\lim_n d_H (P_C(f), P_C(f_n)) =0,$$ where $d_H$ denote the Hausdorff distance between $ P_C(f)$ and $P_{C}(f_n).$
By Theorem \[cont1\] $$\lim_n (\sup\{ dist(d, P_C(f)): d \in P_C(f_n)\}) =0.$$ We need to show that $$\lim_n (\sup\{ dist(c, P_{C}(f_n)): c \in P_C(f)\}) =0.$$ By our assumption $ P_C(f) = \{ c \}.$ Hence, the above inequality reduces to $$\lim_{n\rightarrow\infty} dist(c, P_{C}(f_n))=0.$$ Assume that this is not true. Reasoning as in Theorem \[cont1\] we can choose a subsequence $\{ c_{n_k} \} \subset C$ such that $ c_{n_k} \in P_C(f_{n_k})$ and $d>0$ such that for any $k\in\mathbb{N}$, $$\| c_{n_k} - c\| > \frac{d}{2}.$$ Notice that $$dist(f,C) \leq \|f-c_{n_k}\| \leq dist(f_{n_k},C) + \| f- f_{n_k}\| .$$ Hence, $$\| f-c_{n_k}\| \rightarrow dist(f,C).$$ Reasoning as in Theorem \[cont1\] passing to a convergent subsequence if necessary we can assume that $ \| c_{n_k} - u\| \rightarrow 0$ for some $ u \in C.$ It is clear that $ u \in P_C(f).$ Since $ P_C(f) = \{ c \},$ $u=c,$ which leads to a contradiction.
Let $X = \mathbb{R}$ and let $ C= \{ -1,1\}.$ Observe that $ P_C(0) = C$ and $ P_C(\frac{1}{n}) = \{1 \}.$ Hence $$\sup\{ dist(c, P_{C}(\frac{1}{n})): c \in P_C(0)\} = |-1-1| =2,$$ which shows that the assumption that $ P_C(f)$ is a singleton in Theorem \[cont2\] is essential.
As an immediate consequence of Theorem \[cont2\] we get the following result.
\[cont3\] Let $X$, $\{f_n\}\subset X,$ $f \in X$ and $C \subset X$ be such as in Theorem \[cont1\]. Assume that for any $g \in X$ $ P_C(g)$ is a singleton. Then the mapping $ g \rightarrow P_C(g) \in C$ is continuous.
Now we present in view of Theorem \[cont1\] the full criteria of the continuity of the metric projection in Fréchet function space with the Kadec-Klee property for local convergence in measure with respect to the Hausdorff distance.
\[cont4\] Let $X$ , $ \{f_n\}\subset X,$ $f \in X$ and $C \subset X$ be such as in Theorem \[cont1\]. Then $$\lim_n d_H (P_C(f), P_C(f_n)) =0,$$ if and only if for any $\epsilon>0$ there exists $\delta_\epsilon>0$ such that for any $c\in P_C(f)$ if $ \| f-f_n\| <\delta_\epsilon$ then $ P_C(f_n) \cap B(c,\epsilon) \neq \emptyset $, where $ B(c,\epsilon) = \{ x \in X: \|x-c\| \leq\epsilon\}.$
Let $\epsilon>0$. Then there exists $\delta_\epsilon>0$ such that for any $c\in P_C(f)$ we have $dist(c,P_C(f_n))\leq\epsilon$ whenever $ \| f-f_n\| <\delta_\epsilon$. Since $\|f-f_n\|\rightarrow{0}$, there exists $n_0\in\mathbb{N}$ such that for any $n\geq{n_0}$ we get $\|f-f_n\|<\delta_\epsilon$ and consequently $dist(c,P_C(f_n))\leq\epsilon$ for all $c\in{P_C(f)}$. Therefore, for all $n\geq{n_0}$ we get $$\sup\{dist(c,P_C(f_n)):c\in P_C(f)\}\leq\epsilon.$$ Thus, by Theorem \[cont1\] we conclude that $$\lim_{n\rightarrow\infty}d_H(P_C(f),P_C(f_n))=0.$$ The converse implication follows immediately by definition of the Hausdorff distance. Indeed, we have $$\lim_{n\rightarrow\infty}\sup\{dist(c,P_C(f_n)):c\in P_C(f)\}=0.$$ Now, taking $\epsilon>0$ we may find $n_\epsilon\in\mathbb{N}$ such that for any $n\geq{n_\epsilon}$ and $c\in{P_C(f)}$ we get $$dist(c,P_C(f_n))\leq\epsilon.$$ Consequently, by assumption that ${\ensuremath{\left\Vertf-f_n\right\Vert_{}^{}}}\rightarrow{0}$ there exists $\delta_\epsilon>0$ such that if ${\ensuremath{\left\Vertf-f_n\right\Vert_{}^{}}}<\delta_\epsilon$ then $n\geq{n_\epsilon}$ and for all $c\in{P_C(f)}$ we obtain $$B(c,\epsilon)\cap{}P_C(f_n))\neq\emptyset.$$
Properties of metric projection onto one-dimensional subspaces of sequence Lorentz spaces
=========================================================================================
First we recall a well-known result for a sake of completeness and convenience of the reader.
\[classical\] Let $X$ be a Banach space, $ x \in X$ and let $ Y \subset X$ be a linear subspace. Assume that $dist(x,Y) >0.$ Then $y \in P_Y(x)$ if and only if there exists $ f \in S_{X^*}$ such that $f(x-y) = dist(x,Y)$ and $f(y) =0.$
If $Y$ is a finite-dimensional subspace of a real Banach space $X$ a stronger version of Theorem \[classical\] holds.
\[singer\] (see [@Singer] , p.170) Let $Y$ be an n-dimensional subspace of a real Banach space $ X$ an let $ x \in X \setminus Y.$ Then $y \in P_Y(x)$ if and only if there exists $ j \in \{ 1,...,n+1\}, $ $ f_1,...,f_j \in ext(S(X^*)),$ positive numbers $ \lambda_1,...,\lambda_j$ such that for $i\in\{1,...,j\}$, $$f_i(x-y) = \| x-y\|$$ and $$\sum_{i=1}^j (\lambda_i f^i)_Y =0.$$
By ([@SudWoj], Th. 5.8) we can immediately deduce
\[sudwoj\] Let $Y$ be an n-dimensional subspace of a real Banach space $ X$ an let $ x \in X \setminus Y$ and $y_o \in P_Y(x).$ If the minimal number $j$ satisfying the requirements of Theorem \[singer\] is equal to $ n+1$ then there there exists $ r > 0$ such that for any $y \in Y$ $$\| x -y \| \geq \|x-y_o\| + r\|y -y_o\|.$$ By the Freud Theorem (see [@Chen], p. 82) if $x_n \in X$ and $ \|x_n-x\| \rightarrow 0,$ and $ y_n \in P_{Y}(x_n),$ then $$\| y_n -y\| \leq \frac{2\|x_n-x\|}{r}$$
In the sequel we also need
\[deutsch\] (see [@Deutsch], Th. 4.5). Let $X$ be a Banach space and $Y \subset X$ a one-dimensional subspace. Then $Y$ does not admit a continuous metric selection if and only if there exists $ x \in X$ such that $ 0 \in P_Y(x)$ and disjoint compact intervals $ I_1, I_2$ and two sequences $ \{ x_n\} $ and $ \{ y_n\}$ converging to $x$ such that for any $ n \in \mathbb{N}$ $P_Y(x_n) \subset I_1$ and $P_Y(w_n) \subset I_2.$
\[general\] Let $X$ and $Y$ be two Banach spaces and let $X_1 \subset X$ be a closed subspace of $X$ and $Y_1 \subset Y$ be a closed subspace of $Y.$ Assume that $ T: X \rightarrow Y$ is a linear surjective isometry such that $ T(X_1) = Y_1.$ Then for any $ x \in X,$ $ P_{Y_1}(Tx) = T(P_{X_1}(x)).$
Follows immediately by definitions of $P_Y$ and $T.$
\[Czebyszew\] Let $Y$ be a closed subset of a Banach space $X$ such that $dim(Span(Y))$ is finite. Assume $x\in X$ and $P_Y(x) =\{y\}.$ If $ x_n \in X$ and $ \|x_n-x\| \rightarrow 0,$ then for any $ y_n \in P_Y(x_n),$ we have $\|y_n-y\| \rightarrow 0.$
Assume to the contrary, that there exist $ \{ x_n\} \subset X,$ $y_n \in P_Y(x_n)$ and $x \in X$ such that $ P_Y(x) =\{y\},$ $ x_n \rightarrow x$ and $\{y_n\}$ does not converge to $y.$ Passinng to a subsequence, if necessary, we can assume that there exists $ d> 0$ such that $ \| y_n -y\| > d.$ Since $ x_n \rightarrow x,$ $\{ y_n\} $ is bounded. Since $dim(Span(Y))<\infty$ and $Y$ is closed, passing to a convergent subsequence, if necessary, we can assume that $ y_n \rightarrow z \in Y.$ By the continuity of the function $ x \rightarrow dist(x,Y)$ we get $ \| x -z\| = dist(x,Y).$ Since $P_Y(x) = \{y\},$ $ y = z,$ which leads to a contradiction.
Observe that Lemma \[Czebyszew\] cannot be generalized to the case of $Y$ being closed subspaces of Banach spaces. In fact, in [@Brown1] A. L. Brown constructed a strictly convex reflexive space having a Chebyshev subspace $Y$ of codimension two such that the metric projection operator $ x \rightarrow P_Y(x) \in Y$ is not continuous. (Since $ Y$ is a Chebyshev subspace we treat $ P_Y(x)$ as an element from $Y.)$
\[Izbior\] Let $ X$ be a Banach space and let $ Y$ be a proper $n$-dimensional subspace of $X.$ Assume that $Y $ is not contained in the intersection of kernels of $n$ linearly independent functionals from the set $ext(S(X^*)$, where $ext(S(X^*)$ denotes the set of all extreme point of the unit sphere in $X^*.$ Then for any $x \in X$ there exists $ y_x \in Y$ and $ r_x >0$ depending only on $x$ such that for any $y \in Y,$ $$\| x - y \| \geq \| x - y_x\| + r\|y-y_x\|.$$
Follows from Theorem \[sudwoj\].
Now, motivated by ([@Deutsch], Th. 6.3 and Cor. 6.6) we restrict ourselves to the case of one-dimensional subspaces of Lorentz sequence spaces. Let $w$ be a weight function, i.e. $w=(w(1),..w(n),...)$ is a decreasing sequence of positive numbers such that $$\label{sum}
\sum_{n=1}^{\infty} w(n) = + \infty.$$ The Lorentz sequence space $d(w,1)$ is the collection of all real sequences $ x=\{ x(n)\},$ such that $$\| x\|_{w,1} = \sum_{n=1}^{\infty} x^*(n)w(n) < \infty,$$ where $x^*$ denotes the decreasing rearrangement of $x.$ Notice that by (\[sum\]), $$\label{c0}
d(w,1) \subset c_o.$$ It is well-known that $d(w,1)$ is a Banach space under the norm $ \| \cdot \|_{w,1}.$ The Marcinkiewicz sequence space $d^*(w,1)$ consists of all real sequences $x = \{ x(n)\}$ such that $$\|x\|_W =\sup_n \frac{ \sum_{j=1}^n x^*(j)}{W(n)} < \infty,$$ where $W(n) = \sum_{j=1}^n w(j).$ It is well-known that $d^*(w,1)$ is the dual space of $d(w,1).$ For more details of the Lorentz and Marcinkiewicz spaces the reader is referred to [@KamLee; @KPS; @LinTza]. Now we characterize $1$-dimensional subspaces of $d(w,1)$ under which there exists a continuous metric selection. By Lemma \[Czebyszew\] and Lemma \[Izbior\] if $Y = span[y]$ is not contained in the kernel of an extreme functional from $S_{d(w,1)^*}$ then there exists a continuous metric selection onto $Y.$ Hence to the end of this section we assume that there exists $ f \in ext(S_{d^*(w,1)})$ such that $ Y \subset ker(f).$ By the proof of \[[@HKL], Th. 2.2\] we have
\[hkl\] If $ f \in ext(S_{d^*(w,1)})$ then $ f^* =w.$ Moreover if $ \lim_n w(n) =0, $ then $ f \in ext(S_{d^*(w,1)})$ if and only if $ f^* =w.$
We start with a crucial for our considerations lemma.
\[selection\] Assume that $w$ is a strictly decreasing weight function. Let $$y =(y(1),...y(n),...) \in d(w,1) \setminus \{0\}$$ and let $ Y = span[y] \subset ker(w).$ Assume that there exists $ x=x^* \in d(w,1)$ such that $P_Y(x) = [-1,1]y.$ Set $ z = (z(1),...,z(n),...), $ where $$\label{szereg}
z(j) = \sum_{l=j}^{\infty} |y(l+1) -y(l)|.$$ Then $z \in d(w,1),$ (in particular $z(j) \in \mathbb{R}$ for any $ j \in \mathbb{N}$) and $$\label{zj}
z(j) \geq |y(j)| \hbox{ for } j \in \mathbb{N}.$$ Moreover, if there exists a subsequence $ (n_k)$ such that $$\label{strict}
(y(n_k) - y(n_k +1))(y(n_{k+1}) - y(n_{k+1}+1)<0.$$ then there exist compact intervals $ I_1,$ $I_2,$ with $ I_1 \cap I_2 = \emptyset$ and two sequences $ (x^k)$ and $(w^k)$ converging to $z$ such that for any $ k \in \mathbb{N}$ $P_Y(x^k) \subset I_1$ and $ P_Y(w^k) \subset I_2.$ If (\[strict\]) is not satisfied, $ x=x^* \in d(w,1)$ and $P_Y(x) = [-1,1]y,$ then for any sequences $ (x^k)$ and $(w^k)$ converging to $x$ and any compact intervals $ I_1,$ $I_2,$ satisfying $P_Y(x^k) \subset I_1$ and $ P_Y(w^k) \subset I_2$ for any $ k \in \mathbb{N},$ we have $ I_1 \cap I_2 \neq \emptyset.$
Since $x=x^*,$ $ \| x\|_{w,1} = \sum_{j=1}^\infty w(j)x(j).$ Since $\sum_{j=1}^{\infty} w(j)y(j)=0,$ and $ P_Y(x) = [-1,1]y,$ for any $ a \in [-1,1]$ $$\| x +ay\|_{w,1} = dist(x,Y) = \|x\|_{w,1} = \sum_{j=1}^\infty w(j)x(j) = \sum_{j=1}^\infty w(j)(x(j)+ay(j)).$$ In particular, since w is strictly decreasing, $ (x+y)^* = x+y$ and $ (x-y)^* = x-y.$ Hence, for any $j \in \mathbb{N},$ $$\label{ineq1}
x(j) \geq |y(j)|,$$ and $$\label{ineq2}
x(j) - x(j+1) \geq |y(j) - y(j+1)|.$$ By (\[ineq2\]), for $ m < n,$ $$\label{ineq3}
x(m) - x(n) \geq \sum_{j=m}^{n-1} |y(j+1) - y(j)|.$$ Since $ d(w,1) \subset c_o,$ by (\[ineq3\]), for any $ m \in \mathbb{N},$ $$x(m) \geq \sum_{j=m}^{\infty} |y(j+1) - y(j)| =z(m).$$ Since $ x \in d(w,1), $ $z =(z(1),...,)$ defined by (\[szereg\]) also belongs to $ d(w,1).$ Now we show that $z(j) \geq |y(j)|$ for $ j \in \mathbb{N}.$ Assume to the contrary that $z(j_o)<|y(j_o)|$ for some $j_o\in\mathbb{N}.$ First suppose that $ |y(j_o)| = y(j_o).$ Since $ z(j) - z(j+1) = |y(j)-y(j+1)|,$ for any $ j > j_o$ we get $$0> z(j_o) -y(j_o) \geq z(j) - y(j).$$ Hence for $j\geq j_o,$ $$| z(j) -y(j)| \geq |z(j_o) - y(j_o)| > 0.$$ This shows that $ z-y \notin c_o.$ But $ z-y \in d(w,1) \subset c_o$, so a contradiction. If $ y(j_o)= - |y(j_o)|,$ the proof goes in the same way. Now assume that there exists a subsequence $ (n_k)$ satisfying (\[strict\]). Define for $k \in \mathbb{N},$ $$\label{discont}
z^k =(z(1),...,z(k-1),z(k)-2|y(k)-y(k+1)|,z(k+1),...,),$$ $$\label{xk}
x^k = z^{n_{2k}}$$ and $$\label{wk}
w^k = z^{n_{2k+1}}$$ Without loss of generality we can assume that $ y(n_{2k}) > y(n_{2k}+1).$ Now we claim that $-y \in P_Y(x^k)$ and $ ay \notin P_Y(x^k)$ for $ a >- 1.$ Notice that by (\[zj\]) $$x^k(j) + y(j) = z^{n_{2k}}(j) + y(j) \geq 0$$ for $ j \neq n_{2k}.$ By definition of $z,$ for $j < n_{2k}-1$ and $j\geq n_{2k}+1$ $$x^k(j) + y(j)= z^{n_{2k}}(j) + y(j) \geq z^{n_{2k}}(j+1) + y(j+1) = x^k(j+1)+ y(j+1).$$ Moreover, we have $$\begin{aligned}
x^k(n_{2k}-1) + y(n_{2k}-1) &= z^{n_{2k}}(n_{2k}-1) + y(n_{2k}-1) = z(n_{2k}-1)+y(n_{2k}-1)\\
&\geq z(n_{2k}) + y(n_{2k}) \geq z^{n_{2k}}(n_{2k}) + y(n_{2k}) \\
&= x^k(n_{2k}) + y(n_{2k}).\end{aligned}$$ Finally, notice that for any $ a \in [-1,1]$ $$\begin{aligned}
x^{k}(n_{2k}) + ay({n_{2k}})&=z^{n_{2k}}(n_{2k}) + ay(n_{2k})\geq z^{n_{2k}}(n_{2k}+1) + ay(n_{2k}+1)\\
&=x^{k}(n_{2k}+1) + ay(n_{2k}+1)\end{aligned}$$ if and only if $$z(n_{2k}+1) - |y(n_{2k}+1)-y(n_{2k})| + ay(n_{2k}) \geq z(n_{2k}+1) +ay(n_{2k}+1).$$ Since $ y(n_{2k}) > y(n_{2k}+1)$ the last inequality is satisfied only for $ a=1.$ Hence, $$\| x^{k} + y\|_{w,1} = \| z^{n_{2k}} + y\|_{w,1} = \sum_{j=1}^{\infty} (z^{n_{2k}}(j) - (-y(j)))w(j).$$ Since $ \sum_{j=1}^{\infty}y(j)w(j)=0,$ by Theorem \[classical\] we get $-y \in P_Y(z^{n_{2k}}),$ as required. By the above calculations, for any $ a \in [-1,1)$ we conclude $$x^{k}(n_{2k}) + ay_{n_{2k}} < x^{k}(n_{2k}+1) + ay(n_{2k}+1).$$ Since $w$ is strictly decreasing, for any $a \in [-1,1)$ we have $$\begin{aligned}
dist(x^k,Y)=&\|x^k +y\|=\sum_{j=1}^{\infty}(x^k(j) + y(j))w(j)=\sum_{j=1}^{\infty}(x^k(j)+ay(j))w(j)\\
<&\sum_{j=1}^{n_{2k}-1}(x^k(j) + ay(j))w(j) + w(n_{2k})(x^{k}(n_{2k}+1) + ay(n_{2k}+1))\\
&+w(n_{2k}+1)(x^{k}(n_{2k})+ay(n_{2k})) + \sum_{j=n_k+2}^{\infty}(x^k(j)+ay(j))w(j)\\
\leq&\|x^k+ay\|_{w,1},\end{aligned}$$ which shows our claim. Consequently, for any $ k \in \mathbb{N},$ $ P_Y(x^k)\subset I_1y=[m,-1]y$ for some $ m <-1.$ Since $ y(n_{2k}) > y(n_{2k}+1),$ by (\[strict\]), reasoning exactly in the same way, we can show that for any $ k \in \mathbb{N},$ $ P_Y(w^k) \subset I_2y=[1,p]y$ for some $ p >1.$ By (\[c0\]), $\|x^k-z\|_{w,1}\rightarrow 0$ and $\|w^k-z\|_{w,1}\rightarrow 0$ as $k\rightarrow\infty$. Since $ I_1 \cap I_2 = \emptyset,$ we get our claim. Now assume that there exists $n_o\in\mathbb{N}$ such that either ${y(n)\geq{y(n+1)}}$ for any $n\geq{n_o}$ or $-y(n)\geq{-y(n+1)}$ for any $n\geq{n_o}$. Fix $ x=x^* \in d(w,1)$ with $P_Y(x) = [-1,1]y$ and $(x^n)\subset{d(w,1)}$ with ${\ensuremath{\left\Vertx^n-x\right\Vert_{d(w,1)}^{}}}\rightarrow{0}$. Without loss of generality, replacing $y$ by $ -y$ if necessary, we can assume that for $ n \geq n_o$ $$\label{case1}
y(n)\geq y(n+1).$$ To get our claim, it is enough show that for any compact interval $I\subset\mathbb{R}$ such that $P_Y(x_n)\subset{Iy}$ for every $n\in\mathbb{N},$ $-1 \in I.$ First suppose that $$\label{strict1}
y(n) \neq y(n+1)
\hbox{ for any } n \in \mathbb{N}.$$ Let $P_Y(x_n)=[a_n,b_n]y$ where $a_n \leq b_n$. Let $b=\liminf_{n\in\mathbb{N}}\{a_n\}$. We can assume without loss of generality, that $a_n$ converges to $b$. Since $ P_Y(x) = [-1,1]y,$ $|b|\leq 1$. We show that $b=-1.$ Assume on the contrary that $ b > -1.$ Since $a_ny\in{}P_Y(x^n)$, there exists a supporting functional $f^n \in S_{d^*(w,1)}$ such that $$\label{equ:**:cs}
f^n(x^n-a_ny)={\ensuremath{\left\Vertx^n-a_ny\right\Vert_{d(w,1)}^{}}}=\operatorname{dist}(x^n,Y)$$ and also $$\label{equ:4:cs}
\sum_{i=1}^\infty{f^n(i)y(i)}=0.$$ By the Banach-Alaoglu Theorem the set $ \{ f^n: n \in \mathbb{N} \}$ has a cluster point $f \in d^*(w,1)$ with respect to the weak$^*$ topology in $d^*(w,1)$ and $ \|f\|_W \leq 1.$ Now assume that there exists a subsequece $ \{n_k\} $ such that for any $k$, $f^{n_k}=f=w.$ Hence, since $ w$ is strictly decreasing, for $ k \in \mathbb{N},$ and $ j \in \mathbb{N} $ $$x^{n_k}(j) - a_{n_k}y(j) = (x^{n_k}(j) - a_{n_k}y(j))^*.$$ Fix $ m \in (-1,b).$ Then $ a_{n_k} - m >0$ for $k \geq k_o.$ Hence by (\[strict1\]) for $ j \geq n_o $ and $ k \geq k_o,$ $$(a_{n_k} -m) y(j) > (a_{n_k} -m) y(j+1)$$ and consequently, $$x^{n_k}(j) - my(j) > (x^{n_k}(j+1) - my(j+1)) >0.$$ Since $ m \in (-1,b)$ and $ |b| \leq 1,$ by (\[ineq2\]), $$x^{n_k}(j) - my(j) > (x^{n_k}(j+1) - my(j+1)) >0.$$ for $ k \geq k_1$ and $ j=1,2,...n_o.$ Since $ \sum_{j=1}^{\infty} w(j)y(j) =0,$ by Theorem \[classical\], $my \in P_Y(x^{n_k})$ for $ k \geq k_o.$ Since $ m <b, $ we get a contradiction with defintion of $b.$ So to end the proof under assumption (\[strict1\]), we construct a subsequence $ \{n_k\}$ such that for any $k$ $ f^{n_k} = f =w.$ Since $ f$ is a cluster point of $ \{ f^n: n \in \mathbb{N} \}$ with respect to the weak$^*$ topology in $d^*(w,1),$ applying the diagonal argument, we can choose a subsequence $ \{n_k\}$ such that $ f^{n_k}(j) \rightarrow f(j)$ for $ j \in \mathbb{N},$ $f^{n_k}(y) \rightarrow f(y)$ and $ f_{n_k}(x) \rightarrow f(x).$ Since $ f^n(y) =0, $ for any $ n \in \mathbb{N},$ $f(y) =0.$ Moreover, since $\| x^n- x\|_{w,1} \rightarrow 0,$ and $ a_n \rightarrow b,$ by (\[equ:\*\*:cs\]) $$\begin{aligned}
|f^{n_k}(x^{n_k}-a_{n_k}y) - f(x- by)|&\leq |f^{n_k}(x^{n_k}-a_{n_k}y-(x-by))|+|(f-f^{n_k})(x- by)|\\
&\leq \|x^{n_k}-a_{n_k}y-(x-by)\|_{w,1} + |f-f^{n_k}(x)|.\end{aligned}$$ Consequently, $$\lim_k f^{n_k}(x^{n_k}-a_{n_k}y) - f(x- by) = \lim_k dist(x^{n_k},Y) - f(x-by)=0.$$ Hence $$f(x)= f(x-by) = \|x-by\|_{w,1} = dist(x,Y) = \|x\|_{w,1}$$ By (\[case1\]) and \[ineq2\]), $ x(j) > x(j+1) $ for any $j,$ which means that $x$ has the only one supporting fuctional $w.$ Hence $ f =w.$ By Theorem \[singer\], we can assume that there are $g^k$ and $h^k$ extreme functionals of $S_{d(w,1)^*}$ such that $f^{n_k}=\alpha_n g^{k}+(1-\alpha_k)h^{k}$ for some $\alpha_k \in [0,1]$. By Theorem \[hkl\], $(h^{k})^* = (g^{k})^* =w.$ Since $ w$ is strictly decreasing and $ f^{n_k}(j) \rightarrow f(j)=w(j)$ for $ j =1,...,n_o,$ $ g^k(j)\rightarrow w(j) $ and $h^k(j) \rightarrow w(j)$ for $ j=1,...,n_o+1.$ Hence $g^{k}(j) = w(j)$ and $h^k(j) = w(j)$ for $j=1,...n_o+1$ and $ k \geq k_o.$ Consequently, for $ k \geq k_o,$ $ f^{n_k}(j) = w(j). $ Moreover, since $\sum_{i=1}^{\infty}w(i)y(i)=0$ and by condition (\[equ:4:cs\]) we conclude for $ k \in \mathbb{N}$ $$\sum_{i=n_o+1}^{\infty}f^{n_k}(i)y(i)=\sum_{i=n_o+1}^{\infty} w(i)y(i).$$ Notice that $ |f^{n_k}(j)| \leq w(j)$ for any $k.$ By (\[case1\]) and (\[strict1\]), we get that $ w(j) = f^{n_k}(j) $ for $ j \geq n_o +1 $ and $ k \in \mathbb{N}$ as required. Now assume that (\[strict1\]) is not satisfied and $$\label{nonzero}
y(n) > 0 \hbox{ for } n \geq n_o.$$ By (\[ineq1\]), $$\label{general1}
x(n) = x^*(n) >0 \hbox{ for } n \in \mathbb{N}.$$ The proof goes in a similar way like under assumption (\[strict1\]), so we indicate only necessary modifications. Also the same notation will be used. Ressoning as above, we can choose a subsequence $ \{n_k\}$ such that $ f^{n_k}(j) \rightarrow f(j)$ for $ j \in \mathbb{N},$ $f^{n_k}(y) \rightarrow f(y),$ $ f_{n_k}(x) \rightarrow f(x)$ and $$f(x) = \| x\|_{w,1} = \sum_{j=1}^{\infty} x(j)w(j).$$ Let $ j_1 = 1 $ and for $ n \geq 2$ $$j_n =\min \{ j > j_{n-1}: x(j_{n-1}) > x(j)\}.$$ Since $ x \in d(w,1) \subset c_o,$ by (\[general1\]) our definition is correct. Fix $ n_1$ such that $ j_{n_1} > n_o.$ Reasoning like under assumption (\[strict1\]), we can show that for $k$ sufficiently large $ f^{n_k}(j_l) = w(j_l) $ for $ 2 \leq l \leq n_1.$ Moreover, if $ l \in \{2,...,n_1\}$ is such that $$j_l+ 1 < j_{l+1},$$ then for any $j \in \{j_l+1, j_{l+1}-1\} $ and $k$ sufficiently large $$f^{n_k}(j) \in \{ w(j_{l}+1),..., w(j_{l+1}-1)\}.$$ Also if $ l=1,$ then $f^{n_k}(j) \in \{ w(1),..., w(j_{2}-1)\}$ for $j=1,...,j_2-1.$ Since $$f^{n_k}(y) = \sum_{j=1}^{\infty} f^{n_k}(j)y(j) = w(y) =\sum_{j=1}^{\infty} w(j)y(j) =0,$$ by the above reasoning, $$\sum_{j=j_{n_1}}^{\infty} f^{n_k}(j)y(j) =\sum_{j=j_{n_1}}^{\infty} w(j)y(j).$$ Observe that for any $ j,k \in \mathbb{N},$ $ |f^{n_k}(j) | \leq w(j).$ Hence by (\[nonzero\]) and (\[case1\]) $ f^{n_k}(j) = w(j)$ for $ j > j_{n_1}$ Moreover, since for $k$ sufficiently large $ f^{n_k}(j_l)= f(j_l) = w(j_l)$ for $l=1,...,n_1,$ if $j \in \{j_l+1, j_{l+1}-1\} $ then $$f^{n_k}(j) =f(j) \in \{ w(k_{l}+1),..., w(k_{l+1}-1)\}.$$ (compare with the proof under assumption (\[strict1\])). Observe that, by (\[ineq2\]), if $ x(j) = x(j+1),$ then $ y(j) = y(j+1).$ Since $ m \in (-1,b),$ $ f^{n_k}(y) =0$ and $ f^{n_k}(x^{n_k} - my) = \|x^{n_k} - my\|_{w,1}$ for $ k$ sufficiently large. By Theorem \[classical\], $my \in P_Y(x^{n_k})$ for $ k \geq k_o.$ Since $ m <b, $ we get a contradiction with defintion of $b,$ which finishes the proof under assumption (\[nonzero\]). If $ y(n) =0$ for some $n \geq n_o,$ then by (\[case1\]), $y(j) = 0 $ for $j \geq n.$ Hence for any $ k \in \mathbb{N},$ and $j \geq n$ $$f^{n_k}(j)(x(j)-my(j)) =f^{n_k}(j)(x(j)-a_{n_k}y(j)).$$ Hence reasoning as under assumption (\[nonzero\]) we get that $my \in P_Y(x^{n_k})$ for $ k \geq k_o,$ which completes the proof.
\[Czeb\] Assume that $w$ is a strictly decreasing weight. Let $$y =(y(1),...y(n),...) \in d(w,1) \setminus \{0\}$$ and let $ Y = span[y].$ Then $Y$ is not a Chebyshev subspace in $d(w,1)$ if and only if there exists $ M \subset \mathbb{N},$ $card(M)= \infty$ such that $supp(y) \subset M$ and a bijection $ p: \mathbb{N} \rightarrow M,$ $\sigma \in \{ -1,1\}^{\mathbb{N}}$ such that $$\label{cheb1}
\sum_{j=1}^{\infty} w(j) \sigma(j)y(p(j)) =0,
\sum_{j=1}^{\infty}|\sigma(j)y(p(j))-\sigma(j+1)y(p(j+1))| < \infty$$ and $$\label{cheb2}
z=(z(1),z(2),...) \in d(w,1),$$ where for $j \in \mathbb{N}$ $$\label{cheb3}
z(j) = \sum_{l=j}^{\infty} |\sigma(l)y(p(l))-\sigma(l+1)y(p(l+1))|.$$
Since $Y$ is not a Chebyshev subspace, there exists $x \in d(w,1)$ such that $ P_Y(x) =[-1,1]y$. By Theorem \[classical\], there exists $ f \in S_{d(w,1)^*}$ such that $f(x) = \|x\|_{w,1} $ and $ f(y)=0.$ By Lemma \[Izbior\], $ f \in ext(S_{d(w,1)^*}).$ By Theorem \[hkl\], $ f^* =w.$ Put $M = supp(f).$ Since $w$ is strictly decreasing, $M$ is infinite. Now we show that $ supp(y) \subset M$ and $ supp(x) \subset M.$ Assume to the contrary that $ y(i) \neq 0$ for some $ i \notin M.$ Then, $ |x(i) - y(i)| > 0$ or $ |x(i) + y(i)| > 0.$ Assume that $ |x(i) - y(i)| > 0.$ Since $ x-y \in c_o,$ $ |x(j) -y(j)| < |x(i) -y(i)| $ for $ j \geq j_o.$ Fix $j_1 \in M, $ $j_1 \geq j_o.$ Since $ P_Y(x) =[-1,1]y,$ and $ i \notin M,$ $$\begin{aligned}
\| x-y\|_{w,1} &= f(x-y) = \sum_{j \in M} f(j)x(j)\\
&=\sum_{j=1}^{j_1-1} f(j) (x(j)-y(j)) + \sum_{j =j_1}^{\infty} f(j) (x(j)-y(j))\\
&<\sum_{j=1}^{j_1-1}f(j)(x(j)-y(j))+|f(j_1)||x(i)-y(i)|+\sum_{j=j_1+1}^{\infty}f(j)(x(j)-y(j)) \\
&\leq\|x-y\|_{w,1},\end{aligned}$$ so a contradiction. Since $ f(x) = \|x \|_{w,1},$ reasoning in the same way, we get that $ supp(x) \subset M.$ Since $ supp(y) \subset M,$ and $supp(x) \subset M,$ without loss of generality we can assume that $M = \mathbb{N}.$ Let $p: \mathbb{N} \rightarrow M$ be defined by $ |f(p(j))| = w(j).$ Since $w$ is strictly decreasing and $f^*=w,$ our defintion is correct. Let $ \sigma(j) = sgn f(p(j)).$ Applying Lemma \[selection\] to $ y^1 = (\sigma(j)y(p(j)))_{j \in \mathbb{N}}$ and $ x^1 = (\sigma(j)x(p(j)))_{j \in \mathbb{N}}$ we get (\[cheb1\] - \[cheb3\]). Now assume that (\[cheb1\] - \[cheb3\]) are satisfied. Observe that for any $j \in \mathbb{N},$ $$z(j) - z(j+1) = |\sigma(j)y(p(j))-\sigma(j+1)y(p(j+1))|.$$ Hence for any $ j \in \mathbb{N},$ $$z(j) \pm \sigma(j)y(p(j)) \geq z(j+1) \pm \sigma(j+1)y(p(j+1))\geq{0}.$$ In consequence, by Theorem \[classical\], $ P_{Y_1}(z) = [-1,1]y^1$. Notice that a mapping $
T: d_{w,1} \rightarrow d_{w,1} $ defined by $$\label{isometry}
Tu = (\sigma(j) u(a(j))): j \in \mathbb{N},$$ where $ a= p^{-1},$ is a linear, surjective isometry. Define $ x \in d(w,1),$ by $ x(j) = \sigma(j)z(a(j)),$ for $j \in M.$ Observe that $ Tz = x$ and $Ty^1 =y.$ By Lemma \[general\] applied to $T$ and (\[cheb1\] - \[cheb3\]), $ P_Y(x) =[-1,1]y, $ as required.
Now we prove the main result of this section.
\[selgen\] Assume that $w$ is a strictly decreasing weight. Let $$y =(y(1),...y(n),...) \in d(w,1) \setminus \{0\}$$ and let $ Y = span[y].$ Then $Y$ admits a continuous metric selection if and only if $Y$ is a Chebyshev subspace or for any $p$ and $M$ satisfying the requirements of Theorem \[Czeb\] there exists $n_o \in \mathbb{N}$ such that for $ n \geq n_o$ $$\label{final}
y(p(n)) \geq y(p(n+1))$$ or for $ n \geq n_o$ $$\label{final1}
-y(p(n)) \geq -y(p(n+1)).$$
If $Y$ is a Chebyshev subspace then by Lemma \[Czebyszew\] $Y$ admits a continuous metric selection. If $Y$ is not a Chebyshev subspace, fix $ x \in d{(w,1)}$ such that $ P_Y(x) = [a,b]y,$ where $ a <0<b.$ Assume that (\[final\]) or (\[final1\]) is satisfied. We show that for arbitrary sequences $w^k \rightarrow x$ and $x^k \rightarrow x$ if $I_1 $ and $I_2$ are compact intervals such that $P_Y(w^k) \subset I_1y$ for any $ k \in \mathbb{N}$ and $P_Y(x^k) \subset I_2y$ for any $ k \in \mathbb{N}$ then $I_1 \cap I_2 \neq \emptyset.$ Without loss of generality we can assume that $ a=-1$ and $b=1.$ Observe that, by Theorem \[classical\], there exists $ f \in S_{d^*(w,1)}$ such that $f(x) = \|x\|_{w,1} $ and $ f(y)=0.$ By Lemma \[Izbior\], $f \in ext(S_{d(w,1)^*}).$ By Theorem \[hkl\], $ f^* =w.$ If $f =w,$ then the mapping $p: \mathbb{N} \rightarrow M$ defined in the proof of Theorem \[Czeb\] is the identity on $ \mathbb{N}.$ Without loss of generality we can assume that (\[final\]) is satisfied. Hence for $n \geq n_o,$ $y(n) \geq y(n+1). $ By Lemma \[selection\] $-y \in I_1y \cap I_2y, $ which shows our claim. If $ f \neq w,$ we reduce our proof to the case $f=w.$ To do that, put $ M = supp(f).$ Since $ w$ is strictly decreasing, $M$ is infinite. Reasoning as in the proof of Theorem \[Czeb\] we can show that $ supp(y) \subset M$ and $ supp(x) \subset M.$ Since $f^*=w$ it follows that $M = \mathbb{N}.$ Let $p: \mathbb{N} \rightarrow M$ be defined by $ |f(p(j))| = w(j)$ and let $ \sigma(j) = sgn f(p(j)).$ (Compare with the proof of Theorem \[Czeb\]). Define $ y^1 = (\sigma(j)y(p(j))_{j \in \mathbb{N}},$ $ Y^1 = span[y^1]$ and $ x^1 \in d(w,1),$ by $ x^1(j) = \sigma(j)x(p(j)),$ for $j \in M,$ where $ a= p^{-1}.$ By Lemma \[selection\] and by (\[final\]) applied to $x^1$ and $ y^1$ for any sequence $z^k \in d(w,1)$ converging to $x^1$ and a compact interval $ I$ such that $ P_{Y^1}(z^k) \subset I y^1,$ for any $ k \in \mathbb{N},$ $ -1 \in I. $ Observe that $Tx^1=x$ and $Ty^1=y,$ where $T$ is defined by (\[isometry\]). By Lemma \[general\], apllied to $T$, $x^1$ and $ y^1$ for any sequence $x^k \in d(w,1)$ converging to $x$ and a compact interval $ I$ such that $ P_{Y}(x^k) \subset I y^1,$ for any $ k \in \mathbb{N},$ we get $ -1 \in I, $ which gives our claim. To prove the converse, assume that there are $p$ and $M$ satisfying the requirements of Theorem \[Czeb\] such that (\[final\]) and (\[final1\]) are not satisfied. First suppose that $ p = id_{\mathbb{N}}.$ This implies that $y$ satisfies (\[strict\]) for some subsequence $ \{ n_k\}.$ Let $ z \in d(w,1)$ be defined by (\[cheb2\]) and (\[cheb3\]). By the proof of Lemma \[selection\] there exist two sequences $ x^k$ and $w^k$ converging to $z$ such that there exist compact intervals $ I_1,$ $ I_2,$ with $ I_1 \cap I_2 = \emptyset$ having the property that for any $ k \in \mathbb{N}$ $P_Y(x^k) \subset I_1$ and $ P_Y(w^k) \subset I_2.$ By Theorem \[deutsch\], $Y$ does not admit the continuous metric selection. Applying the isometry $T$ defined by (\[isometry\]) and Lemma \[general\], we can reduce the proof in general case to the above reasoning, which gives our claim.
Now we show four examples illustrating possible applications of Theorem \[selgen\].
Let $ 1= w(1) > w(2)> ..$ be a strictly decreasing weight such that $ \sum_{n=1}^{\infty}w(n) = +\infty .$ Let $ y=(1, y(2),0,0...)$ be so chosen $ y(2) <0$ and $ y(2) \neq \frac{-w(i)}{w(j)}$ for any $ i,j \in \mathbb{N},$ $i \neq j.$ Since $ \mathbb{R} $ is not countable, such an $y(2)$ exists. By Theorem \[hkl\] for any $f \in ext(S(d^*(w,1)))$, $f(y) \neq 0.$ Hence, by Lemma \[Izbior\] any $ x \in d(w,1)$ possesses strongly unique best approximation in $ Y = span[y].$ By the Freud Theorem (see [@Chen], p. 82) the projection operator (which is single-valueds in this case) satisfies the local Lipschitz condition.
Let $w(1)=1$, $ w(n) = \frac{1}{n} ,$ $y(n) = \frac{1}{n},$ for $n \geq 2$ and $ y(1) = - \sum_{n=2}^{\infty}y(n)w(n).$ Observe that in this case $ \sum_{n=1}^{\infty} |y(n+1) - y(n)| < \infty . $ By Theorem \[Czeb\], $Y$ is not a Chebyshev subspace of $d(w,1).$ By Theorem \[selgen\], $Y$ admits a continuous metric selection.
\[no-cont-sele\] Let $ w(n) = \frac{1}{n} $ and for $ k \geq 1$ $y(2k) = \frac{1}{(2k+1)^2},$ $y(2k+1) = \frac{1}{(2k)^2}$ and $ y(1) = - \sum_{n=2}^{\infty}y(n)w(n).$ Observe that in this case $ \sum_{n=1}^{\infty} |y(n+1) - y(n)| < \infty . $ By Theorem \[Czeb\] $Y$ is not a Chebyshev subspace of $d(w,1).$ By Theorem \[selgen\] $Y$ does not admit a continuous metric selection.
Now we consider Example \[no-cont-sele\] with $y(1)=-\sum_{n=2}^{\infty}y(p(n))w(n)$, where $p:\mathbb{N}\rightarrow\mathbb{N}$ is a function given by $p(2k)=2k+1$ and $p(2k+1)=2k$ for any $k\in\mathbb{N}$. Notice that $\sum_{n=1}^{\infty}|y(p(n+1))-y(p(n))|<\infty.$ By Theorem \[Czeb\] $Y$ is not a Chebyshev subspace of $d(w,1).$ On the other hand, by Theorem \[selgen\] $Y$ admits a continuous metric selection.
[99]{}
C. Bennett and R. Sharpley, *Interpolation of operators*, Pure and Applied Mathematics Series 129, Academic Press Inc.,1988.
A. L. Brown, *A rotund reflexive space having a subspace of codimension two with a discontinuous metric projection,* Michigan Math. J. (21), (1974), 145 - 151.
A. L. Brown *Set valued mappings, continuous selections and metric projections*, J. Approx. Theory, (57), (1989), 48 -68.
A. L. Brown, F. Deutsch, V. Indumathi and P. S. Kenderov *Lower semicontinuity concepts, continuous selections and set valued metric projections*, J. Approx. Theory, (115), (2002), 120 - 143.
A. L. Brown *Continuous selections for metric projections in spaces of continuous functions and a disjoint leaves condition*, J. Approx. Theory, (141), (2006), 29 -62.
A. L. Brown *On lower semi-continuous motric projections onto finite-dimensional subspaces of spaces of continuous functions*, J. Approx. Theory, (166), (2013), 85 -105.
A. P. Calderón, *Intermediate spaces and interpolation, the complex method*, Studia Math. 24 (1964), 113-190.
S. Chen, X. He, H. Hudzik, *Monotonicity and best approximation in Banach lattices*, Acta Math. Sinica 25 (5) (2009) 785-794.
E. W. Cheney, *Introduction to Approximation Theory,* McGraw-Hill, New York, 1966.
V. I. Chilin, P. G. Dodds, A. A. Sedaev, and F. A. Sukochev, *Characterizations of Kadec-Klee properties in symmetric spaces of measurable functions*, Trans. Amer. Math. Soc. 348 (12) (1996), 4895-4918.
M. Ciesielski, A. Kamińska, P. Kolwicz and R. Płuciennik, *Monotonicity and rotundity of Lorentz spaces $\Gamma_{p,w}$*, Nonlinear Anal. 75 (2012), 2713-2723.
M. Ciesielski, A. Kamińska and R. Płuciennik, *Gâteaux derivatives and their applications to approximation in Lorentz spaces $\Gamma_{p,w}$*, Math. Nachr. **282** (2009) no. 9, 1242–-1264.
M. Ciesielski, P. Kolwicz and A. Panfil, *Local monotonicity structure of symmetric spaces with applications*, J. Math. Anal. Appl. 409 (2014) 649-662.
M. Ciesielski, P. Kolwicz, R. Płuciennik, *Local approach to Kadec-Klee properties in symmetric function spaces*, J.Math. Anal. Appl. 426 (2015) 700-726.M. C
F. Deutsch, V. Indumathi and K. Schnatz, *Lower semicontinuity, almost lower semicontinuity and continuous selections for set-valued mappings,* Journ. Approx. Theory, 53, (1988), 266 - 294.
A. Kamińskaand H. J. Lee, *$M$-ideal properties in Marcinkiewicz spaces*, Comment. Math., Special volume for $75$th birthday of Julian Musielak, (2004) 123-144.
A. Kamińska, H. J. Lee and G. Lewicki, *Extreme and smooth points in Lorentz and Marcinkiewicz spaces with applications to contractive projections,* Rocky Mountain Journal of Mathematics, (39,5), (2009), 1533-1572.
H. Hudzik and A. Kamińska, *Monotonicity properties of Lorentz spaces*, Proc. Amer. Math. Soc. **123.9**, (1995), 2715-2721.
*Monotonicity properties of Musielak-Orlicz spaces and dominated best approximant in Banach lattices,* J. Approx. Theory 95 (3) (1998), 353-368.
A. Kamińska, *Extreme points in Orlicz-Lorentz spaces*, Arch. Math. (Basel) 55 (2) (1990) 173-180.
A. Kamińska, *Some remarks on Orlicz-Lorentz spaces*, Math. Nachr. 147 (1990), 29-38.
A. Kamińska and L. Maligranda, *On Lorentz spaces* $\Gamma _{p,w}$, Israel J. Math. 140 (2004), 285-318.
S. G. Krein, Yu. I. Petunin and E. M. Semenov, *Interpolation of linear operators*, Nauka, Moscow, 1978 (in Russian).
W. Kurc, *Strictly and uniformly monotone Musielak-Orlicz spaces and applications to best approximation*, J. Approx. Theory **69.2** (1992), 173-187.
A. J. Lazar, D. E. Wulbert and P. D. Morris, *Continuous selections for metric projections,* Journal of Funct. Anal. (3), (1969), 193 - 216.
J. Lindenstrauss and L. Tzafriri, *Classical Banach spaces. II. Function spaces*, Springer-Verlag, Berlin-New York, 1979.
G. G. Lorentz, *On the theory of spaces* $\Lambda $, Pacific J. Math. 1 (1951), 411-429.
E. Sawyer, *Boundedness of classical operators on classical Lorentz spaces*, Studia Math. 96 (2) (1990), 145-158.
I. Singer, *Best Approximation in Normed Linear Spaces by Elements of Linear Subspaces,* Springer, Berlin, 1970.
V. D. Stepanov, *The weighted Hardy’s inequality for nonincreasing functions*, Tran. Amer. Math. Soc. 338 (1993), 173–186.
J. Sudolski and A. Wójcik, *Some remarks on strong uniqueness of best approximation*, J. Approx. Theory, Appl 6, (1990), no. 2, 44 - 78.
$\begin{array}{lr}
\textnormal{\small Maciej CIESIELSKI} & \textnormal{\small Grzegorz Lewicki}\\
\textnormal{\small Institute of Mathematics} & \textnormal{\small Department of Mathematics and Computer Science}\\
\textnormal{\small Pozna\'{n} University of Technology} & \textnormal{\small Jagiellonian University}\\
\textnormal{\small Piotrowo 3A, 60-965 Pozna\'{n}, Poland} & \textnormal{\small \L ojasiewicza 6, 30-348 Krak\'ow, Poland}\\
\textnormal{\small email: [email protected];} & \textnormal{\small email: [email protected]}
\end{array}$
[^1]: $^*$This research is supported by the grant 04/43/DSPB/0086 from Polish Ministry of Science and Higher Education.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the generation of a quasistatic longitudinal electric field by intense laser pulses propagating in a transparent plasma with radiation friction taken into account. For both circular and linear polarization of the driving pulse we develop a 1D analytical model of the process, which is valid in a wide range of laser and plasma parameters. We define the parameter region where radiation friction results in an essential enhancement of the longitudinal field. The amplitude and the period of the generated longitudinal wave are estimated and optimized. Our theoretical predictions are confirmed by 1D and 2D PIC simulations. We also demonstrate numerically that radiation friction should substantially enhance the longitudinal field generated in a plasma by a 10 PW laser such as ELI Beamlines.'
address:
- '$^1$ELI Beamlines, Institute of Physics of the ASCR, v.v.i., Dolni Brezany, Czech Republic'
- '$^2$National Research Nuclear University “MEPhI” (Moscow Engineering Physics Institute), 115409, Kashirskoe sh, 31, Moscow, Russia'
author:
- 'E G Gelfer$^{1,2}$, A M Fedotov$^2$, S Weber$^1$'
title: 'Theory and simulations of radiation friction induced enhancement of laser-driven longitudinal fields'
---
[*Keywords*]{}: radiation friction, radiation pressure, laser-plasma interaction, charge separation, ion acceleration
Introduction
============
The modern laser systems have recently reached the multi-Petawatt power level [@laserreview; @chu; @sung; @zheng]. These facilities already provide extremely strong electromagnetic fields $\simeq 10^9$V/cm [@hercules], but even more powerful 10 PW lasers are now under construction in Czech Republic (ELI Beamlines [@ELI]), Romania (ELI NP [@ELINP]) and France (Apollon [@Apollon]). However, since laser fields oscillate with very high frequency, their application to charged particle acceleration is not straightforward. By separating charges, a laser pulse propagating in plasma creates a wake (longitudinal) wave. As suggested many years ago [@tajima] and demonstrated in a recent experiment [@bellaacc], this wake wave can efficiently accelerate electrons up to several GeVs.
Several models were suggested also for ion acceleration in various laser-plasma setups [@light_sail; @hole_boring], see the reviews [@ion_acc_rev] and the overview of the recent experimental achievements in [@ion_acc_exp]. In all such schemes the laser pulse pushes plasma electrons forward via the ponderomotive force, thus leading to charge separation and creation of a quasistatic longitudinal electric field [@micha2012] capable for particle acceleration. However, as recently demonstrated [@arxiv], besides the usual ponderomotive mechanism (PM) of charge separation in plasma, there also exists a competing radiation friction mechanism (RFM), which for strong and long laser pulses propagating in low density plasmas should dominate over PM.
To clarify its origin, consider an electron in a circularly polarized plane wave. If radiation friction (RF) is neglected, the electron circles with the field frequency, so that its velocity remains perpendicular to the electric field and parallel to the magnetic field [@akhiezer_polovin], with no force acting in the direction of pulse propagation. RF in transverse direction changes the angle between the electron velocity and the magnetic field [@arxiv; @RFpapers], resulting in the longitudinal Lorenz force $[\mathbf{v}\times\mathbf{B}]$[^1], which accelerates the electron in the direction of pulse propagation, thus separating charges in plasma.
The role of RF in laser-plasma interaction was previously a subject of many investigations [@zhidkov2002; @tamburini2010; @tamburini2011; @brady2012; @nakamura2012; @bashinov2013; @stark2016], but mostly for overcritical plasmas. However, if plasma is so dense that it is opaque, then electrons never get deep inside the pulse, where the field is so strong that RF affects the electron transverse motion. As for the overcritical but relativistically transparent case, in a high density plasma RF results in a fast depletion of the laser pulse. Therefore we consider plasma density not exceeding the critical value $n_c=m\omega^2/4\pi e^2$, where $\omega$ is the laser frequency, $m$ and $e$ are the electron mass and charge.
In such a regime charge separation and generation of the quasistatic longitudinal electric field by an intense laser pulse in a plasma with immobile[^2] ions are perfectly described by the following 1D model [@arxiv]. A pulse entering a plasma accelerates the electrons forward via RFM and/or PM, piling them up into a moving density spike. Thereupon the longitudinal electrostatic field of the naked immobile ions is gradually growing, and eventually starts decelerating the electrons. Finally a breakdown occurs, when a part of electrons from the spike penetrates through the pulse backward partially screening the ion field, see Figure \[fig\_spike\]. The amplitude and the period of the thus generated longitudinal field are defined by the position of the spike at the moment of breakdown, and to find it we consider the motion of the leftmost particle in the spike.
The model developed in [@arxiv] provides estimations for the amplitude and the period of the longitudinal field generated by a circularly polarized laser pulse via both PM and RFM under the additional assumption of ultrarelativistic electron longitudinal motion, that are in a rather good agreement with PIC simulations. Here we discuss this model in more detail and generalize it to the cases of arbitrary polarization of the laser pulse and arbitrary electron longitudinal velocities, thus extending its validity to a wider range of laser parameters. We formulate and study the limits of applicability of the model (in particular establish the conditions required to consider the ions immobile), derive the estimations for various regimes of longitudinal field generation, and identify an optimal regime with generation of the highest attainable longitudinal field. Furthermore, by 2D PIC simulations we demonstrate that our 1D model is valid for wide pulses (of comparable length and width), and that RFM can dominate over PM also for tightly focused pulses with the parameters of the upcoming facilities such as ELI Beamlines [@ELI].
1D model for arbitrary laser polarization
=========================================
Let us start with a generalization of the model suggested in [@arxiv] to describe longitudinal electric field generation by a strong laser pulse propagating in plasma from pure circular to arbitrary polarization of the pulse. The 4-dimensional form of the equation of motion for the leftmost particle in the electron spike (see Figure \[fig\_spike\]) reads: $$\label{eqll}\fl
\frac{du^i}{ds}=\frac{e}{m}F^{ij}u_j+\frac{2e^3}{3m^2}\frac{\partial F^{ij}}{\partial x^l}u_ju^l-\frac{2e^4}{3m^3}F^{il}F_{jl}u^j+\frac{2e^4}{3m^3}(F_{jl}u^l)(F^{jm}u_m)u^i+\mathcal{F}^i.$$ Here $F_{ij}=\partial_i A_j-\partial_j A_i$ is the laser field strength tensor, $\mathcal{F}^i=-4\pi e^2 n x\gamma/m\{v_x,1,0,0\}$ is the electrostatic 4-force from the naked ions; $s$, $\gamma$, $\mathbf{v}$ and $u^i=\{\gamma,\gamma\mathbf{v}\}$ are the proper time, gamma factor, 3- and 4-velocities of the electron, and we use the Landau-Lifshitz form [@LL] for RF force. Assume that the laser field is of the form of a plane wave, $A_i=A_i(\varphi)$, where $\varphi=k^ix_i$ is the phase and $k^i=\omega\{1,1,0,0\}$ is the wave 4-vector. By contracting (\[eqll\]) with $k_i$ and taking into account $F_{ij}=k_iA_j'(\varphi)-k_jA_i'(\varphi)$ and $k_ik^i=k_iA^i=k_iF^{ij}=0$, we arrive at $$\label {eq_ku}
\omega\frac{du_-}{ds}=\frac{2e^4\omega^3}{3m^3}\left(\frac{dA_i}{d\varphi}\frac{dA^i}{d\varphi}\right)u_-^3 + k_i\mathcal{F}^i,$$ where $u_-=k^iu_i/\omega=\gamma-u_x$. From $u_iu^i=1$ and $dt=\gamma\,ds=(\gamma/\omega u_-)\,d\varphi$ one can easily derive the formulas $$\frac{du_-}{ds}=u_-\left(\frac{\omega}{2\gamma}\frac{du_\perp^2}{d\varphi}-\frac{du_x}{dt}\right),\quad u_-^2=(1+u_\perp^2)\frac{(1-v_x)}{(1+v_x)},$$ using which follows $$\label{dux}
\frac{du_x}{dt}=\frac{\omega}{2\gamma}\frac{d(u_\perp^2)}{d\varphi}-\frac{2e^4\omega^2}{3m^3}\left(\frac{dA^i}{d\varphi}\right)^2(1+u_\perp^2)\left(\frac{1-v_x}{1+v_x}\right)-\frac{4\pi e^2n}{m}x.$$ An advantage of the above derivation is its generality with respect to polarization of the driving laser field, on which no assumptions have been made yet. We now consider the most important cases of circular (CP) and linear (LP) polarizations.
Circular polarization
---------------------
If the laser pulse is circularly polarized $\mathbf{A}(\varphi)=(m/e)a_0(\varphi)\{\cos\varphi,\sin\varphi,0\}$ then $(dA^i/d\varphi)^2=-m^2a_0^2/e^2$, where $a_0(\varphi)$ is a slowly varying dimensionless pulse envelope. Denoting the dimensionless variables $\tau=\omega t$, $\xi=\omega x$ and parameters $\tilde{n}=n/n_c$, $\mu=2\omega r_e/3c\simeq 1.18\cdot 10^{-8}$, where $r_e=e^2/mc^2$ is the classical electron radius, we can cast (\[dux\]) into the form $$\label{dux_c}
\frac{du_x}{dt}=\frac1{2\gamma}\frac{d(u_\perp^2)}{d\varphi}+\mu a_0^2(1+u_\perp^2)\frac{1-v_x}{1+v_x}-\tilde{n}\xi.$$ Equations (\[dux\]) and (\[dux\_c\]) are the exact implications of (\[eqll\]) and hence up to this moment no approximations except for the Landau-Lifshitz approximation and the assumption of a dilute plasma have been made. But if we further assume that damping due to RF is small (see Sec. \[sec\_ca\] for discussion) then $u_\perp\approx a_0$ and, neglecting also $1$ against $a_0^2$, we can further reduce (\[dux\_c\]) to the form $$\label{eqcp}
\frac{du_x}{d\tau}=\frac1{2\gamma}\frac{d(a_0^2)}{d\varphi}+\mu a_0^4\frac{1-v_x}{1+v_x}-\tilde{n}\xi.$$ This is precisely the 1D model suggested in [@arxiv]. As mentioned in the Introduction, there are two mechanisms of electron longitudinal acceleration and charge separation in a plasma: PM \[corresponding to the first term in the RHS of (\[eqcp\])\], and RFM \[corresponding to the second term proportional to $\mu$\]. In order to further simplify (\[eqcp\]), consider two opposite limiting cases of the electron longitudinal motion: nonrelativistic ($u_x\ll a_0$) and ultrarelativistic ($u_x\gg a_0$).
### Nonrelativistic longitudinal electron motion.
In the nonrelativistic case $v_x\ll1$ we have $\gamma\approx a_0$, $u_x\approx v_x a_0$, $\varphi\approx \tau$, and (\[eqcp\]) is the same as for a damped harmonic oscillator with slowly varying parameters and shifted equilibrium position: $$\label{eqnr}\fl
\ddot{\xi}+2\varkappa\dot{\xi}+\Omega^2\xi=f,\quad \varkappa(\tau)=\mu a_0^3+\frac{\dot{a}_0}{2a_0},\quad \Omega(\tau)=\sqrt{\frac{\tilde{n}}{a_0}},\quad f(\tau)=\mu a_0^3+\frac{\dot{a}_0}{a_0},$$ where the dots abbreviate the derivatives with respect to $\tau$. In dilute plasmas under consideration here this effective oscillator is always over-damped ($\varkappa\gtrsim\Omega$) and very rapidly (within $\tau\lesssim \varkappa/\Omega^2$) approaches its equilibrium $$\label{xi}
\xi_0(\tau)\simeq \frac{f(\tau)}{\Omega^2(\tau)}=\frac{\mu a_0^4(\tau)+\dot{a}_0(\tau)}{\tilde{n}}.$$ Assuming that $$\label{nrcond}
T\gg \frac{\varkappa}{\Omega^2}\simeq \max\left\{\frac{\mu a_0^4}{\tilde{n}},\frac{a_0}{\tilde{n}T}\right\},$$ where $T=\omega t_{pulse}$ is the dimensionless pulse duration, we can assume that the oscillator is settled at the equilibrium (by occasion, exactly the same condition validates the nonrelativistic regime which is under assumption).
Hence the amplitude of longitudinal charge separation field $a_\parallel^{(nr)}=\tilde{n}\xi^{max}$ can be estimated as $$\label{anr}
a_\parallel^{(nr)}=\tilde{n}\xi^{max}\simeq \max_{\varphi<0}\left[\mu a_0^4(\varphi)+\dot{a}_0(\varphi)\right]
=\left\{\begin{array}{cc}\mu a_0^4,&\mu a_0^3 T\gg 1,\\0.858\, a_0/T, &\mu a_0^3 T\ll 1.\end{array}\right.$$ For a gaussian envelop $a_0\propto e^{-\varphi^2/T^2}$ used in this paper the final maximization cannot be done analytically, but only numerically, because of occurrence of a transcendental equation. The limiting cases above are also given for this particular case. Note that in the nonrelativistic case RF becomes dominant if $$\label{rfeffnr}
\mu a_0^3T\gtrsim 1,$$ i.e. for intensities $I\gtrsim 10^{22}$ W/cm${}^2$ if we assume the FWHM pulse duration $150$ fs as announced for ELI Beamlines [@ELI].
### Ultrarelativistic electrons: RF induced charge separation.
From now on assume that longitudinal electron motion is ultrarelativistic, $\gamma\approx u_x\gg a_0$. Then we have: $1-v_x\approx a_0^2/2u_x^2$, $1+v_x\approx2$, $\xi\approx \tau$, so that (\[eqcp\]) reduces to $$\label{eqcp_ur}
\frac{du_x}{d\tau}=\frac1{2 u_x}\frac{d(a_0^2)}{d\varphi}+\frac{\mu a_0^6}{4u_x^2}-\tilde{n}\tau.$$
Let us consider first the RF mechanism (RFM) of charge separation and neglect the first (ponderomotive) term in the RHS of (\[eqcp\_ur\]). The process obviously splits into two stages. During the [*acceleration stage*]{} $0<\tau<\tau_{acc}$ the last electrostatic term in the RHS remains small and can be also neglected, hence $du_x/d\tau\approx \mu a_0^6/4u_x^2$, or $u_x\sim a_0^2(\mu\tau)^{1/3}$. The acceleration stage ends up at about the moment $$\label{tacc_rf}
\tau_{acc}^{(RFM)}\sim(\mu a_0^6/\tilde{n}^3)^{1/5},$$ when the electrostatic term grows to the same order as the RF term. Next, during the [*deceleration stage*]{} $\tau_{acc}<\tau<\tau_{bd}$, electron steadily decelerates until the [*breakdown*]{} at $\tau=\tau_{bd}$, when it finally leaves the laser pulse and rapidly accelerates backward by the electrostatic field. For the deceleration stage by neglecting the LHS in (\[eqcp\_ur\]) we obtain $u_x\simeq\sqrt{\mu/\tilde{n}\tau}a_0^3/2$. The breakdown time can be estimated from the condition that the total phase acquired by the electron in the pulse equals the pulse duration: $$\label{tbd_cond}
\int\limits_0^{\tau_{bd}}(1-v_x)d\tau=T.$$ Since for the laser and plasma parameters providing an efficient RF induced charge separation we have $\tau_{acc}\ll \tau_{bd}$ (see below), in our estimation we can neglect the duration of the acceleration stage and obtain (see [@arxiv]) $$\label{tau_bd_rfm}
\tau_{bd}^{(RFM)}\simeq\sqrt{\frac{\mu T}{\tilde{n}}}a_0^2.$$ Equation (\[tau\_bd\_rfm\]), together with $$\label{arfm}
a_\parallel^{(RFM)}=\tilde{n}\tau_{bd}^{(RFM)}\simeq \sqrt{\mu\tilde{n}T}a_0^2,$$ define the period and the amplitude of a longitudinal wave generated in a plasma.
### Ultrarelativistic electrons: ponderomotive charge separation.
Now let us assume instead that charges are separated by the ponderomotive mechanism (PM) and neglect the second term in the RHS of (\[dux\]). Estimating the derivative of the envelop [^3] by $d a_0^2/d\varphi\sim a_0^2/T$, we can estimate $u_x\sim a_0\sqrt{\tau/T}$ for the acceleration stage and $u_x\sim a_0^2/(\tilde{n}\tau T)$ for the deceleration stage. The acceleration and breakdown times can be estimated in the same manner as above: $$\label{taupm}
\tau_{bd}^{(PM)}\sim\tau_{acc}^{(PM)}\sim \left(\frac{a_0^{2}}{\tilde{n}^{2}T}\right)^{1/3}.$$ Since in this case the acceleration and breakdown times are of the same order, we can use $\tau_{bd}^{(PM)}$ as a rough order-of-magnitude estimation for the period of a longitudinal wave. The amplitude of the longitudinal field is then given by (see also [@arxiv]) $$\label{apm}
a_\parallel^{(PM)}\simeq \tilde{n} \tau_{bd}^{(PM)}\simeq\left(\frac{a_0^{2}\tilde{n}}{T}\right)^{1/3}.$$
Comparing (\[arfm\]) and (\[apm\]) we conclude that in a transparent plasma longitudinal wave generation is enhanced by RF (RFM dominates over PM) if $$\label{rfeff}
\mu^3\tilde{n}T^5a_0^8\gtrsim 1,$$ i.e. for stronger and longer laser pulses.
Linear polarization
-------------------
Assuming that radiation damping is small, for a linearly polarized laser pulse $\mathbf{A}(\varphi)=(m/e)a_0(\varphi)\{0,\cos\varphi,0\}$ we have $\mathbf{u}\approx\{u_x,a_0(\varphi)\sin\varphi,0\}$. Hence the equation for longitudinal motion (\[dux\]) in ultrarelativistic case takes the form $$\label{eq_lp}
\frac{du_x}{d\tau}\approx\frac{1}{2u_x}\frac{d a_0^2}{d\varphi}\sin^2\varphi+\frac{a_0^2}{2u_x}\sin 2\varphi+\frac{\mu a_0^6\sin^4\varphi\cos^2\varphi}{4u_x^2}-\tilde{n}\tau.$$ Unlike the case of circular polarization, the motion is accompanied by fast oscillations, but their envelope behaves qualitatively very similar to that case. During the deceleration stage an approximate solution to (\[eq\_lp\]) can be represented in the form[^4] $u_x\simeq C(\tau)\sin^2\varphi$. As we will shortly see, for $|\varphi|\sim T\gg 1$ we have $\tau\dot{\varphi}\simeq \varphi\gg 1$, indicating that $\tau$ (and hence $C$) is a slowly varying function of $\varphi$. Hence to find $C(\tau)$ we can average (\[eq\_lp\]) over the laser oscillations. The results are $C(\tau)=a_0^3\sqrt{\mu/(8\tilde{n}\tau)}$ for RFM and $C(\tau)=a_0^2/(2\tilde{n}T\tau)$ for PM. Now we can calculate $\tau(\varphi)$ for both cases from $$\label{dtau}
\frac{d\tau}{d\varphi}\approx \left\langle\frac{2u_x^2}{(a_0\sin\varphi)^2}\right\rangle=\frac{C^2(\tau)}{a_0^2},$$ thus obtaining $\tau(\varphi)\propto \varphi^{1/2}$ for RFM and $\tau(\varphi)\propto \varphi^{1/3}$ for PM. Power dependence of $\tau$ on $\varphi$ validates the assumption of slow variation of $\tau$ for $|\varphi|\sim T\gg1$. From $\tau_{bd}=\tau(T)$ we obtain that after the substitution $a_0\to a_0/\sqrt{2}$ equations (\[tau\_bd\_rfm\])–(\[apm\]) remain valid for linear polarization as well. Hence in a dilute plasma the charge separation field strength, expressed in terms of laser intensity, $$a_\parallel^{(RFM)}\simeq\sqrt{\mu\tilde{n}T}\frac{I}{I_0},\quad a_\parallel^{(PM)}=\left(\frac{\tilde{n}}{T}\frac{I}{I_0}\right)^{1/3},$$ where $I_0=\pi m^2c^5/(e^2\lambda^2)\approx 2.74\cdot 10^{18}$ W/cm${}^2$ (for $\lambda=1\mu$m), is independent of the laser polarization.
Analysis of used approximations {#sec_ca}
-------------------------------
Let us discuss in more detail and summarize the conditions of validity of the approximations we made while deriving the estimations (\[arfm\]) and (\[apm\]):
- The laser field is strong, $a_0\gg1$.
- Longitudinal motion is ultrarelativistic, $u_x(\tau_{bd})\gg a_0$, or $$\label{ur}
\left(\frac{\mu}{\tilde{n}T}\right)^{1/4}a_0\gg1,\quad \mathrm{or}\quad \left(\frac{a_0}{\tilde{n}T^2}\right)^{1/3}\gg1\$$ for RFM and PM cases, respectively \[as they should, these conditions are opposite to the ones in (\[nrcond\])\]. Note that while the period of the created quasistatic longitudinal wave $\simeq T$ in the regime of nonrelativistic longitudinal electron motion, under the conditions (\[ur\]) according to (\[tau\_bd\_rfm\]) and (\[taupm\]) it is $\gg T$.
- Transverse radiation damping is small, i.e. the dimensionless transverse RF force $F_{RF}\simeq\mu a_0^2\gamma^2(1-v_x)v_\bot$ is much smaller than the dimensionless transverse Lorenz force $F_L=a_0$. In the ultrarelativistic case it is equivalent to $\mu a_0^4/u_x\ll1$ or, taking into account that $u_x\gtrsim u_x(\tau_{bd}^{(RFM)})$, to $$\label{damping}
(\mu^3\tilde{n} T)^{1/4} a_0^2\ll1.$$
- For RFM, the acceleration time is much smaller than the breakdown time. In fact, this assumption turns out to be literally equivalent to the condition (\[rfeff\]) of dominance of RFM over PM, see (\[tacc\_rf\]) and (\[tau\_bd\_rfm\]).
- Depletion of the transverse laser field is neglected, i.e. we assume that the (dimensionless) energies $\varepsilon_e$ gained by plasma electrons and $\varepsilon_\parallel$ accumulated in the quasistatic longitudinal field are much smaller than the (dimensionless) energy of the laser pulse $\varepsilon_L\simeq a_0^2T$. The number of accelerated electrons can be estimated as $\tilde{n}\tau_{bd}$ and the (dimensionless) average energy of a single electron as $u_x(\tau_{bd})$, hence for RFM we have $\varepsilon_e\simeq (\mu^3\tilde{n}T)^{1/4}a_0^4$ and $\varepsilon_\parallel\simeq a_\parallel^2\tau_{bd}\simeq \sqrt{\mu^3T^3\tilde{n}}a_0^6$. Note that: if the condition (\[rfeff\]) of RFM dominance is fulfilled, then $\varepsilon_\parallel\gtrsim \varepsilon_e$, meaning that RFM is an efficient mechanism for longitudinal field generation; the condition (\[damping\]) of weak damping is equivalent to $\varepsilon_L\gg\varepsilon_\parallel$ of negligibility of the laser field depletion.
- Immobility of the ions. In reality, the generated longitudinal electric field not only decelerates the electrons, but also accelerates the ions. We can neglect the ion motion in the estimations (\[tau\_bd\_rfm\])–(\[apm\]) only if ions remain nonrelativistic until the electron breakdown. In such a case the equation of ion motion is of the form $\ddot{\xi}_i(\tau)=\xi_i(\tau)/\tau_i^2$ and its solution reads $\xi_i(\tau)\simeq \xi_0 e^{\tau/\tau_i}$, where $$\tau_i=\sqrt{\frac{Am_p}{Z_*m_e\tilde{n}}},$$ $\xi_0$ is the initial position of an ion, $m_p$ is the nucleon mass, $A$ and $Z_*e$ are the ionic weight and charge. An estimation of the ionic charge which we used in this paper is discussed in the Appendix. Since ion acceleration is initially exponential, the ions become relativistic within the time of order $\tau_i$. Hence the condition that ions remain nonrelativistic (and can be considered immobile) is written as $\tau_i\gg \tau_{bd}$ or, explicitly, $$\label{nrions}
a_0\ll\left(\frac{m_p}{m_e}\frac{A}{Z_*}\frac{1}{\mu T}\right)^{1/4}.$$ When the ions become relativistic the charge separation and the longitudinal field should get saturated.
The resulting domains of validity of all the independent assumptions discussed above for two plasma densities $n=0.01 n_c$ and $n=0.1 n_c$ are illustrated in Figure \[fig2\] together with the line separating the dominance of RFM over PM. For clarity, the target region between this line and the topmost line of strong pulse depletion is shaded. The values of intensities from the right axis are converted to the equivalent values of $a_0$ for circular polarization (CP) on the left axis. One can see that for relatively long pulses ($t_{pulse}\gtrsim 100$ fs) RF is dominant already at intensities exceeding $10^{22}$ W/cm${}^2$. Since ionization degree depends on the field strength but not on the intensity, the bottom orange line without markers, separating the relativistic and nonrelativistic regimes of ion motion for circular polarization, is drawn with respect to both left and right axis, but for linear polarization with respect to the right axis should be replaced with the top one.
Optimal density and maximally attainable charge separation field
----------------------------------------------------------------
According to (\[arfm\]) and (\[apm\]), for a dilute plasma and given laser parameters $a_0$ and $T$, the charge separation field is growing monotonically with the plasma density. However, in virtue of (\[nrcond\]) and/or (\[ur\]), for $$\label{nopt}
\tilde{n}\gtrsim\tilde{n}_{opt}\simeq \mu a_0^4/T$$ the electron longitudinal motion becomes nonrelativistic and hence for higher densities the charge separation field is density independent[^5], see (\[anr\]). Even though further increase of plasma density can no more affect the maximal charge separation field, it may enhance the laser pulse damping. Hence the longitudinal charge separation field generated in a plasma is maximal for the optimal density (\[nopt\]) and is bounded by $$\label{amax}
a_\parallel\lesssim a_\parallel^{(max)}\simeq \mu a_0^4\ll a_0,$$ where we used that in the case of optimal density (\[nopt\]) the condition of negligible damping (\[damping\]) reduces to $\mu a_0^3\ll 1$. The upper limit $a_\parallel\lesssim a_0$ could be probably approached when damping is moderate $\varepsilon_\parallel\lesssim \varepsilon_L$, i.e. beyond the applicability of our current model.
PIC simulations results
=======================
In order to confirm our theoretical findings we performed one and two dimensional PIC simulations using the code EPOCH [@EPOCH] with the classical radiation friction included in the form of Landau and Lifshitz [@arxiv]. The numerical approach is similar to the one developed in [@zhidkov2002; @tamburini2010].
The results for the dependence of the longitudinal field amplitude on laser intensity from 1D simulations are presented in Figure \[fig3\]. We used 100 cells per wavelength and 20 particles per cell. The laser pulse envelop was chosen Gaussian with FWHM duration $150$ fs, as expected at ELI Beamlines [@ELI]. The initial plasma density was $n=0.01 n_c$ and the ions were considered immobile. As one can observe, when expressed in terms of intensity, the longitudinal field generation is insensitive to laser polarization, and that the simulation data is in excellent agreement with the approximation (\[anr\]) and the estimation (\[arfm\]). At the same time, neglecting RF results in substantial (up to an order of magnitude) underestimation of the longitudinal field already for the intensities $I\gtrsim 2\cdot 10^{22}$ W/cm${}^2$. Note that for such long pulses electron acceleration in longitudinal direction when RF is neglected (i.e. solely due to PM) remains nonrelativistic for the whole shown laser intensity range. However, by considering shorter pulses we have previously justified in [@arxiv] the estimation (\[apm\]) for PM for ultrarelativistic case as well.
In order to strengthen the reliability of our results, we performed also a series of more realistic 2D EPOCH simulations[^6] with linearly polarized driving pulses (for 2D simulations with circularly polarized pulses see [@arxiv]) and the heavy mobile ions $_{238}^{92}U^{80+}$ (the average ionization degree is estimated according to the Appendix). As in [@arxiv], in order to reduce transverse expulsion of electrons by the ponderomotive force of the pulse, in all simulations we use the pulses with a symmetric bimodal Gaussian transverse profile shown in the inset of the Figure \[fig4\] (a), with the distance $d$ between the peaks fixed so that the transverse field amplitude on the $x$-axis coincides with the amplitude of a single pulse, $d=2w\sqrt{\ln{2}}\approx 1.7w$, where $w$ is the waist of each peak.
Our first simulation (see Figure \[fig4\]) corresponds to the case of a wide (weakly focused of comparable width and length, $w=t_{pulse}=22\lambda$, where the wavelength $\lambda=1\mu$m) driving pulse. Since in this particular case we are mainly interested in stronger validation of our 1D model, we determined the simulation parameters according to Figure \[fig2\] to maintain the longitudinal electron motion ultrarelativistic, while keeping ions nonrelativistic: plasma density $n=0.1 n_c$, laser intensity $I=2.5\cdot10^{23}$ W/cm${}^2$ and FWHM $60$ fs – see the cross between the unmarked orange solid line and the blue line marked with diamonds in the Figure \[fig2\] (b). \[param\_fig4\] From Figures \[fig4\] (a) and (b), where we compare the longitudinal fields computed with and without accounting for RF, respectively, one can observe that for these parameters the effect of RF-induced enhancement is extremely well pronounced in 2D. Moreover, for such wide pulses the longitudinal field distribution on the $x$-axis is in a rather good agreement with both the 1D simulations and the estimations of the preceding section, see Figure \[fig4\] (c). The most notable 2D effect is that some electrons bypass the ion bubble [@bubbles], getting inside from its rear side \[see Figure \[fig4\] (d)\], and in this way screening the quasistatic longitudinal field. Its slight decrease (as compared to the 1D simulation) at the rear side of the resulting longitudinal wave in Figure \[fig4\] (c) is explained partially by this effect, and partially by the nonrelativistic ion motion.
\
Since the laser pulses discussed in Figure \[fig4\] are both intense and wide, their total power is extremely high. In order to substantiate the effect in a more realistic setup, we also present the results of simulations with the pulses tightly focused (waist radius $w\simeq\lambda$) at the left plasma boundary, see Figure \[fig5\]. Here we assume the parameters announced for ELI Beamlines [@ELI]\[param\_tight\]: $I=1.7\cdot 10^{23}$ W/cm${}^2$ (total power $\simeq 10$ PW) and FWHM $150$ fs. In order to suppress strong diffraction by selffocusing, keeping longer the field strong enough for a noticeable effect of RF, we choose here plasma density $n=n_c$. The resulting longitudinal fields averaged over laser wavelength – with and without RF – are shown in Figures \[fig5\] (a) and (b) \[comparison of a root mean square (RMS) of the transverse field distribution with the case $n=0.1n_c$ is shown in Figures \[fig5\] (c) and (d)\]. According to the figures, the effect of RF induced enhancement of the longitudinal field generation in a plasma could be observed on the forthcoming 10 PW laser facilities of the near future.
\
Discussion
==========
A laser pulse propagating in plasma separates the charges thus generating a quasistatic longitudinal electric field, which, as discussed in this paper, under certain conditions can be substantially enhanced due to modification of the electron transverse oscillations by even weak radiation friction. To describe the effect, we develop a 1D model and propose explicit analytical estimates of an amplitude and a period for the enhanced longitudinal field according to a particular regime of the process (nonrelativistic versus ultrarelativistic longitudinal electron motion, and RFM versus PM dominance). In particular we show that, being expressed in terms of the laser intensity, these parameters are in fact independent of polarization of the driving pulse. Our theoretical analysis is limited by the assumptions that the damping of the driving laser pulse is small and that ions are immobile, and we provide a detailed analysis of their validity for a wide range of laser and plasma parameters. Due to energy conservation, the amplitude of the generated longitudinal field can never exceed the amplitude of a driving laser field and, according to our theory, for given pulse intensity and duration, the highest longitudinal field is achieved with such an optimal plasma density that the electron longitudinal motion is mildly relativistic.
We have validated our predictions by comparing them to PIC simulations with both immobile and mobile high-Z ions in 1D, as well as with 2D wide driving pulses. However, consideration of concurrently strong and wide pulses implies unrealistically huge laser power, therefore in order to realize the proposed regime of RF-induced enhancement of the longitudinal field generation within the limitations imposed by the current or foreseeable experimental capabilities, tight focusing is required. For tightly focused pulses our model is no longer valid literally on a quantitative level, because such purely 2D effects as fast diffraction, various plasma instabilities, and the alternating longitudinal selffield of a tightly focused pulse (which can in fact be as strong as the transverse field itself), are vital. Increase of the plasma density, in addition to boosting the magnitude of the generated quasistatic longitudinal field as explained above, also facilitates to suppress pulse diffraction by selffocusing, but is limited by laser radiative damping and by the condition of plasma transparency, as in an opaque plasma electrons cannot penetrate deep inside the pulse to experience a strong enough field for a noticeable RF. According to simulations, by certain tuning of the parameters, the effect of the RF-induced longitudinal field enhancement can be still revealed at the ELI Beamlines facility in the near future [@ELI].
We are grateful to M. Grech and M. Vranic for a useful advise on data visualization and to S.V. Popruzhenko for an extremely valuable discussion of the idea underlying our approach for estimation of the ionization degree and for the relevant references. The research was performed using the code EPOCH (developed under the UK EPSRC grants ``, ``, and ``) and the resources of the ELI Beamlines Eclipse cluster, and was partially supported by the MEPhI Academic Excellence Project (Contract No. ``), the Russian Fund for Basic Research (Grants `` and ``), projects ELITAS (ELI Tools for Advanced Simulation) `` and HiFI (High-Field Initiative) `` from European Regional Development Fund.
Estimation of ionization degree
===============================
In order to deal with the ion motion, we need first to estimate the ionization degree (average ionic charges) for a field of a given strength. Even though now EPOCH contains a subroutine for realistic simulation of ionization, for highly-charged ions it generates a large number of species, and correct account for them requires significant computer resources and can seriously slow down simulations. A simplified approach is also needed if we want to make estimations analytically. Hence here we used for this purpose the following heuristic approach.
In a high intensity laser field ionization occurs via the tunneling mechanism, for which general theory is rather well developed (see, e.g. [@poprz]). One can use the known formulas for ionization probability $w_i(a_0)$ to estimate the ionization degree from the condition $w_i(a_0)\cdot t_{pulse}\simeq 1$. Since $w_i(a_0)$ depends on the field $a_0$ exponentially, and also because the laser pulse duration is large enough on the atomic scale, this criterion can be roughly reformulated as that ionization of an atomic level takes place when the laser field strength exceeds about $10\%$ of the atomic field strength at the corresponding outer shell \[the actual fraction weakly (logarithmically) depends on the parameters and can be specified more accurately if needed, see below\].
Let us consider an atom with atomic number $Z$ and denote its residual ionic charge by $Z_*$, meaning that $Z-Z_*$ electrons remain, while $Z_*$ electrons have escaped due to ionization. We can roughly estimate the principle quantum number $n_{max}$ of the outer shell by using the Pauli principle and the known degeneracies of a Hydrogen-like ion: $$Z-Z_*=\sum\limits_{n=1}^{n_{max}} 2n^2\simeq \frac23n_{max}^3.$$ Assuming further that the corresponding fraction of the nuclei charge is completely screened by residual electrons, we can estimate the outer shell radius as $R_{out}\sim (\hbar^2/m Z_*e^2) n_{max}^2$ and the corresponding atomic field strength as $Z_*e/R_{out}^2=m^2e^5Z_*^3/(\hbar^4n_{max}^4)$. With all that our criterion can be formulated as $$\label{i-degree}
a_0\simeq 0.1\frac{m}{\hbar\omega}\frac{(\alpha Z_*)^3}{\left[\frac32(Z-Z_*)\right]^{4/3}}.$$ The solution to (\[i-degree\]) for Uranium ($Z=92$) and Xenon ($Z=54$) ions that were used in the work on this paper are plotted in Figure \[fig6\].
A more advanced approach could include: more accurate consideration of the mentioned large logarithmic factor (resulting from the known pre-exponential factor of the tunneling formula), an account for relativistic corrections, and a two-stage procedure of first determining an ionization potential from the tunneling formula, and next selecting the corresponding ion using the tabular data (e.g. from [@nist]). We have checked, however, that all these complications do not change our heuristic results within $\simeq 10\%$, which is enough for the rough estimates presented in the paper.
References {#references .unnumbered}
==========
[99]{} Danson C, Hillier D, Hopps N and Neely D 2015 [*High Power Laser Sci. Eng.*]{} **3** 2109–2114. Chu Y, Liang X, Yu L, Xu Y, Xu L, Ma L, Lu X, Liu Y, Leng Y, Li R and Xu Z 2013 [*Opt. Express*]{} **21** 29231–39. Sung J H, Lee H W, Yoo J Y, Yoon J W, Lee C W, Yang J M, Son Y J, Jang Y H, Lee S K and Nam C H 2017 [*Opt. Lett.*]{} **42** 2058–61. Zeng X, Zhou K, Zuo Y, Zhu Q, Su J, Wang Xiao, Wang Xiaodong, Huang X, Jiang X, Jiang D, Guo Y, Xie N, Zhou S, Wu Z, Mu J, Peng H and Jing F 2017 [*Opt. Lett.*]{} [**42**]{} 2014–17. Bahk S W, Rousseau P, Planchon T A, Chvykov V, Kalintchenko G, Maksimchuk A, Mourou G A and Yanovsky V 2004 [*Opt. Lett.*]{} **29**(24) 2837; Yanovsky V, Chvykov V, Kalinchenko G, Rousseau P, Planchon T, Matsuoka T, Maksimchuk A, Nees J, Cheriaux G, Mourou G and Krushelnick K 2008 [*Opt. Express*]{} **16** 2109–14. http://www.eli-beams.eu http://www.eli-np.ro/ http://portail.polytechnique.edu/luli/en/cilex-apollon/apollon Tajima T and Dawson J M 1979 **43** 267. Leemans W P, Gonsalves A J, Mao H S, Nakamura K, Benedetti C, Schroeder C B, Toth C, Daniels J, Mittelberger D E, Bulanov S S, Vay J L, Geddes C G R, Esarey E 2014 **113**(24) 245002. Esirkepov T, Borghesi M, Bulanov S V, Mourou G and Tajima T 2004 175003; Macchi A, Veghini S and Pegoraro F 2009 085003; Bulanov S S, Esarey E, Schroeder C B, Bulanov S V, Esirkepov T Zh, Kando M, Pegoraro F and Leemans W P 2015 **114**, 105003. Wilks S C, Kruer W L, Tabak M and Langdon A B 1992 **69** 1383; Macchi A, Cattani F, Liseykina T V and Cornolti F 2006 , 165003; Qiao B, Zepf M, Borghesi M and Geissler M 2009 145002; Naumova N, Schlegel T, Tikhonchuk V T, Labaune C, Sokolov I V and Mourou G 2009 025002; Schlegel T, Naumova N, Tikhonchuk V T, Labaune C, Sokolov I V and Mourou G 2009 [*Phys. Plasmas*]{} **16** 083103. Daido H, Nishiuchi M and Pirozhkov A S 2012 [*Rep. Prog. Phys.*]{} **75** 056401; Macchi A, Borghesi M and Passoni M 2013 , 751. Wagner F, Deppert O, Brabetz C, Fiala P, Kleinschmidt A, Poth P, Schanz V A, Tebartz A, Zielbauer B, Roth M, Stöhlker T and Bagnoud V 2016 **116** 205002; Scullion C, Doria D, Romagnani L, Sgattoni A, Naughton K, Symes D R, McKenna P, Macchi A, Zepf M, Kar S and Borghesi M, 2017 **119** 054801. Siminos E, Grech M, Skupin S, Schlegel T and Tikhonchuk V T, E 2012 **86** 056404. Gelfer E G, Elkina N V, Fedotov A M 2017 Unusual face of radiation friction: enhancing production of longitudinal plasma waves arXiv:1710.09253. Akhiezer A and Polovin R 1956 [*Sov. Phys. JETP*]{} [**3**]{} 696–705. Voronin B S and Kolomenskii A A 1965 [*Sov. Phys. JETP*]{} **65**, 1027; Zeldovich Y B 1975 [*Sov. Phys. Usp.*]{} **18** 79; Fradkin D M 1979 **42** 1209; Di Piazza A 2008 [*Lett. Math. Phys.*]{} **83** 305. Zhidkov A, Koga J, Sasaki A and Uesaka M 2002 **88** 185002. Tamburini M, Pegoraro F, Di Piazza A, Keitel C H and Macchi A 2010 **12** 123005. Tamburini M, Pegoraro F, Di Piazza A, Keitel C H, Liseykina T V and Macchi A 2011 [*Nucl. Instr. Methods Phys. Research*]{} A **653** 181. Brady C S, Ridgers C P, Arber T D, Bell A R and Kirk H G 2012 **109** 245006. Nakamura T, Koga J K, Esirkepov T Z, Kando M, Korn G and Bulanov S V, **108**, 195001 (2012). Bashinov A V and Kim A V 2013 [*Phys. Plasmas*]{} **20** 113111. Stark D J, Toncian T and Arefiev A V 2016 **116** 185003. Landau L D and Lifshitz E M 1975 [*The classical theory of fields*]{} (Oxford: Pergamon Press). Brady C S and Arber T A 2011 **53** 015001. Bulanov S V, Pegoraro F, Pukhov A M and Sakharov A S 1997 **78** 4205; Pukhov A and Meyer-ter-Vehn J 2002 [*Appl. Phys.*]{} B **74** 355. Karnakov B M, Mur V D, Popruzhenko S V and Popov V S 2015 **58** 3. Kramida A, Ralchenko Yu, Reader J and NIST ASD Team (2017). NIST Atomic Spectra Database (ver. 5.5.1), \[Online\]. Available: https://physics.nist.gov/asd \[2018, January 5\]. National Institute of Standards and Technology, Gaithersburg, MD.
[^1]: We use CGS-like units but with speed of light $c=1$.
[^2]: For discussion of the validity of this approximation see Section \[sec\_ca\] below.
[^3]: Here for simplicity we drop all the numerical factors.
[^4]: Unlike the case of circular polarization, we can neglect in this stage the derivative of the envelop $C$, but not of the remaining rapidly oscillating factor.
[^5]: Note, however, that for densities (\[nopt\]) the plasma effects ignored in our model can be important.
[^6]: For 2D PIC simulations we use 50 cells per wavelength and 10 particles per cell.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An alternative to the standard cold dark matter model has been recently proposed in which a significant fraction of the energy density of the universe is due to a dynamical scalar field ($Q$) whose effective equation-of-state differs from that of matter, radiation or cosmological constant ($\Lambda$). In this paper, we determine how the $Q$-component modifies the primordial inflation gravitational wave (tensor metric) contribution to the cosmic microwave background anisotropy and, thereby, one of the key tests of inflation.'
address: |
Department of Physics and Astronomy\
University of Pennsylvania\
Philadelphia, PA 19104
author:
- 'R. R. Caldwell and Paul J. Steinhardt'
title: The Imprint of Gravitational Waves in Models Dominated by a Dynamical Cosmic Scalar Field
---
epsf
Introduction {#sectionintro}
============
A key prediction of inflationary cosmology is a nearly scale-invariant spectrum of initial fluctuations composed of energy density (scalar metric)[@GP; @Hawking; @Starobinskii; @BST] and gravitational wave (tensor metric)[@Rub; @Staro; @AW] perturbations. The two components leave imprints on the cosmic microwave background (CMB) anisotropy that are potentially detectable and distinguishable. The spectrum of each component is characterized by a spectral index ($n$) which determines how the perturbation amplitude varies with wavelength. Each spectrum is also predicted to be nearly scale-invariant; using the standard notational convention, this corresponds to $n_S \approx 1$ for the scalar metric fluctuations and $n_T \approx 0$ for the tensor metric fluctuations. The spectral indices and the ratio of tensor to scalar amplitudes on any given scale are determined by the equation-of-state during inflation or, equivalently, the inflaton potential. A prediction of inflation is a simple relation between $n_S$, $n_T$ and the ratio of tensor to scalar amplitudes that applies to nearly all inflationary models.[@Davisetal] Unlike other features of inflation, such as flatness or a nearly Harrison-Zel’dovich spectra, the predicted relation between the spectral indices and amplitudes was unanticipated prior to the development of the inflationary model and is, in this sense, a unique stamp of inflation.
It is hoped that forthcoming measurements of CMB anisotropy[@KnoxTurner; @Jungman], complemented by measurements of CMB polarization[@CMBgravTest], can be used to detect a tensor component and to test the inflationary scenario. The CMB temperature anisotropy power spectrum can be expressed in multipole moments $C_{\ell}$, where each $C_{\ell}$ can be divided into a sum of independent tensor ($C_{\ell}^{(T)}$) and scalar ($C_{\ell}^{(S)}$) subcomponents. While the individual tensor and scalar contribution to the temperature anisotropy cannot be observed independently, the presence of a strong tensor component may be measured through correlations of polarization and temperature.[@Selj; @Kamion; @Sperg] The amplitude of these correlations, characterized by the ratio of tensor-to-scalar multipole moments, $r_{\ell} \equiv C_{\ell}^{(T)}/C_{\ell}^{(S)}$, can be used to test the inflationary scenario. However, the moments do not depend on the primordial relation between spectral indices and amplitudes alone. The power spectrum also depends on the evolutionary history since inflation ended. For example, in the standard cold dark matter (sCDM) model in which $\Omega_m =1$, the inflationary prediction[@Davisetal; @Oth] is well-known to be $$r_2 \approx 7(1-n_S).$$ This prediction is modified for open CDM models or CDM models with a cosmological constant ($\Lambda$), in which the post-inflation evolution is different, as has been studied previously.[@KnoxPRD]
The purpose of this paper is to discuss the gravitational wave contribution to the CMB anisotropy in a new class of cosmological models where a significant fraction of the energy density of the universe takes the form of a cosmic scalar field ($Q$) with an equation-of-state different from that of matter, radiation or cosmological constant.[@CDS] The scalar field component of the cosmic fluid has been dubbed “quintessence" and cosmological models based on a combination of quintessence and cold dark matter components are known as QCDM models. In a recent paper[@CDS] (henceforth, referred to as Paper I), we computed the background evolution, the CMB power spectrum and the mass spectrum for QCDM models. We showed that models of this type result in a significantly better fit than sCDM to the CMB temperature power spectrum, the mass power spectrum, early structure formation, and distant supernovae and gravitational lens count measurements. In Paper I we restricted our attention to the case of strictly scale-invariant ($n_S=1$) spectra of purely scalar metric perturbations. In the present paper, we discuss how the tensor contribution to the CMB anisotropy is modified by the presence of a $Q$-component. We compute the full, scalar plus tensor CMB power spectrum for QCDM models based on inflationary initial conditions and determine how the relation between primordial spectral indices and amplitudes predicted by inflation is modified in QCDM models. The result is a generalization of a key test of inflationary cosmology.
The organization of this paper is as follows. In Section \[section2\] we briefly review the predictions of inflation for the spectrum of initial tensor and scalar perturbations and the resultant contributions to the CMB anisotropy. In Section \[section3\] we discuss how the computation of the tensor contribution is modified in QCDM models. (The computation of the scalar spectrum has been described in Paper I.) We then present the computed tensor contribution to the CMB power spectrum for QCDM models and compare with an sCDM model. In Section \[section4\] we present our key result, the generalization of the inflationary relation between spectral index and amplitude for QCDM models.
Review of Inflationary Predictions for the sCDM Model {#section2}
=====================================================
Inflation predicts a nearly scale-invariant power spectrum of energy density (scalar metric) and gravitational wave (tensor metric) fluctuations. For both subcomponents, the power spectrum as a function of Fourier mode $k$ can be expressed to lowest order in terms of the Hubble constant and its time-derivatives evaluated when the Fourier mode $k$ was stretched beyond the horizon during inflation, at $k=aH$. Here $a$ is the Friedmann-Robertson-Walker scale factor and $H=H(\phi)$ is the Hubble parameter, which depends on the expectation value of the inflaton field, $\phi$. Expanding around some given wavenumber $k_0$, the power spectrum can be parameterized by spectral amplitudes $A_{(S,T)}$ and spectral indices $n_{(S,T)}$: $$\begin{aligned}
P_S(k) &=& A_S^2 \left(\frac{k}{k_0}\right)^{n_S -1}
= 4 \left(\frac{H^2}{m_p^2 |H'|} \right)^2 {\large|}_{k=aH}
\\
P_T(k) &=& A_T^2 \left(\frac{k}{k_0}\right)^{n_T} =
\frac{16}{\pi} \left( \frac{H}{m_p}
\right)^2 {\large |}_{k=aH},\end{aligned}$$ where $H' = dH/d\phi$ and $m_p$ is the Planck mass.
A signature of inflation is the series of relations between spectral indices and amplitudes.[@Davisetal] The tensor spectral index, $H(\phi)$ and the power spectrum amplitudes obey the relation[@Davisetal; @Oth] $$n_T = - \frac{m_p^2}{2 \pi} \left( \frac{H'}{H}\right)^2
= -\frac{1}{8} \frac{A_T^2 }{A_S^2 }.$$ The ratio of power spectrum amplitudes can be related to the ratio of tensor-to-scalar contributions to the CMB anisotropy quadrupole moment $$\label{key1}
r_2 \equiv \frac{C_2^{(T)}}{C_2^{(S)}} \approx - 7 n_T.$$ We will comment on the coefficient below. The scalar spectral index satisfies a more complicated relation: $$n_S -1 = n_T - \frac{m_p^2}{2 \pi} \,
\left(\frac{H'}{H}\right)'.$$ However, for all but an exceptional set of inflaton potentials, the Hubble parameter evolves so slowly during inflation that the second term is negligible and $n_S-1 \approx n_T$. In this case, $$\label{key}
r_2 \approx 7 (1-n_S).$$ Hence, the relations arise because all parameters associated with the perturbation spectrum are ultimately determined only by $H$ and its derivatives during inflation.
The coefficient in Eqs. (\[key1\]) and (\[key\]) is crucial for the purposes of this paper. In the sCDM model, the coefficient is set largely by the Sachs-Wolfe effect, the variation of the gravitational potential on the surface of last scattering, which is only very weakly dependent on model parameters such as the Hubble constant or the baryon density. Hence, Eq. (\[key\]) has been presented in the literature[@Davisetal; @Oth] as a robust prediction of inflation. However, this prediction is valid [*only*]{} if $\Omega_m=1$. In models in which the matter density is less than unity, such as the QCDM models considered in this paper or models with a cosmological constant, there is a large, integrated Sachs-Wolfe contribution to the large angle CMB anisotropy, which changes the predicted relationship between the spectral indices and the tensor-to-scalar ratio. Because the integrated Sachs-Wolfe effect depends on the time variation of the gravitational potential along the line-of-site to the last scattering surface, the amplitude of the effect is sensitive to the present value of $\Omega_m= 1-\Omega_Q$, the equation-of-state $w$, and the variation of $w$ with time. Consequently, the coefficient in Eq. (\[key\]) is modified by a model-dependent function (Section \[section4\]), the central result of this paper.
Tensor Contribution to the CMB Anisotropy in QCDM Models {#section3}
========================================================
In this section we discuss the QCDM models for which we have computed the gravitational wave contribution to the CMB anisotropy. We present the equations necessary to evolve the background and tensor perturbations. We discuss the properties of the Q-matter and their dependence on the effective potential for $Q$, $V(Q)$, and identify two broad categories of models. This classification simplifies our survey, in the following section, of the imprint of gravitational waves in QCDM models. Finally, we present some sample results of our computation of the tensor contribution to the CMB anisotropy spectrum.
Background Equations
--------------------
The QCDM models are constructed from spatially flat, Friedmann-Robertson-Walker (FRW) space-times containing baryons, cold dark matter, neutrinos, radiation and a cosmic scalar field or Q-component. The space-time metric is given by $ds^2 = -dt^2 + a^2(t) d\vec x^2$ where $a$ is the expansion scale factor and $t$ is the cosmological time. The background equation of motion, the energy density, and the pressure for the Q-matter are $$\begin{aligned}
&& \ddot Q + 3 H \dot Q = - {\partial V \over \partial Q},
\label{Qbackground1} \\
&& \rho_Q = {1 \over 2 }\dot Q^2 + V, \qquad
p_Q = {1 \over 2}\dot Q^2 - V ,
\label{Qbackground2}\end{aligned}$$ where the dot represents $\partial/\partial t$. Hence, the relations in (\[Qbackground1\]-\[Qbackground2\]) supplement the usual background equations to specify the evolution of all components of the cosmological fluid.
Gravitational Wave Equations
----------------------------
Fluctuations about the background space-time can be represented in the synchronous gauge, where the linearized space-time metric is $$ds^2 = -dt^2 + a^2(t)(\gamma_{ij} + h_{ij})dx^i dx^j.$$ Here, $\gamma_{ij}$ is the unperturbed spatial metric and $h_{ij} = h^S_{ij} +
h^T_{ij}$ is the metric perturbation, which includes scalar (S) and tensor (T) perturbations. The equations of motion for the scalar perturbations and the fluctuations in the Q-component have been described elsewhere. The transverse, traceless gauge constraints and equations of motion of the gravitational wave perturbation are $$\begin{aligned}
&& \widetilde\nabla^i {h^T}_{ij} = \gamma^{ij} {h^T}_{ij} = 0 , \cr\cr
&& \Big[ {\partial^2 \over \partial t^2} + 3 H {\partial \over \partial t }
- {1 \over a^2} \widetilde\nabla^m \widetilde\nabla_m \Big]{h^{T\,i}}_j = 0,\end{aligned}$$ where $\widetilde\nabla_j$ denotes the covariant derivative with respect to the spatial metric $\gamma_{ij}$. Because we have modeled the Q-matter as a scalar field, the explicit form of the equation of motion for the tensor perturbations is unchanged; there are no additional inhomogeneous tensor fields or anisotropic tensor sources (as would occur if the Q-component were modeled as a tangled web of non-intercommuting cosmic strings,[@PenSperg] for example). Hence, the only effect on the evolution of the gravitational wave amplitude is through the background expansion. It is then straightforward to adopt the standard algorithms for evolving the Boltzmann equations with tensor fluctuations for the problem at hand. We have modified a series of standard Boltzmann CMB codes based on the synchronous gauge, maintaining a fixed relationship between the initial amplitudes of the scalar and tensor power spectra, for the computations described below.[@Crit; @MaandB; @Sel]
Classification of QCDM Models
-----------------------------
The CMB anisotropy spectrum, both scalar and tensor subcomponents, is sensitive to the detailed time evolution of the Hubble parameter $H$ since last scattering. In the case of the QCDM models, these subcomponents are affected by the time evolution of the equation-of-state of the Q-component, $w \equiv
p_Q/\rho_Q$, which in turn reflects the form of the potential, $V(Q)$. We find that for a wide range of potentials, however, there are two broad categories by which we may classify the behavior of the models. Hence, we will focus our attention on representative models from each of these categories.
The two categories correspond to cases where $w$ increases monotonically versus cases where $w$ has begun to oscillate about an asymptotic value by the present epoch. In either case, the initial value of the scalar field $Q$ is assumed to be set at a particular value by some conditions ([*e.g.*]{}, inflation) in the early universe. So long as $V''(Q) \ll H^2$, $Q$ remains fixed because the Hubble red shift term ($3 H \dot{Q}$) dominates the equation-of-motion. Consequently, the energy density is nearly constant during the early history of the universe so that $w\approx -1$. Once $H$ reduces sufficiently that $Q$ begins to evolve down the potential, the balance of kinetic and potential energy in $Q$ changes and $w$ begins to grow. Depending on $V(Q)$ and the initial conditions, $w$ may continue to grow up to the present epoch or it may begin to oscillate around some asymptotic value. The first possibility corresponds to the “monotonic" class of models. Here we find that the imprint of $Q$ on the CMB anisotropy and the mass power spectrum is well-approximated by the result obtained if $w$ is held constant at roughly the mean value during the period when $\Omega_Q$ is non-negligible. See Figure \[figure1\]. The oscillatory class of models corresponds to cases where $Q$ evolves to a point in its potential where $w$ begins to oscillate around some asymptotic value. One example is a potential in which $Q$ evolves towards, and then begins to oscillate about, a minimum of $V(Q)$ before the present epoch. Correspondingly, $w$ begins to oscillate around a mean value determined by the shape of the potential, [*e.g.*]{} $w \rightarrow
0$ for a quadratic potential ($V \propto Q^2$), $w \rightarrow 1/3$ for a quartic potential ($V \propto Q^4$). See Figure \[figure2\]. Another example is an exponential potential of the form $V = m^4 \, {\rm exp}(-\beta Q)$; as the potential energy begins to dominate the universe, the equation-of-state first overshoots, then relaxes towards $w \rightarrow (\beta^2/24 \pi) -1$ for $\beta \le \sqrt{48 \pi}$. We will use this exponential example to represent the oscillatory-$w$ category.
Note that a given potential can belong to either the monotonic or oscillatory class depending on parameters and the initial conditions for $Q$. Namely, for some choices of the initial value of $Q$, $w$ may have begun to oscillate around its asymptotic value by the present epoch. For other choices, $Q$ may not have evolved so far by the present epoch and $w$ has been increasing monotonically. For example, while an exponential potential is illustrated both in Fig. \[figure1\] (under monotonic) and Fig. \[figure2\] (under oscillatory), different initial conditions, and values of $m$ and $\beta$, have been selected in the two cases.
Scalar and Tensor Anisotropy Spectra in QCDM Models
---------------------------------------------------
We have computed the CMB anisotropy spectrum due to the scalar and tensor subcomponents for a number of QCDM models, which we now present.
Figure \[figure3\] shows the full CMB power spectrum predicted by inflation for tilted sCDM and a QCDM model with a constant $w=-1/3$ equation-of-state and $\Omega_Q=0.7$, each with $n_S=0.9$. In both panels the total spectrum is broken down into scalar and tensor components. It it noticeable that in the range $2
\le \ell \le 10$ the fractional contribution of the tensor spectrum to the total power is different in the two models.
Figure \[figure4\] indicates how the $Q$-component changes the shape of the tensor subcomponent of the power spectrum depending on $w$ for a set of constant equation-of-state models. The tensor power spectra in the first panel have been artificially normalized to $C^{(T)}_2=1$ to compare the shapes, demonstrating that the shape is not strongly affected by the change in $w$. In the second panel, the curves have been properly normalized with respect to COBE. Here we see that the main difference is in the overall amplitude; for increasing $w$, the fractional contribution of the tensor spectrum decreases. In Fig. \[figure5\], a similar set of panels demonstrates the effect of changes in $\Omega_Q$ on the tensor subcomponent. While the shape is not strongly affected, we see that for increasing $\Omega_Q$, the fractional contribution of the tensor spectrum decreases.
In Paper I, it was already noted that, even for purely scalar metric fluctuations, the CMB power spectrum at large angular scales (low $\ell$) in QCDM models exhibits unusual features that do not occur in sCDM or other conventional models. This is owing to a combination of the modification of cosmic expansion caused by $Q$ (that is, an integrated Sachs-Wolfe effect) and the direct effect of fluctuations in $Q$. Adding a tensor component can add to further features at low $\ell$.
Generalization of the Inflationary Prediction for QCDM Models {#section4}
=============================================================
We have discussed in Section \[section3\] how a very large class of QCDM models can be divided into two categories: (1) models in which $w$ is constant or monotonically increasing; and, (2) models in which $w$ overshoots and then approaches or oscillates about an asymptotic value. We have computed the scalar and tensor components of the CMB anisotropy for representative models of each type for a wide range of parameters. We have used the numerical results to obtain a revision of the inflationary relation between spectral amplitudes and spectral indices.
Our results are expressed in terms of an empirical relation between the scalar spectral index $n_S$ and the ratios $r_2 \equiv C^{(T)}_2/C^{(S)}_2$ and $r_{10} \equiv C^{(T)}_{10}/C^{(S)}_{10}$. The quadrupole is a conventional choice; we have also chosen $\ell=10$ because $C_{10}$ is only weakly-dependent on the cosmological model (compared to $C_2$) when the predicted spectra are COBE-normalized, and so our relations can be applied more simply. These differences between QCDM and sCDM results have been expressed in terms of correction factors, $f_2$ and $f_{10}$, multiplying the known sCDM relations:[@Davisetal; @Oth] $$\begin{aligned}
r_2 &\approx& \Big[ 7(1 - n_S) \Big]
\times f_2(\Omega_Q,n_S) \cr\cr
r_{10} &\approx& \Big[ 4.8 (1 - n_S) \Big]
\times f_{10}(\Omega_Q,n_S) .\end{aligned}$$ We have not included the higher order corrections, such as those proportional to $d n_S/d\, {\rm ln}\, k$, which are negligibly small for most models.[@Wangetal] The functions $f_{2, 10}$ are defined in the following subsections. In all cases, $f\to 1$ as $\Omega_Q \to 0$. The dependence on cosmological parameters $h,\,\Omega_b$ is very weak (which is why the relations are considered to be model-independent tests of inflation) and has been ignored.
Monotonic Evolution of $w$
--------------------------
Models in which $w$ evolves monotonically leave an imprint on the CMB that is well-approximated by the constant $w$ models in which $w$ is set to the average value during which $\Omega_Q$ is non-negligible. The ratios $r_2$ and $r_{10}$ for the constant equation-of-state models are shown in Fig. \[figure7\] for the case $n_S=0.9$. Based on plots of this sort for a range of $n_S$, we have obtained the following correction factors for QCDM models. For strictly constant $w$ models, empirical fits are valid to within $10\%$ for $0\le w \le
-1$, $0.7 \le n_S \le 1$, and $0 \le \Omega_Q \le 0.7$. $$\begin{aligned}
f_{2} &=& \Big[ 1 + {10 \over 9}
(1 - n_S )(2 + w) \Omega_Q^2 \Big] \,
(1 - {\Omega_Q /x})^{g_2(\Omega_Q/x) }
\cr\cr
&& g_2(y) = -0.21 + 2.35 y - 1.03 y^2, \qquad
x = {8 \over 5}[1 - {1 \over 2} w - {1 \over 20}(1+w)^5]
\label{rwapproxeqn} \\
f_{10} &=& \Big[ 1 + {1 \over 10}
(1 - n_S )((8 + 7 w) \Omega_Q^2 + 3) \Big] \,
(1 - {\Omega_Q /x})^{g_{10}(\Omega_Q/x) }
\cr\cr
&& g_{10}(y) = 0.18 + 0.84 y^2, \qquad
x = {3 \over 4}[1 - {2 \over 3} w + {5 \over 3}w^2 - {1 \over 2}(1+w)^5]
\label{dwapproxeqn}\end{aligned}$$ This somewhat complicated expression is needed to have a formula that fits the dependence of all three parameters ($w$, $\Omega_Q$, and $n_S$) over the stated range. In the case of $\Lambda$CDM, where $w=-1$, the function $f_2$ agrees with the $\Lambda$CDM result.[@KnoxPRD].
Oscillatory Evolution of $w$
----------------------------
Models in which $w$ overshoots and then approaches or oscillates about an asymptotic value are well-represented by exponential potentials in which $V(Q)$ dominates the energy density of the universe by the present epoch. The ratios $r_2$ and $r_{10}$ for the exponential potential models are shown in Fig. \[figure8\]. Empirical fits to the correction factors, for the same range of models and with the same accuracy as described in the previous subsection, are given below. $$\begin{aligned}
f_{2} &=& \Big[ 1 + {10 \over 9}
(1 - n_S )(2 + w) \Omega_Q^2 \Big] \,
(1 - {\Omega_Q /x})^{g_2(\Omega_Q/x) }
\cr\cr
&& g_2(y) = -0.21 + 2.35 y - 1.03 y^2, \qquad x = 1 - {3 \over 5 }w
\label{rxapproxeqn} \\
f_{10} &=& \Big[ 1 + {1 \over 10}
(1 - n_S )((8 + 7 w) \Omega_Q^2 + 3) \Big] \,
(1 - {\Omega_Q /x})^{g_{10}(\Omega_Q/x) }
\cr\cr
&& g_{10}(y) = 0.18 + 0.84 y^2, \qquad x = 1 - {3 \over 2 }w
\label{dxapproxeqn}\end{aligned}$$ In these expressions, $w$ represents the present value of the equation-of-state, $w = w(t_0)$. In all cases, the initial conditions correspond to $w \rightarrow -1$. In the expressions above, the limit $w=-1$ corresponds to $w=-1$ throughout, which coincides with the standard $\Lambda$CDM result.[@KnoxPRD] We note that these equations do not apply to extreme cases in which the oscillations in the Q-matter are strong enough and begin recently enough to leave a distinct feature at large angular scales in the CMB power spectrum (such as a sharp peak at low $\ell$). Because the scalar and tensor subcomponents for such a model depend sensitively on the detailed evolution of the Q-component, a correction factor must be computed model-by-model.
Conclusion {#sectionend}
==========
We have described how Boltzmann codes to compute the CMB anisotropy power spectrum and the mass power spectrum in QCDM models can be simply modified to incorporate the contribution of tensor metric fluctuations, as predicted by inflationary cosmology. We have demonstrated that a $Q$-component has two important effects on the tensor component. First, by modifying the expansion history of the universe and, hence, producing an integrated Sachs-Wolfe contribution, a $Q$-component changes the shape of the tensor anisotropy power spectrum (see Fig. \[figure3\]). Secondly, since the same effect modifies the scalar component, but by a different factor, the ratio of tensor-to-scalar contributions to the CMB anisotropy is changed. The net result for any given tilt is to reduce the tensor contribution compared to sCDM when spectra are COBE normalized (see Fig. \[figure4\]). Since the mass power spectrum is normalized by the scalar contribution to the CMB anisotropy, a consequence is that, for a given tilt, including the tensor contribution does not reduce the COBE normalization of the mass power spectrum as much as in sCDM models. (However, as pointed out in Paper I, the shape of the mass power spectrum in QCDM is changed and the small-scale power is reduced by other effects that do not occur in sCDM models.)
Finally, the inflationary relations linking the ratio of tensor-to-scalar multipole moments to the tilt ($n_S$) must be modified. The most important results presented here are the generalized relations shown in the previous section. These relations provide the key test for inflation in future CMB anisotropy and polarization measurements, now extended to include QCDM models. Assuming inflation is correct, the relations are also used in the fitting procedure to determine the Hubble parameter, baryon density, and other cosmic parameters from the CMB anisotropy.[@Jungman] The determination of cosmic parameters from CMB anisotropy measurements in QCDM models, including the effect of the new relations between spectral amplitudes and tilt, will be discussed in a forthcoming publication.
This research was supported by the Department of Energy at Penn, DE-FG02-95ER40893.
A.H. Guth and S.-Y. Pi. [*Phys. Rev. Lett.*]{} [**49**]{}, 1110, (1982).
S.W. Hawking. [*Phys. Lett. B*]{} [**115**]{}, 295 (1982).
A.A. Starobinskii. [*Phys. Lett. B*]{} [**117**]{}, 175 (1982).
J.M. Bardeen, P.J. Steinhardt, and M.S. Turner. [*Phys. Rev. D*]{}[**28**]{}, 679 (1983).
V.A. Rubakov, M.V. Sazhin, A.V. Veryaskin, [*Phys. Lett. B*]{}[**115**]{}, 189 (1982).
A.I. Starobinskii, [*Sov. Astron. Lett.*]{} [**11**]{}, 133 (1985).
L.F. Abbott and M. Wise, [*Nucl. Phys. B*]{} [**244**]{}, 541 (1984).
R.L. Davis, H.M. Hodges, G.F. Smoot, P.J. Steinhardt, and M.S. Turner, [*Phys. Rev. Lett.*]{} [**69**]{}, 1856 (1992).
L. Knox and M.S. Turner, [*Phys. Rev. Lett.*]{} [**73**]{}, 3347 (1994).
G. Jungman, M. Kamionkowski, A. Kosowsky, D.N. Spergel, [*Phys. Rev. D*]{}[**54**]{}, 1332 (1996).
David Spergel and Matias Zaldarriaga, “CMB polarization as a direct test of inflation”, astro-ph/9705182 (1997).
U. Seljak and M. Zaldarriaga, [*Phys. Rev. Lett.*]{} [**78**]{}, 2054 (1997).
M. Kamionkowski, A. Kosowsky, and A. Stebbins, [*Phys. Rev. Lett.*]{} [**78**]{}, 2058 (1997).
D. N. Spergel and M. Zaldarriaga, astro-ph/9705182.
F. Lucchin, S. Matarrese, and S. Mollerach, [*Ap. J. Lett.*]{} [**401**]{}, 49 (1992); D. Salopek [*Phys. Rev. Lett.*]{} [**69**]{}, 3602 (1992); A. Liddle and D. Lyth, [*Phys.Lett.B*]{} [**291**]{}, 391 (1992); V. Sahni and T. Souradeep, [*Mod. Phys. Lett.*]{} [**A7**]{}, 3541i (1992); J.E. Lidsey and P. Coles, [*Mon. Not. Roy. astr. Soc.*]{} [**258**]{}, 57P (1992); L. Krauss and M. White, [*Phys. Rev. Lett.*]{}[**69**]{}, 869 (1992).
Lloyd Knox, Phys. Rev. [**D 52**]{}, 4307 (1995).
R.R. Caldwell, R. Dave, and P.J. Steinhardt, astro-ph/9708069.
D. N. Spergel and U.-L. Pen. astro-ph/9611198.
R. Crittenden, J.R. Bond, R.L. Davis., G. Efstathiou, and P. J. Steinhardt, [*Phys. Rev. Lett.*]{}[**71**]{}, 324 (1993).
C.-P. Ma and E. Bertschinger, [*Ap. J.*]{} [**455**]{}, 7 (1995).
U. Seljak and M. Zaldarriaga, astro-ph/9603033.
L. Wang, V. Mukhanov, and P.J. Steinhardt, to appear in [*Phys. Lett.*]{} (1997).
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
=1.5cm =1.5cm 2.0in
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, operator quantum error-correcting codes have been proposed to unify and generalize decoherence free subspaces, noiseless subsystems, and quantum error-correcting codes. This note introduces a natural construction of such codes in terms of Clifford codes, an elegant generalization of stabilizer codes due to Knill. Character-theoretic methods are used to derive a simple method to construct operator quantum error-correcting codes from any classical additive code over a finite field, which obviates the need for self-orthogonal codes.'
---
**Clifford Code Constructions of\
Operator Quantum Error-Correcting Codes**\
Andreas Klappenecker and Pradeep Kiran Sarvepalli\
Department of Computer Science, Texas A&M University
#### Introduction.
One of the main challenges in quantum information processing is the protection of the quantum information against various sources of errors. A possible remedy is given by encoding the quantum information in a subspace $C$ of the state space $H$ of the quantum system. If such a quantum error-correcting code $C$ is well-chosen, then many errors can be corrected through active recovery operations. A more recent development is the encoding of quantum information into a subsystem $A$ of the state space [@kribs05; @kribs05b]. This means that $C$ is further decomposed into a tensor product of vector spaces $A$ and $B$ such that $$H = C\oplus C^\perp = (A\otimes B) \oplus C^\perp.$$ One refers to $C$ as an operator quantum error-correcting code with subsystem $A$ and co-subsystem $B$. Some authors refer to the co-subsystem as the gauge subsystem. One advantage is that errors affecting the co-subsystem $B$ alone do not require any active error-correction. Furthermore, one can detect all errors that map the encoded information into the orthogonal complement $C^\perp$ of $C$.
The operator quantum error-correcting codes generalize and unify the main methods of passive and active quantum error-correction: decoherence free subspaces, noiseless subsystems, and quantum error-correcting codes. More background on operator quantum error-correcting codes can be found, for example, in references [@bacon06; @knill06; @kribs05; @kribs05b; @kribs05c; @poulin05].
The purpose of this paper is to introduce a natural method for constructing such operator quantum error-correcting codes. Our approach is based on an elegant formalism to construct quantum error-correcting codes that has been introduced in 1996 by Knill as a generalization of the stabilizer code concept. At the heart of this quantum code construction is a famous theorem by Clifford concerning the restriction of irreducible representations of finite groups to normal subgroups, so we referred to these codes as “Clifford codes” in [@klappenecker031; @klappenecker033], although “Knill codes” is perhaps a more appropriate name. Unexpectedly, it turned out that Clifford codes are in many cases stabilizer codes, so this construction did not become as widely known as it should.
In our approach, we construct a Clifford code $C$ and give conditions that ensure that this code decomposes into a tensor product $C=
A\otimes B$. The Clifford codes allow us to control the dimensions of $A$ and $B$, and we get a simple characterization of the detectable errors of the operator quantum error-correcting code. Since there may exist many different ways to construct the same Clifford code $C$, we should note that these constructions can lead to different tensor product decompositions. In fact, even if one is just interested in the tensor decomposition of a stabilizer code $C$, then the Clifford codes can provide a natural way to induce an operator quantum error-correcting code on $C$.
[*Notation.* If $N$ is a group, then $Z(N)$ denotes the center of $N$. We denote by $\operatorname{Irr}(N)$ the set of irreducible characters of $N$. If $\chi$ and $\psi$ are characters of $N$, then $(\chi,\psi)_N=|N|^{-1}\sum_{n\in N} \chi(n)\psi(n^{-1})$ defines a scalar product on the vector space of class functions on $N$, and $\operatorname{Irr}(N)$ is an orthonormal basis of this space. We denote by $\operatorname{supp}(\chi)=\{ n\in N|\, \chi(n)\neq 0\}$. If $\chi\in \operatorname{Irr}(N)$, then $Z(\chi)=\{ n\in N\,|\, \chi(1)=|\chi(n)|\}$ denotes the quasikernel of $\chi$. Suppose that $G$ is a group that contains $N$ as a subgroup. If $\phi\in \operatorname{Irr}(G)$, then $\phi_N$ denotes the restriction of this character to $N$. If $x,y\in N$, then $[x,y]=x^{-1}y^{-1}xy$ is the commutator. If $A$ and $B$ are subgroups of a group, then $[A,B]=\langle [a,b]\,|\, a,\in A \text{ and } b\in
B\rangle$ is the commutator subgroup of $A$ and $B$. In particular, $N'=[N,N]$ denotes the derived subgroup of $N$. The reader can find background material on finite groups in [@robinson95] and on character theory in [@isaacs94].]{}
#### Clifford Codes.
Before introducing the concept of a Clifford code, we need to fix a notion of errors that generalizes the concept of the Pauli group. We say that a finite group $E$ is an abstract error group if it has a faithful irreducible unitary representation $\rho$ of degree $d = |E:Z(E)|^{1/2}$. The irreducibility of the representation ensures that one can express any error acting on ${\mathbf{C}}^d$ as a linear combination of the matrices $\rho(g)$, with $g\in E$. The fact that the representation is faithful and has the largest possible degree ensures that the set of matrices $\{ \rho(g)\,|\, g\in T\}$, where $T$ is a set of representatives of $E/Z(E)$, forms a *basis* of the vector space of $d\times d$ matrices.
A Clifford code is constructed with the help of a normal subgroup $N$ of the error group $E$ and an irreducible character $\chi$ of $N$. Let $\phi$ denote the irreducible character corresponding to the representation $\rho$ of the group $E$, that is, $\phi(g)=\operatorname{Tr}\rho(g)$ for $g\in E$. Suppose that $N$ is a normal subgroup of $E$ and that $\chi$ is an irreducible character of $N$ such that $(\chi,\phi_N)_N>
0$. Then the Clifford code $C$ corresponding to $(E,\rho,N,\chi)$ is defined as the image of the orthogonal projector $$P = \frac{\chi(1)}{|N|} \sum_{n\in N} \chi(n^{-1})\rho(n),$$ see [@klappenecker033 Theorem 1]. We emphasize that if we refer to a Clifford code with data $(E,\rho,N,\chi)$, then it is assumed that $(\chi,\phi_N)>0$, as this condition ensures that $\dim C>0$.
Recall that an error $e$ in $E$ is detectable by the quantum code $C$ if and only if $P\rho(e)P=\lambda_e P$ holds for some $\lambda_e\in {\mathbf{C}}$.
The image of $P$ is the homogeneous component that consists of the direct sum of all irreducible ${\mathbf{C}}N$-submodules with character $\chi$ that are contained in the restriction of $\rho$ to $N$. The elements $e$ in $E$ that satisfy $\rho(e)C= C$ form a group known as the inertia group $I_E(\chi)= \{ g\in E\,|\, \chi(gxg^{-1})=\chi(x)
\text{ for all } x\in N\}.$ We note that $C$ is an irreducible ${\mathbf{C}}[I_E(\chi)]$-module. Let $\vartheta$ be the irreducible character corresponding to this module.
\[fact:clifford\] Let $C$ be a Clifford code with data $(E,\rho,N,\chi)$. Then the dimension of the code is given by $\dim C = |Z(E)\cap N| |E:
Z(E)|^{1/2}\chi(1)^2/|N|$. An error $e$ in $E$ can be detected by $C$ if and only if $e$ is in $E-(I_E(\chi)-Z(\vartheta))$.
For a proof of this fact see [@klappenecker033] and for more background on Clifford codes see [@klappenecker031] and the seminal papers [@knill96a; @knill96b].
#### Operator Quantum Error-Correcting Codes.
We are now concerned with the construction of a decomposition of the Hilbert space $H$ in the form $$H = (A\otimes B) \oplus C^\perp.$$ Put differently, we seek a decomposition of the Clifford code $C$ as a tensor product $A\otimes
B$.
The next theorem gives a construction of operator quantum error-correcting codes when one can express the inertia group $I_E(\chi)$ as a central product $I_E(\chi)=LN$, where $L$ is a subgroup of $E$ such that $[L,N]=1$.
\[th:first\] Suppose that $C$ is a Clifford code with data $(E,\rho, N,\chi)$. If the inertia group $I_E(\chi)$ is of the form $I_E(\chi)=LN$, where $L$ is a subgroup of $E$ such that $[L,N]=1$, then $C$ is an operator quantum error-correcting code $C= A\otimes B$ such that
$\dim A = |Z(E)\cap N| |E: Z(E)|^{1/2}\chi(1)/|N|$,
$\dim B=\chi(1)$.
The subsystem $A$ is an irreducible ${\mathbf{C}}L$-module with character $\chi_A\in \operatorname{Irr}(L)$. An error $e$ in $E$ is detectable by subsystem $A$ if and only if $e$ is contained in the set $E-(I_E(\chi)-Z(\chi_A)N)$.
Since the Clifford code $C$ is an irreducible ${\mathbf{C}}[I_E(\chi)]$-module and $I_E(\chi)=LN$, with $[L,N]=1$, there exists an irreducible ${\mathbf{C}}L$-module $A$ and an irreducible ${\mathbf{C}}N$-module $B$ such that $C\cong
A\otimes B$, see [@GLS2 Proposition 9.14]. If $\chi_A\in \operatorname{Irr}(L)$ is the character associated with the module $A$, $\chi_B\in \operatorname{Irr}(N)$ the character associated with $B$, and $\vartheta\in \operatorname{Irr}(I_E(\chi))$ the character associated with $C$, then $\vartheta$ is of the form $\vartheta(\ell n)=\chi_A(\ell)\chi_B(n)$ with $\ell \in L$ and $n\in
N$.
As the restriction of $C$ to a ${\mathbf{C}}N$-module contains an irreducible ${\mathbf{C}}N$-module $W$ with character $\chi$, we must have $$\begin{array}{lcl}
\displaystyle
(\vartheta_N , \chi)_N = \frac{1}{|N|} \sum_{n\in N}
\vartheta(1,n^{-1}) \chi(n) &=&
\displaystyle \frac{1}{|N|} \sum_{n\in N}
\chi_A(1)\chi_B(n^{-1}) \chi(n) \\
&=& \chi_A(1) (\chi_B, \chi)_N > 0.
\end{array}$$ Since $\operatorname{Irr}(N)$ forms an orthonormal basis with respect to $(\,\cdot\,,\,\cdot\,)_N$, we can conclude that the irreducible character $\chi_B$ must be equal to $\chi$. It follows that $C\cong A\otimes W$.
The dimension of $W\cong B$ is $\chi(1)$, and by Fact \[fact:clifford\] the dimension of $C$ is given by $$\operatorname{Tr}P = |Z(E)\cap N| |E:Z(E)|^{1/2} \chi(1)^2/|N|.$$ The dimension of $B$ follows from the formula $\dim C=\dim A\dim B$.
An error $e\in E-I_E(\chi)$ maps $C$ to an orthogonal complement, so the errors are detectable. An error $e$ in $Z(\chi_A)N$ acts by scalar multiplication on $A$ and arbitrarily on $B$, so these errors are by definition detectable (harmless would be a better word). Therefore, all errors in $E-(I_E(\chi)-Z(\chi_A)N)$ are detectable. Conversely, an error $e$ in $I_E(\chi)-Z(\chi_A)N$ cannot be detectable, since $e$ does not act by scalar multiplication on $A$, and thus does not preserve the encoded quantum information.
The data given in the previous theorem can be easily computed, especially with the help of a computer algebra system such as GAP or MAGMA.
We will now consider some important special cases. Recall that most abstract error groups that are used in the literature satisfy the constraint $E'\subseteq Z(E)$ (put differently, the quotient group $E/Z(E)$ is abelian). In that case, we are able to obtain a characterization of the resulting operator quantum error-correcting codes that does not depend on the choice of the character $\chi$.
\[th:second\] Suppose that $E$ is an abstract error group such that $E'\subseteq Z(E)$. Suppose that $C$ is a Clifford code with data $(E,\rho, N,\chi)$. In this case, the inertia group is given by $I_E(\chi)=C_E(Z(N))$. If $C_E(Z(N))=LN$ for some subgroup $L$ of $E$ such that $[L,N]=1$, then $C$ is an operator quantum error-correcting code $C= A\otimes B$ such that
$\dim A = |Z(E)\cap N| |E: Z(E)|^{1/2}|N:Z(N)|^{1/2}/|N|$,
$\dim B=|N:Z(N)|^{1/2}$.
An error $e$ in $E$ is detectable by subsystem $A$ if and only if $e$ is contained in the set $E-(C_E(Z(N))-Z(L)N)$.
Since the abstract error group $E$ satisfies the condition $E'\subseteq Z(E)$, the inertia group of the character $\chi$ in $E$ can be fully determined; it is given by $T:=I_E(\chi)=C_E(Z(N))$, see [@klappenecker033 Lemma 5].
Suppose that $$P_1 = \frac{\chi(1)}{|N|}\sum_{n\in N} \chi(n^{-1})\rho(n)$$ is the orthogonal projector onto $C$. The assumption $E'\subseteq Z(E)$ implies that there exists a linear character $\varphi$ of $\operatorname{Irr}(Z(N))$ such that $$P_2 = \frac{1}{|Z(N)|}\sum_{n\in Z(N)} \varphi(n^{-1})\rho(n)$$ satisfies $P_1=P_2$, see [@klappenecker033 Theorem 6].
Let $\phi$ be the character of the representation $\rho$, that is, $\phi(g)=\operatorname{Tr}\rho(g)$ for $g\in E$. We have $ \operatorname{Tr}P_1 = \chi(1)^2
\phi(1)|N\cap Z(E)|/|N|$ and $\operatorname{Tr}P_2 = \phi(1)|N\cap
Z(E)|/|Z(N)|$. Since $P_1=P_2$ project onto the code space $C$, and $\dim C>0$, we have $\operatorname{Tr}P_1/\operatorname{Tr}P_2=1$, which implies $\chi(1)^2=|N\colon Z(N)|$. Therefore, the claims i) and ii) follow from Theorem \[th:first\].
Let $\vartheta\in \operatorname{Irr}(T)$ be the character associated with the ${\mathbf{C}}[T]$-module $C$; put differently, $\vartheta$ is the unique character in $\operatorname{Irr}(T)$ that satisfies $(\vartheta_N,\chi)_N>0$ and $(\phi_T,\vartheta)_T>0$. Since $Z(E)\le T$ and $(\phi_T,\vartheta)_T>0$, it follows from Lemma \[l:support2\] that $\operatorname{supp}(\vartheta)=Z(T)$.
Since the inertia group $T$ is a central product given by $T=LN$ with $[L,N]=1$, there exist characters $\chi_A\in \operatorname{Irr}(L)$ and $\chi_B=\chi\in \operatorname{Irr}(N)$ such that $\vartheta(\ell
n)=\chi_A(\ell)\chi(n)$ for $\ell\in L$ and $n\in N$. By Lemma \[l:center\], we have $Z(T)=Z(L)Z(N)$; thus, $\operatorname{supp}(\vartheta)=Z(L)Z(N)$. This implies that $\operatorname{supp}(\chi_A)=L\cap
Z(L)Z(N)=Z(L)$; hence $Z(\chi_A)=Z(L)$. The characterization of the detectable errors is obtained by substituting these facts in Theorem \[th:first\].
In the previous theorem, we still need to check whether $C_E(Z(N))$ decomposes into a central product of $N$ and some group $L$. In the case of extraspecial $p$-groups (which is arguably the most popular choice of abstract error groups) the decomposition of the inertia group into a central product is always guaranteed, as we will show next.
Recall that a finite group $E$ whose order is a power of a prime $p$ is called extraspecial if its derived subgroup $E'$ and its center $Z(E)$ coincide and have order $p$. An extraspecial $p$-group is an abstract error group. The quotient group $\overline{E}=E/Z(E)$ is the direct product of two isomorphic elementary abelian $p$-groups. Therefore, one can regard $\overline{E}$ as a vector space ${\mathbf{F}}_p^{2n}$ over the finite field ${\mathbf{F}}_p$.
Let $\zeta$ be a fixed generator of the cyclic group $Z(E)$. As the commutator $[x,y]$ depends only on the cosets $\overline{x}=xZ(E)$ and $\overline{y}=yZ(E)$, one can determine a well-defined function $s\colon
\overline{E}\times \overline{E}\rightarrow {\mathbf{F}}_p$ by $[x,y]=\zeta^{s(\overline{x},\overline{y})}$. The function $s$ is a nondegenerate symplectic form. We note that two elements $x$ and $y$ in $E$ commute if and only if $s(\overline{x},\overline{y})=0$. We write $\overline{x}\, {{\perp_s}}\, \overline{y}$ if and only if $s(\overline{x},\overline{y})=0$.
For a subgroup $G$ of $E$, we will use $\overline{G}$ to denote $G/Z(E)$.
\[l:inertia\] If $E$ is an extraspecial $p$-group and $N$ a normal subgroup of $E$, then $C_E(Z(N))=NC_E(N)$.
Since $Z(E)\le NC_E(N)\le C_E(Z(N))$, it suffices to show that the dimensions of the ${\mathbf{F}}_p$-linear vector spaces $$\overline{NC_E(N)} \quad \text{and}\quad \overline{C_E(Z(N))}$$ are the same. Suppose that $z=\dim\overline{Z(N)}$ and $k=\dim\overline{N}$. Then $$\begin{array}{lcl}
\dim {\overline}{NC_E(N)}&=& \dim ({\overline}{N}+{\overline}{N}^{{\perp_s}})
= \dim{\overline}{N}+\dim{\overline}{N}^{{\perp_s}}-\dim({\overline}{N}\cap {\overline}{N}^{{\perp_s}})\\
&=& \dim{\overline}{N}+\dim{\overline}{N}^{{\perp_s}}-\dim({\overline}{Z(N)})\\
&=& k+ (2n-k)-z=2n-z,
\end{array}$$ which coincides with $\dim
{\overline}{C_E(Z(N))}=\dim{\overline}{Z(N)}^{{\perp_s}}=2n-z$, and this proves our claim.
The next theorem shows that it suffices to choose a normal subgroup $N$ of the extraspecial $p$-group $E$, and this choice determines the parameters of an operator quantum error-correcting code provided by a Clifford code $C$.
\[th:third\] Suppose that $E$ is an extraspecial $p$-group. If $C$ is a Clifford code with data $(E,\rho, N,\chi)$, with $N\neq 1$, then $C$ is an operator quantum error-correcting code $C= A\otimes B$ such that
$\dim A = |Z(E)\cap N| |E: Z(E)|^{1/2}|N:Z(N)|^{1/2}/|N|$,
$\dim B=|N:Z(N)|^{1/2}$.
An error $e$ in $E$ is detectable by subsystem $A$ if and only if $e$ is contained in the set $E-(NC_E(N)-N)$.
The inertia group $I_\chi(E)=C_E(Z(N))$, since $E'\subseteq Z(E)$, see [@klappenecker033 Lemma 5]. By Lemma \[l:inertia\], we have $I_E(\chi)=LN=NL$ with $L=C_E(N)$. Thus, $C$ is an operator quantum error-correcting code and the statements i) and ii) follow from Theorem \[th:second\]. Furthermore, Theorem \[th:second\] shows that an error $e$ in $E$ is detectable if and only if $e\in
E-(NC_E(N)-Z(L)N)$. Since $E$ is a $p$-group and $N\neq 1$, we have $N\cap Z(E)\neq 1$; hence $Z(E)\le N$. We note that ${\overline}{Z(L)}\subseteq {\overline}{L}\cap {\overline}{L}^{{\perp_s}}=
{\overline}{N}^{{\perp_s}}\cap{\overline}{N}\subseteq {\overline}{N}$; therefore, $N\subseteq Z(L)N\subseteq Z(N)N=N$, forcing $Z(L)N=N$.
#### Classical Codes.
We conclude this note by showing how the previous results can be related to classical coding theory.
Let $a$ and $b$ be elements of the finite field ${\mathbf{F}}_q$ of characteristic $p$. We define unitary operators $X(a)$ and $Z(b)$ on ${\mathbf{C}}^q$ by $$X(a){|x\rangle}={|x+a\rangle},\qquad Z(b){|x\rangle}=\omega^{\operatorname{tr}(bx)}{|x\rangle},$$ where $\operatorname{tr}$ denotes the trace operation from the extension field ${\mathbf{F}}_q$ to the prime field ${\mathbf{F}}_p$, and $\omega=\exp(2\pi i/p)$ is a primitive $p$th root of unity. Let $\mathbf{a}=(a_1,\dots, a_n)\in
{\mathbf{F}}_q^n$. We write $ X(\mathbf{a}) = X(a_1)\otimes\, \cdots \,\otimes
X(a_n)$ and $Z(\mathbf{a}) = Z(a_1)\otimes\, \cdots \,\otimes Z(a_n)$ for the tensor products of $n$ error operators. One readily checks that the group $$E = \langle X(a), Z(b)\,|\, a, b\in {\mathbf{F}}_q^n\rangle$$ is an extraspecial $p$-group of order $pq^{2n}$. As a representation $\rho$, we can take the identity map on $E$. We have $E/Z(E)\cong {\mathbf{F}}_q^{2n}$.
We need to introduce a notion of weights of errors. Recall that an error in $E$ can be expressed in the form $\alpha X(a)Z(b)$ for some nonzero scalar $\alpha$. The weight of $\alpha X(a)Z(b)$ is defined as $|\{ i\,|\, 1\le i\le n, a_i\neq 0 \text{ or } b_i\neq 0\}|$, that is, as the number of quantum systems that are affected by the error. Similarly, we can introduce a weight on vectors of ${\mathbf{F}}_q^{2n}$ by $$\operatorname{swt}(a|b)=\{ i\,|\, 1\le i\le n, a_i\neq 0 \text{ or } b_i\neq 0\}|$$ for $a,b\in {\mathbf{F}}_q^n$.
Theorem \[th:third\] suggests the following approach to construct operator quantum error-correcting codes.
\[th:four\] Let $X$ be a classical additive subcode of ${\mathbf{F}}_q^{2n}$ such that $X\neq \{0\}$ and let $Y$ denote its subcode $Y=X\cap X^{{\perp_s}}$. If $x=|X|$ and $y=|Y|$, then there exists an operator quantum error-correcting code $C=
A\otimes B$ such that
$\dim A = q^n/(xy)^{1/2}$,
$\dim B = (x/y)^{1/2}$.
The minimum distance of subsystem $A$ is given by $d=\operatorname{swt}((X+X^{{\perp_s}})-X)=\operatorname{swt}(Y^{{\perp_s}}-X)$. Thus, the subsystem $A$ can detect all errors in $E$ of weight less than $d$, and can correct all errors in $E$ of weight $\le \lfloor (d-1)/2\rfloor$.
Let $E$ be the extraspecial $p$-group of order $pq^{2n}$, and let $N$ be the full preimage of ${\overline}{N}=X$ in $E$ under the canonical quotient map. Therefore, we can apply Theorem \[th:third\]. The remainder of the proof justifies how the parameters given in Theorem \[th:third\] can be expressed in terms of the code sizes $x$ and $y$.
Then ${\overline}{Z(N)}=X\cap X^{{\perp_s}}=Y$. By definition, $N$ contains $Z(E)$; hence, $Z(E)\le Z(N)$. It follows that $|N:Z(N)|=|\overline{N}:{\overline}{Z(N)}|=x/y$, so ii) follows from Theorem \[th:third\]. For the claim i), we remark that $x=|X|=|N|/p$, which implies that $\dim A =(p/|N|)
|E:Z(E)|^{1/2}|N:Z(N)|^{1/2}=q^n(x/y)^{1/2}/x$.
The minimum distance of subsystem $A$ is the weight of the smallest nondetectable error, so it is the minimum weight of an error in the set $NC_E(N)-N=C_E(Z(N))-N$. Since the quotient map $E\rightarrow
{\overline}{E}$ maps an error $e$ of weight $w$ onto a vector ${\overline}{e}$ such that $w=\operatorname{swt}{{\overline}{e}}$, the claim about the minimum distance follows from the observations that ${\overline}{NC_E(N)-N}=(X+X^{{\perp_s}})-X$ and ${\overline}{C_E(Z(N))-N}=Y^{{\perp_s}}-X$.
An operator quantum error-correcting code with parameters $((n,a,b,d))_q$ is a subspace $C=A\otimes B$ of a $q^n$-dimensional Hilbert space $H$ such that $a=\dim A$, $b=\dim B$, and the subsystem $A$ has minimum distance $d$. The above theorem constructs an $((n,q^n/(xy)^{1/2},(x/y)^{1/2},d))_q$ operator quantum error-correcting code given a classical $(n,x)_q$ code $X$ and its $(n,y)_q$ subcode $Y=X\cap X^{{{\perp_s}}}$. We write $[[n,k,r,d]]_q$ for an $((n,q^k,q^r,d))_q$ operator quantum error-correcting code.
Sometimes one would like to characterize the minimum distance in terms of the familiar Hamming weight. For this purpose, we reformulate the above result in terms of codes of length $n$ over ${\mathbf{F}}_{q^2}$.
Let $(\beta,\beta^q)$ be a fixed normal basis of ${\mathbf{F}}_{q^2}$ over ${\mathbf{F}}_q$. We can define a bijection $\phi$ from ${\mathbf{F}}_q^{2n}$ onto ${\mathbf{F}}_{q^2}^n$ by setting $$\phi((a|b))=\beta a+\beta^q b\quad \text{for}\quad (a|b)\in
{\mathbf{F}}_q^{2n}.$$ The map is chosen such that a vector $(a|b)$ of symplectic weight $x$ is mapped to a vector $\phi((a|b))$ of Hamming weight $x$. If we define a trace-alternating form $\langle v|w\rangle_a$ for vectors $v$ and $w$ in ${\mathbf{F}}_{q^2}^n$ by $$\langle v|w\rangle_a=\operatorname{tr}_{q/p}\left(\frac{v\cdot w^q-v^q\cdot
w}{\beta^{2q}-\beta^q}\right),$$ then it is easy to show that $
\langle c|d\rangle_s = \langle \phi(c)|\phi(d)\rangle_a$ holds for all $c,d\in {\mathbf{F}}_q^{2n}$, see [@pre3 Lemma 14]. Specifically, we have $c\, {{\perp_s}}\, d$ if and only if $\phi(c)\, {{\perp_a}}\, \phi(d)$. Therefore, the previous theorem can be reformulated terms of codes of length $n$ over ${\mathbf{F}}_{q^2}$ as follows:
\[th:five\] Let $X$ be a classical additive subcode of ${\mathbf{F}}_{q^2}^{n}$ such that $X\neq \{0\}$ and let $Y$ denote its subcode $Y=X\cap X^{{\perp_a}}$. If $x=|X|$ and $y=|Y|$, then there exists an operator quantum error-correcting code $C=
A\otimes B$ such that
$\dim A = q^n/(xy)^{1/2}$,
$\dim B = (x/y)^{1/2}$.
The minimum distance of subsystem $A$ is given by $$d=\operatorname{wt}((X+X^{{\perp_a}})-X)=\operatorname{wt}(Y^{{\perp_a}}-X),$$ where $\operatorname{wt}$ denotes the Hamming weight. Thus, the subsystem $A$ can detect all errors in $E$ of Hamming weight less than $d$, and can correct all errors in $E$ of Hamming weight $\lfloor (d-1)/2\rfloor$ or less.
This follows from Theorem \[th:four\] and the definition of the isometry $\phi$.
The above connections of Clifford operator quantum error-correcting codes to classical code allow one to explore a plethora of code constructions.
#### Conclusions.
We have introduced a method for constructing operator quantum error-correcting codes. We have seen that a Clifford codes $C$ offers naturally a tensor-product decomposition $C= A\otimes B$, where the dimensions of the subsystems are controlled by the choice of the normal subgroup $N$ and its character $\chi$.
Our construction in terms of classical codes is fairly simple: Any classical (additive) code over a finite field can be used to construct an operator quantum error-correcting code. In particular, we do not require any self-orthogonality conditions as in the case of stabilizer code constructions.
The most prominent open problem concerning operator quantum error-correcting codes is whether one can achieve better error correction that by means of a quantum error-correcting code. The construction given in Theorem \[th:four\] allows one to compare the parameters of Clifford codes with the parameters of stabilizer codes. One should note that a fair comparison should be made between $[[n-r,k,d]]$ stabilizer codes and $[[n,k,r,d]]$ Clifford codes. It would be helpful to have bounds on the best possible minimum distance $d$ of Clifford codes to answer this question.
#### Acknowledgments.
This research was supported by NSF grant CCF-0218582, NSF CAREER award CCF-0347310, and a TITF project.
Appendix
========
In this appendix, we prove some simple technical results on groups and characters.
\[l:support\] Let $E$ be a finite group such that $E'\subseteq Z(E)$, and let $H$ be a subgroup of $E$. If $\chi\in \operatorname{Irr}(H)$ satisfies $Z(E)\cap \ker \chi
= \{1\}$, then $\operatorname{supp}\chi = Z(H)$.
Let $h\in \operatorname{supp}(\chi)$. Seeking a contradiction, we assume that $h\in
H-Z(H)$. Since $E'\subseteq Z(E)$, there exists an element $g\in H$ such that $ghg^{-1}=zh$ with $z\in Z(E)$ such that $z\neq 1$. Since $zh\in H$ and $h\in H$, we have $z\in H\cap Z(E)$. As $\chi$ is irreducible, the element $z\in H\cap Z(E)$ is represented by $\omega
I$ for some $\omega\in {\mathbf{C}}$ by Schur’s lemma; furthermore, $\omega \neq
1$, since $Z(E)\cap \ker \chi=\{1\}$. We note that $ \chi(h)=\chi(ghg^{-1}) = \chi(zh)= \omega\chi(h)$, with $\omega\neq 1$, forcing $\chi(h)=0$, contradiction.
The elements of $Z(H)$ belong to the support of $\chi$, since they are represented by scalar invertible matrices.
\[l:support2\] Let $E$ be a finite group such that $E'\subseteq Z(E)$, and let $\phi\in \operatorname{Irr}(E)$ be a faithful character of degree $\phi(1)=|E:Z(E)|^{1/2}$. Let $T$ be a subgroup of $E$ such that $Z(E)\le T$. If $\vartheta\in \operatorname{Irr}(T)$ and $(\phi_T,\vartheta)_T>0$, then $\operatorname{supp}(\vartheta)=Z(T)$.
Let $Z=Z(E)$. We have $\operatorname{supp}(\phi)=Z$ by [@isaacs94 Lemma 2.29]. Since the support of $\phi$ equals $Z$, it follows from the definitions that $$0<(\phi_T,\vartheta)_T = \frac{1}{|T:Z|}(\phi_Z,\vartheta_Z)_Z.$$ Clearly, $\phi_Z=\phi(1)\varphi$ and $\vartheta_Z=\vartheta(1)\theta$ for some linear characters $\varphi$ and $\theta$ of $Z$. As $(\phi_Z,\vartheta_Z)_Z=\phi(1)\vartheta(1)(\varphi,\theta)_Z>0$, we must have $\theta=\varphi$. Since $\phi$ is faithful, it follows that $\varphi=\theta$ is faithful; hence, $\ker \vartheta \cap Z(E)=\{1\}$. Thus, $\operatorname{supp}\vartheta=Z(T)$ by Lemma \[l:support\].
\[l:center\] Suppose that $T$ is a group with subgroups $L$ and $N$ such that $T=LN$ and $[L,N]=1$. Then $Z(T)=Z(L)Z(N)$.
Since $T=LN$, an arbitrary element $z$ of $Z(T)$ can be expressed in the form $z=ln$ for some $l\in L$ and $n\in N$. For $n'$ in $N$, we have $lnn'=n'ln=ln'n$, where the latter equality follows from $[L,N]=1$. Consequently, $nn'=n'n$ for all $n'$ in $N$, so $n$ is an element of $Z(N)$. Similarly, $l$ must be an element of $Z(L)$. It follows that $Z(T)=Z(L)Z(N)$.
[10]{} D. Bacon. Operator quantum error correcting subsystems for self-correcting quantum memories. , 73(012340), 2006.
D. Gorenstein, R. Lyons, and R. Solomon. , volume 40 of [*Mathematical Surveys and Monographs*]{}. AMS, 1994.
I.M. Isaacs. . Dover, 1994.
A. Ketkar, A. Klappenecker, S. Kumar, and P.K. Sarvepalli. Nonbinary stabilizer codes over finite fields. To appear in IEEE Trans. Inform. Theory, November, 2006.
A. Klappenecker and M. R[ötteler]{}. Beyond stabilizer codes [II]{}: [C]{}lifford codes. , 48(8):2396–2399, 2002.
A. Klappenecker and M. R[ö]{}tteler. Clifford codes. In R. Brylinski and G. Chen, editors, [*Mathematics of Quantum Computing*]{}, pages 253–273. Chapman & Hall/CRC Press, 2002.
E. Knill. Group representations, error bases and quantum codes. Los Alamos National Laboratory Report LAUR-96-2807, 1996.
E. Knill. Non-binary unitary error bases and quantum codes. Los Alamos National Laboratory Report LAUR-96-2717, 1996.
E. Knill. On protected realizations of quantum information. Eprint: quant-ph/0603252, 2006.
D. W. Kribs. A brief introduction to operator quantum error correction. Eprint: math/0506491, 2005.
D. W. Kribs, R. Laflamme, and D. Poulin. Unified and generalized approach to quantum error correction. , 94(180501), 2005.
D. W. Kribs, R. Laflamme, D. Poulin, and M. Lesosky. Operator quantum error correction. Eprint: quant-ph/0504189, 2005.
D. Poulin. Stabilizer formalism for operator quantum error correction. , 95(230504), 2005.
D.J.S. Robinson. . Springer, 2nd edition, 1995.
| {
"pile_set_name": "ArXiv"
} |
HU-EP-05/35\
hep-th/0508071
[**D-branes in overcritical electric fields** ]{}
Harald Dorn, Mario Salizzoni and Alessandro Torrielli
Humboldt–Universität zu Berlin, Institut für Physik\
Newtonstraße 15, D-12489 Berlin\
`dorn,sali,[email protected]`
[**Abstract**]{}
We collect some arguments for treating a D-brane with overcritical electric field as a well-posed initial condition for a D-brane decay. Within the field theoretical toy model of Minahan and Zwiebach we give an estimate for the condensates of the related infinite tower of tachyonic excitations.
Introduction
============
Tachyon dynamics in the context of D-brane decay was the subject of intense studies in recent years [@s]. The main focus has been on the field theoretical or string field theoretical aspects of tachyons as excitations indicating some instability leading to the decay of certain field configurations. This field theoretical point of view is by far more adequate than discussions in terms of a particle language including its conceptual problems with superluminal velocities.
Open strings in the presence of an overcritical electric field mainly have been considered as an ill-defined setting (there are however interesting connections with S-branes, see for instance [@sbranes1; @sbranes2]). Classically the endpoints of open strings have to move with superluminal velocity and at the quantum level there appears a singularity at the critical field strength in the string analog of the Schwinger pair creation [@b; @bp]. On the other hand, looking at String Field Theory in the presence of an overcritical electric field presents a picture conceptually not so different from that with undercritical or zero electric field. Instead of one tachyonic excitation, accompanied by a zero mass and an infinite tower of stable excitations, one is faced with a stable excitation corresponding to the former tachyon and the fact that the infinite tower has become tachyonic. At this level the difference between undercritical and overcritical electric field seems to be only technical.
In the present paper we want to sketch this point of view in some more detail. Section 2 is devoted to a discussion of several aspects supporting the conjecture that the decay of D-branes due to the presence of an overcritical electric field can be a well posed problem in string theory. Of course any attempt to implement this picture has to handle the serious technical problems caused by the infinite number of tachyonic excitations driving the D-brane decay. The level truncation, which turned out to be very effective in the undercritical case, is no longer applicable. Some comments on the treatment of the overcritical case within Boundary Conformal Theory are also added. In Section 3 we then restrict ourselves to a discussion within a toy model introduced by Minahan and Zwiebach in [@z; @mz], accordingly modified by the presence of the overcritical electric field. As a quantitative result we will get an estimate of the values of all the fields in the infinite tachyonic tower after condensation.
General considerations
======================
Spectrum
--------
Propagation of open bosonic strings in the background of a constant electromagnetic field has been analyzed in a series of papers [@b; @bp; @ft; @n]. The spectrum of the masses of a neutral string in the presence of a purely electric background has been derived as a particular case in [@b; @n], where a classical instability has been shown to arise above a critical value of the electric field. In this case, string modes develop negative squared masses, and a tachyonic contribution to the mass comes also from the motion of the string in transverse directions. This classical instability, which is indeed present both for neutral and for charged strings, has no analogue in particle mechanics[^1]. Its appearance is advocated in [@b] as a signal of the fact that, in a second quantized treatment of the theory, the string field would evolve away from the chosen unstable configuration. The spectrum can be reproduced by the formula $$\label{spectrum}
k_{\mu} G^{\mu \nu} k_{\nu} = - (n - 1)$$ where we set $\alpha' = 1$. The metric to be used is the so called “open string metric” $$\label{openmetric}
G^{\mu\nu} = \left( \frac{1}{g + B} g \frac{1}{g - B} \right)^{\mu\nu},$$ where $g, B$ are the sigma-model (closed string) constant backgrounds [@sw]. This, in turn, is the natural formula arising from the analysis of the conformal dimensions of the vertex operators of the worldsheet conformal field theory [@sw], and consistently from evaluation of the singularities of string scattering amplitudes (see [@btv] and references therein). For our purposes we will take the closed string metric to be $g = \eta = (- 1, 1, ..., 1)$.
In the case of a purely electric background, without loss of generality one can choose the only non-zero components of the antisymmetric tensor to be $$\label{Bfield}
B_{0 1} = - B_{1 0} = E .$$ The open string metric becomes $G_{\mu \nu} = (- (1 - E^2),
(1 - E^2), 1, ..., 1)$. One can notice the appearance of a critical value for the electric field: when $E = 1$, the open string metric (\[openmetric\]) changes its signature in the $(0,1)$ block. Formula (\[spectrum\]) can be written as $$\label{spectrumel}
- k_0^2 + k_1^2 + (1 - E^2) k_{\perp}^2 = - (1 - E^2) (n - 1),$$ or, equivalently, $$\label{spectrumel2}
k_{\mu} \eta^{\mu \nu} k_{\nu} = - (1 - E^2) (n - 1) + E^2 k_{\perp}^2,$$ $k_{\perp}$ indicating the momentum component transverse to the electric field. If one sets $k_{\mu} \eta^{\mu \nu} k_{\nu} = - M^2$, this formula coincides with what is obtained for example in Sec. 4 of [@n]. Generically, the distance between equidistant levels of the operator $M^2$ is smaller by a factor $(1 - E^2)$ [@n]. In the overcritical case, then one has $M^2<0$ (up to the first level due to presence of the zero-point energy of the oscillators), and a tachyonic contribution coming from the motion of the string in the transverse direction [^2].
What we can read out of (\[spectrumel\]), is that what was a time-like momentum (positive squared mass) in the absence of the background, becomes in the presence of an overcritical electric field a space-like one (tachyon). The spectrum of the bosonic open string has therefore been reversed, and it contains an infinite tower of tachyonic modes[^3].
It is also important to notice that a similar situation, in the presence of an electric background, would occur in superstring theory [@b; @btv], whose spectrum does not originally contain a tachyon. The appearance of an infinite tower of tachyons has no analog in the absence (or in the presence of an undercritical) electric field. Therefore, in the following we will regard the ensuing classical evolution away from the chosen configuration (which we will suggest to interpret as a decay of the electrified D25-brane) as *due to the overcritical electric field [@b].*
D-brane Decay
-------------
The above mentioned instability is accompanied by another effect. When interpreting bosonic open string theory as a description of the dynamics of a space-filling D25-brane, the tension of such a brane is naturally derived from the Dirac-Born-Infeld action [@l]. Such a tension, in the presence of a constant electric field background, is therefore proportional to the Born-Infeld factor $\sqrt{1 - E^2}$. This factor becomes imaginary in the overcritical case, which is to say that the squared mass of the brane becomes negative.
We can interpret this occurrence in the open string picture. Classically, an electric field stretches open strings against their internal tension [@sst]. At the critical value the electric force counterbalances the tendency of the string to oscillate, and stretches it to infinity. Beyond the critical value the classical tachyonic instability is generated [^4]. This is a signal that the nonperturbative string field theory vacuum, whose dynamics is determined by the open string theory defined on such a background, acquires a tachyonic instability. The fact that the D25-brane becomes tachyonic has another manifestation in the superluminal velocity of the T-dual D24-brane [@bach]. Equivalently, this is known to be a (tilted) S-brane, a spacelike object which arises in the description of the standard D-brane decay *via the rolling tachyon picture [@sbranes1; @sbranes2]. In particular, in [@sbranes1] an effective action has been derived for S-branes, which turns out to be equal to the Dirac-Born-Infeld action times a factor $i$.*
We suggest to interpret the features of the first quantized spectrum discussed in the previous section in the light of this physical picture. While it is by now clear how one should interpret the presence of the usual bosonic open string tachyon, which is responsible for driving the D-brane decay towards the closed string vacuum [@s], here the situation is complicated by the fact that the brane itself behaves like a tachyonic soliton background of String Field Theory. This is precisely because, while the nature of the bosonic string tachyon is related to string zero point quantum oscillations without classical analogue, the tachyonic tower described in the previous section reflects the classical instability of the tachyonic brane. What we propose is therefore that, still, the idea of describing the decay through the rolling of the tachyonic excitations living on the worldvolume is applicable, with the natural difference being represented by the presence of an infinite number of tachyonic states. The practical way to do this will be described in the next section.
String Field Theory and BCFT
----------------------------
The main tool to study the rolling of the tachyon has been Cubic String Field Theory [@w], in connection with methods of Boundary Conformal Field Theory [@s]. On one side, the presence of a minimum of the String Field Theory potential corresponding to the absence of perturbative open string excitations has been firmly established in the level truncation scheme. On the other side, an exactly marginal boundary perturbation to the world-sheet action for the tachyon profile of the form $$\label{bcft}
\lambda \int_{\partial \Sigma} \cosh [X_0 (t)] dt$$ has been used [@s] to describe the time evolution of the system and to derive the related stress-energy tensor during the decay. In the presence of an undercritical electric field, the procedure has been generalized in [@mukho]. The final products of the decay include in this case the additional presence of fundamental string charges in the tachyon vacuum.
The idea is therefore to make use of the available formulation of String Field Theory in the presence of antisymmetric backgrounds, this time evaluated for an overcritical electric field. We still expect it to be the instrument which describes the nonperturbative decay of the original unstable configuration, through the evolution of the infinite tower of tachyons. However, if one tries to apply the above mentioned strategy to this new situation, one soon realizes the scarce suitability of the level truncation scheme[^5]. Its good approximation relies on the fact that all the tower of positive mass excitations is stable around the origin (perturbative vacuum), and even if the higher spin fields assume non-zero vacuum expectation values at the closed string minimum [@ks], it is conceivable that these values will not differ too much from the original stable point at zero (this appears also as a feature of the toy model used in [@z; @mz] on which we will elaborate in the next section). Now instead, all of them are tachyonic at the origin, and one will expect all of them to substantially move away. Neglecting them all but a finite set, or in other words setting them to zero starting from a certain level on, does not look *a priori like a good approximation.*
On the other hand, since all the tower of states is rolling down, one would need like an infinite sum of boundary perturbations of the type (\[bcft\]) for the higher spin fields, with the suitable modifications in order to account for the electric field.
The boundary perturbation (\[bcft\]) was originally introduced by inspection of the linearized String Field Theory equations of motion. If one would like to proceed in analogous way, one has to look at solutions of the linearized equations of motion for the higher spin fields of the tower, in the background of an overcritical electric field. The main problem is the inclusion of all the tower, and we will sketch here only the proposal for the first state. For this purpose, we can follow the treatment reported in [@lv] to determine the condition for conformal invariance of the boundary coupling to the spin two field at the linearized level in the background field. Adopting their parameterization and choosing the gauge where the Stueckelberg field is set to zero, one obtains the term $$\label{labasbcft}
\int_{\partial \Sigma} A_{\mu \nu} (X) \, \partial^a X^{\mu} \,
\partial_a X^{\nu} \, dt$$ with the following conditions on the symmetric tensor $A_{\mu \nu} (X)$: $$\label{transverse}
(\partial^{\sigma} \partial_{\sigma} - 1) A_{\mu \nu} = 0 \, \, \, ; \, \, \,
\partial^{\mu} A_{\mu \nu}
= 0 \, \, \, ; \, \, A_{\mu}^{\mu} = 0,$$ where all indices are contracted with the open string metric. Taking this into account, and making a spatially independent ansatz, one finds that $$\label{solution}
A_{\mu \nu} (X^0) = a_{\mu \nu} \cosh (X^0 \sqrt{E^2 - 1})$$ is a solution, provided $a_{\mu \nu}$ is symmetric, purely spatial and traceless. Inserted in (\[labasbcft\]), the solution (\[solution\]) provides a generalization of the boundary deformation (\[bcft\]) suitable for the new case. The polarization tensor $a_{\mu \nu}$ would have the same *status as the parameter $\lambda$, the initial value of the tachyon.*
This will have to be supplemented with all the remaining boundary couplings for the higher spin states. However, it can already be studied as a prototype in order to gain insight about the features of the boundary state resulting from such kind of perturbations, which will in turn determine the type of vacuum obtained at the end of the decay.
A Toy Model
===========
In [@z; @mz] toy models for the standard bosonic string tachyon condensation were studied, whose features simulate the behaviour found in String Field Theory for the true rolling tachyon. The idea is to study the lump solution of an effective field theory for the tachyon field. Fluctuations around this nonperturbative solution are determined by solving a Schrödinger equation of a known type, whose spectrum can be exactly computed resulting in an infinite tower of scalar excitations of increasing mass.
This situation has a close resemblance with the case of the String Field Theory obtained by quantization around the vacuum provided by a D-brane. In particular, the presence of a tachyon signals the instability of such a configuration, and the decay towards a stable minimum can be studied by looking for the minima of the obtained multiscalar potential. The main advantage is that in this simplified case one already knows the minimum of the original potential, and one can fix the values of the infinite set of scalar fields in the static case by simply requiring that summing the fluctuations to the lump profile produces the minimum of the original tachyon field. One can therefore treat simultaneously the [*whole*]{} tower of states, which in the case of an overcritical electric field is a strict requirement.
The technical procedure, in the case of absence of electric field, is briefly summarized in what follows. One considers the action for the tachyon field $\phi$ of a D25-brane $$\label{cubic}
S = {{1}\over{g_o^2}} \int d^{25}y \, dx \bigg[-{{1}\over{2}} \partial_{\mu}
\phi \partial^{\mu}
\phi - {{1}\over{2}} {(\partial_x \phi )}^2 - V(\phi)\bigg],$$ where $y^\mu=(t,\vec{y})$ and $g_o$ is the open string coupling constant. The potential has the form shown in Fig.1:
{width="4.8cm"} {width="4.8cm"}
$$\label{unosue}
V(\phi) = - {{1}\over{4}} \phi^2 {\mathop{\rm ln}} \phi^2.$$
It has a maximum at $e^{- 1/2}$, corresponding to the D25-brane, and a local minimum at $0$, corresponding to the closed string vacuum, where the curvature diverges. The last fact beautifully mimics the idea of decoupling of the open string degrees of freedom in the stable vacuum, which is a feature of the standard tachyon condensation. The potential admits a Gaussian lump solution independent of the $y$ variables $$\label{lump}
\bar{\phi} (x) = \exp\Big[- {{x^2}\over{4}}\Big],$$ shown in Fig.1, that represents a D24-brane. After $\phi \rightarrow \bar{\phi} + \phi$, making the ansatz $$\label{expa}
\phi = \sum_n \bar{\phi}_n (y) \psi_n (x),$$ the computation of the spectrum of fluctuations around the lump is effectively reduced to the one dimensional Schrödinger equation for the harmonic oscillator
$$\label{schroe}
- {{d^2 \psi}\over{dx^2}} +
\bigg[ -\frac32 +\frac14 x^2 \bigg] \psi (x) = m^2 \psi (x),$$
that determines in this way the “open string” spectrum $m^2 = n-1$ for $n\geq0$.
The advantage of the model is that we know the minimum to be at the value where the original tachyon field $\phi$ is equal to $0$, therefore the exact values of the condensates for the whole tower are obtained by solving the following equation:
$$\label{determ1}
0 = \bar\phi (x) + \sum_n \bar\phi_n \psi_n (x) .$$
Since the Gaussian itself is the first eigenfunction of the harmonic oscillator, the solution of the equation is immediately found. It simply amounts to fixing the expectation value $-1$ for the tachyon ($n=0$), and zero for all the other excitations. This means that only the tachyon condenses.
We want to use this model to describe the decay of a D-brane due to the overcritical electric field. The decay of the D-brane is now driven by the infinite tower of former massive fields now being tachyonic. We will see that it is still possible to study the decay of a D24-brane using the potential for the tachyon of a D25-brane.
Modifying the model in the presence of an electric field, simply amounts to use for the $y$ coordinates the open string metric, and simultaneously to introduce for them the star-product[^6].
Then, instead of (\[cubic\]), the action becomes
$$\label{overcubic}
S = {{1}\over{g_o^2}} \sqrt{1-E^2} \int d^{25}y \,
dx \bigg[-{{1}\over{2}} G^{\mu \nu} \partial_{\mu}
\phi \partial_{\nu}
\phi - {{1}\over{2}} {(\partial_x \phi )}^2 - V(\phi)\bigg],$$
where we included the factor $\sqrt{1-E^2}$ in order to recover the correct D-brane tension. Technically it would be equivalent to take into account the factor $\sqrt{-\mbox{det}G}=\vert 1-E^2\vert$ in the measure and to replace $g_o^2$ by $g_o^2\sqrt{-\mbox{det}(g+B)}=g_o^2\sqrt{1-E^2}$, as is done for the DBI action in ref.[@sw]. Note that we insist on keeping the string coupling real for use in full String Field Theory. The action (\[overcubic\]) is thought to be an effective one for the lowest string excitation. Having this in mind, the presence of an imaginary factor in (\[overcubic\]) is no obstacle and in agreement with the DBI analysis.
The “transverse” part in the $x$ variable is not touched by the electric field. In particular, the same lump (\[lump\]) is still a solution of the new equations of motion. With $G_{\mu\nu}$ from Sec.2, the action (\[overcubic\]) can be written as $$\begin{aligned}
\label{explovercubic}
&&S = {{1}\over{g_o^2}} {{1}\over{\sqrt{1-E^2}}}
\int d^{25}y \,
dx \bigg[{{1}\over{2}} {(\partial_{0}
\phi)}^2 - {{1}\over{2}} {(\partial_1 \phi )}^2
\nonumber \\
&&- {{1}\over{2}} (1-E^2)\sum_a {(\partial_a \phi )}^2 - {{1}\over{2}}
(1-E^2) {(\partial_x
\phi )}^2 - (1-E^2)V(\phi) \bigg].\end{aligned}$$ Here, $a$ labels directions parallel to the brane, but perpendicular to the electric field.
Referring ourselves to the action (\[explovercubic\]), we first collected an overall factor in front, in order to make the following analysis clearer. We are interested in determining the solutions of the classical equations of motion, and to discuss their stability. The overall factor does not change this kind of analysis, since it does not alter the equations of motion and it does not affect the issue of stability around the extrema of the classical potential. We can therefore disregard this factor as far as this analysis is concerned. We will return later on its actual role.
From inspection of (\[explovercubic\]), one can see that the time derivative is normalized as in the $E = 0$ case, but the potential has a factor in front with respect to that case, which changes sign in the overcritical case. We realize therefore that, together with the already known effect of the extra contribution coming from transverse motion[^7], the net effect of the overcritical electric field is to reverse the sign of the potential. This potential is drawn in Fig.2.
{width="6cm"}
Such an effect is consistent with the general argument concerning the sign of the squared masses in the overcritical electric background. The reversed tachyon potential for static configurations has a stable minimum at $e^{-1/2}$, that corresponds to the D25-brane. This means that the tachyon is now a stable (positive mass) state in the spectrum of the D25-brane[^8].
The model is therefore suitable to mimic the situation in the full String Field Theory, and one can be further convinced of this by examining the expanded action around the lump solution by replacing $\phi$ with $\bar{\phi} + \phi$. The action reads now $$\begin{aligned}
\label{overcubicexp}
S&=&
{{1}\over{g_o^2}} \sqrt{1-E^2} \int d^{25}y \, dx ~\Big[ - {{1}\over{2}}
{\Big({{d\bar\phi}\over{dx}}\Big)}^2 - V(\bar\phi)\Big] \nonumber\\
&& + {{1}\over{g_o^2}}
\frac{1}{\sqrt{1-E^2}} \int d^{25}y \, dx ~\Big[{{1}\over{2}} {(\partial_{0}
\phi)}^2 - {{1}\over{2}} {(\partial_1 \phi )}^2
- {{1}\over{2}} (1-E^2)\sum_a {(\partial_a \phi )}^2 \nonumber \\
&& - {{1}\over{2}} (1-E^2) \phi \big( - \partial^2_x + V''(\bar\phi)\big)\phi
+ \ldots
%- {{1}\over{3}} (1-E^2) \phi^3
\Big].\end{aligned}$$ The lump solution is still interpreted as a codimension $1$ brane. Its asymptotes at $x=\pm \infty$ correspond to the absence of the D25-brane. The first two terms correctly reproduce the D24-brane tension with the Born-Infeld factor which accounts for the electric background. This factor becoming imaginary was the object of the discussion in Sec.2.
The remaining part gives just the effective action for the fluctuations, which is used to determine the dynamics of the decay through the equations of motion and the analysis of stability, which are not influenced by the overall factor. The effective one dimensional Schrödinger problem is untouched by the presence of the electric field, the spectrum of fluctuations is therefore the same. But when we plug back the mode expansion in the action, we see that the eigenvalues of the Schrödinger equation contribute with a multiplicative factor $(1-E^2)$. This is precisely the modification of the squared mass for the open string spectrum according to the general treatment in Sec.2. We see here that the theory living on the worldvolume of the lump suitably describes the infinite tachyonic tower of states of an electrified D-brane. After integration over $x$, the resulting multiscalar potential can be taken as a toy model for the behaviour of the full String Field Theory in the presence of an overcritical electric field.
We can now predict inside this toy model the value of the condensate for the infinite tachyonic tower. The final stage representing the end of the rolling manifests itself in this toy model as a stable minimum of the reversed potential (corresponding to the old maximum) [^9].
We can apply the same technique as in the case without electric field to compute the expectation values, only requiring that the lump profile plus the fluctuations reduces the original tachyon to its true minimum at $e^{- 1/2}$.
The overcritical counterpart of the equation (\[determ1\]) is therefore $$\begin{aligned}
\label{master}
e^{- 1/2} & = & \bar{\phi}(x) + \sum_{n=0}^{\infty} \bar\phi_n \psi_n(x)\\
& = & \exp[-{{x^2}/{4}}] + \sum_{n=0}^{\infty} \bar\phi_n \psi_n(x), \nonumber \end{aligned}$$ where the eigenfunctions $\psi_n (x)$ are the usual harmonic oscillator basis constructed in terms of Hermite polynomials $$\label{Herm}
\psi_n (x) = {{1}\over{2^{n/2} \sqrt{n!}}}\, H_n \Big({{x}\over{\sqrt{2}}}\Big) \,
\exp[-{{x^2}/{4}}].$$ If we absorb the Gaussian profile in a redefinition of the coefficient $\bar\phi_0$, then Eq.(\[master\]) represents an expansion of a constant in terms of a Hilbert space basis. A constant is certainly not in the Hilbert space. Still it is possible to compute the coefficients $\bar\phi_n$ by performing the scalar product with any elements of the basis[^10]. From $$\label{scalpro}
\bar\phi_n = {{1}\over{\sqrt{2\pi}}} \int_{-\infty}^{\infty} dx \, \psi_n (x)
\Big(e^{- 1/2} -
\exp \Big[-{{x^2}\over{4}}\Big]\Big),$$ one gets $$\label{condens}
\bar\phi_{2n+1} = 0,~~~~~~ \bar\phi_{2n} = - \delta_{n,0} + {{2^{- n}}\over{n!}} \sqrt{{{2 \,(2 n)!}\over{e}}}.$$ This is the value of the condensates for the whole tower of states at the true vacuum. At large level number $n$, using Stirling’s approximation, one gets a behaviour $$\label{behav}
\bar\phi_{2n} \longrightarrow \sqrt{{{2}\over{e}}} {(\pi n)}^{- {{1}\over{4}}}.$$
Due to the slow decrease of $\bar{\phi}_{2n}$ in (\[behav\]), the resulting series is not absolutely and not uniformly (for all $x$) convergent. However, the oscillations with $n$ of the sign of the Hermite polynomials enforce uniform convergence in finite $x$-intervals. It is straightforward to check this at least numerically.
Although the expectation values for higher modes go to zero with $n \rightarrow \infty$ the decrease is not as fast as in the case without electric field [@z; @mz]. In accordance with our previous discussion we take the [*slow*]{} decrease as an indication that level truncation would be a bad approximation. Indeed, we performed some explicit checks in level truncation, and they confirmed these expectations. Taking into account up to level $n=2$, and expanding the reversed effective potential up to the fifth power in these scalars, a minimum is found at ${\bar\phi}_0 =
-0.226$, ${\bar\phi}_1 = 0$, $\bar\phi_2 = 0.555$. This has to be compared with the exact result (\[condens\]), namely ${\bar\phi}_0 =
-0.142$, ${\bar\phi}_1 = 0$, $\bar\phi_2 = 0.607$. At first sight this may seem not so bad an approximation, but we notice that, first, truncating at even powers up to the fourth, sixth and eighth ones produce instead no minimum, and, second, truncation to seventh and ninth powers produce a worse minimum. For example, with the seventh powers included the minimum is at ${\bar\phi}_0 =
-0.0831$, ${\bar\phi}_1 = 10^{-8}$, $\bar\phi_2 = 0.350$, and with the ninth powers included the minimum is at ${\bar\phi}_0 =
-0.0463$, ${\bar\phi}_1 = 10^{-4}$, $\bar\phi_2 = 0.259$. If convergence of level truncation is eventually to be obtained, it is done with wide oscillating behaviour, similar to the convergence of the sum in (\[master\]).
A last comment concerns the nature of the endpoint of the condensation process. The toy model setup is designed to simulate the decay of a D24-brane as a lump solution of the D25 tachyon field $\phi$. For the case of undercritical electric field the minimum of the potential for $\phi$ corresponds to the absence of the D25. Now in the overcritical case $\phi$ takes a value as in the presence of a D25. Within the lump-based model one has no possibility to decide whether this value for $\phi$ indicates a D25 or the true vacuum in which $\phi$ could assume just the same value ($\phi$ does not drive a D25 decay). However, beyond the model one should expect that it will be the duty of the tower of unstable D25 modes to prevent a formation of a D25.
From the open string point of view, the analogy with pair production of pointlike charges suggests that the endpoint of the condensation process corresponds to the discharge of the overcritical electric field: dipoles nucleating from the vacuum easily stretch to infinity until they screen the electric field making it critical or undercritical [@sbranes2]. When looking at the D-brane decay, this means that the fluctuations around the unstable configuration are going to lower the value of the electric background itself towards critical or undercritical values. In the T-dual picture, one expects therefore that the T-dual superluminal brane (S-brane) would lower its velocity to (under-) luminal, thereby becoming a usual timelike D-brane. This would correspond to what is observed for example in [@sbranes1], where the final stage of the S-brane evolution is represented by a flattening of its profile. This gives a hint that a remnant could be left after condensation, in accord to the general expectation of fundamental string fluxes as final products of the decay in the presence of electric fields [@mukho]. In order to get a full description of this phenomenon one would have to allow dynamics for the electric field as well and properly consider backreaction of the D-brane. In this respect we think that a closer connection to the S-brane picture could again be fruitful.
Acknowledgments
===============
We thank Antonio Bassetto, Loriano Bonora, Giancarlo De Pol, Matthias Gaberdiel, Carlo Maccaferri, Adriano Parodi, Rodolfo Russo, Roberto Valandro and Hyun Seok Yang for useful discussions. We would like to thank Koji Hashimoto for pointing out to us the connection with S-branes, and for very interesting e-mail exchange. DFG supported H.D. and A.T. within the “Schwerpunktprogramm Stringtheorie 1096” and M.S. under the project SA 1356/1.
[99]{}
A. Sen, “Rolling tachyon,” JHEP [**0204**]{} (2002) 048 \[arXiv:hep-th/0203211\].
A. Sen, “Tachyon matter,” JHEP [**0207**]{} (2002) 065 \[arXiv:hep-th/0203265\].
A. Sen, “Tachyon dynamics in open string theory,” arXiv:hep-th/0410103.
K. Hashimoto, P. M. Ho and J. E. Wang, “S-brane actions,” Phys. Rev. Lett. [**90**]{} (2003) 141601 \[arXiv:hep-th/0211090\].
K. Hashimoto, P. M. Ho, S. Nagaoka and J. E. Wang, “Time evolution via S-branes,” Phys. Rev. D [**68**]{} (2003) 026007 \[arXiv:hep-th/0303172\].
B. Durin and B. Pioline, “Aspects of Dirichlet S-branes,” arXiv:hep-th/0507059.
C. P. Burgess, “Open String Instability In Background Electric Fields,” Nucl. Phys. B [**294**]{} (1987) 427.
C. Bachas and M. Porrati, “Pair creation of open strings in an electric field,” Phys. Lett. B [**296**]{} (1992) 77 \[arXiv:hep-th/9209032\].
E. S. Fradkin and A. A. Tseytlin, “Nonlinear Electrodynamics From Quantized Strings,” Phys. Lett. B [**163**]{} (1985) 123.
A. Abouelsaood, C. G. . Callan, C. R. Nappi and S. A. Yost, “Open Strings In Background Gauge Fields,” Nucl. Phys. B [**280**]{} (1987) 599.
V. V. Nesterenko, “The Dynamics Of Open Strings In A Background Electromagnetic Field,” Int. J. Mod. Phys. A [**4**]{} (1989) 2627.
N. Seiberg and E. Witten, “String theory and noncommutative geometry,” JHEP [**9909**]{} (1999) 032 \[arXiv:hep-th/9908142\].
A. Bassetto, A. Torrielli and R. Valandro, “One-loop unitarity of string theories in a constant external background and their Seiberg-Witten limit,” JHEP [**0401**]{} (2004) 040 \[arXiv:hep-th/0311120\].
N. Seiberg, L. Susskind and N. Toumbas, “Strings in background electric field, space/time noncommutativity and a new noncritical string theory,” JHEP [**0006**]{} (2000) 021 \[arXiv:hep-th/0005040\].
R. G. Leigh, “Dirac-Born-Infeld Action From Dirichlet Sigma Model,” Mod. Phys. Lett. A [**4**]{} (1989) 2767.
J. Polchinski, “Lectures on D-branes,” arXiv:hep-th/9611050.
C. Bachas, “D-brane dynamics,” Phys. Lett. B [**374**]{} (1996) 37 \[arXiv:hep-th/9511043\].
E. Witten, “Noncommutative Geometry And String Field Theory,” Nucl. Phys. B [**268**]{} (1986) 253.
P. Mukhopadhyay and A. Sen, “Decay of unstable D-branes with electric field,” JHEP [**0211**]{} (2002) 047 \[arXiv:hep-th/0208142\].
C. Maccaferri, R. J. Scherer Santos and D. D. Tolla, “Time-localized projectors in string field theory with E-field,” Phys. Rev. D [**71**]{} (2005) 066007 \[arXiv:hep-th/0501011\].
V. A. Kostelecky and S. Samuel, “On A Nonperturbative Vacuum For The Open Bosonic String,” Nucl. Phys. B [**336**]{} (1990) 263.
A. Sen and B. Zwiebach, “Tachyon condensation in string field theory,” JHEP [**0003**]{} (2000) 002 \[arXiv:hep-th/9912249\].
B. Zwiebach, “A solvable toy model for tachyon condensation in string field theory,” JHEP [**0009**]{} (2000) 028 \[arXiv:hep-th/0008227\].
J. A. Minahan and B. Zwiebach, “Field theory models for tachyon and gauge field string dynamics,” JHEP [**0009**]{} (2000) 029 \[arXiv:hep-th/0008231\].
J. M. F. Labastida and M. A. H. Vozmediano, “Bosonic Strings In Background Massive Fields,” Nucl. Phys. B [**312**]{} (1989) 308.
[^1]: For charged strings, another kind of instability due to pair-production dominates at the quantum level for weak fields [@b; @bp]. This is the analog of the Schwinger phenomenon for particle electrodynamics. The derived rate is however zero for neutral strings, preventing this kind of mechanism from screening the growth in absolute value of the electric field, until it reaches the critical value at which the classical instability appears. On the other side, raising slowly $E$ exposes the bosonic D-brane to the usual tachyon decay (which is absent in the similar case of open superstrings). We assume an overcritical electric field as a given initial condition.
[^2]: We keep in mind that this last contribution to the squared mass is present even for undercritical fields, while only for overcritical fields it gives a tachyonic contribution to the energy, according to the dispersion relation $k_0^2 = k_1^2 + (1 - E^2) k_{\perp}^2 + (1 - E^2) (n - 1)$ derived from (\[spectrumel\]). The sign of the “bare” squared mass term depending on the level $n$ is always reversed in overcritical fields.
[^3]: We remark that a possible interpretation of this change of signature from the point of view of the full target space could be ascribed to an interchange of the role of space and time between the directions $0$ and $1$. However, if one demands now $x_1$ to assume the role of time, thus recovering a standard positive mass spectrum for the open string tower, then the closed string sector of the theory, whose masses are determined *via the closed string metric as $\kappa_{\mu} g^{\mu \nu} \kappa_{\nu}
= - 4 (n - 1)$, would have infinite tachyons.*
[^4]: At the same time the endpoints of the strings classically acquire a superluminal velocity [@n].
[^5]: See [@carlo] for considerations on critical electric fields in Vacuum SFT.
[^6]: We will again always consider here either effective one dimensional $x$-variable problems, or static or linearized ones in the $y$ variables, therefore the star product will never play any role, and we will omit to write it.
[^7]: Compare the derivatives in the direction $y^a$ in (\[explovercubic\]) with the contribution from $k_\perp$ in (\[spectrumel\]) and (\[spectrumel2\]).
[^8]: The D25-brane is affected as well as the D24-brane (lump) by the electric field, which is switched on in directions common to both of them.
[^9]: Of course it is open whether this feature is shared by the real String Field Theory potential. At this stage we take it as a conjecture.
[^10]: The basis is normalized such that $\int dx \, \psi_n (x) \, \psi_m (x)
= \sqrt{2\pi} \delta_{n,m}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This article studies the scaling limit of a class of shot-noise fields defined on an independently marked stationary Poisson point process and with a power law response function. Under appropriate conditions, it is shown that the shot-noise field can be scaled suitably to have a non degenerate $\alpha$-stable limit, as the intensity of the underlying point process goes to infinity. More precisely, finite dimensional distributions are shown to converge and the finite dimensional distributions of the limiting random field have i.i.d. stable random components. We hence propose to call this limit the $\alpha$- stable white noise field. Analogous results are also obtained for the extremal shot-noise field which converges to a Fréchet white noise field. Finally, these results are applied to the modeling and analysis of interference fields in large wireless networks.
[**AMS subject classifications:**]{} (Primary) 60D05, 44A10, (Secondary) 60G55, 82B43,
[**Keywords:**]{} Poisson point process, shot-noise process, percolation, $\alpha$-stable white noise, scaling limits.
author:
- |
François Baccelli and Anup Biswas\
\[2mm\] Department of Electrical and Computer Engineering,\
University of Texas, Austin.
date: 'November 19, 2014'
title: 'On Scaling Limits of Power Law Shot-noise Fields [^1]'
---
Introduction
============
The present paper is focused on scaling limits for a class of shot-noise fields [@SKM] associated with stationary Poisson point processes and with a power law response function. It is motivated by the modeling of ultra-dense wireless networks and the analysis of interference fields that arise in this type of stochastic networks [@Baccelli-Bartek]. By scaling limits, we understand the analogue of a functional central limit theorem, namely a rescaling of the field such that, in the limit, this field has non-degenerate joint distributions. To give a first example, let $L$ be a bounded, non-negative, radial function (i.e. $L(x)=L(\|x\|)$) on ${\mathbb{R}}^d$ that is integrable w.r.t. to Lebesgue measure $m(dy)$ on ${\mathbb{R}}^d$. Now if $\Phi_\lambda$ is a stationary Poisson point process on ${\mathbb{R}}^d$ of intensity $\lambda$, the corresponding shot-noise field is given by $${\mathcal{I}}_\lambda(L;z)=\sum_{\{x_i\}\in\Phi_\lambda}L(\|x_i-z\|).$$ This field is translation invariant. It turns out that ${\mathcal{I}}_\lambda(L;0)$ has finite moments. In fact, ${\mathbb{E}}[{\mathcal{I}}_\lambda(L;0)]=\lambda\int_{{\mathbb{R}}^d}L(\|y\|)\, m(dy)$. Therefore we can define the second order approximation of ${\mathcal{I}}_\lambda$ as follows $$\hat{{\mathcal{I}}}_\lambda(L;z)=\frac{1}{\sqrt{\lambda}}\Big({\mathcal{I}}_\lambda(L;z)-{\mathbb{E}}[{\mathcal{I}}_\lambda(L;0)]\Big).$$ Then using the Laplace transform, it is not difficult to show that as $\lambda\to\infty$, the scaled field $\hat{{\mathcal{I}}}_\lambda$ converges to a Gaussian random field. In particular, given two points $z_1, z_2\in {\mathbb{R}}^d$, $(\hat{{\mathcal{I}}}_\lambda(L;z_1), \hat{{\mathcal{I}}}_\lambda(L;z_2))$ converges to a $2$-dimensional Gaussian random variable with covariance matrix given by $$\begin{pmatrix}
\int_{{\mathbb{R}}^d}L^2(\|z_1-y\|)\, m(dy) & \int_{{\mathbb{R}}^d}L(\|z_1-y\|)L(\|z_2-y\|)\, m(dy)\\
\int_{{\mathbb{R}}^d}L(\|z_1-y\|)L(\|z_2-y\|)\, m(dy) & \int_{{\mathbb{R}}^d}L^2(\|z_2-y\|)\, m(dy)
\end{pmatrix}.$$
It turns out that the response function that is the most commonly used in the wireless literature is a power law [@Baccelli-Bartek] i.e., $L(x)=\frac{1}{\|x\|^\beta}, \beta>d$, so that the central limit scale is not the appropriate one. Two interesting limits come from the identification of the appropriate scaling: the [*stable white noise field*]{} and the [*Fréchet white noise field*]{}. A particularly nice property of these fields is their independence property: all their finite dimensional distributions are those of random vectors with independent and identically distributed components. In general, this property is not shared by fields for which the right scale is the central limit one (see e.g. [@bierme; @breton; @kaj]). The convergence of the random fields are defined in terms of the convergence of their finite dimensional distributions. In the present paper, the underlying point process is a Poisson point process (PPP) and we use the Laplace transform to establish convergence. Scaling limits of random fields fed by a PPP have already been thoroughly studied in the literature. In the $1$-dimensional setting, scaling properties of Poisson shot-noise processes were thoroughly studied in [@klup-kuhn-04; @klup-mik-03; @klup-mik-95]. For other $1$-dimensional point processes than Poisson, see also [@doney-obrien; @hsing-teugels]. In higher dimension, [@hein-molcha] studies limits of union random fields associated with a stationary point process and defined by certain classes of functions that are regularly varying at zero. Properties of extremes of shot-noise fields defined by a bounded slowly varying function are studied in [@lebedev]. [@dombry] establishes various qualitative properties of extremal random fields. Let us mention that the main difference whith these papers is that the class of functions that are used in the present paper to analyze high density shot-noise (or extremal shot-noise) fields, are singular at $0$ and not integrable on ${\mathbb{R}}^d$. It is shown below that this singularity at $0$ is instrumental to get the independence (white noise) properties of the limiting random fields, which are one of the main findings of the present paper.
These limit results are discussed in Section \[secmr\]. Section \[secapp\] discusses a few applications of these limit results to the modeling of communication rates in ultra-dense wireless networks where interference is treated as noise. For the detailed context, see [@Baccelli-Bartek]. The networks considered feature wireless transmitters located according to some realization of a homogeneous PPP of intensity $\lambda$ in ${\mathbb{R}}^2$. The interference field is modeled as the shot-noise field of this Poisson point process for the response function $r^{-\beta}$. Each transmitter maintains a wireless link (an information theoretic channel) to its receiver, which is assumed to be located at distance 1 from it, in some random and uniform direction. It is shown in Section \[secapp\] that the results of Section \[secmr\] can be used to obtain the speed of decrease of the SIR (Signal to Interference Ratio) of a typical link in such a network when $\lambda$ tends to infinity. The interest in this question stems from the following result of information theory: when treating interference as noise, the (Shannon) communication rate obtained by this typical link is proportional to $\log (1+{\mathrm{SIR}})$ [@CoT91]. Hence, in the first place, these limiting results allow one to predict the speed of decrease of the Shannon rate of the typical link when the network is densified (i.e. $\lambda$ tends to infinity).
These limiting results also allow one to estimate the speed at which the SIR decreases with $\lambda$ in a network with density 1 when all links now have length $\lambda$. More precisely, it is shown in Section \[secapp\] that in such a scenario, it is possible to transmit $1$ bit over distance $\lambda$ with a delay $D_\lambda$ such that the ration $\frac{D_\lambda}{\lambda^{\beta}}$ has a non-degenerate limit with high probability when $d$ tends to infinity (Theorem \[theo-2\]). To compare this result with the existing ones, we recall a result from [@Baccelli-Bartek-Omid] which can be expressed as follows: for all fixed positive SIR threshold[^2], for all possibly single or multi-hop strategies, the expected *delay*, say ${\mathbb{E}}[D_\lambda]$, to transmit $1$ bit over distance $\lambda$ in a Poisson wireless network of the type described above and with density $1$ grows faster than $\lambda$, i.e., $\frac{{\mathbb{E}}[D_\lambda]}{\lambda}\to\infty$ as $\lambda\to\infty$. Finally, we use the joint scaling limit results of Section \[secmr\] to derive percolation properties of the SINR graph [@Fran-Mees; @Baccelli-Bartek]. The Signal to Interference and Noise Ratio (SINR) differs from the SIR defined above as some constant or random term, called the thermal noise power, is added to the interference in the denominator of the ratio of SINR [@Baccelli-Bartek]. The SINR graph is a random geometric graph with an edge between two nodes if the SINR from one to the other is above some threshold. The results of Section \[secmr\] are used to estimate the speed of decrease to 0 of the SINR threshold that makes this graph percolate when the network density tends to infinity (Theorems \[theo-3\] and \[theo-4\]).
Model and Scaling Limits {#secmr}
========================
As mentioned above, one of the most common response functions used in wireless networking (for $d=2$) is $\frac{1}{\|x\|^\beta}, \beta>2,$ and this is neither bounded nor globally integrable. Nevertheless, the additive and the extremal shot-noise fields generated by this type of response function are finite a.s. We are interested in their scaling limits. Let $m(dy)$ denote the Lebesgue measure on ${\mathbb{R}}^d$. By $\|\cdot\|$ we denote the Euclidean norm on ${\mathbb{R}}^d$. We start by defining the Poisson point process of intensity $\lambda$ on ${\mathbb{R}}^d$.
A Poisson point process (PPP) $\Phi_\lambda$ of intensity $\lambda$ on ${\mathbb{R}}^d$ is a stationary point process such that for any bounded disjoint collection of Borel sets $A_i, i=1,\ldots, k,$ of ${\mathbb{R}}^d$ we have $${\mathbb{P}}\big(\Phi(A_1)=n_1, \Phi(A_2)=n_2,\ldots, \Phi(A_k)=n_k\big)=\Pi_{i=1}^k\Big[e^{-\lambda m(A_i)}\frac{\big(\lambda m(A_i)\big)^{n_i}}{n_i!}\Big].$$
By independently marked PPP we mean a marked point process $\tilde\Phi=\{(x_i, {\mathbf{p}}_i)\}_i$ where the locations $\Phi=\{x_i\}$ are given by a PPP and the markings $\{{\mathbf{p}}_i\}$ are independent of the PPP, i.e., ${\mathbb{P}}({\mathbf{p}}\in A|\Phi)=\int_A F(dp)$, $A\subset (0, \infty)$ Borel set, is independent of $\Phi$. Marks are assumed to take values in $(0, \infty)$. We also assume that ${\mathbb{E}}[{\mathbf{p}}]=\int p\, F(dp)<\infty$. By the intensity of a independently marked PPP $\tilde\Phi$ we refer to the intensity of the underlying PPP $\Phi$.
Now we recall the following classical result on the Laplace functional of an independently marked Poisson point process (see e.g. [@Baccelli-Bartek; @daley-vere]) which will play a central role in our analysis.
\[laplace\] Let $\tilde\Phi_\lambda$ be an independently marked PPP of intensity $\lambda$ and $\mathcal{L}_{\tilde\Phi_\lambda}$ be its Laplace functional. Then for any non-negative measurable $g$ we have $$\mathcal{L}_{\tilde\Phi_\lambda}(g)=\exp\Big\{-\lambda\int_{{\mathbb{R}}^2}\Big(1-\int e^{-g(y, p)} F(dp)\Big) m(dy)\Big\},$$ where $F(\cdot)$ denotes the distribution of the marks.
Now we define the *shot-noise* random field corresponding to a marked point process. Given a marked point process $\tilde\Phi=\{(x_i, p_i)\}$ on ${\mathbb{R}}^d\times (0, \infty)$ and a response function $L:{\mathbb{R}}^d\times {\mathbb{R}}^d\times (0, \infty)\to [0, \infty)$, the corresponding shot-noise field at $y\in{\mathbb{R}}^d$ is defined by $${\mathcal{I}}(L;z)=\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi}L(z, x_i, {\mathbf{p}}_i).$$ Let $f:[0, \infty)\to [0, \infty]$ be a measurable function with the property that for some positive $\varrho>0$, $$\label{gen-fn}
f(r)=\frac{1}{r^\beta}, \quad \text{for}\ 0\leq r\leq \varrho,\; \; \sup_{r\geq \varrho} f(r)<\infty,
\quad\text{and},\; \int_\varrho^\infty r^{d-1}f(r)dr<\infty,$$ for some $\beta>d$. All the response functions considered in this article have the form $L(z, x, p)=p\cdot f(\|z-x\|)$ which is commonly used in the wireless network literature as already explained. We will denote by ${\mathcal{I}}(f; \cdot)$ the shot-noise corresponding to the response function $L(z, x, p)=p\cdot f(\|x-z\|)$. A special response function that will occur often in this article is $$\label{res-fn}
L(z,x, p)=\frac{p}{\|x-z\|^\beta},$$ for some $\beta>d$. We use the notation ${\mathcal{I}}$ to denote the shot-noise corresponding to the response function , i.e., $${\mathcal{I}}(z)=\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi}\frac{{\mathbf{p}}_i}{\|x_i-z\|^\beta}.$$ Now if $\tilde\Phi$ is an independently marked PPP, then for any $f$ given by , the following facts are well known [@Baccelli-Bartek], Part I, Chapter 2:
- ${\mathcal{I}}(f; 0)$ is almost surely finite,
- ${\mathcal{I}}(f; \cdot)$ is stationary.
The finiteness property follows from the integrability condition of $f$ (last condition of (\[gen-fn\])). In fact, if $B$ denotes the unit ball around $0$ then $${\mathbb{E}}\Bigl[\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi[B^c]}{\mathbf{p}}_if(\|x_i\|)\Bigr]
= \int_{B^c}f(x)m(dx)\int pF(dp)=
\omega(d)\int_{1}^\infty r^{d-1}f(r) dr\int pF(dp)<\infty,$$ where $\tilde\Phi[A]$ denotes the independently marked PPP with Poisson point process restricted to the Borel set $A$ and $\omega(d)$ denotes the surface area of $B$. Therefore finiteness of ${\mathcal{I}}(f; 0)$ would follow if one has $$\label{101}
{\mathbb{P}}\Bigl(\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi[B]}{\mathbf{p}}_if(\|x_i\|)<\infty\Bigr)=1.$$ Let $\Phi[B]$ be the restriction of PPP to $B$. Then $$\begin{aligned}
\label{102}
{\mathbb{P}}\Bigl(\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi[B]}{\mathbf{p}}_if(\|x_i\|)<\infty\Bigr)
&=\sum_{k=1}^\infty
{\mathbb{P}}\Bigl(\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi[B]}{\mathbf{p}}_i f(\|x_i\|)<\infty\big|
\Phi[B]=k\Bigr) {\mathbb{P}}(\Phi[B]=k).\end{aligned}$$ Using the property of Poisson point process for every $k\geq 1$, we can have a i.i.d. sequence $\{X_1,\ldots, X_k\}$, uniformly distributed on $B$, such that $$\label{103}
{\mathbb{P}}\Bigl(\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi[B]}{\mathbf{p}}_i f(\|x_i\|)<\infty\big|
\Phi[B]=k\Bigr)
= {\mathbb{P}}\Bigl(\sum_{i=1}^k{\mathbf{p}}_i f(\|X_i\|)<\infty\Bigr)=1.$$ Hence follows from and . Stationarity of ${\mathcal{I}}(f, \cdot)$ is obvious from the stationary behavior of the underlying Poisson point process.
Scaling Limits of Shot-noise Fields {#subsec-shot}
-----------------------------------
We are interested in the limiting behavior of the shot-noise field when the intensity of the underlying PPP goes to infinity. By ${\mathcal{I}}_\lambda(f;\cdot)$ (resp. ${\mathcal{I}}_\lambda$) we denote the shot-noise field w.r.t. to an independently marked PPP of intensity $\lambda$ and response function given by $L(z, x, p)=p\cdot f(\|x-z\|)$ where $f$ satisfies (resp (\[res-fn\])). Let $\kappa=\frac{\beta}{d}\in(1, \infty)$. Define $$\hat{{\mathcal{I}}}_\lambda(f;\cdot)=\frac{1}{\lambda^\kappa}{\mathcal{I}}_\lambda(f;\cdot) \quad (\mbox{resp. }
\hat{{\mathcal{I}}}_\lambda=\frac{1}{\lambda^\kappa}{\mathcal{I}}_\lambda).$$ For $n\geq 1$, and $(x, t)\in {\mathbb{R}}^d\times[0, \infty)$ we define $$\mathcal{L}_\lambda(f; x, t)={\mathbb{E}}\Big[e^{-t\,\hat{{\mathcal{I}}}_\lambda(f;x)}\Big],\quad
\left(\mbox{resp. } \mathcal{L}_\lambda(x, t)={\mathbb{E}}\Big[e^{-t\, \hat{{\mathcal{I}}}_\lambda(x)}\Big]
\right).$$ We recall that marks are independently distributed with distribution $F(dp)$ and ${\mathbb{E}}[{\mathbf{p}}]=\int p\, F(dp)<\infty$.
\[lem1\] Let $x\in{\mathbb{R}}^d$. Define $\alpha=\frac{1}{\kappa}=\frac{d}{\beta}$. Then, for all $x\in {\mathbb{R}}^d$ and $t\ge 0$, as $\lambda\to\infty$, for all $t\geq 0$, we have $$\label{21}
\mathcal{L}_\lambda(f; x, t)\longrightarrow \exp\Big(-{\mathbb{E}}[{\mathbf{p}}^{\alpha}]t^{\alpha}C(d,\beta)\Big),$$ where $C(d, \beta)=\frac{\omega(d)}{\beta}\int_0^\infty (1-e^{-s})s^{-1-\alpha}ds$ and $\omega(d)$ is the surface area of the unit ball in ${\mathbb{R}}^d$.
Note that $C(d, \beta)$ is finite because $(1-e^{-s})\leq s$ for all $s\geq 0$.
[**Proof:**]{} Because of stationarity it is enough to prove the result for $x=0$. Also there is nothing to prove if $t=0$. Assume $t>0$. First we observe from that $$\begin{aligned}
\label{105}
&\mathcal{L}_\lambda(f; 0, t)\nonumber
\\
&=\exp\Big[-\int_{{\mathbb{R}}_+}\omega(d)\int_{0}^\varrho\Big(1-e^{-\frac{\lambda^{-\kappa}t\, p}{r^\beta}}\Big)\lambda r^{d-1}\; dr F(dp)
-\int_{{\mathbb{R}}_+}\int_{\|y\|\geq \varrho}\Big(1-e^{-{\lambda^{-\kappa}tp}{f(\|y\|)}}\Big)\lambda\, m(dy) F(dp)\Big].\end{aligned}$$ Now for the first term on the r.h.s. we use the change of variable $pt r^{-\beta} \lambda^{-\kappa}=s$ to obtain $$\begin{aligned}
\int_{0}^\varrho\Big(1-e^{-\frac{n^{-\kappa}t\, p}{r^\beta}}\Big)\lambda r^{d-1}\; dr
= \frac{\omega(d)}{\beta}p^\alpha\, t^\alpha\int_{\lambda^{-\kappa}\varrho^{-\beta}tp}^\infty
(1-e^{-s})s^{-1-\alpha}ds.\end{aligned}$$ Thus by dominating convergence theorem we get as $\lambda\to \infty$, $$\int_{{\mathbb{R}}_+}\omega(d)\int_{0}^\varrho\Big(1-e^{-\frac{\lambda^{-\kappa}t\, p}{r^\beta}}\Big)\lambda r^{d-1}\; dr F(dp)\to
{\mathbb{E}}[{\mathbf{p}}^\alpha]t^\alpha C(d, \beta).$$ Thus we only need to show from that the second term inside the exponential goes to $0$ with $\lambda\to\infty$. Since $(1-e^{-x})\leq x,$ for $x\geq 0,$ we have $$\begin{aligned}
\int_{{\mathbb{R}}_+}\int_{\|y\|\geq \varrho}\Big(1-e^{-{\lambda^{-\kappa}tp}{f(\|y\|)}}\Big)\lambda\, m(dy) F(dp)
&\leq\; \int_{{\mathbb{R}}_+}\int_{\|y\|\geq \varrho}{\lambda^{-\kappa}tp}{f(\|y\|)}
\lambda\, m(dy) F(dp)
\\
&\leq\; t\, \lambda^{1-\kappa}\int pF(dp) \Big[\omega(d) \int_{\varrho}^\infty rf(r)dz\Big]
\\
&\to\; 0,\end{aligned}$$ as $\lambda\to\infty$ where we use that $\kappa>1$. Hence the proof. $\Box$
It is interesting to note that Lemma \[lem1\] holds without the assumption that $\sup_{r\geq\varrho}f(r)<\infty$.
A non-degenerate random variable $X$ is said to be stable if for any $a, b>0$ and two independent copies $(X_1, X_2)$ of $X$, $aX_1+b X_2$ has the same distribution as $cX+d$ for some $c>0$ and $d\in{\mathbb{R}}$ [@Em-Kl-Mi]. Stable random variables are characterized by their characteristic functions and Laplace transforms. $X$ is said to be an $\alpha$-stable random variable ($0<\alpha<1$) if its Laplace transform is given by ${\mathbb{E}}[\exp(-tX)]=\exp(-\eta|t|^\alpha)$ for $t\geq 0$ and some constant $\eta$.
The fact that stable laws show up in the context of (\[res-fn\]) is not new (see e.g. [@Baccelli-Bartek; @haenggi]). What is new, to the best of our knowledge, is the fact that the distribution of the finite dimensional marginals of the limiting random field is of product form:
\[theo-1\] Let $x_i, i=1,2,\ldots, k$ be $k$-distinct points in ${\mathbb{R}}^d$. Also let $t_i, i=1,2,\ldots, k$ be positive real numbers. By $\mathcal{L}_\lambda(f; x_1,\ldots, x_k, t_1,\ldots, t_k)$ we denote the Laplace transform of $\big(\hat{{\mathcal{I}}}_\lambda(f; x_1),
\ldots, \hat{{\mathcal{I}}}_\lambda(f; x_k)\big)$ at the point $(t_1, \ldots, t_k)$. Then $$\lim_{\lambda\to\infty}\mathcal{L}_\lambda(f; x_1,\ldots, x_k, t_1,\ldots, t_k)\;=\;\exp\Big(-{\mathbb{E}}[{\mathbf{p}}^{\alpha}]C(d, \beta)\sum_{i=1}^k t^{\alpha}_i\Big), \quad \alpha=\frac{d}{\beta}.$$
In view of Theorem \[theo-1\] we see that $\big(\hat{{\mathcal{I}}}_\lambda(f; x_1), \ldots, \hat{{\mathcal{I}}}_\lambda(f; x_k)\big)$ converges in distribution to $(\xi_1, \ldots, \xi_k)$ as $\lambda\to\infty$ where $(\xi_1, \ldots, \xi_k)$ is an i.i.d. sequence with $${\mathbb{E}}[e^{-t \xi_1}]\;=\;\exp\big(-C(d, \beta){\mathbb{E}}[{\mathbf{p}}^{\alpha}]t^{\alpha}\big),$$ i.e., $\xi_1$ has a $\alpha$-stable distribution. Therefore the limiting random field of $\hat{{\mathcal{I}}}_\lambda(f;\cdot)$ is a stationary random field and its finite dimensional distributions are given by independent and identically distributed $\alpha$-stable variables. It makes sense to call the limiting field a *$\alpha$-stable white noise field*.
[**Proof of Theorem \[theo-1\]:**]{} Recall that $L(x, y, p)=p\, f(\|x-y\|)$. Using Proposition \[laplace\] we have $$\begin{aligned}
\mathcal{L}_\lambda(x_1,\ldots, x_k; t_1,\ldots, t_k)\;=\;
\exp\Big\{-\lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}\Big(1-e^{-\lambda^{-\kappa}\sum_{i=1}^k t_i L(x_i, y, p)}\Big)m(dy)F(dp)\Big\}.\end{aligned}$$ Now for each $p$ we write $$\Big(1-e^{-\lambda^{-\kappa}\sum_{i=1}^k t_i L(x_i, y, p)}\Big)=\sum_{i=1}^k
e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big).$$ Hence $$\begin{aligned}
\label{002}
&\quad \lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}\Big(1-e^{-\lambda^{-\kappa}\sum_{i=1}^k t_iL(x_i, y, p)}\Big)m(dy)F(dp)\nonumber
\\
&=\;\sum_{i=1}^k \lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d} e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)m(dy) F(dp).\end{aligned}$$ Note that the first term of the summation has already been calculated in Lemma \[lem1\]. So we care for $i$-th terms for $i\geq 2$. Fix $i\geq 2$. Then
$$\begin{aligned}
\label{003}
&\quad \lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d} e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp)\nonumber
\\
&\leq\;
\lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)m(dy) F(dp)\nonumber
\\
&\to \; C(d, \beta){\mathbb{E}}[{\mathbf{p}}^{\alpha}]t_i^{\alpha},\end{aligned}$$
where the last line follows from Lemma \[lem1\] (see for example ). Let $\delta=\frac{1}{2} \min_{i\neq j}\|x_i-x_j\|$. Define $B_i=\cup_{l=1}^{i-1} B_\delta(x_l)$ where $B_\delta(x_l)$ denotes the $\delta$-open ball around $x_l$ in ${\mathbb{R}}^d$. Thus $B_i$ is the union of $\delta$-neighborhoods of $x_1,\ldots, x_{i-1}$. By definition $x_i\notin B_i$. Then for any ${\varepsilon}>0$ and $M>0$ we get
$$\begin{aligned}
\label{04}
& \lambda\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d} e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp)\nonumber
\\
&\geq\, \lambda\int_{{\mathbb{R}}_+}\int_{B_i^c} e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp)\nonumber
\\
&\geq\, (1-{\varepsilon})\lambda\int_{\{p\leq M\}}\int_{B_i^c}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp),\end{aligned}$$
for all large $\lambda$ where in the last line we use the fact that $L(x_l, y, p)\leq {M}\sup_{z\geq \delta} f(z)$ for $(y, p)\in B_i^c\times\{p\leq M\}, l\in\{1, \ldots, i-1\},$ and $\lambda^{-\kappa}\to 0$ as $\lambda\to\infty$.
Again on $B_i$, we have $L(x_i, y, p)\leq p \sup_{z\geq \delta} f(z)$. Hence using the fact $(1-e^{-x})\leq x$ for all $x\geq 0$, we get $$\begin{aligned}
\lambda\int_{\{p\leq M\}}\int_{B_{i}}\Big(1-e^{-\lambda^{-\kappa}t_iL(x_i, y, p)}\Big)
m(dy) F(dp)\leq
\lambda\, [\sup_{z\geq \delta} f(z)]\, m(B_\delta(0))\int_{\{p\leq M\}}t_i\, p \lambda^{-\kappa}F(dp).\end{aligned}$$ Since $\kappa=\frac{\beta}{d}>1$ we see that the r.h.s. of the above expression tends to $0$ as $\lambda\to\infty$ for every fixed $M$. Thus combining with we see that $$\begin{aligned}
&\liminf_{\lambda\to\infty}\; \lambda\;\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}
e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp)
\\
&\geq\; \liminf_{\lambda\to\infty}\quad (1-{\varepsilon})
\lambda\int_{\{p\leq M\}}\int_{{\mathbb{R}}^d}\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp),
\\
&=\;(1-{\varepsilon})\int_{\{p\leq M\}}p^{d/\beta}C(d, \beta)t_i^{d/\beta} F(dp),\end{aligned}$$ for any ${\varepsilon}, M>0$ where for the last equality we can follow the same computation as in Lemma \[lem1\]. Now using the fact that ${\mathbb{E}}[{\mathbf{p}}^{d/\beta}]<\infty$, we let $M\to\infty$ and ${\varepsilon}\to 0$ to obtain $$\begin{aligned}
\label{004}
\liminf_{\lambda\to\infty}\; \lambda\; \int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}
e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
m(dy) F(dp)
\;\geq\; {\mathbb{E}}[{\mathbf{p}}^{d/\beta}]C(d, \beta)t_i^{d/\beta}.\end{aligned}$$ Hence combining and we get for $i\geq 2$, $$\begin{aligned}
\lim_{\lambda\to\infty}\; \lambda\;\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^d}
e^{-\lambda^{-\kappa}\sum_{l=1}^{i-1}t_l L(x_l, y, p)}
\Big(1-e^{-\lambda^{-\kappa}t_i L(x_i, y, p)}\Big)
dy F(dp)
\;=\; {\mathbb{E}}[{\mathbf{p}}^{\alpha}]C(d, \beta)t_i^{\alpha}.\end{aligned}$$ Thus the result follows from .$\Box$
Scaling Limits of Extremal Shot-noise Fields
--------------------------------------------
This subsection is devoted to the analysis of the *extremal-random field* generated by the response function (and also by , see Remark \[rem3.2\]). Let $\tilde\Phi$ be given independently marked PPP on ${\mathbb{R}}^d\times (0, \infty)$. We define the *extremal-random field* at a point $y$ as follows: $$\label{max-fl}
{\mathcal{M}}(y)=\sup_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi}\frac{{\mathbf{p}}_i}{\|x_i-y\|^\beta}, \quad \beta>d.$$ We see that ${\mathcal{M}}(y)\leq {\mathcal{I}}(y)$. Therefore ${\mathcal{M}}$ is an almost surely finite and stationary random field. When the underlying PPP has intensity $\lambda$ we denote the extremal-random field by ${\mathcal{M}}_\lambda$. Define $$\hat{{\mathcal{M}}}_\lambda=\frac{1}{\lambda^\kappa}{\mathcal{M}},\quad \text{for}\, \,
\kappa=\frac{\beta}{d}.$$ Let $x_1, x_2, \ldots, x_k$ be $k$-distinct points in ${\mathbb{R}}^d$. By ${\mathcal{G}}_\lambda$ we denote the multivariate cumulative distribution of $(\hat{{\mathcal{M}}}_\lambda(x_1), \ldots, \hat{{\mathcal{M}}}_\lambda(x_k))$ , i.e., $${\mathcal{G}}_\lambda(t_1, \ldots, t_k)={\mathbb{P}}\Big[\hat{{\mathcal{M}}}_\lambda(x_1)\leq t_1, \ldots, \hat{{\mathcal{M}}}_\lambda(x_k)\leq t_k\Big],$$ for $t_i\geq 0,\, i=1,\ldots, k$. The proof of the following lemma is similar to that of [@dombry Proposition 2.4].
\[lem-2\] Let $(t_1,\ldots, t_k)\in [0, \infty)^k$. Then $${\mathcal{G}}_\lambda(t_1, \ldots, t_k)=\exp\Big\{-\lambda\int_{{\mathbb{R}}^d}\int_0^\infty\Big(1-\Pi_{i=1}^k 1_{\{\frac{p}{\|y-x_i\|^\beta}\leq \lambda^{\kappa}t_i\}}\Big)
F(dp)\, m(dy)\Big\},$$ where ${\mathcal{G}}_\lambda$ is defined above.
[**Proof:**]{} For simplicity, we prove the lemma for $\lambda=1$. The proof for general $\lambda$ is analogous. First we observe that $$1_{\{{\mathcal{M}}(x_i)\leq t_i\}} =\Pi_{(y_j, {\mathbf{p}}_j)\in\tilde\Phi}1_{\{\frac{{\mathbf{p}}_j}{\|y_j-x_i\|^\beta}\leq t_i\}}.$$ Therefore $$\begin{aligned}
{\mathbb{P}}\Big[{\mathcal{M}}(x_1)\leq t_1, \ldots, {\mathcal{M}}(x_k)\leq t_k\Big]
&= {\mathbb{E}}\Big[\Pi_{(y_j, {\mathbf{p}}_j)\in\tilde\Phi}1_{\{\frac{{\mathbf{p}}_j}{\|y_j-x_1\|^\beta}\leq t_1\}}\cdots
\Pi_{(y_j, {\mathbf{p}}_j)\in\tilde\Phi}1_{\{\frac{{\mathbf{p}}_j}{\|y_j-x_k\|^\beta}\leq t_k\}}\Big]
\\
& =\; {\mathbb{E}}[e^{H}],\end{aligned}$$ where $$H=\sum_{(y_j, {\mathbf{p}}_j)\in\tilde\Phi}\log\big(1_{\{\frac{{\mathbf{p}}_j}{\|y_j-x_1\|^\beta}\leq t_1\}}\big)+\cdots+
\sum_{(y_j, {\mathbf{p}}_j)\in\tilde\Phi}\log\big(1_{\{\frac{{\mathbf{p}}_j}{\|y_j-x_k\|^\beta}\leq t_k\}}\big).$$ Therefore from Proposition \[laplace\] we have $$\begin{aligned}
{\mathcal{G}}_1(t_1, \ldots, t_k)& =\;\exp\Big\{-\int_{{\mathbb{R}}^d}\bigg(1-\int_0^\infty \exp\Big[\sum_{i=1}^k\log(1_{\{\frac{p}{\|y-x_i\|^\beta}\leq t_i\}})\Big]F(dp)\bigg)m(dy)\Big\}
\\
&=\; \exp\Big\{-\int_{{\mathbb{R}}^d}\int_0^\infty \Big(1-\Pi_{i=1}^k1_{\{\frac{p}{\|y-x_i\|^\beta}\leq t_i\}} \Big)F(dp)m(dy)\Big\}.\end{aligned}$$ Hence the proof. $\Box$
\[theo-ex\] Let $x_1, \ldots, x_k$ be $k$-distinct points in ${\mathbb{R}}^d$. Consider ${\mathcal{G}}_\lambda$ as defined above. Let $$\gamma(d,\beta):=\frac{\omega(d)}{\beta}\int_0^\infty {\mathbb{P}}(p>s)s^{-1+d/\beta}ds=\frac{\omega(d)}{d} {\mathbb{E}}[p^{d/\beta}]<\infty,$$ where $\omega(d)$ denotes the surface area of the unit ball in ${\mathbb{R}}^d$. Then $$\lim_{\lambda\to\infty}{\mathcal{G}}_\lambda(t_1, \ldots, t_k)\;=\;\Pi_{i=1}^k \exp\Big(-\gamma(d, \beta)t_i^{-\alpha}\Big),\quad \alpha=\frac{d}{\beta},$$ for $t_i\geq 0, i=1, \ldots, k$.
By Theorem \[theo-ex\] we see that the limiting random field of the extremal field is a max stable random field, i.e., the finite dimensional distribution of the limiting fields are given by collection of i.i.d. *Fréchet* distribution of exponent $\alpha=d/\beta$ ([@Em-Kl-Mi]). We call the limiting field a *$\alpha$-Fréchet white noise field*. This result is similar to that obtained in [@dombry Section 3.2]. The results in [@dombry] are obtained for $f$ that are integrable on $\mathbb{R}^d$ and the finite dimensional distributions of limiting random field are not necessarily independent there.
[**Proof:**]{} The proof is similar to the proof of Theorem \[theo-1\]. However, we add it for clarity. Without loss of generality we assume that $t_i>0$ for all $i=1, \ldots, k$. Now by Lemma \[lem-2\] we have $$\label{31}
{\mathcal{G}}_\lambda(t_1, \ldots, t_k)\;=\;\exp\Big\{-\lambda\int_{{\mathbb{R}}^d}\int_0^\infty\Big(1-\Pi_{i=1}^k 1_{\{\frac{p}{\|y-x_i\|^\beta}\;\leq\; \lambda^{\kappa}t_i\}}\Big)
F(dp)\, m(dy)\Big\}.$$ As earlier (see the display preceding ) we write $$\begin{aligned}
\Big(1-\Pi_{i=1}^k 1_{\{\frac{p}{\|y-x_i\|^\beta}\;\leq\; \lambda^{\kappa}t_i\}}\Big)
&=\;\sum_{i=1}^k \Pi_{l=1}^{i-1}
1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1-1_{\{\frac{p}{\|y-x_i\|^\beta}\;\leq\; \lambda^{\kappa}t_i\}}\Big)
\\
& =\; \sum_{i=1}^k \Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\;
\lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big).\end{aligned}$$ We notice that to prove the theorem we only need to prove the convergence of the exponent in . Thus we write $$\begin{aligned}
\label{32}
&\quad \lambda\int_{{\mathbb{R}}^d}\int_0^\infty\Big(1-\Pi_{i=1}^k 1_{\{\frac{p}{\|y-x_i\|^\beta}\;\leq\; \lambda^{\kappa}t_i\}}\Big)
F(dp)\, m(dy) \nonumber
\\
&=\;\sum_{i=1}^k \lambda\int_{{\mathbb{R}}^d}\int_0^\infty\Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}
\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy).\end{aligned}$$ A simple change of variable in shows that $$\label{33}
\lambda\int_{{\mathbb{R}}^d}\int_0^\infty\Big(1-\Pi_{i=1}^k 1_{\{\frac{p}{\|y-x_i\|^\beta}\;\leq\; \lambda^{\kappa}t_i\}}\Big)
F(dp)\, m(dy)
\;\leq\; \sum_{i=1}^k\gamma(d, \beta) t_i^{-d/\beta}.$$ Now recall the sets $B_i$ from the proof of Theorem \[theo-1\]. Then for any positive $M$ we have, for every $i$, $$\begin{aligned}
&\liminf_{\lambda\to\infty}\; \lambda\;\int_{{\mathbb{R}}^d}\int_0^\infty \Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy)
\\
&\geq\; \liminf_{\lambda\to\infty}\; \lambda\; \int_{B_i^c}\int_0^M \Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy)
\\
&=\; \liminf_{\lambda\to\infty}\; \lambda\; \int_{B_i^c}\int_0^M
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy),\end{aligned}$$ where we used the fact that $\sup_{l\leq i-1}\sup_{y\in B_i^c}\frac{1}{\|y-x_l\|^\beta}\;<\infty$. Since $\text{dist}(x_i, B_i)\geq\frac{1}{2}\min_{i\neq j}\|x_i-x_j\|$ we get $$\begin{aligned}
&\quad \liminf_{\lambda\to\infty} \;\lambda\;\int_{{\mathbb{R}}^d}\int_0^\infty \Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy)
\\
&\geq\; \liminf_{\lambda\to\infty} \;\lambda \;\int_{{\mathbb{R}}^d}\int_0^M
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy),
\\
&=\;\frac{\omega(d)}{\beta}t^{-d/\beta}_i\int_0^\infty {\mathbb{P}}\big(s<{\mathbf{p}}\leq M\big)s^{-1+d/\beta}ds.\end{aligned}$$ Now let $M\to\infty$ to obtain $$\begin{aligned}
\label{34}
&\quad \liminf_{\lambda\to\infty} \lambda\int_{{\mathbb{R}}^d}\int_0^\infty \Pi_{l=1}^{i-1}1_{\{\frac{p}{\|y-x_l\|^\beta}\;\leq\; \lambda^{\kappa}t_l\}}
\Big(1_{\{\frac{p}{\|y-x_i\|^\beta}\;>\; \lambda^{\kappa}t_i\}}\Big)F(dp)\, m(dy)\nonumber
\\
&\geq\; \frac{\omega(d)}{\beta}t^{-d/\beta}_i\int_0^\infty {\mathbb{P}}({\mathbf{p}}\;>\;s)s^{-1+d/\beta}ds=\gamma(d,\beta)t^{-d/\beta}_i.\end{aligned}$$ The proof is completed by combining , and . $\Box$
\[rem3.2\] The result of Theorem \[theo-ex\] extend to extremal random fields defined using any positive $f$ satisfying the conditions in .
Applications to Stochastic Wireless Networks {#secapp}
============================================
Scaling of the Shannon Rate with Distance {#ssec:distan}
-----------------------------------------
In this section we give a first application of the above results to SIR stochastic models [@Baccelli-Bartek; @haenggi]. In the rest of this article we will consider $d=2$. We start by defining the Signal to Interference Ratio (SIR), which finds its root in Shannon’s Channel Coding Theorem [@CoT91]. Let $\Phi$ be a given PPP. The support of $\Phi$ represents the network nodes on the plane. We consider two fixed additional points in ${\mathbb{R}}^2$: $0$ and $e_\lambda=(\lambda, 0)$. Let $\{F_{0\lambda}, F_{i\lambda} ; i\geq 1\}$ be a collection of positive i.i.d. random variables. The variable $F_{i\lambda}$ represents a random perturbation called the *fading* from $x_i\in\Phi$ to $e_\lambda$. $F_{0\lambda}$ denotes the fading between $0$ and $e_\lambda$. We assume that the fading is independent of the PPP. Let $F(dp)$ be the common distribution of the fading variables. Let $\ell:[0, \infty)\to[0, \infty]$ be the *path-loss function* given by $$\ell(r)=r^\beta, \quad \text{for some}\ \beta>2.$$ We also assume that $$\label{10}
\int_{{\mathbb{R}}_+}p\, F(dp)<\infty.$$ Node $e_\lambda$ receives the signal from $x_i\in\Phi$ with power $\frac{F_{i\lambda}}{\ell(\|e_\lambda-x_i\|)}$. Hence the total power received at $e_\lambda$ from $\Phi$ is given by the shot-noise $$\label{11}
{\mathcal{I}}(e_\lambda)=\sum_{x_i\in\Phi} L(e_\lambda, x_i, F_{i\lambda}), \quad \text{with}\quad L(e_\lambda, x, p)=\frac{p}{\ell(\|e_\lambda-x\|)}.$$
Now we define the SIR between $0$ and $e_\lambda$ as $${\mathrm{SIR}}_{0\lambda}=\cfrac{F_{0\lambda}/\ell(\|e_\lambda\|)}{{\mathcal{I}}(e_\lambda)}.$$ It follows from Campbell’s theorem [@SKM] that ${\mathcal{I}}(e_\lambda)$ is finite with probability 1 so that ${\mathrm{SIR}}_{0\lambda}$ is positive with probability $1$. Let $c>0$. From Shannon’s Channel Coding Theorem, when treating interference as Gaussian noise, the transmission from 0 to $e_\lambda$ is possible at rate $\frac{1}{2} \ln (1+c)$ if ${\mathrm{SIR}}_{0\lambda}>c$ and impossible if ${\mathrm{SIR}}_{0\lambda}<c$. This Shannon rate explains the practical importance of the following scaling result:
\[theo-2\] Let $\Phi$ be a stationary PPP of intensity $1$. Then under Condition we have $$\liminf_{c\to 0+}\;\liminf_{\lambda\to\infty}
\;{\mathbb{P}}\Big({\mathrm{SIR}}_{0\lambda}\geq \frac{c}{\lambda^{\beta}}\Big)
\;=\;1.$$
[**Proof:**]{} The proof is based on the following observation: if $\Phi_1$ is a stationary PPP of intensity $1$, then the PPP obtained by the mapping $x\mapsto \frac{x}{\lambda}$ has intensity $\lambda^2$. Let $\Phi_{\lambda^2}$ be the stationary PPP of intensity $\lambda^2$. Therefore we see from that ${\mathcal{I}}(e_n)$ has the same distribution as $\frac{1}{\lambda^\beta}{\mathcal{I}}_{\lambda^2}(e_1)$, where $${\mathcal{I}}_{\lambda^2}(e_1)=\sum_{x_i\in\Phi_{\lambda^2}} L(e_1, x_i, F_{i\lambda}).$$ Let $\xi$ be an $\alpha$-stable random variable such that for any $t\geq 0$, $${\mathbb{E}}[e^{-t \xi}]=\exp(-t^{\alpha} C(\beta)),$$ where $C(\beta)=C(2, \beta)\int_{{\mathbb{R}}_+}p^{\alpha}\, F(dp), \alpha=2/\beta,$ and $C(\beta)$ is given by Lemma \[lem1\]. By Lemma \[lem1\] we know that $ \frac{1}{\lambda^\beta}{\mathcal{I}}_{\lambda^2}(e_1)\to\xi$ in the sense of convergence in distribution. Therefore for any $c>0$, we obtain $$\begin{aligned}
{\mathbb{P}}\Big({\mathrm{SIR}}_{0\lambda}\geq \frac{c}{\lambda^{\beta}}\Big)&=\;{\mathbb{P}}\Big(\frac{F_{0\lambda}/\ell(\|e_1\|)}{{\mathcal{I}}_{\lambda^2}(e_1)}\geq
\frac{c}{\lambda^{\beta}}\Big)
\\
&=\;{\mathbb{P}}\Big(\frac{F_{0\lambda}}{c\ell(\|e_1\|)}\geq \lambda^{-\beta}{\mathcal{I}}_{n^2}(e_1)\Big)
\\
&\geq\; {\mathbb{P}}\Big(\frac{F_{0\lambda}}{c\ell(\|e_1\|)}\geq \lambda^{-\beta}{\mathcal{I}}_{\lambda^2}(e_1), F_{0\lambda}>\delta\Big)
\\
&\geq\; {\mathbb{P}}\Big(\frac{\delta}{c\ell(\|e_1\|)}\geq \lambda^{-\beta}{\mathcal{I}}_{\lambda^2}(e_1), F_{0\lambda}>\delta\Big)\end{aligned}$$ for any positive constant $\delta$. Therefore using the independence we have $$\begin{aligned}
{\mathbb{P}}\Big({\mathrm{SIR}}_{0\lambda}\geq \frac{c}{\lambda^{\beta}}\Big) &\geq\; {\mathbb{P}}\Big(\frac{\delta}{c\ell(\|e_1\|)}\geq \lambda^{-\beta}{\mathcal{I}}_{\lambda^2}(e_1)\big){\mathbb{P}}\big( F_{0\lambda}>\delta\Big)
\\
& =\;{\mathbb{P}}\Big(\frac{\delta}{c\ell(\|e_1\|)}\geq \lambda^{-\beta}{\mathcal{I}}_{\lambda^2}(e_1)\Big)\int_{{\mathbb{R}}_+}1_{\{p>\delta\}}F(dp).\end{aligned}$$ Therefore letting $\lambda\to\infty$, we get $$\liminf_{\lambda\to\infty}\; {\mathbb{P}}\Big({\mathrm{SIR}}_{0\lambda}\geq\;
\frac{c}{\lambda^{\beta}}\Big)\;\geq\;
{\mathbb{P}}\Big(\frac{\delta}{c\ell(\|e_1\|)}> \xi\Big)\int_{{\mathbb{R}}_+}1_{\{p>\delta\}}F(dp),$$ where $\xi$ is an $\alpha$-stable random variable.
Now first let $c\to 0+$ and then $\delta\to 0$ to obtain the result. $\Box$
Another way of rephrasing Theorem \[theo-2\] is that for a Poisson field of interferers with density $\lambda$, the scale at which the ${\mathrm{SIR}}$ decreases for a link of length 1 is $\lambda^{-\kappa}, \, \kappa=\frac{\beta}{2},$ and the Shannon rate on that link scales like $\lambda^{-\kappa}$.
SINR Percolation in Ultra Dense Networks
----------------------------------------
Signal to Interference and Noise Ratio (SINR) Percolation received a lot of attention (see [@Fran-Mees; @Baccelli-Bartek] and the references therein) but was only studied in the case of a bounded response function to the best of our knowledge.
The aim of this subsection is to discuss SINR Percolation for the power law response functions considered here in terms of scaling laws for ultra-dense networks.
Let $v_1, \ldots, v_k$ be $k$-given distinct points on ${\mathbb{R}}^2$. Here we are interested in finding the scale of SINR at these points when the network density tends to infinity. Let $\Phi=\Phi_\lambda$ be a stationary PPP of intensity $\lambda$. Let $(W_1, \ldots, W_k)$ be $k$ non-negative i.i.d. random variables that are independent of $\Phi_\lambda$. We may think of $W_i$ as the power of *thermal noise* at $v_i$. Let $\{F_{l(l+1)}, F_{ij}; i, j, n\geq 1\}$ be a family of non-negative distributions with common cumulative distribution $F(dp)$ where $F(dp)$ satisfies Condition . As earlier we may think of $F_{l({l+1})}$ as the fading variables between nodes at $v_l$ and $v_{l+1}$. Let $F_{ij}$ denote the fading variable between $x_i\in\Phi_\lambda$ and $v_j$. As earlier we assume that the fading is independent of $\Phi_\lambda$ and $\{W_i\}$. We define the SINR between $v_l$ and $v_{l+1}$ as follows $$S_{l(l+1)}(\lambda)={\mathrm{SINR}}_{l(l+1)}=\frac{F_{l(l+1)}/\ell(\|v_l-v_{l+1}\|)}{W_{l+1}+{\mathcal{I}}_\lambda(v_{l+1})},$$ where $${\mathcal{I}}_\lambda(v_{l+1})=\sum_{x_i\in\Phi_\lambda} L(v_{l+1}, x_i, F_{i(l+1)}) \quad \text{and}\quad L(v, x, p)=\frac{p}{\ell(\|v-x\|)}.$$
\[theo-3\] Let $c>0$ be given. Let $\gamma=\min_{1\leq i\leq k-1}\frac{1}{\ell(\|v_i-v_{i+1}\|)}$. Then $$\begin{aligned}
&\liminf_{\lambda\to\infty}{\mathbb{P}}(S_{12}(\lambda)\geq c\lambda^{-\kappa},\ldots,
S_{(k-1)k}(\lambda)\geq c\lambda^{-\kappa})
\\
&\geq
\Big[{\mathbb{P}}(\xi<\frac{\gamma}{2\sqrt{c}})\big(\int_{\{p\geq\sqrt{c}\}}F(dp)\big){\mathbb{P}}(W_2\leq\frac{1}{\sqrt{c}})\Big]^{k-1},\end{aligned}$$ where $\xi$ is a $\alpha$-stable variable such that ${\mathbb{E}}[e^{-t\xi}]=\exp(-C(\beta) t^{\alpha})$ for $t\geq 0$ and $C(\beta)=C(2, \beta)$ is given by Theorem \[theo-2\]. In particular, $$\lim_{c\to 0+}\liminf_{\lambda\to\infty}{\mathbb{P}}(S_{12}(\lambda)\geq c \lambda^{-\kappa},\ldots, S_{(k-1)k}(\lambda)\geq c\lambda^{-\kappa})=1.$$
[**Proof:**]{} Let $\gamma=\min_{1\leq i\leq k-1}\frac{1}{\ell(\|v_i-v_{i+1}\|)}$. Then for any $c>0$ $$\begin{aligned}
& {\mathbb{P}}\Big(S_{12}(\lambda)\geq c\lambda^{-\alpha},\ldots, S_{(k-1)k}(\lambda)
\geq c\lambda^{-\alpha}\Big)
\\
& \geq {\mathbb{P}}\Big(\lambda^{-\alpha}{\mathcal{I}}_\lambda(v_2)\leq \big(\frac{\gamma}{c}F_{12}-\lambda^{-\alpha}W_2\big), \ldots,
\lambda^{-\alpha}{\mathcal{I}}_\lambda(v_k)\leq \big(\frac{\gamma}{c}F_{(k-1)k}
-\lambda^{-\alpha}W_k\big)\Big)
\\
&\geq {\mathbb{P}}\Big(\lambda^{-\alpha}{\mathcal{I}}_\lambda(v_2)\leq \big(\frac{\gamma}{c}F_{12}-\lambda^{-\alpha}W_2\big), \ldots,
\lambda^{-\alpha}{\mathcal{I}}_\lambda(v_k)\leq \big(\frac{\gamma}{c}F_{(k-1)k}-
\lambda^{-\alpha}W_k\big), A(\delta)\Big)\end{aligned}$$ for any positive $\delta$ where $A(\delta)=\{ F_{i(i+1)}\geq \delta, \forall\ i=1,\ldots, k-1\}\cap\{W_{i}<1/\delta, \forall\ i=2, \ldots , k\}$. Therefore $$\begin{aligned}
& {\mathbb{P}}\big(S_{12}(\lambda)\geq c\lambda^{-\kappa},\ldots, S_{(k-1)k}(\lambda)\geq
c\lambda^{-\kappa}\big)
\\
&\geq {\mathbb{P}}\Big(\lambda^{-\kappa}{\mathcal{I}}_\lambda(v_2)\leq \big(\frac{\gamma}{c}\delta
-\lambda^{-\kappa}\frac{1}{\delta}), \ldots,
\lambda^{-\kappa}{\mathcal{I}}_\lambda(v_k)\leq \big(\frac{\gamma}{c}\delta-\lambda^{-\kappa}\frac{1}{\delta}\big), A(\delta)\Big)
\\
&\geq {\mathbb{P}}\Big(\lambda^{-\kappa}{\mathcal{I}}_\lambda(v_2)\leq \frac{\gamma}{2c}\delta, \ldots,
\lambda^{-\kappa}{\mathcal{I}}_\lambda(v_k)\leq \frac{\gamma}{2c}\delta\Big){\mathbb{P}}\Big(A(\delta)\Big),\end{aligned}$$ for all large $\lambda$, where we used independence in the last line. Therefore, applying Theorem \[theo-1\], we get $$\liminf_{\lambda\to\infty}{\mathbb{P}}\big(S_{12}(\lambda)\geq c\lambda^{-\kappa},\ldots, S_{(k-1)k}(\lambda)\geq c\lambda^{-\kappa}\big)
\geq {\mathbb{P}}(\xi_1<\frac{\gamma}{2c}\delta, \ldots, \xi_{k-1}<\frac{\gamma}{2c}\delta){\mathbb{P}}\Big(A(\delta)\Big),$$ where $(\xi_i, \ldots, \xi_{k-1})$ is an i.i.d. sequence of $\alpha$-stable random variables. The proof follows by choosing $\delta=\sqrt{c}$ and using the independence property of $\{F_{i(i+1)}, i=1, \ldots, k-1\}$ and $\{W_i, i=2, \ldots, k\}$.$\Box$
This is a much stronger result which leverages the independence of the limiting fields to assess the joint scale of decrease of the SINR threshold or the joint Shannon rate on a [*collection of links*]{} as above.
Finally we use our scaling limit result to produce results on percolation. By ${\mathbb{Z}}^2$ we denote the integer lattice on ${\mathbb{R}}^2$. The nodes of ${\mathbb{Z}}^2$ will be considered as sites. By $(z_1, z_2)$ we denote the coordinate of points in ${\mathbb{Z}}^2$. Let $\{g_{z\bar{z}}\; :\; z\in{\mathbb{Z}}^2, \bar{z}\in{\mathbb{Z}}\}$ be a collection of positive i.i.d. random variables where $g_{z\bar{z}}$ denotes the fading between the sites $z$ and $\bar{z}$. Now we consider an independently marked PPP $\tilde\Phi_\lambda$ on ${\mathbb{R}}^2$ of intensity $\lambda$. We assume that the marks are distributed according to distribution $F(dp)$ and $$\int_0^\infty p F(dp)<\infty.$$ Let $f$ be a measurable function with compact support and satisfying the conditions in . We define the shot-noise at location $z$ by $${\mathcal{I}}_\lambda(f;z)=\sum_{(x_i, {\mathbf{p}}_i)\in\tilde\Phi_\lambda} {\mathbf{p}}_i\, f(\|x_i-z\|).$$ Note that $f$ is not bounded on $[0, \infty)$. The SINR between $z$ and $\bar{z}$ is defined by $$S_\lambda(z, \bar{z}):={\mathrm{SINR}}_{z\bar{z}}=\frac{g_{z\bar{z}}/f(1)}{W_{\bar{z}}
+{\mathcal{I}}_\lambda(f;\bar{z})},$$ where $\{W_{z}\; : z\in{\mathbb{Z}}^2\}$ is an independent sequence of non-negative i.i.d. random variables representing thermal noise. We also assume that $\{g_{z\bar{z}}\; :\; z\in{\mathbb{Z}}^2, \bar{z}\in{\mathbb{Z}}^2\}$ is independent of $\tilde\Phi_\lambda$ for all $\lambda$. Let $\tau=\tau_\lambda>0$ denote the SINR threshold. As explained above, only link that experience a SINR above this threshold are operational. We hence say the site $z$ is connected to the site $\bar{z}$ if $S_n(z, \bar{z})\geq \tau_\lambda$. We construct a random graph on ${\mathbb{Z}}^2$ as follows: we put an edge between $z$ and $\bar{z}$ if and only if $z, \bar{z}$ are connected to each other and $z, \bar{z}$ are neighbors to each other. Note that this forms a random graph of ${\mathbb{Z}}^2$. We denote by $\mathcal{C}^\lambda_0$ the largest connected component in the random graph containing $0$. Let also $|\mathcal{C}^\lambda_0|$ denote the number of sites in $\mathcal{C}^\lambda_0$. In the next theorem we show that, for large $\lambda$ and suitably chosen $\tau$, the graph percolates.
\[theo-4\] Let $f$ be a function with compact support satisfying the conditions in . Let $\tau_\lambda=\frac{c}{\lambda^\kappa}$ for $\kappa=\beta/2$. Then there exist positive constants $\lambda_0, c_0,$ such that, for any $\lambda\geq \lambda_0, c\in (0, c_0]$, we have ${\mathbb{P}}(|\mathcal{C}^\lambda_0|=\infty)>0$.
By the above theorem we see that if we choose $c$ small enough, then in a sufficiently dense network the site $0$ can send/receive *information* to/from infinitely many sites with positive probability. One should also compare this result with [@Fran-Mees Theorem 2.7.4] where percolation is established for a bounded $f$.
[**Proof:**]{} We say a site $z\in{\mathbb{Z}}^2$ is occupied if and only if $$\frac{\min(g_{z^{-1}z}, g_{z^{+1}z}, g_{z_{-1}z}, g_{z_{+1}z})/f(1)}
{W_{z}+{\mathcal{I}}_n(f;z)}\geq \frac{c}{\lambda^\kappa},$$ where $z^{\pm}=(z_1, z_2)^{\pm}=(z_1\pm 1, z_2)$ and $z_{\pm}=(z_1, z_2)_{\pm}=(z_1, z_2\pm 1)$. Define $$p\;=\;p_\lambda\;=\;{\mathbb{P}}\Big(\frac{\min(g_{z^{-1}z}, g_{z^{+1}z}, g_{z_{-1}z}, g_{z_{+1}z})/f(1)}
{W_{z}+{\mathcal{I}}_\lambda(f;z)}\geq \frac{c}{\lambda^\kappa}\Big).$$ Therefore every site is occupied with probability $p_\lambda$. We put an edge between two neighboring sites if and only if the sites are occupied. It is easy to see that the probability that a site is occupied is the same for all sites. It is also important to see that the random graph is stationary. But two neighboring sites might not be independent due to the dependencies through the shot-noise field. Let the support of $f$ lie in a box of size $m$. By box of size $m$ we mean collection of all vertices with graph distance less than $m$ from $0$. By $D_m(z)$ we denote the box of size $m$ with center at $z\in{\mathbb{Z}}^2$. Therefore we see that $${\mathcal{I}}_\lambda(f;z)=\sum_{(x_i, p_i)\in\tilde\Phi, x_i\in D_m(z)}p_if(|x_i-z|).$$ Therefore, if $z$ and $\bar{z}$ are such that $D_m(z)\cap D_m(\bar{z})=\emptyset$, we have $({\mathcal{I}}_\lambda(f;z), {\mathcal{I}}_\lambda(f;\bar{z}))$ i.i.d. since the underlying point process is Poisson. Therefore the states of sites that are at distance more than $(2m+2)$ are independent. Now from [@Fran-Mees Theorem 2.3.1] we have a constant $p=p(m)<1$ such that if $p_\lambda>p(m)$, then ${\mathbb{P}}(|\bar{\mathcal{C}}^\lambda_0|=\infty)>0$, where $\bar{\mathcal{C}}^\lambda_0$ denotes the maximal connected component in our new random graph containing $0$. We note that $\bar{\mathcal{C}}_0^\lambda\subset\mathcal{C}^\lambda_0$ and therefore $${\mathbb{P}}\Big(|\bar{\mathcal{C}}^\lambda_0|=\infty\Big)>0\;\Rightarrow\;
{\mathbb{P}}\Big(|\mathcal{C}^\lambda_0|=\infty\Big)>0.$$ Thus to complete the proof we only need to find $n$ and $c$ so that that $p_\lambda>p(d)$. We assume that $f(1)>0$ otherwise there is nothing to prove. We note that for $c>0$, (from calculation as in Theorem \[theo-3\]) $$\begin{aligned}
p_\lambda &=\;{\mathbb{P}}\Big(\frac{\min(g_{0^{-1}0}, g_{0^{+1}0}, g_{0_{-1}0}, g_{0_{+1}0})/f(1)}
{W_{0}+{\mathcal{I}}_\lambda(f;0)}\geq \frac{c}{\lambda^\kappa}\Big)
\\
& \geq\; {\mathbb{P}}(\hat{{\mathcal{I}}}_\lambda(f;0)<\frac{1}{2f(1)\sqrt{c}}){\mathbb{P}}(\min(g_{0^{-1}0}, g_{0^{+1}0}, g_{0_{-1}0}, g_{0_{+1}0})\geq \sqrt{c})
{\mathbb{P}}(W_0\leq \frac{\lambda^\kappa}{2f(1)\sqrt{c}}).\end{aligned}$$ Therefore, using Lemma \[lem1\], we can find $\lambda_0, c_0$ such that, for $\lambda\geq \lambda_0, c\in (0, c_0]$, we have $p_\lambda>p(m)$.$\Box$
Scaling of the Delay with Distance and an Open Question
-------------------------------------------------------
Another way of looking at the result of Subsection \[ssec:distan\] is in terms of *delay*. For an $o(1)$ SIR, the number of bits transmitted in an $o(1)$ interval is $o(1)$. We will then say that an $o(1)$ SIR has an $o(1)$ delay. Theorem \[theo-2\] tells us that for a link of length $\lambda$ in a Poisson field of interferers with density 1, the Shannon rate tends to 0 like $\lambda^{-\beta}$. Hence, this theorem can be rephrased by saying that the delay $D_\lambda$ to transmit $o(1)$ bits over distance $\lambda$ in one hop (in the above scheme, one sends these $o(1)$ bits directly from 0 to $\lambda$ in one hop) scales like $\lambda^{\beta}$ as $\lambda\to\infty$. The general case (including the possibility of multiple hops) was studied in [@Baccelli-Bartek-Omid] where it was shown that for all possible schemes within this framework (single or multi-hop), $\frac{{\mathbb{E}}[D_\lambda]}{\lambda}\to\infty$ as $\lambda\to\infty$. This raises the following question: does there exists a scheme that allows one to transmit $o(1)$ bits with a delay $D_\lambda$ such that $\frac{D_\lambda}{\lambda^\gamma}$ has a non-degenerate limit for some $\gamma\in(1, \beta)$?
**Acknowledgment:** Authors are grateful to the referees for their fruitful comments and pointing out some important references from literature.
[99]{} Baccelli, F. and B[ł]{}aszczyszyn, B.: [*Stochastic geometry and wireless networks*]{}, Volume I-II, Foundations and Trends in Networking. New York: Now, 2009.
Baccelli, F.; B[ł]{}aszczyszyn, B. and Haji-Mirsadeghi, M. O.: Optimal paths on the space-time SINR random graph. Adv. in Appl. Probab. 43, no. 1, 131–150, 2011.
Biermé, H.; Estrade, A. and Kaj, I.: Self-similar random fields and rescaled random balls models, J. Theoret. Probab. 23, no. 4, 1110–1141, 2010.
Breton, J.-C. and Dombry, C.: Functional macroscopic behavior of weighted random ball model, ALEA Lat. Am. J. Probab. Math. Stat. 8, 177–196, 2011.
Cover, T. M. and Thomas, J.A.: [*Elements of Information Theory*]{}, J. Wiley & Sons, New York, 1991.
Daley, D. J. and Vere-Jones, D.: [*An introduction to the theory of point processes*]{}. Vol. I. Elementary theory and methods. Second edition. Probability and its Applications (New York). Springer-Verlag, New York, 2003.
Dombry, C.: Extremal shot noises, heavy tails and max-stable random fields. Extremes 15, no. 2, 129–158, 2012.
Doney, R. A. and O’Brien, George L. Loud shot noise. Ann. Appl. Probab. 1, no. 1, 88–103, 1991.
Embrechts, P.; Klüppelberg, C. and Mikosch, T.: [*Modelling extremal events. For insurance and finance*]{}. Applications of Mathematics (New York), 33. Springer-Verlag, Berlin, 1997.
Franceschetti, M. and Meester, R.: [*Random networks for communication*]{}. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2007.
Haenggi, M.: [*Stochastic geometry for wireless networks*]{}, Cambridge University Press, 2013.
Heinrich, L.; Molchanov, I. S.: Some limit theorems for extremal and union shot-noise processes. Math. Nachr. 168, 139–159, 1994.
Hsing, Tailen and Teugels, J. L. Extremal properties of shot noise processes. Adv. in Appl. Probab. 21, no. 3, 513–525, 1989.
Kaj, I.; Leskelä, L.; Norros, I. and Schmidt, V.: Scaling limits for random fields with long-range dependence, Ann. Probab. 35 , no. 2, 528–550, 2007.
Klüppelberg, Claudia and Kühn, Christoph: Fractional Brownian motion as a weak limit of Poisson shot noise processes with applications to finance. Stochastic Process. Appl. 113, no. 2, 333–351, 2004.
Klüppelberg, Claudia; Mikosch, Thomas; and Schärf, Anette: Regular variation in the mean and stable limits for Poisson shot noise. Bernoulli 9 , no. 3, 467–496, 2003.
Klüppelberg, Claudia and Mikosch Thomas: Explosive Poisson shot noise processes with applications to risk reserves. Bernoulli 1, no. 1-2, 125–147, 1995.
Lebedev, A. V.: Extremes for shot-noise fields in the case of regularly varying tails. Theory Probab. Appl. 47, no. 4, 702–706, 2003.
Stoyan, D.; Kendall, W.; and Mecke, J.: [*Stochastic Geometry and its Applications*]{}, Wiley, Chichester, 1995.
[^1]: This work was supported by an award from Simons Foundation ([\#]{}197982) to the University of Texas at Austin.
[^2]: Here it is assumed that there is threshold on the SIR of a link above (resp. below) which communication is assumed successful (resp. unsuccessful).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Yun Wang[^1], & Peter M. Garnavich'
title: |
Measuring Time-Dependence of Dark Energy Density\
from Type Ia Supernova Data
---
\#1\#2[3.6pt]{}
astro-ph/0101040\
ApJ, 552, May 10 (2001)\
[*submitted on Aug.2, 2000*]{}\
**Abstract**
> Observations of high redshift supernovae imply an accelerating Universe which can only be explained by an unusual energy component such as vacuum energy or quintessence. To assess the ability of current and future supernova data to constrain the properties of the dark energy, we allow its density to have arbitrary time-dependence, $\rho_X(z)$. This leads to an equation of state for the dark energy, $w_X(z)=p_X(z)/\rho_X(z)$, which is a free function of redshift $z$. We find that current type Ia supernova (SNe Ia) data are consistent with a cosmological constant, with large uncertainties at $z\ga 0.5$. We show that $\rho_X(z)/\rho_X(z=0)$ can be measured reasonably well to about $z=1.5$ using type Ia supernova data from realistic future SN Ia pencil beam surveys, provided that the weak energy condition (energy density of matter is nonnegative for any observer) is imposed.
>
> While it is only possible to differentiate between different models (say, quintessence and k-essence) at $z \la 1.5$ using realistic data, the correct trend in the time-dependence of the dark energy density can be clearly detected out to $z=2$, even in the presence of plausible systematic effects. This would allow us to determine whether the dark energy is a cosmological constant, or some exotic form of energy with a time-dependent density.
Introduction
============
Most of the present energy content of our universe is unknown ([@neta99]). Distance-redshift relations derived from cosmological standard candles at redshifts between zero and a few are the most sensitive probes of the equation of state of the universe, which allows us to constrain the energy content of the universe.
Type Ia supernovae (SNe Ia) are our best candidates for cosmological standard candles. They can be calibrated to have small dispersion in their intrinsic luminosities ([@Phillips93; @Riess95]). The data from two independent observational teams, the High-$z$ SN Search (Schmidt et al.) and the Supernova Cosmology Project (Perlmutter et al.), seem to suggest that our universe has a significant vacuum energy content ([@Garna98a; @Riess98; @Perl99]).
While a cosmological constant term (vacuum energy) in the Einstein equations provides the simplest explanation of the current SN Ia data, other forms of energy (quintessence, dark energy, etc) have also been studied (e.g. [@White98; @Garna98b; @Stein99; @Efsta99; @Ratra00; @Josh00]) and are consistent with current data as well. Since cosmology has matured into a phenomenological science at the turn of the new millennium, observational data will dominate aesthetics in the selection of cosmological models.
So far, most cosmologists have assumed time-independent equations of state, i.e., power-law dark energy density (see §2), in the context of constraining the energy content of the universe. In this paper, we allow the dark energy density of the universe to be a free function of redshift, i.e., arbitrary equation of state. We derive constraints on the dark energy density from current SN data, and assess future prospects of measuring the dark energy density using simulated data from realistic future SN Ia surveys. We expect a model-independent measurement of the dark energy density as a function of redshift to be very useful in our quest of the unknown energy contents of the universe.
Parametrization of the dark energy density
==========================================
In a smooth Friedmann-Robertson-Walker (FRW) universe, the metric is given by $ds^2=dt^2-a^2(t)[dr^2/(1-kr^2)+r^2 (d\theta^2
+\sin^2\theta \,d\phi^2)]$, where $a(t)$ is the cosmic scale factor, and $k$ is the global curvature parameter. The cosmological redshift $z$ is given by $1+z=1/a$.
To make model-independent measurements of the equation of state, we replace the vacuum energy density $\rho_{\Lambda}$ with $\rho_X=\rho_X^0 f(z)$ in the total matter density of the universe: (z) & = & \_m\^0(1+z)\^3 +\_k\^0(1+z)\^2 +\_X\^0 f(z)\
& = & \_c\^0 , where the superscript “0” indicates present values, $f(z=0)=1$, and $\Omega_k=1-\Omega_m-\Omega_X=-k/H_0^2$. If the unknown energy is due to a cosmological constant $\Lambda$, $f(z)=1$. Clearly, the function $f(z)$ is a very good probe of the nature of the unknown energy.
The comoving distance $r$ is given by ([@Weinberg72]) \[eq:r(z)\] r(z)=cH\_0\^[-1]{} , | \_k |\^[1/2]{}, (z;\_m,\_X, f)=\_0\^zdz’ , E(z) \^[1/2]{}, where S(x)&=&(x), \_k>0\
&=&x, \_k=0\
&=&(x), \_k<0. The angular diameter distance is given by $d_A(z)=r(z)/(1+z)$, and the luminosity distance is given by $d_L(z)=(1+z)^2 d_A(z)$.
Einstein’s equations in a FRW metric, together with the first law of thermodynamics give us \[eq:energycons\] (1+z) = 3(+p). For unknown energy $\rho_X(z)$, we find \_X(z) &=& \_X\^0 f(z)\
p\_X(z) &=& \_X\^0 . Now we can write the equation of state of the unknown energy as w(z) =(1+z) -1. A constant equation of state corresponds to $f(z) \propto (1+z)^{\alpha}$, where $\alpha$ is a constant. The values $\alpha=0$, $\alpha=3$, and $\alpha=4$ correspond to a cosmological constant, matter, and radiation respectively.
To obtain accelerated expansion, we need $\rho+3p<0$. Since $\rho_X+3p_X=\rho_X^0\,
\left[ (1+z) f'-2f\right]$, this implies $\alpha<2$ for $f(z) \propto (1+z)^{\alpha}$.
The weak energy condition states that for all physically reasonable classical matter, the energy density of matter as measured by any observer is nonnegative ([@Wald84]). For a perfect fluid, the weak energy condition will be satisfied if and only if +p 0. This leads to (see Eq.\[\[eq:energycons\]\]) f’(z) 0. The weak energy condition imposes strong constraints on the jointly estimated cosmological parameters $\Omega_m$, $\Omega_{\Lambda}$, and the dark energy density $f(z)$.
Note that the comoving distance $r(z)$ depends on the equation of state of X through (z) \_0\^z dz’\^[-1/2]{}=H\_0 \_0\^r , Hence \[eq:derivf\] ’(z) = \^[-1/2]{}= . To measure $f(z)$ directly from data, we need to evaluate the derivative of the distance $r(z)$ with respect to redshift $z$.
To avoid taking derivatives of noisy data, we can parametrize $f(z)$ with its values at $n$ equally spaced redshifts $z_i$, and assume that $f(z)$ is given by linear interpolations at other values of $z$. We write \[eq:f(z)\] &&f(z)=( ) f\_[i-1]{}+ ( ) f\_i, z\_[i-1]{} < z z\_i,\
&& z\_0=0, z\_[n]{}=z\_[max]{}; f\_0=1, f\_n=f\_[n-1]{} where $f_i$ ($i=1,2,...,n-1)$ are independent variables to be estimated from data.
Parameter estimation
====================
The measured distance modulus for a SN Ia is \_0\^[(l)]{}= \_p\^[(l)]{}+\^[(l)]{} where $\mu_p^{(l)}$ is the theoretical prediction \[eq:mu0p\] \_p\^[(l)]{}= 5( )+25, and $\epsilon^{(l)}$ is the uncertainty in the measurement, including observational errors and intrinsic scatters in the SN Ia absolute magnitudes.
Denoting all the parameters to be fitted as [**s**]{}, we can estimate [**s**]{} using a $\chi^2$ statistic, with (Riess et al. 1998) \^2()= \_l \_l , where $\sigma_{\mu_0}$ is the estimated measurement error of the distance modulus, and $\sigma_{mz}$ is the dispersion in the distance modulus due to the dispersion in galaxy redshift, $\sigma_z$, due to peculiar velocities and uncertainty in the galaxy redshift (for the Perlmutter et al. data, the dispersion due to peculiar velocities is included in $\sigma_{m_B^{eff}}$, i.e., $\sigma_{\mu_0}$). Since \[eq:sigmamz\] \_[mz]{}= ( ) \_z, $\sigma_{mz}$ depends on the parameters [**s**]{}. The probability density function (PDF) for the parameters [**s**]{} is p() ( - ). The normalized PDF is obtained by dividing the above expression by its sum over all possible values of the parameters [**s**]{}.
In order to impose the weak energy condition $f'(z) \ge 0$, we compute the PDFs on a $N$-dimensional grid for $N$ parameters. The PDF of a given parameter $s_i$ is obtained by integrating over all possible values of the other $N-1$ parameters. To reduce the computation time, we can integrate over the Hubble constant $H_0$ analytically, and define a modified $\chi^2$ statistic, with \[eq:chi2mod\] \^2 \_\*\^2 - ( C\_1+ 10 ), where \_\*\^2 && \_l ( \_[p]{}\^[\*(l)]{}- \_[0]{}\^[(l)]{} )\^2,\
C\_1 && \_l ( \_[p]{}\^[\*(l)]{}- \_[0]{}\^[(l)]{} ),\
C\_2 && \_l , where \_p\^\* \_p(h=h\^\*)=42.384-5h\^\*+ 5. We take $h^*=0.65$. Our results are independent of the choice of $h^*$.
After experimenting with a number of different techniques, we developed an adaptive iteration method of estimating $f(z)$ based on the requirement that $f'(z)\ge 0$ (i.e., the weak energy condition is satisfied). Starting with the initial guess of $f(z)=f(z=0)=1$ (a cosmological constant), we iteratively build up $f(z)$ as parametrized by Eq.(\[eq:f(z)\]) while minimizing $\chi^2$.
Constraints of $\rho_X(z)$ from current SNe Ia data
===================================================
Wang (2000b) has combined the data of the High-$z$ SN Search team ([@Schmidt98; @Garna98a; @Riess98]) and the Supernova Cosmology Project ([@Perl99]), yielding a total of 92 SNe Ia. Using the method described in the previous section, we estimate $\Omega_m$, $\Omega_X$, and $f(z)$ simultaneously by minimizing the modified $\chi^2$ (see Eq.(\[eq:chi2mod\])).
Fig.1 shows the dimensionless dark energy density $f(z)$ (as parametrized by Eq.(\[eq:f(z)\])) measured from this set of 92 SNe Ia. It is consistent with $f(z)=1$ (a cosmological constant), with large uncertainty beyond $z\ga 0.5$. Note that none of the error bars extend beneath $f(z)=1$, because we have imposed the weak energy condition, i.e., $f'(z) \ge 0$, which implies that $f(z) \ge f(z=0)=1$. The estimated values of $\Omega_m=.3$ (0, .9); $\Omega_X=1.7$ (.4, 2.2) are consistent with previous results ([@Garna98a; @Riess98; @Perl99; @Wang00b]). The errors are estimated from the ranges of parameters for which $\chi^2=\chi^2_{min}+1$.
Wang (2000b) found that when fit to a model with a cosmological constant as the dark energy, flux-averaging changes the best fit model to this data set of 92 SNe Ia. Without flux-averaging, the best fit model is a closed universe with $\Omega_m=0.7\pm0.4$, and a vacuum energy density fraction of $\Omega_{\Lambda}=1.2\pm0.5$, consistent with the estimated values in Fig.1. The flux-averaged data yield $\Omega_m=0.3\pm0.6$, and $\Omega_{\Lambda}=0.7\pm0.7$. This difference may have resulted from the data containing large redshift dependent uncertainties, which would have caused the results from the unbinned data to be biased. The effect of flux-averaging on the best fit model assuming an arbitrary dimensionless dark energy density $f(z)$ will be studied elsewhere.
Future cosmic microwave background (CMB) space missions MAP ([@Bennett97]) and Planck ([@DeZotti99]), together with the galaxy redshift surveys SDSS ([@Gunn99]) and 2df ([@Dalton00]), will give us exquisitely accurate measurements of the geometry of the universe and the matter density in the universe ([@Eisen99; @Mike99; @Wang99a]). SN data can provide the unique probe on the nature of dark energy by allowing us to measure how the dark energy density varies with time.
Current cosmic microwave background (CMB) anisotropy measurements seem to indicate that we live in a flat universe ([@deBernardis00; @Balbi00]). Cluster abundances strongly suggest a low matter density universe ([@neta95; @Carlberg96; @neta98]). $\Omega_m=0.3$ and $\Omega_{\Lambda}=0.7$ is the best fit model to current observational data. We will use $\Omega_m=0.3$ and $\Omega_X=0.7$ for our simulated data in the rest of this paper.
Measuring $\rho_X(z)$ from future SNe Ia data
=============================================
A large number of SNe Ia at $z\ga 1$ is critical in resolving the important systematic uncertainties of SNe Ia as cosmological standard candles, such as dust ([@Aguirre99]), gravitational lensing ([@Kantow95; @Wamb97; @Holz98; @Metca99; @Wang99b; @Barber00]), and luminosity evolution ([@Drell99; @Riess99; @Wang00b]), and in making SNe Ia useful probes of the dark energy content of the universe. The most efficient method of obtaining a large number of SNe Ia at $z>1$ is conducting a supernova pencil beam survey on a dedicated large aperture telescope with a square degree field of view ([@Wang00a]).
To study how well SN data can probe the dark energy density, let us consider two hypothetical dimensionless dark energy densities $f_q(z)$ and $f_k(z)$, given by \[eq:f(z)1\] f\_q(z)&=& , w\_q(z)= -1+0.5z\
f\_k(z) &=& , w\_k(z)= 0.3 (1+z) e\^[-z]{} -1 We have chosen $f_q(z)$ and $f_k(z)$ to represent quintessence ($dw_q/dz >0$) and k-essence ($dw_k/dz <0$) models respectively ([@Caldwell98; @Armenda00]). Note that $f_q(z)$ and $f_k(z)$ satisfy the weak energy condition $f'(z)\ge 0$; they give an accelerating universe for $z\la 1.33$ and $z\la 2$ respectively.
A feasible SN pencil beam survey (either from ground[^2] or from space), with a square degree field of view and for an effective observational period of one year, can yield almost 2000 SNe Ia out to $z=2$ ([@Wang00a]). Let us combine the data from the SN pencil beam survey with SN data at smaller redshifts, so that there are a minimum of 50 SNe Ia per 0.1 redshift interval at any redshift. This yields a total of 1966 SNe Ia for the quintessence and 1898 SNe Ia for the k-essence model, up to $z=2$ and for $\Omega_m=0.3$, $\Omega_X=0.7$. We simulate the data by placing perfect standard candles at random redshifts, with the total number per 0.1 redshift interval given as described above. Then we add intrinsic and observational dispersions which are Gaussian with zero mean and a variance of 0.20 magnitudes. A systematic shift in $\mu_0$ of $dm_{sys}\,z$ magnitudes is also added to mimic possible systematic errors as one goes to larger redshifts.
Realistic SN Ia data should contain gravitational lensing noise. Wang (2000a,b) has shown that flux averaging is important in reducing the bias due to lensing. The effect of gravitational lensing and of flux averaging in the context of a general equation of state will be investigated elsewhere.
Fig.2 illustrates how well we can recover the dark energy densities $f_q(z)$ and $f_k(z)$ in the absence of lensing noise and systematic shifts, when we apply our adaptive iteration method to 100 random data sets with a realistic dispersion of 0.2 magnitudes. To study the dependence of our results on the parametrization of $f(z)$ \[see Eq.(\[eq:f(z)\])\], we show results for (a) $n=10$, and (b) $n=6$. The thick and thin solid lines are the assumed true $f_q(z)$ and $f_k(z)$ respectively. We have assumed that we know $\Omega_m+\Omega_X=1$. The mean and 1-$\sigma$ errors of $f(z)$ and $\Omega_m$ are estimated from averaging over the 100 random samples. The error of estimated $\Omega_m$ reflects the resolution of the program. For a given data set, the recovered $f(z)$ and $\Omega_m$ are expected to fall within the errors with 68% probability. Clearly, the quintessence and k-essence models can be differentiated marginally for $z\la 1.5$. For $z \ga 1.5$, the errors increase significantly, while the estimates become more biased, making it impossible to differentiate between the two models. However, the correct trend in the time variation of the dark energy density can be clearly detected out to $z=2$.
Fig.2 shows that the parametrization of $f(z)$ with $n=10$ yields less biased estimates than $n=6$. This is as expected, since $f(z)$ is more accurately parametrized as one increases $n$. However, the errors in the estimates increase with $n$ as well. One must choose an optimal $n$ such that $f(z)$ is adequately parametrized, while the errors on the estimated amplitudes of $f(z)$ are not too large to be useful. We’ve experimented with $n>10$ parametrizations of $f(z)$, and found that $n=10$ is a good choice.
The biased estimates of the dark energy density $f_q(z)$ and $f_k(z)$ in Fig.2 is mainly due to the bias in the estimate of $\Omega_m$. Fig.3 shows $f_q(z)$ and $f_k(z)$ estimated assuming that we know $\Omega_m=0.3$, with the same line types as in Fig.2. Even with this ideal assumption, it is only possible to marginally differentiate between the two models.
Fig.4 shows the effect of adding a systematic shift of $dm_{sys}\,z$ to 100 random data sets with a realistic dispersion of 0.2 magnitudes. The line types are the same as in Fig.2. We have added a systematic shift in $\mu_0$ of (a) $0.01\,z$ magnitudes; (b) $0.05\,z$ magnitudes. Clearly, systematic shifts can significantly increase the bias in the estimate of $\Omega_m$ and the estimates of $f(z)$ for $z\ga 1.2$, while having little effect on the estimates of $f(z)$ at $z\la 1.2$. Systematic errors as a function of $z$ may arise from intrinsic properties of the supernovae varying with $z$ such as progenitor evolution or dust characteristics changing with metalicity of the Universe. Accuracy may also be limited by the observations themselves in the case of k-corrections or selection biases. The estimated errors from these sources in the current surveys are between 2% and 10% ([@Schmidt98]). It is perhaps not realistic to expect to measure both $\Omega_m$ and $f(z)$ to high redshift accurately from SN Ia data alone.
It is clear from Fig.4 that even in the presence of plausible systematic effects, we can expect to measure the time-variation in the dark energy density $f(z)$ with reasonable accuracy to a redshift of about 1.2. The bias and the errors in the estimated $f(z)$ increase substantially beyond $z=1.2$.
We only applied our adaptive iteration method to 100 random samples, and with a resolution of $\Delta\Omega_m=0.02$, because this method takes several hours per sample on a fast Sun work station in finding the best fit $f(z)$ and $\Omega_m$. It is work planned for the future to improve this promising method for application to much larger number of random samples, as well as adding $\Omega_X$ as an estimated parameter.
Implication for dark energy models
==================================
Recently, there has been a great deal of activity in exploring the possibilities of the existence of exotic dark energy ([@Peebles88; @Frieman95; @Caldwell98; @Sahni00]) in the universe. While the present observational data are consistent with the dark energy being a cosmological constant, they do not rule out alternatives in the form of various scalar fields.
It is important that we measure the time dependence of the dark energy density. If the dark energy density is measured to be constant in time within reasonable uncertainties, a cosmological constant should be favored, and more theoretical efforts should be directed toward the derivation of a cosmological constant from first principles. At the very least, this places strong constraint on the classes of scalar-field models for the dark energy. On the other hand, if the time-dependence of the dark energy density is established by the observational data, we would come to the exciting discovery of new physics in the universe.
We have studied two dark energy models $f_q(z)$ and $f_k(z)$ (see Eq.(\[eq:f(z)1\])), representing two general classes of dark energy models, quintessence ([@Caldwell98]) and k-essence ([@Armenda00]). This allows us to examine how well different models can be differentiated by realistic data, as well as the robust determination of the time-dependence of the dark energy density.
Realistic future SN data, as described in the previous section, has the potential of determining the time-dependence of the dark energy density. This will clearly have a dramatic impact on models of the dark energy.
Conclusions
===========
To access the prospects of measuring the time variation in the equation of state, we have developed a promising adaptive iteration method that is powerful in extracting the dark energy density $f(z)=\rho_X(z)/\rho_X(z=0)$ from realistic data. This method is based on the requirement that the weak energy condition (energy density of matter is nonnegative for any observer) is satisfied.
We have found that current type Ia supernova (SNe Ia) data are consistent with a cosmological constant, with large uncertainties at $z\ga 0.5$. We show that $\Omega_m$ (assuming a flat universe) and the dimensionless dark energy density $f(z)=\rho_X(z)/\rho_X(z=0)$ can be measured reasonably well to about $z=1.5$ using type Ia supernova data from realistic future SN Ia pencil beam surveys, provided that the weak energy condition (energy density of matter is nonnegative for any observer) is imposed. For $z \ga 1.5$, the errors increase significantly, while the estimates become more biased, making it impossible to differentiate between different models (say, quintessence and k-essence). However, the correct trend in the time-dependence of the dark energy density can be clearly detected out to $z=2$, even in the presence of plausible systematic effects. This would allow us to determine whether the dark energy is vacuum energy, or some exotic form of energy with a time-dependent density.
The simulated data we used are for a SN pencil beam survey (either from ground or from space) with a square degree field of view and for an effective observational period of one year ([@Wang00a]), combined with SN data at smaller redshifts, so that there are a minimum of 50 SNe Ia per 0.1 redshift interval at any redshift. Although the dispersion of 0.20 magnitudes (intrinsic plus observational) assumed in our simulated data is appropriate for ground-based surveys, we expect our results to apply qualitatively to space based SN pencil beam surveys as well, because our assumed dispersion of 0.20 magnitudes is dominated by intrinsic dispersion (about 0.17 magnitudes).
At the completion of this lengthy numerical study of the feasibility of measuring the time-dependence of the dark energy density from realistic SN Ia data, we became aware of the recent paper by Maor, Brustein, & Steinhardt (2000). They claimed that distance-redshift relations derived from SNe Ia and similar classical measures are poor methods for resolving the time-dependence or measuring the amplitude of the equation of state $w_X(z)=p_X(z)/\rho_X(z)$, and consequently no useful information can be obtained about the future fate of the universe. Our work has confirmed the difficulty of measuring the properties of the dark energy from realistic SN Ia data. However, we have found their assessment to be overly pessimistic. Their work indicates that it is impossible to tell whether the equation of state $w_X(z)$ varies in time (see also, [@Barger01]), but knowing that the dark energy density $\rho_X(z)$ varies in time is sufficient to rule out vacuum energy as dark energy, thus giving support to exotic dark energy models. Our work has shown that one can indeed clearly detect the time-dependence of $\rho_X(z)$ using realistic future SN Ia data.
The main problem reported by Moar et al. was the “smearing” effect of the multi-integral relation between the luminosity distance $d_L(z)$ and the equation of state $w_X(z)$. Instead of $d_L(z)$ and $w_X(z)$, our analysis uses the time derivative of the comoving distance $r'(z)$ and the dimensionless dark energy density $f(z)$. Thus our results are less affected by the smearing effect. Our method will be refined and made more efficient, and should become quite useful in analyzing future SN data.
In view of our results, it is important that reasonable yet substantial observational efforts are devoted to future SN Ia surveys, for example, a dedicated 4m telescope for a SN pencil beam survey ([@Wang00a]), combined with surveys of nearby SNe Ia. The total cost of such surveys would be modest compared to the great scientific return, the determination of the systematic uncertainties of SNe Ia as cosmological standard candles, and the measurement of the time-dependence in the dark energy density of the universe to constrain fundamental physics.
It is a pleasure for us to thank Grant Mathews for helpful discussions, and the referee for useful suggestions. PMG acknowledges support from NASA LTSA grant NAG-9364.
Aguirre, A.N. 1999, ApJ, 525, 583
Armendariz-Picon, C., Mukhanov, V., Steinhardt, P.J. 2000, Phys. Rev. Lett. 85, 4438
Bahcall, N.A. & Fan, X. 1998, ApJ 504, 1; Nat. Acad. Sci. Proc. 95, 5956
Bahcall, N.A., Lubin, L.M., and Dorman, V. 1995, ApJ 447, L81
Bahcall, N.A., Ostriker, JP, Perlmutter,S., and Steinhardt, P., 1999, Science, 284, 1481
Balbi, A., et al. 2000, ApJ, 545L, 1
Barber, A.J., Thomas, P.A., Couchman, H.M.P., Fluke, C.J., MNRAS, 319, 267
Barger, V., Marfatia, D. 2001, Phys. Lett. B498, 67
Bennett, C.L., et al., AAS, 191, 8701
Caldwell, R.R., Dave, R., Steinhardt, P.J. 1998, Phys. Rev. Lett. 80, 1582
Carlberg, R.G., Yee, H.K.C. et al. 1996, , 462, 32
Dalton, G. B.; 2dF Galaxy Redshift Survey Team 2000, AAS, 196, 5605
de Bernardis, P., et al. 2000, Nature, 404, 955
De Zotti, G., et al., in proceedings of “3K Cosmology”, AIP, vol 476, P204.
Drell, P.S.; Loredo, T.J.; & Wasserman, I. 2000, ApJ, 530, 593
Efstathiou 1999, , 310, 842
Eisenstein, D.J., Hu, W., & Tegmark, M. 1999, , 518, 2
Frieman, J.A.; Hill, C.T.; Stebbins, A.; & Waga, I. 1995, Phys. Rev. Lett. 75, 2077
Garnavich, P.M. et al. 1998, , 493, L53
Garnavich, P.M. et al. 1998, , 509, 74
Gunn, J. E. 1999, AAS, 195, 3402
Holz, D.E. 1998, ApJ, 506L, 1
Kantowski, R., Vaughan, T., & Branch, D. 1995, , 447, 35
Maor, I., Brustein, R., & Steinhardt, P.J. 2001, Phys. Rev. Lett. 86, 6
Metcalf, R.B.; Silk, J. 1999, ApJ, 519L, 1
Peebles, P. J. E.; & Ratra, B. 1988 ApJ, 325L, 17 U
Perlmutter, S., et al. 1999, , 517, 565
Phillips, M.M. 1993, , 413, L105
Podariu, S., & Ratra, B. 2000, ApJ, 532, 109
Riess, A.G., Press, W.H., & Kirshner, R.P. 1995, , 438, L17
Riess, A.G., et al 1998, AJ, 116, 1009
Riess, A.G., et al. 1999, AJ, 118, 2675
Sahni, V., & Wang, L. 2000, Phys. Rev. D., 62, 103517
Schmidt, B.P., et al. 1998, , 507, 46
Steinhardt, P.J., Wang, L., & Zlatev, I. 1999, Phys. Rev. D59, 123504
Turner, M.S., & Tyson, J.A. 1999, RvMPS, 71, 145
Waga, I., & Frieman, J.A. 2000, Phys. Rev. D 62, 043521
Wald, R.M. 1984, General Relativity, The University of Chicago Press, Chicago and London
Wambsganss, J., Cen, R., Xu, G., & Ostriker, J.P. 1997, , 475, L81
Wang, Y. 1999, , 525, 651
Wang, Y. 2000a, ApJ, 531, 676 (astro-ph/9806185)
Wang, Y. 2000b, ApJ, 536, 531
Wang, Y., Spergel, D.N., & Strauss, M.A. 1999, , 510, 20
Weinberg, S. 1972, Gravitation and Cosmology, John Wiley & Sons, New York
White, M. 1998, , 506, 495
[^1]: current address: Department of Physics & Astronomy, University of Oklahoma, Norman, OK 73019. email: [email protected]
[^2]: The SNe Ia at $z \ga 1.5$ will likely require follow up spectroscopy from space.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the critical behavior of the three-dimensional planar magnet model in which each spin is considered to have three components of which only the $x$ and $y$ components are coupled. We use a hybrid Monte Carlo algorithm in which a single-cluster update is combined with the over-relaxation and Metropolis spin re-orientation algorithm. Periodic boundary conditions were applied in all directions. We have calculated the fourth-order cumulant in finite size lattices using the single-histogram re-weighting method. Using finite-size scaling theory, we obtained the critical temperature which is very different from that of the usual $XY$ model. At the critical temperature, we calculated the susceptibility and the magnetization on lattices of size up to $42^3$. Using finite-size scaling theory we accurately determine the critical exponents of the model and find that $\nu$=0.670(7), $\gamma/\nu$=1.9696(37), and $\beta/\nu$=0.515(2). Thus, we conclude that the model belongs to the same universality class with the $XY$ model, as expected.'
address: |
Department of Physics and Center for Research and Technology,\
Florida State University, Tallahassee, Florida 32306-4350
author:
- Kwangsik Nho and Efstratios Manousakis
title: Critical behavior of the planar magnet model in three dimensions
---
Introduction {#sec0}
============
Our understanding of critical phenomena has been significantly advanced with the development of the renormalization-group (RG) theory[@Wilson]. The RG theory predicts relationships between groups of exponents and that there is a universal behavior. In a second order phase transition, the correlation length $\xi$ diverges as the critical point is approached, and so the details of the microscopic Hamiltonian are unimportant for the critical behavior. All members of a given universality class have identical critical behavior and critical exponents.
The three-dimensional classical XY model is relevant to the critical behavior of many physical systems, such as superfluid $^{4}He$, magnetic materials and the high-Tc superconductors. In the pseudospin notation, this model is defined by the Hamiltonian $$H=-J\sum_{<ij>}(S_{i}^{x}S_{j}^{x}+S_{i}^{y}S_{j}^{y}),
\label{XY}$$ where the summation is over all nearest neighbor pairs of sites $i$ and $j$ on a simple cubic lattice. In this model one considers that the spin has two components, $\vec S_{i}= (S_{i}^x,S_{i}^y)$ and $S_i^{x 2}+S_i^{y 2}=1$.
In this paper we wish to consider a three component local spin $\vec S_{i}= (S^x_i,S^y_i,S^z_i)$ and the same Hamiltonian as given by Eq. (\[XY\]) (namely, with no coupling between the z-components of the spins) in three dimensions. Even though the Hamiltonian is the same, namely, there is no coupling between the z-component of the spins, the constrain for each spin is $(S_i^{x})^2 + (S_i^{y})^2+(S_i^{z})^2=1$, which implies that the quantity $(S_i^{x})^2+(S_i^{y})^2$ is fluctuating. In order to be distinguished from the usual XY model, the name [*planar magnet model*]{} will be adopted for this model.
The reason for our desire to study this model is that it is related directly to the so-called model-F[@Hohenberg] used to study non-equilibrium phenomena in systems, such as superfluids, with a two-component order parameter and a conserved current. In the planar magnet model, the order parameter is not a constant of the motion. A constant of the motion is the $z$ component of the magnetization. Thus, there is an important relationship between the order parameter and the $z$ component of magnetization, which is expressed by a Poisson-bracket relation[@Hohenberg]. This equation is crucial for the hydrodynamics and the critical dynamics of the system. One therefore needs to find out the critical properties of this model in order to study non-equilibrium properties of superfluids or other systems described by the model F. In future work, we shall use model-F to describe the dynamical critical phenomena of superfluid helium. Before such a project is undertaken, the static critical properties of the planar magnet model should be investigated accurately.
Although the static properties of the $XY$ model with $\vec{S}_{i}=
(S^{x}_{i},S^{y}_{i})$ have been investigated by a variety of statistical-mechanical methods[@Guillou; @Albert; @high; @Li; @MC; @S2; @S3; @Janke; @Hasenbusch], the system with $\vec{S}_{i}=
(S^{x}_{i},S^{y}_{i},S^{z}_{i})$ has been given much less attention. So far the critical behavior of this model has been studied by high temperature expansion[@Ferer] and Monte Carlo(MC) simulation methods[@Costa; @Oh]. High temperature expansion provides the value for the critical temperature and the critical exponents. In these recent MC calculations[@Costa; @Oh], only the critical temperature is determined. These MC calculations were carried out on small size systems and thus only rough estimates are available.
In this paper we study the three-dimensional planar magnet model using a hybrid Monte Carlo method (a combination of the cluster algorithm with over-relaxation and Metropolis spin re-orientation algorithm) in conjuction with single-histogram re-weighting technique and finite-size scaling. We calculate the fourth order cumulant, the magnetization, and the susceptibility (on cubic lattices $L\times L \times L$ with $L$ up to $42$) and from their finite-size scaling behavior we determine the critical properties of the planar magnet model accurately.
Physical Quantities and Monte Carlo Method {#sec1}
==========================================
Let us first summarize the definitions of the observables that are calculated in our simulation. The energy density of our model is given by $$<e>=E/V=\frac{1}{V}\sum_{<ij>}<S_{i}^{x}S_{j}^{x}+S_{i}^{y}S_{j}^{y}>,$$ where $V=L^{3}$ and the angular brackets denote the thermal average. The fourth-order cumulant $U_{L}(K)$[@Binder] can be written as $$U_{L}(K)=1-\frac{<m^{4}>}{3<m^{2}>^{2}},$$ where $m=\frac{1}{V}(M_{x}^{2}+M_{y}^{2}+M_{z}^{2})^{1/2}$ is the magnetization per spin, $\vec M = \sum_{i} \vec S_i$ and $K=J/(k_{B}T)$ is the coupling, or the reduced inverse temperature in units of $J$. The fourth-order cumulant $U_{L}(K)$ is one important quantity which we use to determine the critical coupling constant $K_{c}$. In the scaling region close to the critical coupling, the fourth-order cumulant $U_{L}(K)$ as function of $K$ for different values of $L$ are lines which go through the same point.
The magnetic susceptibility per spin $\chi$ is given by $$\chi = VK(<m^{2}>-<\vec{m}>^{2}),$$ where $\vec{m}$ is the magnetization vector per spin.
The three-dimensional planar magnet model with ferromagnetic interactions $J>0$ has a second-order phase transition. In simulations of systems near a second-order phase transition, a major difficulty arises which is known as critical slowing down. The critical slowing down can be reduced by using several techniques and what we found as optimal for our case was to use the hybrid Monte Carlo algorithm as described in Ref. [@Landau]. Equilibrium configurations were created using a hybrid Monte Carlo algorithm which combines cluster updates of in-plane spin components[@Wolff] with Metropolis and over-relaxation[@Brown] of spin re-orientations. After each single-cluster update, two Metropolis and eight over-relaxation sweeps were performed[@Landau]. The $K$ dependence of the fourth-order cumulant $U_{L}(K)$ was determined using the single-histogram re-weighting method[@Ferr]. This method enables us to obtain accurate thermodynamic information over the entire scaling region using Monte Carlo simulations performed at only a few different values of $K$. We have performed Monte Carlo simulation on simple cubic lattices of size $L\times L\times L$ with $6\leq L\leq 42$ using periodic boundary conditions applied in all directions and $10^{6}$ MC steps. We carried out of the order of 10000 thermalization steps and of the order of 20000 measurements. After we estimated the critical coupling $K_{c}$, we computed the magnetization and the magnetic susceptibility at the critical coupling $K_{c}$.
Results and Discussion {#sec2}
======================
In this section, we first have to determine the critical coupling $K_{c}$, and then to examine the static behavior around $K_{c}$. Binder’s fourth-order cumulant[@Binder] $U_{L}(K)$ is a convenient quantity that we use in order to estimate the critical coupling $K_{c}$ and the correlation length exponent $\nu$.
Near the critical coupling $K_{c}$, the cumulant is expanded as $$U_{L}=U^{\ast}+U_{1}L^{1/\nu}(1-\frac{T}{T_{c}})+\cdot\cdot\cdot\cdot.$$ Therefore, if we plot $U_{L}(K)$ versus the coupling $K$ for several different sizes $L$, it is expected that the curves for different values of $L$ cross at the critical coupling $K_{c}$. In order to find the $K$ dependence of the fourth-order cumulant $U_{L}(K)$, we performed simulations for each lattice size from $L=6$ to $L=42$ at $K$=0.6450 which is chosen to be close to previous estimates for the critical inverse temperature[@Ferer; @Oh]. The $U_{L}(K)$ curves were calculated from the histograms and are shown in Fig. \[fi-1\] for $L$=12, 24, and 32.
=[3.375in]{}
If one wishes to obtain higher accurary, then one needs to examine Fig. 1 more carefully and to see that the points where each pair of curves cross are slightly different for different pairs of lattices; in fact the points where the curves cross move slowly to lower couplings for larger system sizes. For the pair which corresponds to our largest lattice sizes $L$=32 and 42, the point where they cross is $K_{c}\approx 0.64455$. In order to extract more precise critical coupling $K_{c}$ from our data, we compare the curves of $U_{L}$ for the two different lattice sizes $L$ and $L'=bL$ and then find the location of the intersection of two different curves $U_{L}$ and $U_{L'}$. As a result of the residual corrections to the finite size scaling [@Binder], the locations depend on the scale factor $b=L'/L$. We used the crossing points of the $L$=12, 14, and, 16 curves with all the other ones with higher $L'$ value respectively. Hence we need to extrapolate the results of this method for (ln$b$)$^{-1} \longrightarrow 0$ using $(U_{bL}/U_{L})_{T=T_{c}}=1$. In Fig. \[fi-2\] we show the estimate for the critical temperature $T_{c}$. Our final estimate for $T_{c}$ is $$T_{c}=1.5518(2), K_{c}=0.6444(1).$$ For comparison, the previous estimates are $T_{c}$=1.54(1)[@Costa; @Oh] obtained using Monte Carlo simulation and $T_{c}$=1.552(3)[@Ferer] obtained using high-temperature series. The latter result obtained with an expansion is surprisingly close to ours.
=[3.375in]{}
In order to extract the critical exponent $\nu$, we performed finite-size scaling analysis of the slopes of $U_{L}$ versus $L$ near our estimated critical point $K_{c}$. In the finite-size scaling region, the slope of the cumulant at $K_{c}$ varies with system size like $L^{1/\nu}$, $$\frac{dU_{L}}{dK} \sim L^{1/\nu}.$$ In Fig. \[fi-3\] we show results of a finite-size scaling analysis for the slope of the cumulant. We obtained the value of the static exponent $\nu$, $$\nu = 0.670(7).$$ For comparison, the field theoretical estimate[@Guillou] is $\nu$=0.669(2) and a recent experimental measurement gives $\nu$=0.6705(6)[@Goldner].
=[3.375in]{}
In order to obtain the value of the exponent ratio $\gamma/\nu$, we calculated the magnetic susceptibility per spin $\chi$ at the critical coupling $K_{c}$. The finite-size behavior for $\chi$ at the critical point is $$\chi \sim L^{\gamma/\nu}.$$ Fig. \[fi-4\] displays the finite-size scaling of the susceptibility $\chi$ calculated at $K_{c}$=0.6444. >From the log-log plot we obtained the value of the exponent ratio $\gamma/\nu$, $$\gamma/\nu=1.9696(37).$$ >From the hyperscaling relation, $d\nu=\gamma+2\beta$, we get the exponent ratio $\beta/\nu$, $$\beta/\nu=0.515(2).
\label{betanu}$$
=[3.375in]{}
The equilibrium magnetization $m$ at $K_{c}$ should obey the relation $$m \sim L^{-\beta/\nu}$$ for sufficiently larger $L$. In Fig. \[fi-5\] we show the results of a finite-size scaling analysis for the magnetization $m$. We obtain the value of the exponent ratio $\beta/\nu$, $$\beta/\nu=0.515(2).$$ This result agrees very closely to that of Eq. (\[betanu\]) obtained from the susceptibility and the fourth-order cumulant.
=[3.375in]{}
$L$
----- ------------ -------------
12 82.39(28) 0.26195(55)
14 111.88(36) 0.24219(43)
16 145.12(59) 0.22567(55)
18 182.91(52) 0.21241(35)
20 224.08(85) 0.20072(49)
22 272.23(60) 0.19163(23)
24 322.35(98) 0.18308(32)
32 571.0(4.0) 0.15833(66)
42 972.0(4.8) 0.13749(40)
: \[t1\] Results for the magnetization and the susceptibility
In conclusion, we determined the critical temperature and the exponents of the planar magnet model with three-component spins using a high-precision MC method, the single-histogram method, and the finite-size scaling theory. Our simulation results for the critical coupling and for the critical exponents are $K_{c}$=0.6444(1), $\nu$=0.670(7), $\gamma/\nu$=0.9696(37), and $\beta/\nu$=0.515(2). Our calculated values for the critical temperature and critical exponents are significantly more accurate that those previously calculated. Comparison of our results with results of MC studies of the 3$D$ $XY$ model with two-component spins[@MC; @S3; @Janke; @Hasenbusch] shows that both the system with $\vec{S}_{i}=
(S^{x}_{i},S^{y}_{i})$ and the planar magnet system with $\vec{S}_{i}=(S^{x}_{i},S^{y}_{i},S^{z}_{i})$ belong to the same universality class.
acknowledgements
================
This work was supported by the National Aeronautics and Space Administration under grant no. NAG3-1841.
K.G. Wilson, Phys. Rev. B [**4**]{}, 3174 (1971). P.C. Hohenberg and B.I. Halperin, Rev. Mod. Phys. [**49**]{}, 435 (1977). J.C. Le Guillou and J. Zinn-Justin, Phys. Rev. B [**21**]{}, 3976 (1980). D.Z. Albert, Phys. Rev. B [**25**]{}, 4810 (1982). P. Butera, M. Comi, and A. J. Guttmann, Phys. Rev. B [**48**]{}, 13987 (1993); R. G. Bowers and G. S. Joyce, Phys. Rev. Lett. [**19**]{}, 630 (1967). Y.-H.Li and S. Teitel, Phys. Rev. B [**40**]{}, 9122 (1989). A. P. Gottlob and M. Hasenbusch, Physica A [**201**]{}, 593 (1993); A. P. Gottlob, M. Hasenbusch and S. Meyer, Nucl. Phys. B (Proc. Suppl.) [**30**]{}, 838 (1993). N. Schultka and E. Manousakis, Phys. Rev. B [**51**]{}, 11712 (1995). N. Schultka and E. Manousakis, Phys. Rev. B [**52**]{}, 7528 (1995). W. Janke, Phys. Lett. A [**148**]{}, 306 (1990). Martin Hasenbusch and Steffen Meyer, Phys. Lett. B [**241**]{}, 238 (1990). M. Ferer, M.A. Moore, and M. Wortis, Phys. Rev. B [**8**]{}, 5205 (1973). B.V. Costa, A.R. Pereira, and A.S.T. Pires, Phys. Rev. B [**54**]{}, 3019 (1996). S.K. Oh, C.N. Yoon, and J.S. Chung, Phys. Rev. B [**56**]{}, 13677 (1997). K. Binder, Z. Phys. B [**43**]{}, 119 (1981). P. Peczak and D.P. Landau, Phys. Rev. B [**47**]{}, 14260 (1963). U. Wolff, Phys. Rev. Lett. [**62**]{}, 361 (1989). F.R. Brown and T.J. Woch, Phys. Rev. Lett. [**58**]{}, 2394 (1987). A.M. Ferrenberg and R.H. Swendsen, Phys. Rev. Lett. [**61**]{}, 2635 (1988); $ibid$, [**63**]{}, 1195 (1989). L.S. Goldner, N. Mulders, and G. Ahlers, J. Low Temp. Phys. [**93**]{}, 131 (1992).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using examples [@13] of compact complex 3-manifolds which arise as twistor spaces, we show that the class of compact complex manifolds bimeromorphic to Kähler manifolds is not stable under small deformations of complex structure.'
author:
- ' [*and*]{} '
title: |
Twistors, Kähler Manifolds, and\
Bimeromorphic Geometry\
II
---
****
Ø[[O]{}]{}
A well-known theorem of Kodaira and Spencer [@11][@15] states that any small deformation of the complex structure of a compact Kähler manifold again yields a complex manifold of Kähler type. The question has been therefore been raised [@7] [@19] as to whether a similar stability result holds for compact complex manifolds which are [*bimeromorphically equivalent*]{}[^1] to Kähler manifolds— that is, for manifolds of Fujiki’s class $\cal C$ [@6]. In this article, we will analyze the twistor spaces obtained in the previous article [@13] as small deformations of the Moishezon twistor spaces discovered in [@12], and show that they are generically [*not*]{} spaces of class $\cal C$, even though they are obtained as small deformations of spaces which [*are*]{}. In short, the bimeromorphic analogue of the Kodaira-Spencer stability theorem is false.[^2]
In an attempt to make this article as self-contained as possible, we begin with a brief introduction to the subject, including a quick review of the essential results of the preceding article [@13].
Our focus here will be on the following class of complex manifolds:
A [*twistor space*]{} will herein mean a compact complex 3-manifold $Z$ with the following properties:
- There is a free anti-holomorphic involution $\s : Z\to Z$, $\s^2=$ identity, called the [*real structure*]{} of $Z$;
- There is a foliation of $Z$ by $\s$-invariant holomorphic curves $\cong \bcp_1$, called the [*real twistor lines*]{}; and
- Each real twistor line has normal bundle holomorphically isomorphic to $\O (1)\+\O (1)$, where $\O (1)$ is the degree-one line bundle on $\bcp_1$.
The space $M$ of real twistor lines is thus a compact real-analytic 4-manifold, and we have real-analytic submersion $\wp :Z\to M$ known as the [*twistor projection*]{}. By a construction discovered by Roger Penrose [@16], the complex structure of $Z$ induces a half-conformally-flat conformal Riemannian conformal metric on $M$, and every such metric conversely arises in this way [@1]; however, we will never explicitly need this in the sequel.
We will only concern ourselves here with the class of twistor spaces admitting hypersurfaces of the following type:
An [*elementary divisor*]{} $D$ on a twistor space $Z$ is a complex hypersurface $D\subset Z$ whose homological intersection number with a twistor line is +1, and such that $D\cap \s (D)\neq \emptyset$.
An elementary divisor is necessarily a smooth hypersurface. The existence of such a divisor $D$ is a powerful hypothesis indeed, for it follows ([@13], Proposition 6) that $D$ is an $n$-fold blow-up of $\bcp_2$, that $M$ is diffeomorphic to an $n$-fold connected sum $\bcp_2\#\cdots\# \bcp_2$, and the map $\wp|_D:D\to M$ contracts a projective line to a point, but is elsewhere an orientation-reversing diffeomorphism.
In fact, these conclusions are quite sharp.
Let $X$ be any compact complex surface obtained from $\bcp_2$ by blowing up distinct points. Then there exists a twistor space $Z$ which contains an elementary divisor $D$ such that $D\cong X$ as a complex surface. Moreover, given a smooth (respectively, real-analytic) 1-parameter family $X_t$ of surfaces obtained from $\bcp_2$ by blowing up distinct ordered points, there is a smooth (respectively, real-analytic) family $(Z_t, D_t)$ of twistor spaces with elementary divisors such that $D_t\cong X_t$.\[def\]
In [@12] it was shown that, given an arbitrary blow-up $D$ of $\bcp_2$ at $n$ collinear points, there is a twistor space $Z$ containing a degree 1 divisor isomorphic to $D$. In fact, such twistor spaces $Z$ may be explicitly constructed from conic bundles over $\bcp_1\times\bcp_1$ by a process of blowing subvarieties up and down, and thus may be taken to be [*Moishezon*]{} in this case. In the accompanying article [@13], the deformation theory of these twistor spaces was studied, with the following conclusion. Let $p_1:=(0,0)$ and $p_2:=(1,0)$ in $\bc^2$, and let ${\cal W}\subset [\bc^2]^{n-2}$ denote the set $$\{ (p_3, \ldots , p_n)|p_j\in \bc^2, p_j\neq p_k, j,k=1,\ldots , n\};$$ let ${\cal L}\subset {\cal W}$ denote the subset $p_3, \ldots , p_n\in (\bc \times \{ 0\})$ of collinear configurations. It was shown there ([@13], [**Theorem 3**]{}) that there exists a (versal) family $({\cal Z}, {\cal D})$ of twistor spaces with elementary divisors over a ${\cal U}$ neighborhood of $[{\cal L}\times
(\br^+)^n]
\subset [{\cal W}\times (\br^+)^n]$ such that the divisor $D$ associated with a configuration of points $$p_1,\ldots , p_n\in {\Bbb C}^2\subset\bcp_2$$ and arbitrary collection of positive weights $$m_1,\ldots , m_n\in \br^+$$ is isomorphic to $\bcp_2$ blown up at $p_1,\ldots , p_n$.
Now suppose we are given an arbitrary compact complex surface $X$ obtained from $\bcp_2$ by blowing up $n$ distinct points $q_1, \ldots , q_n$. There is a line $L\subset\bcp_2$ which misses $q_1, \ldots , q_n$, and now identify $\bcp_2-L$ with $\bc^2$ in such a way that $q_1=(0,0)$ and $q_2=(1,0)$. Assign all the points, say, weight 1. By making a linear transformation, we may also take the points $q_1, \ldots , q_n$ to be as close as we like to the $z_1$-axis, so that our configuration becomes a point of ${\cal U}$. The corresponding fiber of our family $({\cal Z}, {\cal D})$ then comes equipped with an elementary divisor isomorphic to the given $X$.
On the other hand, suppose we are instead given an arbitrary smooth family $X_t$ of surfaces obtained by blowing up $n$ distinct, ordered points in $\bcp_2$, where $t$ ranges over $\br$. Let ${\cal X}\to \br$ denote the family with fibers $\{ X_t\}$. There is a bundle ${\cal P}\to B$ of $\bcp_2$’s from which ${\cal X}\to \br$ is obtained by blowing up $n$ sections $q_1, \ldots , q_n$; let ${\cal P}^{\ast}\to \br$ denote the bundle of dual planes, in which the $q_1, \ldots , q_n$ define $n$ complex hypersurfaces. The complement of these hypersurfaces in ${\cal P}^{\ast}$ has real codimension 2, so, by transversality, a generic smooth (respectively real-analytic) section of ${\cal P}^{\ast}$ will miss them, and we may therefore smoothly (respectively real-analytically) choose a projective line $L_t$ in each fiber $P_t$ of ${\cal P}$ which misses the points ${q_1}_t, \ldots {q_n}_t$. Using $q_1$ as the zero section, the complement of these chosen lines becomes a vector bundle over $\br$ and so may be trivialized in such a manner that $q_2\equiv (1,0)$. Our family of surfaces may therefore be thought of as associated with a family of point configurations $(q_1, \ldots , q_n)_t$ in $\bc^2$, where $q_1\equiv (0,0)$ and $q_2\equiv (1,0)$. Again, let us assign each point a positive weight, say 1. Now there is a positive real-analytic function $F(\zeta_3,\ldots , \zeta_n)$ such that a weighted configuration $((0,0,1), (0,1,1), (\zeta_3, \eta_3, 1),
\ldots , (\zeta_n, \eta_n, 1))$ is in ${\cal U}$ provided that $$\sum | \eta_j|^2 < F(\zeta_3,\ldots , \zeta_n).$$ Setting $(q_3, \ldots , q_n)_t = ((\zeta_3(t), \eta_3(t)),
\ldots , (\zeta_n(t), \eta_n(t)))$, define $$(p_1, \ldots , p_n)_t:=
\left[ \begin{array}{cc}1&0\\0&\sqrt{\frac{F(\zeta_3(t),
\ldots , \zeta_n(t))}{
(1+\sum|\zeta_j(t)|^2)}}
\end{array} \right](q_1, \ldots , q_n)_t~.$$ The family $((p_1,1) , \ldots , (p_n,1))_t$ of weighted configurations then takes values within the parameter space ${\cal U}$ of the family $({\cal Z}, {\cal D})$. Pulling back $({\cal Z}, {\cal D})$ now yields the desired family of twistor spaces with elementary divisors.
It might be emphasized, incidentally, that the twistor space $Z$ is by no means determined by the intrinsic structure of a elementary divisor $D$. Nonetheless, we will presently see that the intrinsic structure of such a divisor [*does*]{} tell us a great deal about a twistor space, and is, in particular, sufficient to determine its [*algebraic dimension*]{}.
Let us recall that the algebraic dimension $a(Z)$ of a compact complex manifold $Z$ is by definition the degree of transcendence its the field of meromorphic functions, considered as an extension of the field $\bc$ of constant functions. Equivalently, the algebraic dimension of $Z$ is precisely the maximal possible dimension of the image of $Z$ under a meromorphic map to $\bcp_N$; in particular, $a(Z)\leq \mbox{dim}_{\bc}(Z)$. When equality is achieved in the latter inequality, $Z$ is said to be a [*Moishezon manifold*]{} [@14], and a suitable sequence of blow-ups of $Z$ along complex submanifolds will then result in a projective variety.
The following lemma of F. Campana will be of critical importance:
[[@2].]{} A twistor space $Z$ is bimeromorphic to a Kähler manifold iff it is Moishezon.\[camp\]
Let $p$ and $q$ be distinct points of a real twistor line $L$ in a twistor space $Z$, and let $S_p$ (respectively, $S_q$) denote the space of rational curves through $p$ (respectively, $q$) which are deformations of $L$. Assume that $Z$ is in the class $\cal C$. Because the components of the Chow variety of $Z$ are therefore compact, the correspondence space $$Z':=\{ (r, C_1, C_2)\in
Z\times S_p\times S_q~|~r\in C_1\cap C_2\}$$ is thus a compact complex space; by blowing up any singularities, we may assume that $Z'$ is smooth. But since a real twistor line has the same normal bundle as a projective line in $\bcp_3$, a generic point of $Z$ is joined to either $p$ or $q$ only by a discrete set of curves of the fixed class. The correspondence space $Z'$ is therefore generically a branched cover of $Z$, and is, in particular, a 3-fold. On the other hand, we have a canonical map $$\phi:
Z'\to \bp (T_pZ)\times \bp (T_pZ)
\cong \bcp_2\times\bcp_2$$ obtained by taking the tangent spaces of curves at their base-points $p$ or $q$. Let $r$ be a point of $Z$ which is not on $L$, but close enough to $L$ so that $r$ is joined to $p$ and $q$ by small deformations $C_1$ and $C_2$ of $L$, both of which are $\bcp_1$’s with normal bundle $\O (1)\oplus\O (1)$. Then $(r,C_1, C_2)$ is a point of $Z'$ at which the derivative of $\phi$ has maximal rank. Pulling back meromorphic functions from $\bcp_2\times\bcp_2$ to $Z'$ must therefore yield 3 algebraically independent functions, and $Z'$ is therefore a Moishezon space. But since the projection $Z'\to Z$ is surjective, and since the class of Moishezon manifolds is closed under holomorphic surjections [@14], it follows that $Z$ is also a Moishezon manifold.
The converse is, of course, trivial.
On the other hand, the following lemma allows one to determine the algebraic dimension of a twistor space:
[[@17].]{} Any meromorphic function on a simply-connected twistor space $Z$ can be expressed as the ratio of two holomorphic sections of a sufficiently large power $\k^{-m}$ of the anti-canonical line bundle $\k^{-1}:=\wedge^3TZ$. \[poon\]
We begin by observing that any (compact) twistor space satisfies $h^1(Z,\O )=b_1(Z)$. This is a consequence of the [*Ward correspondence*]{} [@20], which says that the set of holomorphic vector bundles on $Z$ which are trivial on real twistor lines is in 1-1 correspondence with the instantons on $M$; in particular, every holomorphic line bundle on $Z$ with $c_1=0$ is obtained by pulling back a flat $\bc_{\ast}$-bundle from $M$ and equipping it with the obvious holomorphic structure. With the exponential sequence $$\cdots\to H^1(Z,\O )\to H^1(Z,\O_{\ast})\stackrel{c_1}{\to}
H^2(Z,{\Bbb Z})\to \cdots$$ this implies that holomorphic line bundles on a simply-connected twistor space are classified by their Chern classes.
Since we have assumed that $Z$ is simply connected, it follows that $H^2(Z, {\Bbb Z})$ is free. On the other hand, the Leray-Hirsch theorem tells us that $H^2 (Z, {\Bbb Q})={\Bbb Q}c_1(Z)\oplus H^2(M,{\Bbb Q})$. The latter splitting of the cohomology is exactly the decomposition of $H^2 (Z, {\Bbb Q})$ into the $(\mp 1)$-eigenspaces of $\sigma^{\ast}$; a class will be called [*real* ]{} if it is in the $(-1)$-eigenspace, and a complex line-bundle will be called real if its first Chern class is real. There is thus a unique “fundamental” holomorphic line bundle $\xi$ on $Z$ such that any real holomorphic line bundle is a power of $\xi$ and such that the restriction of $\xi$ to a twistor line is positive; in particular, $\k=\xi^k$ for some $k$. While we will not need to know this explicitly, it can in fact be shown [@9] that $k=4$ if $M$ is spin, and $k=2$ otherwise.
Now suppose that we are given a meromorphic function $f$ on such a $Z$. The function $f$ can [*a priori*]{} be expressed in the form $f=g/h$, where $g$ and $h$ are holomorphic sections of a line-bundle $\eta\to Z$; for example, we could take $\eta$ to be the divisor line bundle of the polar locus of $f$. The pull-back $\sigma^{\ast}\overline{\eta }$ of the conjugate line-bundle of $\eta$ is automatically holomorphic, and $\sigma^{\ast}\overline{g}$ and $\sigma^{\ast}\overline{h}$ are holomorphic sections of this bundle. The holomorphic bundle $\eta\otimes\sigma^{\ast}
\overline{\eta}$ is now [*real*]{} and has sections, and so must be the form $\xi^m$ for some positive integer $m$. Thus $$f=\frac{gh^{k-1}\sigma^{\ast}\overline{h}^k}{h^{k}\sigma^{\ast}\overline{h}^k},$$ expresses our meromorphic function as the quotient of two holomorphic sections of $\k^m$.
We have already seen that there are examples of Moishezon twistor spaces $Z$ containing an elementary divisor $D$ isomorphic to $\bcp_2$ blown up at a collinear configuration of points. We will now see that the situation is dramatically different when the intrinsic structure of $D$ is generic.
Suppose that $Z$ is a twistor space with an elementary divisor isomorphic to the blow-up of $\bcp_2$ at $n$ generic points, $n\geq 7$. Then $Z$ has no non-constant meromorphic functions, and so has algebraic dimension $0$. The set of configurations $(p_1 ,\ldots ,p_n )$ of points in $\bcp_2$ which are generic in this sense is the complement of a countable union of proper algebraic sub-varieties of $(\bcp_2)^n$, and in particular has full measure. \[big\]
Let us begin by considering the case of a configuration of $n$ points in ${\Bbb C}^2$ containing a 6-point configuration of the following type:
(18,5)(0,0) (0,0)[(3,1)[12]{}]{} (18,0)[(-3,1)[12]{}]{} (0,1)[(1,0)[18]{}]{} (3,1) (9,1) (9,3) (6,2) (12,2) (15,1)
[**Figure 1.**]{}
We assume that the other points of the configuration are not on any of three projective lines of the figure. The proper transforms of these three lines are then (-2)-curves $E_j$, $j=1,2,3$. The anti-canonical bundle $\kappa_D^{-1}$ of the surface $D$ thus satisfies $\kappa_D^{-1}|_{E_j}\cong {\cal O}$. On the other hand, since the half-anti-canonical bundle of $Z$ is given by $\kappa^{-1/2}= [D] \otimes [\overline{D}]$, we have $$\kappa^{-1/2}|_{D}=\nu\otimes [L_{\infty}]~,$$ where $\nu$ denotes the normal bundle of $D\subset Z$ and $L_{\infty}\subset D$ is the projective line $D\cap \overline{D}$. Yet the adjunction formula yields $$\kappa^{-1}|_{D}=\nu \otimes \kappa_D^{-1}~,$$ so that $$\nu^{2}\otimes [L_{\infty}]^2= \nu \otimes \kappa_D^{-1}$$ implying that $\nu = \kappa_D^{-1}\otimes [L_{\infty}]^{-2}$ and hence \^[-1/2]{}|\_[D]{}= \_D\^[-1]{}\^[-1]{} .\[snuk\]It follows that $$\kappa^{-1/2}|_{E_j}\cong {\cal O}(-1)~.$$ On the other hand, the normal bundle $N_j$ of $E_j\subset D$ is isomorphic to ${\cal O}(-2)\to {\Bbb CP}_1$. Since (E\_j, [O]{} ((\^[-m/2]{}|\_[E\_j]{})N\_j\^[-k]{}))&=& ([CP]{}\_1, [O]{} (-m+2k))\
&=& 0 it follows that any section of $\kappa^{-m/2}|_{D}$ vanishes along $E_j$ to order $[\frac{m-1}{2}]$. But through the generic point of $D$ we can find a projective line in $D$ passing through a blown-up point not on the diagram, avoiding all other blown-up points, and meeting the $E_j$ in three distinct points. Letting $L$ denote the proper transform of such a line, one has \^[-1/2]{}|\_L&=& (\_D\^[-1]{}\^[-1]{})|\_L\
&& [O]{}(2)(-1)\
&&[O]{}(1) ,so that $\kappa^{-m/2}|_L\cong {\cal O}(m)$. Yet any holomorphic section of of $\kappa^{-m/2}|_{D}$ must have 3 zeroes on $L$ of multiplicity $[\frac{m-1}{2}]$ at $L\cap E_j$. Since $3[\frac{m-1}{2}]> m$ for $m>6$, we conclude that such a section must vanish identically on $L$ provided $m$ is sufficiently large. Hence $\Gamma (D, {\cal O} (\kappa^{-m/2}))=0$ for $m$ sufficiently large, and hence, by taking tensor powers of sections, for all $m>0$. Similarly, $\Gamma (\overline{D}, {\cal O} (\kappa^{-m/2}))=0$ for all $m>0$. From the exact sequences 0\_[Z]{} (\^[-(m-1)/2]{})\_[Z]{} (\^[-m/2]{})\_[D]{} (\^[-(m-1)/2]{})0,\[snork\] we conclude by induction that $$\Gamma ({Z}, {\cal O}(\kappa^{-m/2}))=\bc$$ for all $m> 0$. By Lemma \[poon\], any meromorphic function on $Z$ must therefore be constant.
We now examine the case of $D$ obtained from $\bcp_2$ by blowing up $n>6$ generically located points. For each n-tuple of points $(p_1 ,\ldots ,
p_n )_u$ in $\bc^2=\bcp_2-L_{\infty}$, let $D_u$ denote the corresponding blow-up of $\bcp_2$, and consider the behavior of $h^0(D_u , {\cal O} (\kappa_{D_u}^{-m}\otimes [L_{\infty}]^{-m}))$. By the semi-continuity principle [@8] and the above calculation, this vanishes, for $m$ fixed, on a non-empty Zariski-open subset of configurations. The set of $n$-point configurations for which $h^0(D_u , {\cal O} (\kappa_{D_u}^{-m}\otimes [L_{\infty}]^{-m}))\neq 0$ for some $m$ is therefore a countable union of subvarieties, and so has measure $0$. Using the exact sequence \[snork\] and the isomorphism \[snuk\], we conclude that $\Gamma ({Z}, {\cal O}(\kappa^{-m/2}))=\bc ~\forall m\neq 0$ provided that $Z$ contains an elementary divisor $D$ obtained from $\bcp_2$ by blowing up $n>6$ generic points. Again applying Lemma \[poon\], we conclude that, for $n\geq 7$, any meromorphic function on a twistor space $Z$ containing a generic elementary divisor must therefore be constant.
Our main result now follows immediately:
The class $\cal C$, consisting of of compact complex manifolds which are bimeromorphic to Kähler manifolds, is not stable under small deformations.
By Propositions \[def\] and \[big\], there exist 1-parameter families of twistor spaces $Z_t$ for which almost every $Z_t$ has algebraic dimension $0$, whereas $Z_0$ is Moishezon; in fact it suffices to take $Z_0$ to be one of the explicit examples of [@12], with $D_0$ corresponding to a collinear configuration of $n\geq 7$ points, arrange for the curve of configurations $(p_1, \ldots ,
p_n)_t$ to be real-analytic and contain at least one generic configuration. (Actually, one can do better: by taking the elementary divisors $D_t$ to all correspond to configurations containing projective copies of Figure 1 when $t\neq 0$, one can even arrange for $Z_0$ to be the [*only*]{} Moishezon space in the family.) By Lemma \[camp\], the non-Moishezon twistor spaces of the family $Z_t$ are not of class $\cal C$, despite the fact that they are arbitrarily small deformations of the Moishezon space $Z_0$.
[**Remarks.**]{}
- In order to keep this article as short and clear as possible, we have only considered the case of $n\geq 7$, and only presented the extreme cases of $a(Z)=3$ and $a(Z)=0$. In fact, in can be shown that generically $a(Z)<3$ as soon as $n\geq 4$. One can also find simple non-collinear configurations for which $n=1, 2$ as soon as $n\geq 5$. Finally, one can show that the existence of an elementary divisor corresponding to a collinear configuration [*forces*]{} $Z$ to be one of the examples of [@12], and, in particular, Moishezon. For details, see [@18].
- The existence of self-dual metrics on arbitrary connected sums $\BP_2\# \cdots\#
\BP_2$ was first proved abstractly by Donaldson and Friedman [@4] and, using completely different methods, by Floer [@5]. Unlike the methods used here, these methods do not show that the twistor space of some of these metrics are Moishezon. It was nonetheless the Donaldson-Friedman construction which originally gave the authors reason to believe that the generic deformation of the explicit twistor spaces of [@12] should not be of Fujiki-class $\cal C$. For providing this source of inspiration, as well as for their friendly advice and encouragement, the authors would therefore like to gratefully thank Robert Friedman and Simon Donaldson.
[99]{} M. Atiyah, N. Hitchin and I. Singer,“Self-Duality in Four Dimensional Riemannian Geometry,” [**Proc. R. Soc. Lond. A 362**]{} (1978) 425-461. F. Campana, “On Twistor Spaces of the Class ${\cal C}$,” [**J. Differential Geometry 33**]{} (1991) 541-549. F. Campana,“The Class $\cal C$ is not Stable under Deformations,” [*preprint*]{}. S.K. Donaldson and R. Friedman, “Connected Sums of Self-Dual Manifolds and Deformations of Singular Spaces,” [**Nonlinearity 2**]{} (1989) 197-239. A. Floer, “Self-Dual Conformal Structures on $\ell\bcp^2$,” [**J. Diff. Geometry 33**]{} (1991) 551-573. A. Fujiki, “On Automorphism Groups of Compact Kähler Manifolds,” [**Inv. Math. 44**]{} (1978) 225-258. A. Fujiki, “On a Compact Complex Manifold in ${\cal C}$ without Holomorphic 2-Forms,” [**Publ. RIMS 19**]{} (1983). H. Grauert and R. Remmert, [**Coherent Analytic Sheaves**]{}, Springer-Verlag, 1984. N.J. Hitchin, “Kählerian Twistor Spaces,” [**Proc. Lond. Math. Soc. 43**]{} (1981) 133-150. K. Kodaira, “A Theorem of Completeness of Characteristic Systems for Analytic Families of Compact Submanifolds of Complex Manifolds,” [**Ann. Math. 75**]{} (1962) 146-162. K. Kodaira and D. Spencer, “On Deformations of Complex Analytic Structure, I, II, III,” [**Ann. Math. 67**]{} (1958) 281-294. C. LeBrun,“Explicit Self-Dual Metrics on $\BP_2\# \cdots\#
\BP_2$,” [**J. Differential Geometry**]{}, [*to appear*]{}. C. LeBrun, “Twistors, Kähler Manifolds, and BimeromorphicGeometry I,” [*preprint*]{}. B. Moishezon, “On $n$-Dimensional Compact Varieties with $n$ Algebraically Independent Meromorphic Functions,” [**Amer. Math. Soc. Translations 63**]{} (1967) 51-177. J. Morrow and K. Kodaira, [**Complex Manifolds**]{} , Holt-Rhinehart & Winston, 1971. R. Penrose, “Non-linear Gravitons and Curved Twistor Theory,” [**Gen. Rel. Grav. 7** ]{}(1976) 31-52. Y.S. Poon,“On the Algebraic Dimension of Twistor Spaces,” [**Math. Ann. 282**]{} (1988) 621-627. Y.S. Poon, “Algebraic Structure of Twistor Spaces,” [*preprint*]{}, 1990. K. Ueno, [*ed.*]{}, “Open Problems,” [**Classification of Algebraic and Analytic Manifolds**]{}, Birkhäuser, 1983. R. Ward,“On Self-Dual Gauge-Fields,” [**Phys. Lett. 61A**]{} (1977) 81-82.
[^1]: Two connected compact complex m-manifolds $X$ and $Y$ are called [*bimeromorphically equivalent*]{} if there exists a complex m-manifold $V$, and degree 1 holomorphic maps $V\to X$ and $V\to Y$.
[^2]: A technically different proof of this result, incorporating information exchanged in discussions and letters with the first author during the summer of 1990, was found simultaneously by F. Campana [@3], who has chosen to publish his work separately.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using the Ginzburg-Landau theory in very general terms, we develop a simple scaling procedure which allows to establish the temperature dependence of the upper critical field $H_{c2}$ and the value of the superconducting critical temperature $T_{c}$ of type-II superconductors from measurements of the reversible isothermal magnetization. An analysis of existing experimental data shows that the normalized dependencies of $H_{c2}$ on $T/T_{c}$ are practically identical for all families of high-$T_{c}$ superconductors at all temperatures for which the magnetization data are available.'
author:
- 'I. L. Landau$^{1,2}$ H. R. Ott$^{1}$'
title: 'Temperature dependence of the upper critical field of type-II superconductors from isothermal magnetization data. Application to high temperature superconductors'
---
epsf
Introduction
============
Establishing the upper critical field $H_{c2}$ and its temperature dependence from experimental data is not a simple task for high-$T_{c}$ superconductors (HTSCs). The main problem is that the transition to the superconducting state, probed by either measuring the magnetization $M(T)$ of the sample or its resistance $R(T)$, does not reveal any sharp features in $M(T)$ or $R(T)$ around $H_{c2}$. The notorious widths of the transitions are usually attributed to fluctuation effects. One of the ways to overcome this difficulty is to use the temperature dependence of the magnetic moment of the sample $M(T)$, measured in a constant external magnetic field $H$, and to extrapolate the linear part of the $M(T)$ curve to the $M$ value corresponding to the normal state of the sample. [@1]$^-$[@5] This procedure is usually justified by invoking the Abrikosov theory of the mixed state according to which the magnetization per unit volume may by written as $$M(H,T)={1 \over {4\pi }}{{H_{c2}(T)-H} \over {(2\kappa ^2-1)\beta _A}},$$ were $\kappa$ is the Ginzburg-Landau (GL) parameter and $\beta _{A}= 1.16$ for a triangular vortex lattice. [@6; @7] From Eq. (1) it follows naturally that the magnetic moment varies linearly with temperature for a fixed value of the external field, if $H_{c2}(T)$ is a linear function of temperature. The problem is that Eq. (1) is only valid close to $H_{c2}$. Nevertheless, most experimental M(T) curves are practically linear in $T$ for magnetic fields between $0.1H_{c2}$ and $0.6H_{c2}$, [@1] i.e., well below $H_{c2}$. In this range of magnetic fields, the magnetic moment is certainly a non-linear function of $[H_{c2}(T) - H]$. The apparent linearity of the experimental $M(T)$ curves is most likely the result of some non-linearity of $H_{c2}(T)$. In this situation, a simple linear extrapolation of $M(T)$ curves will almost certainly result in wrong $H_{c2}(T)$ curves. As an example of this type of failure, we mention a study of Bi$_{2}$Sr$_{2}$CaCu$_{2}$O$_{8}$ single crystals where the application of this extrapolation procedure resulted in completely unphysical $H_{c2}(T)$ curves. [@8]
Another method for establishing $H_{c2}(T)$ is to use theoretical calculations of $M(H)$ extended to $H << H_{c2}$ in order to evaluate $H_{c2}$ from experimental $M(H)$ or $M(T)$ curves. This approach is again not very reliable. First of all, solving the GL equations in two dimensions for magnetic fields well below $H_{c2}$ represents a formidable mathematical task. To our knowledge, there is only one study in which this problem has been solved numerically for the Abrikosov vortex lattice for a selected set of values of the GL parameter $\kappa$. [@9] However, as far as we are aware, nobody has tried to use the results of Ref. 9 for the interpretation of the experimental data. More often, approximate models for the space dependence of the order parameter in the vortex structures are used. The most popular is the Hao-Clem model,[@10] which has widely been used to derive different parameters of HTSCs from magnetization data. [@11]$^-$[@27] However, as has been pointed out recently, this model is far from being accurate. [@28; @29] Both the Hao-Clem model and the numerical calculations in Ref. 9 assume uniform and isotropic superconductors, i.e., conditions that are definitely not met in the case of HTSC compounds.
Although the dependence of the sample resistance on temperature, $R(T)$, in external magnetic fields is often used for the evaluation of $H_{c2}(T)$, [@30]$^-$[@40] we believe that this approach is even less reliable than the use of magnetization measurements. The transition to the normal state resistance is very gradual and there is no appropriate theory for an interpretation of $R(T)$ curves. It is quite likely that the misinterpretation of the resistance data is the main reason why $H_{c2}(T)$ curves derived from the results of resistance measurements often exhibit an unusual positive curvature.
In order to evaluate $H_{c2}$ from the experimental data in such complicated materials as HTSCs, it is very important to introduce an appropriate definition of the upper critical field. In an ideal type-II superconductor, $H_{c2}$ is the highest value of a magnetic field compatible with superconductivity, i.e., the $H_{c2}(T)$ curve on the $H - T$ phase diagram represents a line of second order phase transitions to the normal state. As is well known for HTSC superconductors, this transition degenerates to a cross-over region because of fluctuation effects and even in magnetic fields $H > H_{c2}(T)$ superconducting features appear in the data of resistivity and magnetization measurements. We also note that small inclusions of another superconducting phase with a higher or lower critical temperature, $T'_c$, than that of the bulk cannot always be excluded in HTSCs. In magnetic fields $H > H_{c2}$ the impact of such inclusions with $T'_c>T_{c}$ on the sample resistance or its magnetization is similar to that arising from superconducting fluctuations. At the same time, in magnetic fields well below $H_{c2}(T)$, the effect of fluctuations and possible inclusions of impurity phases on the sample magnetization is small and the $M(H)$ curves in this magnetic field range must be practically the same as for the perfectly uniform sample without fluctuations. This circumstance provides the possibility to evaluate the temperature dependence of $H_{c2}$, in its traditional sense, from such magnetization measurements.
In this paper we propose a new approach to this problem by scaling the $M(H)$ curves measured at different temperatures. This scaling procedure is based on the application of the GL theory, without assuming any specific magnetic field dependence of the magnetization. In this way one can only establish the temperature dependence of $H_{c2}$, but its absolute values remain unknown. Below we describe the method in detail and apply it to experimental data available in the literature. It turns out that in many cases the extrapolation of the normalized $H_{c2}(T)$ curve to $H_{c2} = 0$ provides reliable values of the superconducting critical temperature $T_{c}$.
Scaling procedure
=================
Our scaling procedure is based on the assumption that the Ginzburg-Landau (GL) parameter $\kappa$ is temperature independent. Although the microscopic theory of superconductivity predicts a temperature dependence of $\kappa$, [@41; @42] this dependence is rather weak and is not expected to change the results significantly. From the GL theory it follows straightforwardly that, if $\kappa$ is temperature independent, the magnetic susceptibility $\chi$ of the sample is a universal function of $H/H_{c2}$, i.e., $\chi (H,T) = \chi (h)$ with $h = H/H_{c2}(T)$, [@6]. The magnetization density is $$M(H,T)=H_{c2}(T)h\chi (h).$$ According to Eq. (2) the sample magnetization, for the same value of $h = H/H_{c2}$, is proportional to $H_{c2}(T)$. This leads to the following relation between the values of $M$ at two different temperatures, $T_{0}$ and $T$, $$M(H,T_0)=M(h_{c2}H,T)/h_{c2}$$ with $h_{c2} = H_{c2}(T)/H_{c2}(T_{0})$. The collapse of individual $M(H)$ curves measured at different temperatures may be achieved by a suitable choice of $h_{c2}(T)$. Of course, the scaling procedure implied by Eq. (3) is valid for ideal type-II superconductors only and in the following we consider the necessary corrections to Eq. (3) that are dictated by some specific features of HTSCs.
Most of the families of HTSCs reveal a weak paramagnetic susceptibility $\chi _{n}$ in the normal state. [@5; @13]$^-$[@19; @43]$^-$[@46] Its influence may be accounted for by replacing Eq. (3) by $$M(H,T_0)=M(h_{c2}H,T)/h_{c2}+c_0(T)H,$$ where $c_{0}(T)= \chi_{n}(T_{0}) - \chi_{n}(T)$.
For many HTSC materials the derivative $dM/dH$ changes its sign when approaching the critical temperature from below. [@8; @14]$^-$ [@22; @43]$^-$[@45; @47]$^-$[@49] Because the field dependence of the magnetization in the mixed state always requires $dM/dH > 0$, [@6] the change of sign of $dM/dH$ cannot be explained by considering the properties of a static mixed state alone. This sign change is usually attributed to fluctuation effects. We assume that the additional contributions to the magnetization arising from fluctuation effects may be described by an effective susceptibility $\chi_{eff}(T)$, which is independent of the applied magnetic field. In this case we can still use Eq. (4) but with $c_{0}(T) = [\chi_{n}(T_{0} -
\chi_{n}(T)] + [\chi_{eff}(T_{0}¥) - \chi_{eff}(T)]$. In the following we use the parameter $c_{0}¥(T)$ in Eq. (4) as an additional adjustable parameter in the scaling procedure. The assumption that $\chi_{eff}(T)$ does not depend on the magnetic field is a simplification and this is why Eq. (4) should not be used in the temperature range where $dM/dH < 0$ and where the fluctuation-induced magnetization dominates the magnetic moment of the sample.
We note that the term $c_{0}H$ in Eq. (4) may also account for any contribution to the magnetization arising from small inclusions of another superconducting phase with a different $T_{c}$. If the value of $T_{c}$ of this minority phase is higher than that for the bulk of the sample, some small regions of the sample will remain in the superconducting state even if $H > H_{c2}(T)$. These superconducting islands also give a non-zero magnetic moment with $dM/dH < 0$. In magnetic fields $H < H_{c2}$, the contribution from these regions, where superconductivity is stronger than in the bulk of the sample, is superimposed onto the contribution to the magnetic moment arising from the mixed state.
At this point, we wish to comment on the physical relevance of $H_{c2}(T)$ and $T_{c}$ obtained in this way. Because our analysis is based on measurements of the magnetization in the mixed state, $H_{c2}(T)$ corresponds to the disappearance of the mixed state rather than to a complete suppression of superconductivity in the sample. In magnetic fields $H > H_{c2}$ superconducting regions may, as mentioned above, still exist in the form of separated islands formed due to either thermal fluctuations or inclusions of minority phases with enhanced values of $T_{c}$ and $H_{c2}$. Similar superconducting islands may also exist in zero magnetic field at $T > T_{c}$. It is only important that the lateral extension of these islands is small enough, such that no mixed state can be established inside the island. The cause of such superconducting islands is not important for our consideration. It may be due to thermal fluctuations or sample inhomogeneities, as well as a combination of both. Note that for an ideal type-II superconductor without fluctuations these definitions of $H_{c2}(T)$ and $T_{c}$ coincide with the values of magnetic fields and temperature fixed be the onset of superconductivity.
Analysis of experimental data
=============================
=0.9
We now apply our scaling procedure to experimental results available in the literature. As it turns out, the relative temperature variations of $H_{c2}$ are identical for practically all HTSC materials. Because this is a completely unexpected and, in our view, rather important result, we describe the analysis in some detail. We have analyzed magnetization data for 29 samples presented in 25 publications. Some information concerning these samples is listed in Tables I to IV. Letters in the sample identification denote the chemical element characterizing the considered family of HTSCs. Because the sample homogeneity is important for the applicability of our method, only single crystals and grain-aligned samples have been chosen. We have also limited our analysis to studies in which the magnetization measurements were extended up to temperatures $T \ge 0.94-0.95 T_{c}$, because only in these cases we may expect a reliable evaluation of $T_{c}$ by extrapolating $H_{c2}(T)$ to $H_{c2} = 0$.
=0.9
In order to make use of Eq. (4), the following procedure was employed. First, the $M(H)$ curve for some temperature $T = T_{0}$ was approximated by $$M(H)=h_{c2}\left\{ {\sum\limits_{i=0}^n {A_i[\ln (H/h_{c2})]^i+c_0H}} \right\}$$ with $h_{c2} = 1$ and $c_0 = 0$. The coefficients $A_{i}$ were used as fit parameters and the number $n$ was chosen such that a further enhancement of its value had no influence on the deviation parameter $\sigma$ of the approximation of $M(H)$. [@50] In the next steps, the coefficients $A_{i}$ were fixed and the parameters $h_{c2}$ and $c_{0}$ were evaluated via the fitting procedure for approximating the available $M(H)$ curves measured at other temperatures $T = T_{i}$. [@51] The result of this scaling procedure, representing the field dependence of the magnetization of sample Y\#1 at $T = T_{0}$, is shown in Fig. 1. It may be seen that a rather perfect overlap of the individual $M(H)$ curves measured at different temperatures, which are displayed in the inset of Fig. 1, is obtained in this way. Because the renormalized field variable $h_{c2}$ enters the denominator of Eq. (4), the magnetization data sets for the highest temperatures are considerably expanded along the vertical axis in comparison with low temperature data,. This is the reason for the somewhat enhanced scatter in the high temperature data.
=1.9
In order to demonstrate the consistency of our procedure, we show typical data sets for $H_{c2}(T)$ and $c_{0}(T)$ in Figs. 2(a) and 2(b), respectively. In this particular case, the scaling procedure was done twice, with $T_{0}$ either at the upper or at the lower end of the covered temperature range. In order to compare the results obtained in these two cases, $H_{c2}(T)$ and $c_{0}(T)$ are normalized by their values at $T = T_{1} = 87$ K. As may be seen, the result is practically independent of the choice of $T_{0}$. The parameter $c_{0}(T)$ in Eq. (4) accounts for only a small correction to $M(H)$. This causes a much enhanced uncertainty in the values of $c_{0}(T)$ than that for the normalized upper critical field, as may easily be seen by comparing Figs. 2(a) and 2(b).
We note that the uncertainty of $h_{c2}(T)$ increases considerably for temperatures close to $T_{c}$ as well as for the lowest temperatures. The loss of accuracy for the highest temperatures is due to the obvious enhancement of the experimental uncertainty of the $M(H)$ data. Although the accuracy is improving with decreasing temperature, the increase of the irreversibility field limits the available magnetic field range as may clearly be seen in the inset to Fig. 1. If the experimental data are collected in a narrow magnetic field range only, our scaling procedure is not reliable.
The temperature dependence of the normalized upper critical field, as shown in Fig. 2(a), may also be used to evaluate the critical temperature $T_{c}$. For this purpose the ratio $H_{c2}(T)/H_{c2}(T_{1})$ was approximated by $${{H_{c2}(T)} \over {H_{c2}(T_1)}}={{1-(T/T_c)^\mu } \over {1-(T_1/T_c)^\mu }},$$ in which $\mu$ and $T_{c}$ are used as fit parameters. Eq. (6) provides a rather good approximation to $h_{c2}(T)$ curves for $T \ge 0.8T_{c}$. The corresponding fit is shown as the solid line in Fig. 2(a). The values of $\mu$ and $T_{c}$ are indicated in Fig. 2. [@52] If the experimental data were obtained up to temperatures rather close to the critical temperature, the extrapolated value of $T_{c}$ is quite accurate. A reliable value of $T_{c}$ is essential for the comparison of the results that were obtained for the samples with different critical temperatures. Using the values of $T_{c}$ evaluated in such a way, we have plotted $H_{c2}(T)/H_{c2}(0.9T_{c})$ versus $T/T_{c}$ as shown in Fig. 3(a). Quite surprisingly, the temperature variations of $H_{c2}$ for different Y-based compounds and different types of samples turn out to be identical. In the inset of Fig. 3(a) we display the low field magnetization curve $M(T)$ of sample Y$\# $5 and indicate the position of $T_{c}$ resulting from our extrapolation procedure with a vertical line.
The temperature variations of $h_{c2}$ for other families of HTSCs are plotted in Figs. 3(b) - 3(d). Similar to what has been found for Y-based compounds, the scaling procedure again leads to an almost perfect merging of all the data onto one single curve for different samples. Furthermore, as may clearly be seen in Fig. 4, the temperature dependencies of the normalized upper critical field for different families of HTSCs are practically identical at all temperatures for which the experimental data are available. We note that the insignificant differences between the $h_{c2}(T/T_{c})$ curves for different samples, visible at the lowest temperatures in Figs. 3(a), 3(b), and 4, are due to small errors in the determination of the critical temperature. For the data presented in Figs. 3(a-d) the relative errors in the determination of the critical temperature, $\Delta T_{c}/T_{c}$, are between 0.001 and 0.003, depending on the quality of the original experimental data. Although this uncertainty is quite small, it is sufficient to explain the observed differences between the $h_{c2}$ values at low temperatures.
=0.9
=0.9
Among the numerous samples listed in Tables I - IV, only for the oxygen deficient sample Y$\# $9, the $h_{c2}(T)$ curve is distinctly different. As may be seen in Fig 5, $h_{c2}(T)$ for this sample is perfectly linear in the entire covered temperature range, in striking difference to two other, over-doped and optimally doped Y-based samples (Y\#7 and Y\#8) investigated in the same study. The magnetization data for sample Y\#9 were collected in a very wide range of magnetic fields and, as may be seen in Fig. 6, our scaling leads again to a nearly perfect merging of the curves. The difference in $h_{c2}(T)$ between Y\#9 and other samples is thus not due to insufficient sample quality but rather reflects the intrinsic difference in properties of under-doped YBa$_{2}$Cu$_{3}$O$_{7-x}$ materials. Only very few magnetization studies of oxygen deficient YBa$_{2}$Cu$_{3}$O$_{7-x}$ single crystals or grain-aligned samples are available in the literature and we could find only two additional publications which are suitable for our analysis (samples Y\#10 and Y\#11). [@5; @53] Unfortunately, as may be seen in Fig. 5, the measurements reported in Refs. 5 and 53 were made at temperatures very close to $T_{c}$ and in this temperature range, the temperature dependence of $H_{c2}$ is linear for all HTSC materials.
=0.9
HTSCs are strongly anisotropic and it is well known that, depending on the orientation of the applied magnetic field, the absolute values of $H_{c2}$ differ significantly. This is why it is interesting to compare the results of our analysis for different orientations of the magnetic field. Unfortunately, we have found only two data sets from magnetization measurements that were made on the same samples but with two different orientations of the external magnetic field (Hg$\#$5 and Bi$\#$6). As may be seen in Fig. 3(c), the results for the grain-aligned sample Hg$\#$5 are practically independent of the orientation of the magnetic field. The situation for the single-crystal sample Bi$\# $6 is different. The resulting $h_{c2}(T)$ curves for this sample are shown in Fig. 7. The data perfectly match each other if we assume that the value of $T_{c}$ depends on the orientation of the magnetic field. This at first glance rather strange result may easily be understood if we recall our definition of $H_{c2}$ at the end of Section 2 and we discuss this point in the next paragraph.
=0.9
As has been described above, our procedure provides $H_{c2}$ as it enters the equation for the magnetization in the mixed state, i.e., the resulting $H_{c2}(T)$ curve in the $H$ - $T$ phase diagram represents the upper boundary for the existence of the mixed state. In a perfect type-II superconductor without fluctuations this definition of $H_{c2}$ coincides with the upper limit for the existence of superconductivity. In real HTSC materials the situation is different and superconducting regions may still exist in the sample even above the $H_{c2}(T)$ curve due to, for instance, thermal fluctuations or impurities with a higher $T_{c}$. We consider it as an advantage that such effects have practically no influence on our evaluation of $H_{c2}$. The situation is further complicated by the layered structure of HTSCs. As has previously been established by resistance measurements in zero magnetic field for Bi-based compounds, the superconducting coherence in the Cu-O planes sets in at a somewhat higher temperature than along the direction perpendicular to the planes. [@54; @55; @56; @57; @58] The same conclusion can be gained from results of magnetization measurements in magnetic fields of several Oersteds. [@47; @59; @60] This justifies the introduction of two critical temperatures $T_c^{(ab)}$ and $T_c^{(c)}$. Below $T_c^{(ab)}$, the superconducting phase coherence is established along the $ab$-planes, but only below $T_c^{(c)}$ supercurrents can propagate in the direction of the $c$-axis. In the temperature range $T_c^{(c)}<T<T_c^{(ab)}$, although superconductivity already exists in the $ab$-planes, no mixed state can be created in magnetic fields parallel to the planes. This occurs only at $T\le T_c^{(c)}$, i.e., in this case our evaluation of $T_{c}$ corresponds to $T_c^{(c)}$. If the magnetic field is parallel to the $c$-axis, the mixed state can be created already at $T_c^{(ab)}$. This simple picture gives a natural explanation for the difference in $T_c$ for the different orientations of the magnetic field that was obtained for the sample Bi\#6 (see Fig. 7). The observed difference in $T_{c}$ is quite small ($\Delta T_{c}/T_{c} \approx 0.01$) and may easily be masked, for instance, by grain misalignments in grain-aligned samples. This could be the reason why we do not see this effect in the sample Hg\#4. It is also possible that this difference in $T_{c}$ is of significant magnitude only in Bi-based cuprates due to their very special crystalline structure.
Discussion
==========
The scaling procedure based on Eq. (4) turns out to be rather successful for the analysis of the reversible magnetization of HTSCs. Figs. 1 and 6 demonstrate very well the scaling of isothermal magnetization data resulting in plots of the magnetization at a chosen temperature versus a renormalized magnetic field. The quality of scaling is remarkable for all cases that are listed in Tables I-IV and the mismatch between the $M(H)$ curves measured at different temperatures does not exceed the scatter of the original experimental data.
The most surprising result of our analysis is that for practically all families of HTSCs the $h_{c2}(T/T_{c})$ curves are virtually identical (Fig. 4). It is difficult to imagine that this universality of the $h_{c2}(T/T_{c})$ dependence is just a coincidence. We are of the opinion that the spectacular agreement between the $h_{c2}(T/T_{c})$ data for a great variety of different samples is an unambiguous evidence that our approach captures the essential features of the magnetization process of HTSCs. It does not necessarily mean, of course, that the Ginzburg-Landau parameter $\kappa$ is indeed temperature independent. The universality of $h_{c2}(T/T_{c})$ is preserved if the temperature dependence of $\kappa$ is the same for the different HTSC compounds studied here.
Our analysis is applicable only to reversible magnetization data and therefore, all the results and conclusions are limited to temperatures close to $T_{c}$. The lower limit of validity, $T_{\min}$, is quite different for different families of HTSCs, as may be seen in Fig. 4. The ratio $T_{\min}/T_{c}$, which depends on the strength of the pinning of vortices, is highest for the Y-based compounds that exhibit the strongest pinning forces.
The universality of the normalized temperature dependence of $H_{c2}$ implies that the normalized temperature variations of the thermodynamic critical field, $H_{c}$ for different HTSCs are also identical. Since $H_c^2/8\pi$ is the difference in the free energy densities between the normal and superconducting states, $H_{c}(T)$ also reflects the temperature dependence of the superconducting energy gap $\Delta$. [@6] In other words, our result that the normalized temperature dependence of $H_{c2}$ follows the same universal curve for different families of HTSCs implies that the normalized temperature variations of the energy gap $\Delta (T/T_{c})/\Delta (0)$ for different HTSCs are also identical, at least in the temperature ranges covered in this study.
We note that the temperature dependencies of $H_{c2}$ for HTSCs obtained as a result of our analysis are qualitatively very similar to those of conventional superconductors. They are linear at temperatures close to $T_{c}$ with a pronounced negative curvature at lower temperatures. Apparently, the positive curvature of $H_{c2}(T)$ for HTSCs, which is often reported in the literature, is due to the uncertainty of defining $H_{c2}$ in those studies.
Conclusion
==========
We have developed a scaling procedure that allows to obtain the temperature dependence of the upper critical field from the measurements of the reversible isothermal magnetization. If the magnetization measurements are extended up to temperatures close to the superconducting critical temperature, our procedure also allows for a fairly reliable evaluation of the zero-field critical temperature. We have applied this scaling procedure for the analysis of experimental data for high-$T$ superconductors available in the literature and have shown that the normalized temperature dependencies of $H_{c2}$ are qualitatively the same as those of conventional superconductors and we obtain the same universal curve for different families of HTSCs. This universality is a very strong indication that also the temperature dependence of the superconducting energy gap is the same for all cuprate superconductors. All these statements have been verified to be valid at all temperatures for which data of measurements of the reversible magnetization of different types of cuprate superconductors are available in the literature.
[lcccccc]{} & & & &\
Y\#1 & 46 & YBa$_{2}$Cu$_{3}$O$_{7-x}$ & single crystal & 91.1\
Y\#2 & 61 & YBa$_{2}$Cu$_{3}$O$_{7}$ & single crystal & 88.0\
Y\#3 & 17 & YBa$_{2}$Cu$_{3}$O$_{7-x}$ & grain-aligned & 92.0\
Y\#4 & 16 & YBa$_{2}$Cu$_{4}$O$_{8}$ & grain-aligned & 79.8\
Y\#5 & 62 & YBa$_{2}$Cu$_{3}$O$_{7-x}$ & single crystal & 93.0\
Y\#6 & 63 & (YCa)Pb$_{2}$Sr$_{2}$Cu$_{3}$O$_{8+x}$ & single crystal & 76.0\
Y\#7 & 11 & YBa$_{2}$Cu$_{3}$O$_{6.94}$ & grain-aligned & 92.9\
Y\#8 & 11 & YBa$_{2}$Cu$_{3}$O$_{7}$ & grain-aligned & 88.7\
Y\#9 & 11 & YBa$_{2}$Cu$_{3}$O$_{6.85}$ & grain-aligned & 79.9\
Y\#10 & 5 & YBa$_{2}$Cu$_{3}$O$_{6.65}$ & single crystal & 62.3\
Y\#11 & 53 & YBa$_{2}$Cu$_{3}$O$_{6.5}$ & single crystal & 44.8\
[lcccccc]{} & & & &\
Bi\#1 & 64 & Bi$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{2}$O$_{8+x}$ & single crystal & 84.0\
Bi\#2 & 65 & (BiPb)$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{3}$O$_{x}$ & whisker & 108.7\
Bi\#3 & 26 & (BiPb)$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{2}$O$_{8}$ & single crystal & 91.4\
Bi\#4 & 48 & Bi$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{2}$O$_{8}$ & single crystal & 88.2\
Bi\#5 & 66 & Bi$_{2.1}$Sr$_{1.7}$Ca$_{1.2}$Cu$_{2}$O$_{x}$ & single crystal & 86.7\
Bi\#6 & 8 & Bi$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{2}$O$_{8}$ & single crystal & 80.5\
Bi\#6 & 8 & Bi$_{2}$Sr$_{2}$Ca$_{2}$Cu$_{2}$O$_{8}$ & single crystal & 79.1\
[lcccccc]{} & & & &\
Hg\# 1 & 67 & HgBa$_2$CaCu$_2$O$_{6+x}$ & grain-aligned & 117.1\
Hg\# 2 & 68 & HgBa$_2$Ca$_2$Cu$_4$O$_{10+x}$ & grain-aligned & 123.1\
Hg\# 3 & 24 & Hg$_{0.7}$Pb$_{0.3}$Sr$_2$Ca$_2$Cu$_3$O$_x$ & grain-aligned & 125.5\
Hg\# 5 & 13 & HgBa$_2$Ca$_2$Cu$_3$O$_{8+x}$ & grain-aligned & 131.5\
Hg\# 5 & 13 & HgBa$_2$Ca$_2$Cu$_3$O8+x & grain-aligned & 131.5\
Hg\# 6 & 21 & Hg$_{1-y}$Pb$_{y}$Ba$_{2-z}$Sr$_{z}$Ca$_2$Cu$_3$O$_x$ & grain-aligned & 124.6\
Hg\# 7 & 71 & (HgCu)Ba$_2$CuO$_{4+x}$ & single crystal & 97.4\
[lcccccc]{} & & & &\
Tl\#1 & 15 & Tl$_2$Ba$_2$Ca$_2$Cu$_3$O$_{10}$ & grain-aligned & 114.6\
Tl\#2 & 69 & Tl$_2$Ba$_2$Ca$_2$Cu$_3$O$_{10+x}$ & grain-aligned & 122.8\
Tl\#3 & 70 & Tl$_{0.5}$Pb$_{0.5}$Sr$_2$CaCu$_2$O$_7$ & single crystal & 76.9\
Tl\#4 & 25 & TlBa$_2$Ca$_3$Cu$_4$O$_{11+x}$ & single crystal & 121.2\
[71]{}
U. Welp, W. K. Kwok, G. W. Crabtree, K. G. Vandervoort, and J. Z. Liu, Phys. Rev. Lett. [**62**]{}, 1908 (1989).
U. Welp, W. K. Kwok, G. W. Crabtree, K. G. Vandervoort, and J. Z. Liu, Phys. Rev. B [**40**]{}, 5263 (1989).
U. Welp, M. Grmsditch, H. You, W. K. Kwok, M. M. Fang, G. W. Crabtree, and J. Z. Liu, Physica C [**161**]{}, 1 (1989).
U. Welp, M. Grmsditch, H. You, W. K. Kwok, M. M. Fang, G. W. Crabtree, and J. Z. Liu, Physica B [**163**]{}, 473 (1990).
K. G. Vandervoort, U. Welp, J. E. Kessler, H. Claus, G. W. Crabtree, W. K. Kwok, A. Umezawa, B. W. Veal, J. W. Downey, A. P. Paulikas, and J. Z. Liu, Phys. Rev. B [**43**]{}, 13042 (1991).
A. A. Abrikosov, Zh. Eksp. Teor. Fiz. [**32**]{}, 1442 (1957) \[Sov. Phys. JETP [**5**]{}, 1174 (1957)\].
A. A. Abrikosov, Fundamentals of the theory of metals. North-Holland, Amsterdam, 1988.
W. Kritscha, F. M Sauerzopf, H. W. Weber, G. W. Garbtree, Y. C. Chang, and P. Z. Jiang, Physica C [**179**]{}, 59 (1991).
E. H. Brandt, Phys. Rev. Lett. [**78**]{}, 2208 (1997).
Z. Hao and J. R. Clem, M. W. McElfresh, L. Civale, A. P. Malozemoff, and F. Holtzberg, Phys. Rev. B [**43**]{}, 2844 (1991).
J. G. Ossandon, J. R. Thompson, D. K. Christen, B. C. Sales, H. R. Kerchner, J. O. Thomson, Y. R. Sun, K. W. Lay, and J. E. Tkaczyk, Phys. Rev. B [**45**]{}, 12534 (1992).
Qiang Li, M. Suenaga, T. Kimura, and K.Kishio, Phys. Rev. B [**47**]{}, 11384 (1993).
Y. C. Kim, J. R. Thompson, J. G. Ossandon, D. K. Christen, and M. Paranthaman, Phys. Rev. B [**51**]{}, 11767 (1995).
U. Welp, S. Fleshler, W. K. Kwok, J. Downey, G. W. Crabtree, H. Claus, A. Erb, and G. Müller-Vogt, Phys. Rev. B [**47**]{}, 12369 (1993).
D. N. Zheng, A. M. Campbell, and R. S. Liu, Phys. Rev. B [**48**]{}, 6519 (1993).
Junghyun Sok, Ming Xu, Wei Chen, B. J. Suh, J. Gohng, D.K. Finnemore, M.J. Kramer, L.A. Schwartzkopf, and B. Dabrowski, Phys. Rev. B [**51**]{}, 6035 (1995).
Junho Gohng and D. K. Finnemore, Phys. Rev. B [**46**]{}, 398 (1992).
D. N. Zheng, A. M. Campbell, and R. S. Liu, Phys. Rev. B [**48**]{}, 6519 (1993).
Ming Xu, Junghyun Sok, J. E. Ostenson, D. K. Finnemore, and B. Dabrowski, Phys. Rev. B [**53**]{}, 15313 (1996).
Mung-Seog Kim, Sung-Ik Lee, Seong-Cho Yu, and Nam H. Hur, Phys. Rev. B [**53**]{}, 9460 (1996).
Yi Zhuo, Jae-Hyuk Choi, Mung-Seog Kim, Wan-Seon Kim, Z. S. Lim, Sung-Ik Lee, and Sergey Lee, Phys. Rev. B [**55**]{}, 12719 (1997).
Yi Zhuo, Jae-Hyuk Choi, Mung-Seog Kim, Jin-Nam Park, Myoung-Kwang Bae, and Sung-Ik Lee, Phys. Rev. B [**56**]{}, 8381 (1997).
J. R. Thompson, J. G. Ossandon, D. K. Christen, M. Paranthaman, E. D. Specht, and Y. C. Kim, Phys. Rev. B [**54**]{}, 7505 (1996).
Yi Zhuo, Su-Mi Oh, Jae-Hyuk Choi, Mun-Seog Kim, Sung-Ik Lee, N. P. Kiryakov, M. S. Kuznetsov, and Sergey Lee, Phys. Rev. B [**60**]{}, 13094 (1999).
Lu Zhang, J. Z. Liu, and R. N. Shelton, Solid State Comm. [**109**]{}, 761 (1999).
Lu Zhang, J. Z. Liu, and R. N. Shelton, Phys. Rev. B [**45**]{}, 4978 (1992).
A. A. Nugroho, I. M. Sutjahja, A. Rusydi, M. O. Tjia, A. A. Menovsky, F. R. de Boer, and J. J. M. Franse, Phys. Rev. B [**60**]{}, 15384 (1999).
W. V. Pogosov, K. I. Kugel, and A. L. Rakhmanov, Zh. Eksp. Teor. Fiz. [**118**]{}, 676 (2000) \[JETP [**91**]{}, 588 (2000)\].
W. V. Pogosov, K. I. Kugel, A. L. Rakhmanov, and E. H. Brandt, Phys. Rev. B [**64**]{}, 064517 (2001).
T. T. M. Palstra, B. Batlogg, L. F. Schneemeyer, R. B. van Dover, and J. V. Waszczak, Phys. Rev. B [**38**]{}, 5102 (1988).
M. J. Naughton, R. C. Yu, P. K. Davies, J. E. Fischer, R. V. Chamberlin, Z. Z. Wang, T. W. Jing, N. P. Ong, and P. M. Chaikin, Phys. Rev. B [**38**]{}, 9280 (1988).
Ichiro Matsubara, Hideo Tanigawa, Toru Ogura, Hiroshi Yamashita, Makoto Kinoshita, and Tomoji Kawai, Phys. Rev. B [**45**]{}, 7414 (1992).
A. S. Alexandrov, V. N. Zavaritsky, W. Y. Liang, and P. L. Nevsky, Phys. Rev. Lett [**76**]{}, 983 (1996).
X.-J. Xu, L. Fu, and Y.-H. Zhang, Physica C [**282-287**]{}, 1557 (1997).
M. N. Khlopkin, G. Kh. Panova, N. A. Chernoplekov, A. A. Shikov, and A. V. Suetin, Zh. Eksp. Teor. Fiz. [**112**]{}, 1386 (1997) \[JETP [**85**]{}, 755 (1997)\].
V. F. Gantmakher and G. E. Tsydynzhapov, Zh. Eksp. Teor. Fiz. [**115**]{}, 268 (1999) \[JETP [**88**]{}, 148 (1999).
S. I. Vedeneev, A. G. M. Jansen, E. Haanappel, and P. Wyder, Phys. Rev. B [**60**]{}, 12467 (1999).
Yoichi Ando, G. S. Boebinger, A. Passner, L. F. Schneemeyer, T. Kimura, M. Okuya, S. Watauchi, J. Shimoyama, K. Kishio, K. Tamasaku, N. Ichikawa, and S. Uchida, Phys. Rev. B [**60**]{}, 12475 (1999).
V. F. Gantmakher, G. A. Emelchenko, I. G. Naumenko, and G. E. Tsydynzhapov, PisÕma Zh. Eksp. Teor. Fiz. [**72**]{}, 33 (2000) \[JETP Letters [**72**]{}, 21 (2000)\].
V. N. Zavaritsky, V. V. Kabanov, and A. S. Alexandrov, cond-mat/0204012.
L. P. Gorkov, Zh. Eksp. Teor. Fiz. [**37**]{}, 833 (1959) \[Soviet Phys.-JETP [**10**]{}, 593 (1960)\].
E. Helfand and N. R. Werthamer, Phys Rev. [**147**]{}, 288 (1966).
Yung M. Huh, J. E. Ostenson, F. Borsa, V. G. Kogan, D. K. Finnemore, A. Vietkin, A. Revcolevschi, M.-H. Julien, Phys. Rev. B [**63**]{}, 064512 (2001).
O. Jeandupeux, A. Schilling,H. R. Ott, and A. vanOtterlo, Phys. Rev. B [**53**]{}, 12475 (1996).
Qiang Li, K. Shibutani, M. Suenaga, I. Shigaki, and R. Ogawa, Phys. Rev. B [**48**]{}, 9877 (1993).
J. R. Cooper, J. W. Loram, J. D. Johnson, J. W. Hodby, and Chen Changkang, Phys. Rev. Lett. [**79**]{}, 1730 (1997).
G. Le Bras, L. Fruchter, V. Vulcanescu, V. Viallet, A. Bertinotti, A. Forget, J. Hammann, J.-F. Marucco, and D. Colson, Physica C [**271**]{}, 205 (1996).
E. G. Miramontes, J. A. Campá, A. Pomar, I. Rasines, C. Torrón, J. A. Veira, and Félix Vidal, Physica C [**235-240**]{}, 2931 (1994).
R. Jin, A. Schilling, and H. R. Ott, Phys. Rev. B [**49**]{}, 9218 (1994).
This function has been chosen only because it fits the experimental $M(H)$ data quite well and it may be replaced by any other suitable function.
The measurements at different temperatures usually cover different ranges of the normalized magnetic field $H/h_{c2}$. This is why it is important that only those data points at $T = T_{i}$, for which the values of $H/h_{c2}$ are inside the magnetic field range covered at $T = T_{0}$, are considered in the fitting procedure.
The error margins indicated in this paper are solely due to the approximation. Taking into account experimental errors may enhance the error margins considerably.
B. Rosenstein, B. Ya. Shapiro, R. Prozorov, A. Shaulov, and Y. Yeshurun, Phys. Rev. B [**63**]{}, 134501 (2001).
G. Briceño, M. F. Crommie, and A Zettl, Phys. Rev. Lett. [**66**]{}, 2164 (1991).
J. H. Cho, M. P. Maley, S. Fleshler, A. Lacerda, and L. N. Bulaevskii, Phys. Rev. B [**50**]{}. 6493 (1994).
Xiao-Guang Li, Xuefeng Sun, Wenbin Wu, Qianwang Chen, Lei Shi, Yuheng Zhang, Y. Kotaka, and K. Kishio, Physica C [**279**]{}, 241 (1997).
K. Nakao, Yu. Eltsev, J. G. Wen, S. Shibata, and N. Koshizuka, Physika C, [**322**]{}, 79 (1999).
T. Fujii, I. Terasaki, T. Watanabe, and A. Matsuda, Cond-mat/0205121.
H. Claus, G. W. Crabtree, K. G. Vandervoort, J. Z. Liu, D. Lam, E. D. Bukowski, and D. M. Ginsberg, J. Appl. Phys. [**87**]{}, 5049 (1990).
A. Bertinotti, V. Viallet, D. Colson, J.-F. Marucco, J. Hammann, G. Le Bras, A Forget, Physica C [**268**]{}, 257 (1996).
Terukazu Nishizaki, Kenji Shibata. Takahiko Sasaki, and Norio Kobayashi, Physica C [**341-348**]{}, 957 (2000)
N. Kobayashi, T. Nishizaki, K. Shibata, T. Sato, M. Maki, and T. Sasaki, Physica C [**362**]{}, 121 (2001).
M. Reedyk, C. V. Stager, T. Timusk, J. S. Xue, and J. E. Greedan, Phys. Rev. B [**44**]{}, 4539 (1991).
Kayuo Kadowaki and Kazuhiro Kimura, Phys. Rev. B [**57**]{}, 11674 (1998).
Ichiro Matsubara, Ryoji Funahashi, Kazuo Ueno, Hirishi Yamashita, and Tomoji Kawai, Physica C [**256**]{}, 33 (1996).
A. Junod, K.-W, Wang, T. Tsukamoto, G. Triscone, B. Revaz E. Walker, and J. Muller, Physica C [**229**]{}, 209 (1994).
M. Y. Cheon, G. C. Kim, B. J. Kim, and Y. C. Kim, Physica C [**302**]{}, 215 (1998).
Mun-Seog Kim, Sung-Ik Lee, Seong-Cho Yu, Irina Kuzemskaya, Efim S. Itskevich, and K. A. Lokshin, Phys. Rev. B [**57**]{}, 6121 (1998).
J. R. Thompson, D. K. Christen and J. G. Ossandon, Physica B [**194-196**]{}, 1557 (1994).
A. Maignan, C. Martin, V. Hardy, and Ch. Simon, Physica C [**228**]{}, 323 (1994).
G. Villarad, A. Daignere, D. Pelloquin, A. Maignan, Physica C [**314**]{}, 196 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper is concerned with the existence of the *simultaneous route-and-departure choice dynamic user equilibrium* (SRDC-DUE) in continuous time. The SRDC-DUE problem was formulated as an infinite-dimensional variational inequality in Friesz et al. (1993). In deriving our existence result, we employ the *generalized Vickrey model* (GVM) introduced in Han et al. (2012a, 2012b) to formulate the underlying network loading problem. As we explain, the GVM corresponds to a path delay operator that is provably strongly continuous on the Hilbert space of interest. Finally, we provide the desired SRDC-DUE existence result for general constraints relating path flows to a table of fixed trip volumes without invocation of bounds on the path flows.'
author:
- |
Ke Han$^{a}\thanks{e-mail: [email protected];}$ Terry L. Friesz$^{b}\thanks{e-mail: [email protected];}$ Tao Yao$^{b}\thanks{e-mail: [email protected];}$\
\
$^{a}$*Department of Mathematics,*\
*Pennsylvania State University, PA 16802, USA*\
$^{b}$*Department of Industrial and Manufacturing Engineering,*\
*Pennsylvania State University, PA 16802, USA*
title: Existence of Simultaneous Route and Departure Choice Dynamic User Equilibrium
---
\[Intro\]Introduction
=====================
In this paper we shall consider *dynamic traffic assignment* (DTA) to be the positive (descriptive) modeling of time-varying flows of automobiles on road networks consistent with established traffic flow theory and travel demand theory. *Dynamic User Equilibrium* (DUE) is one type of DTA wherein effective unit travel delay for the same purpose is identical for all utilized path and departure time pairs. The relevant notion of travel delay is effective unit travel delay, which is the sum of arrival penalties and actual travel time. For our purposes in this paper, DUE is modeled for the within-day time scale based on fixed travel demands established on a day-to-day time scale.
In the last two decades there have been many efforts to develop a theoretically sound formulation of dynamic network user equilibrium that is also a canonical form acceptable to scholars and practitioners alike. DUE models tend to be comprised of four essential sub-models:
1. a model of path delay;
2. flow dynamics;
3. flow propagation constraints;
4. a path/departure-time choice model.
Furthermore, analytical DUE models tend to be of two varieties: (1) *route choice* (RC) user equilibrium (Friesz et al. 1989, Merchant and Nemhauser 1978a, 1978b, Mounce 2006, Smith and Wisten 1995, Zhu and Marcotte 2000) ; and (2) *simultaneous route-and-departure* *choice* (SRDC) dynamic user equilibrium (Friesz et al. 1993, 2010, 2012, Ran et al. 1996, Wie et al. 2002). For both types of DUE models, the existence of a dynamic user equilibrium in continuous time remains a fundamental issue. A proof of DUE existence is a necessary foundation for qualitative analysis and computational studies. In this paper, we provide a DUE existence result for the SRDC DUE problem when it is formulated as an infinite-dimensional variational inequality of the type presented in Friesz et al. (1993). In this paper, in order to establish a DUE existence result, we study the network loading problem based on the generalized Vickrey model (GVM) proposed in Han et al. (2012a, 2012b). All of our results presented in this paper are more general than any obtained previously for DUE when some version of the point queue model is employed.
Formulation of the SRDC user equilibrium {#introformulation}
----------------------------------------
There are two essential components within the RC or SRDC notions of DUE: (i) the mathematical expression of Nash-like equilibrium conditions, and (ii) a network performance model, which is, in effect, an embedded network loading problem. The embedded network loading problem captures the relationships among arc entry flow, arc exit flow, arc delay and path delay for any path departure rate trajectory. Note that, by studying the embedded network loading problem based on the GVM, we are not suggesting or employing a sequential solution paradigm for DUE. If the reader believes a sequential perspective is implicit in our work, he/she has not fully understood the mathematical presentation provided in this and subsequent sections.
There are multiple means of expressing the Nash-like notion of a dynamic equilibrium, including the following:
1. a variational inequality (Friesz et al. 1993; Smith and Wisten 1994, 1995)
2. an equilibrium point of an evolution equation in an appropriate function space (Mounce 2006; Smith and Wisten 1995)
3. a nonlinear complementarity problem (Wie et al. 2002; Han et al. 2011)
4. a differential variational inequality (Friesz et al. 2001, 2010, 2012; Friesz and Mookherjee 2006); and
5. a differential complementarity system (Pang et al. 2011).
The variational inequality representation is presently the primary mathematical form employed for both RC and SRDC DUE. The most obvious approach to establishing existence for any of the mathematical representations mentioned above is to convert the problem to an equivalent fixed point problem and then apply Brouwer’s fixed point existence theorem. Alternatively, one may use an existence theorem for the particular mathematical representation selected; it should be noted that most such theorems are derived by using Brouwer’s famous theorem. So, in effect, all proofs of DUE existence employ Brouwer’s fixed point theorem, either implicitly or explicitly. One statement of Brouwer’s theorem appears as Theorem 2 of Browder (1968). Approaches based on Brouwer’s theorem require the set of feasible path flows (departure rates) under consideration to be compact and convex in a Banach space, and typically involve an bound on each path flow.
We also wish to point out that this paper employs much more general constraints relating path flows to a table of fixed trip volumes than has been previously considered when studying SRDC DUE. Moreover, in our study of existence, we do not invoke bounds on the path flows to assure boundedness needed for application of Brouwer’s theorem. That is, a goal of this paper is to investigate the existence of DUE without making the assumption of *a priori* bounds for departure rates. Note should be taken of the following fact: the boundedness assumption is less of an issue for the RC DUE by virtue of problem formulation; that is, for RC DUE, the travel demand constraints are of the following form: $$\sum_{p\in \mathcal{P}_{ij}}h_{p}(t)~=R_{ij}(t),\qquad \forall ~t,\quad
\forall ~(i,\,j)\in \mathcal{W} \label{introeqn1}$$where $\mathcal{W}$ is the set of origin-destination pairs, $\mathcal{P}_{ij}
$ is the set of paths connecting $\left( i,j\right) \in \mathcal{W}$ and $
h_{p}(t)$ is the departure rate along path $p$. Furthermore, $R_{ij}(t)$ represents the rate (not volume) at which travelers leave origin $i$ with the intent of reaching destination $j$ at time $t$; each such trip rate is assumed to be bounded from above. Since (\[introeqn1\]) is imposed pointwise and every path flow $h_{p}$ is nonnegative, we are assured that each $h=\left( h_{p}:\left( i,j\right) \in \mathcal{W},p\in \mathcal{P}_{ij}\right) $ are automatically uniformly bounded. On the other hand, the SRDC user equilibrium imposes the following constraints on path flows: $$\sum_{p\in \mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}h_{p}(t)\,dt~=~Q_{ij},\qquad
\forall ~(i,\,j)\in \mathcal{W} \label{introeqn2}$$where $Q_{ij}\in \mathcal{\Re }_{+}^{1}$ is the volume (not rate) of travelers departing node $i$ with the intent of reaching node $j$. The integrals in (\[introeqn2\]) are interpreted as Legesgue; hence, ([introeqn2]{}) alone is not enough to assure bounded path flows. This observation has been the major hurdle to proving existence without the invocation of bounds on path flows. In this paper, we will overcome this difficulty through careful analysis of the GVM and by investigating the effect of user behavior in shaping network flows, in a mathematically intuitive yet rigorous way.
Importance of the path delay operator
-------------------------------------
Clearly another key component of continuous-time DUE is the path delay operator, typically obtained from *dynamic network loading* (DNL), which is a subproblem of a complete DUE model[^1]. Any DNL must be consistent with the established path flows and link delay model, and DNL is usually performed under the *first-in-first-out* (FIFO) rule. The properties of the delay operator are critical to proving existence of a solution to the infinite-dimensional variational inequality used to express DUE. In Zhu and Marcotte (2000), using the *link delay model* introduced by Friesz et al. (1993), the authors showed weak continuity of the path delay operator under the assumption that the path flows are *a priori* bounded. Their continuity result is superceded by a more general result proven in this paper: the path delay operator of interest is strongly continuous without the assumption of boundedness. Strong continuity without boundedness is central to our proof of existence in the present paper.
In this paper, as a foundation for DNL, we will consider Vickrey’s model of congestion first introduced in Vickrey (1969) and later studied by Han et al. (2012a, 2012b). Vickrey’s model for one link is primarily described by an *ordinary differential equation* (ODE) with discontinuous right hand side. Such irregularity has made it difficult to analyze Vickery’s model in continuous time. Fortunately, in this paper, we will be able to take advantage of the reformulation proposed in Han et al. (2012a, 2012b), then prove the strong continuity of the path delay operator without boundedness of the path flows. This will provide a quite general existence proof for SRDC-DUE based on the GVM.
Organization
------------
The balance of this paper is organized as follows. Section \[Prel\] provides essential mathematical background on the concepts that will be used in the paper. Section \[DUE\] briefly reviews the formal definition of dynamic user equilibrium and its formulation as a variational inequality. Section \[DNL\] recaps the generalized Vickrey model (GVM) originally put forward by Han et al. (2012a, 2012b). Section \[dueexistence\] formally discusses the properties of the effective delay operator. The main result of this paper, the existence of an SRDC-DUE when the GVM informs network loading is established in Theorem \[gvmthm\] of Section \[esec\].
\[Prel\] Mathematical preliminaries
===================================
A *topological vector space* is one of the basic structures investigated in functional analysis. Such a space blends a topological structure with the algebraic concept of a vector space. The following is a precise definition:
\[tvdef\] **(Topological vector space)** A topological vector space $X$ is a vector space over a topological field $\mathbb{F}$ (usually the real or complex numbers with their standard topologies) which is endowed with a topology such that vector addition $X\times X\rightarrow X$ and scalar multiplication $\mathbb{F}\times X\rightarrow X$ are continuous functions.
As a consequence of Definition \[tvdef\], all normed vector spaces, and therefore all Banach spaces and Hilbert spaces, are examples of topological vector spaces. Also important is the notion of a seminorm:
\[seminormdef\] **(Seminorm)** A seminorm on a vector space $X$ is a real-valued function $p$ on $X$ such that
- $p(x+y)~\leq~p(x)+p(y)$
- $p(\alpha\,x)~=~|\alpha|\,p(x)$
for all $x$ and $y$ in $X$ and all scalars $\alpha$.
\[lctvdef\] **(Locally convex space)** A locally convex space is defined to be a vector space $X$ along with a family of seminorms $\{p_i\}_{i\in\mathcal{I}}$ on $X$.
As part of our review we make note of the following essential knowledge:
**Fact 1.** The space of square-integrable real-valued functions on a compact interval $[a,\,b]$, denoted by $\mathcal{L}^{2}([a,\,b])$, is a locally convex topological vector space.
**Fact 2.** The $m$-fold product of the spaces of square-integrable functions $\mathcal{L}^{2}([a,\,b])^{m}$ is a locally convex topological vector space.
\[dualdef\] **(Dual space)** The dual space $X^{\ast }$ of a vector space $X$ is the space of all continuous linear functions on $X$
Another key property we consider without proof is:
**Fact 3.** The dual space of $L^{p}(\mu )$ for $1<p<\infty $ has a natural isomorphism with $L^{q}(\mu )$ where $q$ is such that $1/p+1/q=1$. In particular, the dual space of $\mathcal{L}^{2}([a,\,b])$ is again $\mathcal{L}^{2}([a,\,b])$.
The key foundation for analysis of existence is the following theorem given in Browder (1968):
\[mainthm\] Let $K$ be a compact convex subset of the locally convex topological vector space $E$, $T$ a continuous (single-valued) mapping of $K$ into $E^{\ast }$. Then there exits $u_{0}$ in $K$ such that $$\Big<T(u_{0}),\,u_{0}-u\Big>~\geq ~0$$for all $u\in K$.
Let us now give the formal definition of a variational inequality in a topological setting:
**(Infinite-Dimensional Variational inequality)** Let $V$ be a topological vector space and $F:\,U\times \mathcal{\Re }_{+}^{1}\rightarrow V
$, where $U\subset V$. The infinite-dimensional variational inequality is posed as the following problem $$\left.
\begin{array}{c}
\hbox{find}~u^{\ast }\in U \\
\left\langle F(u),\,u-u^{\ast }\right\rangle ~\geq ~0~~\forall ~u\in U\end{array}\right\} ~VI(F,\,U) \label{videf}$$
**(Compactness of subspaces)** A subset $K$ of a topological space $X$ is called compact if for every arbitrary collection $\left\{ U_{\alpha
}\right\} _{\alpha \in A}$ of open subsets of $X$ such that $$K~\subset ~\bigcup_{\alpha \in A}U_{\alpha }$$there is a finite subset $I$ of $A$ such that $$K~\subset ~\bigcup_{i\in I}U_{i}$$
**(Sequential compactness)** A topological space is sequentially compact if every sequence has a convergent subsequence.
An outgrowth of the concepts and results given above, the following fact is stated without proof:
**Fact 4.** In metric space (hence topological vector space), the notions of compactness and sequential compactness are equivalent.
The final bit of specialized knowledge about topological vector spaces that we shall need is the following:
**(Weak convergence in Hilbert space)** A sequence of points $\left\{
x_{n}\right\} $ in a Hilbert space $\mathcal{H}$ is said to be convergent weakly to a point $x\in \mathcal{H}$, denoted as $x_{n}\rightharpoonup y$ if $$\left\langle x_{n},\,y\right\rangle ~\rightarrow ~\left\langle
x,\,y\right\rangle$$for all $y\in \mathcal{H}$, where $\left\langle \cdot ,\,\cdot \right\rangle
$ is the inner product on the Hilbert space.
\[DUE\] Continuous-time dynamic user equilibrium
================================================
In this section, we will assume the time interval of interest is $$\lbrack t_{0},\,t_{f}]\subset \mathcal{\Re }_{+}^{1}$$The most crucial component of the DUE model is the path delay operator, which provides the time to traverse any path $p$ per unit of flow departing from the origin of that path. The delay operator is denoted by $$D_{p}(t,\,h)\qquad \forall ~p\in \mathcal{P}$$where $\mathcal{P}$ is the set of all paths employed by network users, $t$ denotes the departure time, and $h$ is a vector of departure rates. Throughout the rest of the paper, we stipulate that $$h\in \Big(\mathcal{L}_{+}^{2}([t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$$where $\Big(\mathcal{L}_{+}^{2}[t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$ denotes the non-negative cone of the $|\mathcal{P}|$-fold product of the Hilbert space $\mathcal{L}^{2}[t_{0},\,t_{f}]$ of square-integrable functions on the compact interval $[t_{0},\,t_{f}]$. The inner product of the Hilbert space $\Big(\mathcal{L}^{2}[t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$ is defined as $$\big<u,\,v\big>~\doteq ~\int_{t_{0}}^{t_{f}}\big(u(s)\big)^{T}\,v(s)\,ds
\label{nnorm}$$where the superscript $T$ denotes transpose of vectors. Moreover, the norm $$\big\|u\big\|_{\mathcal{L}^{2}}~\doteq ~\big<u,\,u\big>^{1/2} \label{l2norm}$$is induced by the inner product (\[nnorm\]).
Next, we need to consider a more general notion of travel cost that will motivate on-time arrivals. To this end, for each $p\in \mathcal{P}$, we introduce the effective unit path delay operator $\Psi
_{p}:[t_{0},\,t_{f}]\times \Big(\mathcal{L}_{+}^{2}([t_{0},\,t_{f}])\Big)^{|\mathcal{P}|}\rightarrow \mathcal{\Re }_{++}^{1}$ and define it as follows: $$\Psi _{p}(t,\,h)~\equiv ~D_{p}(t,\,h)+\mathcal{F}\Big(t+D_{p}(t,\,h)-T_{A}\Big) \label{phidef}$$where $\mathcal{F}(\cdot )$ is the penalty for early or later arrival relativerelative to the target arrival time $T_{A}$. Note that, for convenience, $T_{A}$ is assumed to be independent of destination. However, that assumption is easy to relax, and the consequent generalization of our model is a trivial extension. We interpret $\Psi _{p}(t,\,h)$ as the perceived travel cost of driver starting at time $t$ on path $p$ under travel conditions $h$. Presently, our only assumption on such costs is that for each $h\in \Big(\mathcal{L}_{+}^{2}([t_{0},\,t_{f}])\Big)^{|\mathcal{P}|}
$, the vector function $\Psi (\cdot ,\,h):[t_{0},\,t_{f}]\rightarrow
\mathcal{\Re }_{++}^{\left\vert \mathcal{P}\right\vert }$ is measurable and stictly positive. The assumption of measurability was used for a measure theory-based argument in Friesz et al. (1993). Later in this paper, we shall discuss other properties of this operator, such as continuity on a Hilbert space. The continuity of effective delay is crucial for applying the general theorems in Browder (1968), especially Theorem \[mainthm\] stated above.
To support the development of a dynamic network user equilibrium model, we introduce some additional constraints. Foremost among these are the flow conservation constraints $$\sum_{p\in \mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}h_{p}(t)\,dt~=~Q_{ij}\qquad
\forall ~(i,\,j)\in \mathcal{W} \label{flowcons}$$where $\mathcal{P}_{ij}$ is the set of all paths that connect origin-destination (O-D) pair $(i,\,j)\in \mathcal{W}$, while $\mathcal{W}$ is the set of all O-D pairs. In addition, $Q_{ij}$ is the fixed travel demand for O-D pair $(i,\,j)$. Using the notation and concepts we have thus far introduced, the set of feasible solutions for DUE when the effective delay operator $\Psi (\cdot ,\,\cdot )$ is given is $$\Lambda ~=~\left\{ h\in \Big(\mathcal{L}_{+}^{2}([t_{0},\,t_{f}])\Big)^{|\mathcal{P}|}:\quad \sum_{p\in \mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}h_{p}(t)\,dt~=~Q_{ij}\quad \forall ~(i,\,j)\in \mathcal{W}\right\} \label{feasible}$$Using a presentation very similar to the above, the notion of a dynamic user equilibrium in continuous time was first introduced by Friesz et al. (1993), who employ a definition tantamount to the following:
\[duedef\]**(Dynamic user equilibrium)**. A vector of departure rates (path flows) $h^{\ast }\in \Lambda $ is a dynamic user equilibrium if $$h_{p}^{\ast }\left( t\right) >0,p\in \mathcal{P}_{ij}\Longrightarrow \Psi
_{p}\left[ t,h^{\ast }\left( t\right) \right] =v_{ij}\in \Re _{++}^{1}\text{
\ \ \ }\forall (i,\,j)\in \mathcal{W} \label{defdue}$$We denote the dynamic user equilibrium defined this way by $DUE\left( \Psi
,\Lambda ,\left[ t_{0},t_{f}\right] \right) $.
In the analysis to follow, we focus on the following infinite-dimensional variational inequality formulation of the DUE problem reported in Theorem 2 of Friesz et al. (1993).
$$\left.
\begin{array}{c}
\text{find }h^{\ast }\in \Lambda \text{ such that} \\
\sum\limits_{p\in \mathcal{P}}\displaystyle\int\nolimits_{t_{0}}^{t_{f}}\Psi
_{p}(t,h^{\ast })(h_{p}-h_{p}^{\ast })dt\geq 0 \\
\forall h\in \Lambda
\end{array}\right\} VI\Big(\Psi ,\Lambda ,[t_{0},t_{f}]\Big) \label{duevi}$$
The variational inequality formulation $VI\big(\Psi ,\,\Lambda
,[t_{0},\,t_{f}]\big)$ expressed above subsumes almost all DUE models regardless of the arc dynamics or network loading models employed provided $\Psi _{p}(t,h^{\ast })$
The dynamic network loading {#DNL}
===========================
A key ingredient of the variational inequality formulation of the DUE (\[duevi\]) is the effective delay operator $\Psi (t,\,\cdot )$, which maps a vector of admissible departure rates to the vector of strictly travel costs associated with each route-and-departure-time choice. The problem of predicting time-varying network flows consistent with known travel demands and departure rates (path flows) is usually referred to as the *dynamic network loading* (DNL) problem. Since effective path delays are constructed from arc delays that depend on arc activity and performance, DNL is intertwined with the determination of effective delay operators.
In this section we present a continuous-time DNL model. This model is based on a reformulation of Vickrey’s model (Vickrey 1969) that we refer to as the *generalized Vickrey model* (GVM); it was apparently first proposed in Han et al. (2012a, 2012b). The generalized Vickrey model determines arc exit flow and the arc traversal time from arc entry flow. This formulation not only leads to a simple and explicit computational scheme, but also makes it easier to conduct rigorous analyses of the arc delay operator and, hence, of the effective path delay operator $\Psi (t,\,\cdot )$.
The generalized Vickrey model
-----------------------------
First introduced in Vickrey (1968), the Vickrey’s model is based on two key assumptions: (i) vehicles have negligible sizes, and, therefore, any queue is of negligible size; and (ii) link travel time consists of a fixed travel time plus a congestion-related arc-traversal delay. Let us introduce the following notation: $$\begin{aligned}
u(t):& \quad \hbox{the flow profile into the link} \\
M~:& \quad \hbox{the flow capacity of the link} \\
q(t):& \quad \hbox{the queue size} \\
w(t):& \quad \hbox{the link exit flow} \\
T~:& \quad \hbox{the fixed free flow travel time} \\
D(t):& \quad
\hbox{the link traverse time of drivers entering the link
at}~t\end{aligned}$$Then the model is described by the following set of equations. $$w(t)~=~\begin{cases}
\min \left\{ u(t-T),\,M\right\} \qquad & q(t)~=~0 \\
M & q(t)~\neq ~0\end{cases}
\label{pqm1}$$$${\frac{dq(t)}{dt}}~=~u(t-T)-w(t) \label{pqm2}$$$$D(t)~=~T+{\frac{q(t+T)}{M}} \label{pqm3}$$Notice that (\[pqm1\]) and (\[pqm2\]) amount to an ordinary differential equation (ODE) with a right hand side that is discontinuous in the state variable $q(\cdot )$. Such ODEs have been the main hurdle to further analysis and computation of this model in continuous time. In Han et al. (2012a, 2012b), a reformulation of Vickrey’s model as a Hamilton-Jacobi equation was proposed and solved with a version of the Lax-Hopf formula (see Evans 2010 for more detailed on general Hamilton-Jacobi equation and Lax-Hopf formula). As a result, the solution to (\[pqm1\])-(\[pqm3\]) was presented in closed form. Due to the space limitation, we will omit details of the derivation of explicit solution to Vickrey’s model and refer the reader to Han et al. (2012a, 2012b) for details.
Let us next introduce the cumulative entering vehicle count $U(\cdot )$ and the exiting vehicle count $\mathcal{W}(\cdot )$ at the entrance and exit of the link of interest, respectively. Furthermore, $U(\cdot )$ is assumed to be non-decreasing and left continuous. Notice that these latter assumption imply that the link entry flows can be unbounded and possibly contain the dirac delta function. In contrast, the Vickrey’s original model requires that the entry flow to be at least Lebesgue integrable. As such, the GVM is more general than Vickrey’s model.
Using the notation introduced previously, an equivalent statement of ([pqm1]{}) and (\[pqm2\]) is the following: $$\begin{aligned}
\mathcal{W}(t)& ~=~\min_{\tau \leq t-T}\left\{ U(\tau )-M\,\tau \right\}
+M(t-t_{0}) \label{pqm4} \\
q(t)& ~=~U(t)-M\,t-\min_{\tau \leq t-T}\left\{ U(\tau )-M\,\tau \right\} \label{pqm5} \\
D(t)& ~=~T+{\frac{q(t+T)}{M}} \label{pqm6}\end{aligned}$$Note that all the unknowns of system (\[pqm4\])-(\[pqm6\]) may be explicitly stated in terms of the cumulative vehicle count $U(\cdot )$. The system (\[pqm4\])-(\[pqm6\]) will serve as the mathematical formulation of link dynamics in the DNL, as we shall explain shortly. The system ([pqm4]{})-(\[pqm6\]) may also be used for deriving mathematical properties of the effective path delay operator, as is demonstrated in Section [psicont]{}.
The network model
-----------------
It is straightforward to extend the generalized Vickrey model to a network, which is represented as a directed graph $G(N,\,A)$, where $N$ and $A$ are the set of nodes and arcs, respectively. In order to proceed, we introduce some additional notations. In particular, for each node $v\in N$, let $
\mathcal{I}^{v}$ be the set of incoming links, $\mathcal{O}^{v}$ the set of outgoing links. For each arc $a\in A$, let $u_a(t)$, $w_a(t)$ be the entry flow and exit flow, respectively. The arc entry/exit flows are the sum of entry/exit flows associated with individual paths using this arc; that is $$\label{disaggregate1}
u_a(t)~=~\sum_{p\in\mathcal{P}} \delta_{ap}\,u_a^p(t), \qquad w_a(t)~=~\sum_{p\in\mathcal{P}} \delta_{ap}\,w_a^p(t),\qquad \forall~a\in A$$ where $$\delta_{ap}~=~\begin{cases} 1 \qquad &\hbox{if arc}~a~\hbox{belongs to path}~p\\
0\qquad & \hbox{otherwise}
\end{cases}$$ Let us also define the cumulative entering vehicle count $U_a(t)$ and cumulative exiting vehicle count $W_a(t)$, for each arc $a$. Similarly, each one is disaggregated into quantities associated with each path that uses this arc: $$\label{disaggregate2}
U_a(t)~=~\sum_{p\in\mathcal{P}} \delta_{ap}\,U_a^p(t), \qquad W_a(t)~=~\sum_{p\in\mathcal{P}} \delta_{ap}\,W_a^p(t),\qquad \forall~a\in A$$
The arc delay $D_a(t)$ is the time taken to traverse the arc $a$ when the time of entry is $t$. The arc exit time function $\tau_a(t)$ is defined as $\tau_a(t) \doteq t+D_a(t)$, that is, $\tau_a(t)$ represents the exit time on arc $a$ when the time of entry is $t$.
For each group of drivers using the arc, the ratio of their arrival and departure rates must be the same under FIFO. This is expressed as $$w^{p}_{a}\big(\tau_{a}(t)\big)~=~\begin{cases}
\displaystyle w_{a}\big(\tau_{a}(t)\big)\cdot {\frac{u^{p}_{a}(t)}{u_{a}(t)}},\qquad & \hbox{if}~~u_{a}(t)~\neq ~0 \\
0\qquad & \hbox{if}~~u_{a}(t)~=~0\end{cases}
\label{junction3}$$(\[junction3\]) uniquely determines the turning percentages at junctions with more than one outgoing link, and is consistent with the FIFO discipline and established route choices. It remains to express the path delay as the sum of finitely many link delays. If we describe path $p\in\mathcal{P}$ as the following sequence of conveniently labeled arcs: $$p~=~\left\{a_1,\,a_2,\,\ldots,\,a_{i-1},\,a_i,\,a_{i+1}\,\ldots,\,a_{m(p)}\right\}$$ where $m(p)$ is number of arcs in path $p$.
It then follows immediately that path arrival time along $p$ when the departure time at origin is $t$ can be expressed as a composition of arc exit time functions: $$\label{pexit}
\tau^{p}(t)~=~\tau_{a_{m(p)}}\,\circ \,\ldots\, \circ \tau_{a_2} \,\circ\,\tau_{a_1}(t)
\qquad p~=~\left\{ a_{1},\,a_{2},\,\ldots
,\,a_{m(p)}\right\} \in \mathcal{P}$$where the operator $\circ$ means decomposition, that is, $f\circ g(x)\equiv f\big(g(x)\big)$.
Now the complete network loading procedure is given by (\[pqm4\])-([pexit]{}), which is interpreted as a well defined *differential algebraic equation* (DAE) system. Moreover, as well shall see in the next section, the (effective) path delay operator defined in this way is strongly continuous from the subset $\Lambda $ of a Hilbert space into $\Big(\mathcal{L}^{2}[t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$.
Existence of the DUE {#dueexistence}
====================
Existence results for DUE are most general if based on formulation (\[duevi\]). Theorem \[mainthm\] for the existence of solutions of variational inequalities in topological spaces can be applied if the operators $\Psi _{p}(t,\,\cdot )$ can be shown to be continuous and the feasible set $\Lambda $ can be shown to be compact. After Section [psicont]{} addresses the continuity of the effective delay operator, based on the DNL model introduced in previously, the last obstacle to proving existence is the compactness of $\Lambda $, which unfortunately does not generally occur in SRDC DUE. To overcome the aforementioned difficulty, we will consider instead successive finite-dimensional approximations of $\Lambda $, and rely on a topological argument. Such an approach is mathematically rigorous but much more challenging than would be the case if $\Lambda $ were compact in the appropriate Hilbert space. The topological argument and supporting infrastructure for a proof of existence are presented in Section \[bddsec\] and Section \[visec\].
\[psicont\]Continuity of the effective path delay operator
----------------------------------------------------------
In this section, we will establish continuity of the map $h\mapsto \Psi
(\cdot ,\,h)$. These results will be essential for the proof of existence theorem for DUE in Section \[visec\]. Notice that unlike the argument in Zhu and Marcotte (2000) which required *a priori* bound for the path flows, the proof provided here works for unbounded path flows and even distributions, thanks to the generalized Vickrey model.
The next lemma provides the criterion for the continuity of the delay function $D_a(\cdot),\, a\in A$.
\[continuitylemma\] Consider an arc $a\in A$, with inflow $u_a(\cdot)$. Under the generalized Vickrey model expressed in (\[pqm4\])-(\[pqm6\]), the arc delay function $D_a(\cdot)$ is continuous if $u_a(\cdot)\in \mathcal{L}^2[t_0,\,t_f]$.
Assume that $u_a(\cdot)\in\mathcal{L}^2[t_0,\,t_f]$, then $u_a(\cdot)\in\mathcal{L}^1[t_0,\,t_f]$. Therefore the cumulative entering vehicle count $$U_a(t)~\doteq~\int_{t_0}^tu_a(s)\,ds$$ is absolutely continuous. It is straightforward to verify that the following quantity is continuous. $$q_a(t)~\doteq~U_a(t)-M_a\,t-\min_{\tau\leq t-T_a}\left\{U_a(\tau)-M_a\,\tau\right\}$$ where $q_a(t)$ denotes the queue length, $M_a,\, T_a$ denotes flow capacity and free flow time, respectively. By (\[pqm6\]), the delay function $D_a(\cdot)$ is continuous.
The next lemma is a technical result that will facilitate the proof of Theorem \[contthm\].
\[techlemma\] Assume $f(\cdot): [a_2,\,b_2]\rightarrow \mathcal{\Re}^1$ is continuous. Let $g_n(\cdot): [a_1,\,b_1]\rightarrow [a_2,\,b_2],\,n\geq 1$ be a sequence of functions such that $g_n$ converges to $g(\cdot): [a_1,\,b_1]\rightarrow [a_2,\,b_2]$ uniformly. Then $$f\big(g_n(\cdot)\big)~\longrightarrow ~ f\big(g(\cdot)\big),\qquad n~\longrightarrow~\infty$$ uniformly.
According to the Heine-Cantor theorem, $f(\cdot)$ is uniformly continuous on $[a_2,\,b_2]$. It follows that, for every $\varepsilon>0$, there exists $\delta>0$ such that for all $y_1,\,y_2\in[a_2,\,b_2]$, whenever $|y_1-y_2|<\delta$, the inequality $$|f(y_1)-f(y_2)|~\leq~\varepsilon$$ holds. Moreover, by the uniform convergence of $g_n$, there exists some $N>0$ such that, for all $n>N$, we have $$|g_n(x)-g(x)|~<~\delta,\qquad\forall~x\in[a_1,\,b_1]$$ Thus, for every $n> N$, $$\big|f\big(g_n(x)\big)-f\big(g(x)\big)\big|~\leq~\varepsilon,\qquad \forall~x\in[a_1,\,b_1]$$
\[contthm\] Under the network loading model described in Section [DNL]{}, the effective path delay operator $\Psi (t,\,\cdot ):\Lambda
\rightarrow \big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$, $h\mapsto \Psi (\cdot ,\,h)$ is well-defined and continuous.
For each $h\in \Lambda $, the functions $\Psi _{p}(\cdot ,\,h)$ for all $p\in \mathcal{P}$ are uniquely determined by the network loading procedure. To show that the effective path delay operator is well-defined, it remains to show that $\Psi (\cdot ,\,h)\in \Big(\mathcal{L}^{2}[t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$ for each $h\in \Lambda $. Notice that there exists an upper bound for the path delays regardless of the network flow profile: $$D_{p}(t,\,h)~\leq ~\sum_{a\in p}\left\{ {\frac{1}{M_{a}}}\sum_{(i,j)\in
\mathcal{W}}Q_{ij}+T_{a}\right\} ,\qquad \forall ~h\in \Lambda ,\,p\in
\mathcal{P},\,t\in \lbrack t_{0},\,t_{f}]$$where $M_a$, $T_a$ are the flow capacity and free flow travel time respectively, that are associated with the arc $a$. Recall the definition of effective path delay (\[phidef\]): $$\Psi _{p}(t,\,h)~= ~D_{p}(t,\,h)+\mathcal{F}\Big(t+D_{p}(t,\,h)-T_{A}\Big)$$ Since $\mathcal{F}(\cdot)$ is continuous, the uniform boundedness of $D_p(t,\,h)$ thus implies the uniform boundedness of $\Psi _{p}(t,\,h)$ for all $h\in\Lambda, \, p\in \mathcal{P}$ and $t\in[t_0,\,t_f]$. This leads to the conclusion that $\Psi (\cdot ,\,h)\in \Big(\mathcal{L}^{2}[t_{0},\,t_{f}]\Big)^{|\mathcal{P}|}$ for all $h\in \Lambda $. With the preceding as background, the proof of continuity of the effective delay operator may be given in five parts.
**Part 1.** Consider first only a single link with a sequence of entry flows $u_{\nu }$ where$\,\nu \geq 1$ that converge to $u$ in the $\mathcal{L}^{2}$ norm. That is. $$\Vert u_{\nu }-u\Vert _{2}~\doteq ~\left( \int_{t_{0}}^{t_{f}}\left( u_{\nu
}(t)-u(t)\right) ^{2}\,dt\right) ^{1/2}~\longrightarrow ~0\text{ \ \ as \ \ }\nu ~\longrightarrow ~\infty$$Consider the cumulative entering vehicle counts $$\begin{array}{l}
\displaystyle U_{\nu }(t)~\doteq ~\int_{t_{0}}^{t}u_{\nu }(s)\,ds\qquad
\nu~\geq ~1 \\
\displaystyle U(t)~\doteq ~\int_{t_{0}}^{t}u(s)\,ds\end{array}\qquad t\in \lbrack t_{0},\,t_{f}]$$ then $U_{\nu }$ converge to $U$ uniformly on $[t_{0},\,t_{f}]$. This is due to the simple observation $$\left\vert U_{\nu }(t)-U(t)\right\vert ~\leq ~\int_{t_{0}}^{t}\left\vert
u_{\nu }(s)-u(s)\right\vert \,ds~\leq ~\Vert u_{\nu }-u\Vert _{1}~\leq
~(t_{0}-t_{f})^{1/2}\Vert u_{\nu }-u\Vert _{2}~\longrightarrow ~0$$where $\Vert \cdot \Vert _{1}$ is the norm in $\mathcal{L}^{1}[t_{0},\,t_{f}]
$. The last inequality is a version of Jenssen’s inequality.
**Part 2.** Define $R(\tau )\doteq U(\tau )-M\,\tau $, $R_{\nu }(\tau )\doteq U_{\nu }(\tau )-M\,\tau $. We claim the following uniform convergence holds $$\min_{\tau \leq t}\left\{ R_{\nu }(\tau )\right\} ~\longrightarrow
~\min_{\tau \leq t}\left\{ R(\tau )\right\} \qquad \forall ~t\in \lbrack
t_{0},\,t_{f}] \label{contproof1}$$where $M$ is the link flow capacity. Indeed, for any $\varepsilon >0$, by uniform convergence of $U_{\nu }$, we can choose $N$ such that for all $\nu
\geq N$, the following inequality holds $$\left\vert U_{\nu }(t)-U(t)\right\vert ~\leq ~\varepsilon \qquad \forall
~t\in \lbrack t_{0},\,t_{f}]$$For any fixed $t$, if $\nu \geq N$, we have $$\left\vert R_{\nu }(\tau )-R(\tau )\right\vert ~=~\left\vert U_{\nu }(\tau
)-U(\tau )\right\vert ~\leq ~\varepsilon \label{contproof2}$$Let $\hat{\tau}=\hbox{argmin}_{\tau \leq t}\left\{ R(\tau )\right\} $, so that, by (\[contproof2\]), we have $$\min_{\tau \leq t}\left\{ R_{\nu }(\tau )\right\} ~\leq ~R_{\nu }(\hat{\tau})~\leq ~R(\hat{\tau})+\varepsilon ~=~\min_{\tau \leq t}\left\{ R(\tau
)\right\} +\varepsilon \label{contproof3}$$On the other hand, let $\hat{\tau}_{\nu }=\hbox{argmin}_{\tau \leq t}\left\{
R_{\nu }(\tau )\right\} $. Then for each $\nu \geq N$, it must be that $$\min_{\tau \leq t}\left\{ R(\tau )\right\} ~\leq ~R(\hat{\tau}_{\nu })~\leq
~R^{(\nu )}(\hat{\tau}_{\nu })+\varepsilon ~=~\min_{\tau \leq t}\left\{
R_{\nu }(\tau )\right\} +\varepsilon \label{contproof4}$$Taken together, (\[contproof3\]) and (\[contproof4\]) imply $$\left\vert \min_{\tau \leq t}\left\{ R_{\nu }\right\} -\min_{\tau \leq
t}\left\{ R(\tau )\right\} \right\vert ~\leq ~\varepsilon ,\qquad \forall
~\nu \geq N$$Since $t$ is arbitrary, the claim is demonstrated.
**Part 3.** The immediate consequence of [**Part 2**]{} and ([pqm4]{})-(\[pqm6\]) is the following uniform convergence $$W_{\nu }(t)~\longrightarrow ~W(t),\quad q_{\nu }(t)~\longrightarrow ~q(t),\quad
D_{\nu }(t)~\longrightarrow ~D (t),\quad \tau _{\nu
}(t)~\longrightarrow ~\tau (t),\quad \nu ~\longrightarrow ~\infty
\label{contproof5}$$for which we employ notation whose meaning is transparent. The next step is to extend such convergence to the whole network. Consider the sequence of departure rates $h_{\nu }$ converging to $h$ in the $\Vert \cdot \Vert _{\mathcal{L}^{2}}$ norm. By the definition (\[l2norm\]), this implies each path flow $h_{p, \nu}(\cdot )\rightarrow h_{p}$ in the $\Vert \cdot \Vert
_{2}$ norm, for all $p\in \mathcal{P}$. A simple induction based on results established in [**Part 2**]{} yields, as $\nu\longrightarrow \infty$, $$\label{aconv}
U_{a, \nu }(t)~\longrightarrow ~U_{a}(t),\quad W_{a, \nu
}(t)~\longrightarrow ~W_{a}(t),\quad D _{a, \nu
}(t)~\longrightarrow ~D_{a}(t),\quad \tau _{a, \nu
}(t)~\longrightarrow ~\tau_{a}(t),$$uniformly for all $a\in A$.
[**Part 4.**]{} We will show next the uniform convergence of the path delay function $D_p(\cdot,\,h_{\nu})\rightarrow D_p(\cdot,\,h)$, based on (\[aconv\]). Recall the path exit time function (\[pexit\]) $$\label{induction}
\tau^{p}(t)~=~\tau_{a_{m(p)}}\,\circ \,\ldots\, \circ \tau_{a_2} \,\circ\,\tau_{a_1}(t)
\qquad p~=~\left\{ a_{1},\,a_{2},\,\ldots,\, a_{m(p)}\right\}\in\mathcal{P}$$ The desired uniform convergence of path delay follows from the uniform convergence $\tau^p_{\nu}(t)\rightarrow\tau^p(t)$. We start by showing that $\tau_{a_2,\nu}\circ\tau_{a_1, \nu}(t)\rightarrow \tau_{a_2}\circ\tau_{a_1}(t)$ uniformly
For every $\nu\geq 1$, since the inflow of arc $a_2$ is square-integrable, $\tau_{a_2,\nu}(\cdot)$ is continuous by Lemma \[continuitylemma\]. This means that $\tau_{a_2}(\cdot)$ is also continuous since it is the uniform limit of $\tau_{a_2,\nu}(\cdot)$. Lemma \[techlemma\] then implies that $\tau_{a_2}\big(\tau_{a_1,\nu}(\cdot)\big)$ converges uniformly to $\tau_{a_2}\big(\tau_{a_1}(\cdot)\big)$, that is, for any $\varepsilon>0$, there exists an $N_1>0$ such that for all $\nu> N_1$, $$\big|\tau_{a_2}\big(\tau_{a_1,\nu}(t)\big)-\tau_{a_2}\big(\tau_{a_1}(t)\big)\big|~<~\varepsilon/2,\qquad\forall~t\in[t_0,\,t_f]$$ Moreover, there exists some $N_2>0$ such that for all $\nu> N_2$, $$\big|\tau_{a_2,\nu}(t)-\tau_{a_2}(t)\big|~<~\varepsilon/2,\qquad\forall~t\in[t_0,\,t_f]$$ Now let $N_0=\max\{N_1,\,N_2\}$. For any $\nu>n$, and all $t\in[t_0,\,t_f]$, $$\begin{aligned}
&\big|\tau_{a_2,\nu}\big(\tau_{a_1,\nu}(t)\big)-\tau_{a_2}\big(\tau_{a_1}(t)\big)\big|\\
~\leq~& \big|\tau_{a_2,\nu}\big(\tau_{a_1,\nu}(t)\big)-\tau_{a_2}\big(\tau_{a_1,\nu}(t)\big)\big|+\big|\tau_{a_2}\big(\tau_{a_1,\nu}(t)\big)-\tau_{a_2}\big(\tau_{a_1}(t)\big)\big|\\
~<~& \varepsilon/2+\varepsilon/2~=~\varepsilon\end{aligned}$$ This shows the desired uniform convergence $\tau_{a_2,\nu}\circ\tau_{a_1, \nu}(t)\rightarrow \tau_{a_2}\circ\tau_{a_1}(t)$.
The uniform convergence $\tau^p_{\nu}(\cdot)\rightarrow \tau^p(\cdot)$ follows immediately by (\[induction\]) and induction. As a result, we obtain the uniform convergence of path delay $$D_p(\cdot,\,h_{\nu})~\longrightarrow~D_p(\cdot,\,h),\qquad \nu~\longrightarrow~\infty$$
**Part 5**. Finally, recall the definition of the effective delay $$\Psi (t,\,h)~=~D_{p}(t,\,h)+\mathcal{F}\big(t+D_{p}(t,\,h)-T_{A}\big)$$Note that $\mathcal{F}(\cdot)$ is continuous, the following uniform convergence follows by Lemma \[techlemma\] $$\mathcal{F}\big(t+D_p(t,\,h_{\nu})-T_A\big)~\longrightarrow~\mathcal{F}\big(t+D_p(t,\,h)-T_A\big),\qquad \nu~\longrightarrow~\infty$$ We conclude that the effective delay $\Psi_p(\cdot,\,h_{\nu})$ converges uniformly to $\Psi_p(\cdot,\,h)$. The desired convergence in the $\Vert \cdot \Vert _{\mathcal{L}^{2}}$ norm now follows since the interval $[t_{0},\,t_{f}]$ is compact.
Alternative definition of effective path delay {#bddsec}
----------------------------------------------
The integrals employed in defining the feasible domain ([feasible]{}) are not enough to assure bounded path flows $h_{p},\,p\in
\mathcal{P}$. This observation is the fundamental hurdle to providing existence of the DUE solution. One of the main accomplishments of this paper is to address the boundedness of path flows not only for the proof of existence result but also for future analysis and estimation of network flows. In this section, we will present an alternative formulation of the effective path delay $\Psi _{p}(t,\,h)$, where that alternative formulation will facilitate our analysis leading to the proof of our main result, Theorem \[VIthm\].
Recall the effective delay operator $$\Psi _{p}(t,\,h)~\doteq ~D_{p}(t,\,h)+\mathcal{F}\big(t+D_{p}(t,\,h)-T_{A}\big) \label{ed}$$ In order to simplify our analysis, it is convenient to rewrite (\[ed\]) in a slightly different form. In particular, for each O-D pair $(i,\,j)\in \mathcal{W}$, let us introduce the cost functions $\phi _{ij}(\cdot ):[t_{0},\,t_{f}]\rightarrow
\mathcal{\Re }_{+}^{1}$ and $\,\psi _{ij}(\cdot ):[t_{0},\,t_{f}]\rightarrow
\mathcal{\Re }_{+}^{1}$, which measure the travel costs at the origin and destination of the path, respectively. More precisely, $\phi _{ij}(\cdot )$ is a function of departure time, while $\psi _{ij}(\cdot )$ is a function of arrival time. Fix any vector of path flows $h\in \Lambda $, recall the path exit time function $\tau^{p}(t)=t+D_p(t,\,h)$. Then (\[ed\]) can be equivalently written as $$\begin{aligned}
\Psi _{p}(t,\,h)~=~& -t+\tau^p(t)+\mathcal{F}\big(\tau^p(t)-T_{A}\big) \\
~=~& \phi _{ij}(t)+\psi _{ij}\big(\tau _{p}(t,\,h)\big)\end{aligned}$$where $$\phi _{ij}(t)~\equiv ~-t,\qquad \psi _{ij}(t)~\doteq ~t+\mathcal{F}\big(t-T_{A}\big) \label{vppsidef}$$
In Bressan and Han (2011, 2012), the travel costs are measured in terms of $\phi_{ij}(\cdot),\,\psi_{ij}(\cdot)$, in other words, the general effective delay (\[ed\]) can be alternatively evaluated as a sum of costs at the beginning and end of journey of each driver.
In Section \[esec\], we will exploit the alternative representative of effective path delay (cost) to establish existence.
To prepare for the existence proof, consider a general network $G(N, \, A)$. Associate with each O-D pair $(i,\,j)\in \mathcal{W}$ the pair of cost functions $\phi _{ij}(\cdot )$ and $\psi _{ij}(\cdot )$ as defined in (\[vppsidef\]). We make the following two assumptions on $\phi
_{ij}(\cdot ),\,\psi _{ij}(\cdot )$ and the underlying traffic flow model.
1. *For all, $\phi _{ij}(\cdot )$, $\psi
_{ij}(\cdot )$ is continuously differentiable on $[t_{0},\,t_{f}]$. In addition, $${\frac{d}{dt}}\psi _{ij}(t)~>~0,\qquad t\in \lbrack t_{0},\,t_{f}]
\label{a1eqn}$$*
2. *Each link $a\in A$ in the network has a flow capacity $f_{a}^{max}~<~\infty $*
Note that A1 stipulates that each arrival cost $\psi _{ij}(\cdot )$ is strictly increasing, while A2 applies to all traffic flow models that assume a flow capacity for each network link. Notice that by (\[vppsidef\]), A1 amounts to requiring that
1. *$\mathcal{F}(\cdot )$ is continuously differentiable and satisfies $\mathcal{F}^{\prime }(s)>-1$ on $[t_{0},\,t_{f}]$*.
In view of the preceding assumptions, we are prompted to define the following: $$\psi _{min}^{\prime }~\doteq ~\min_{(i,\,j)\in \mathcal{W}}\min_{t\in
\lbrack t_{0},\,t_{f}]}\,{\frac{d}{dt}}\psi _{ij}(t)~>~0 \label{psimindef}$$$$F^{max}~\doteq ~\max_{a\in A}f_{a}^{max}~<~+\infty
\label{fmaxdef}$$Note that $\psi _{min}^{\prime }>0$ follows from (\[a1eqn\]) and the fact that ${\frac{d}{dt}}\psi _{ij}(\cdot )$ is continuous on the compact interval $[t_{0},\,t_{f}]$.
\[esec\]Existence of solution to the variational inequality {#visec}
-----------------------------------------------------------
The classical result explained by Theorem \[mainthm\] will be the key ingredient for the proof of existence of the DUE solution. Using the same notation as in Theorem \[mainthm\], the underlying topological vector space $E$ will be instantiated by $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$, which is a locally convex topological vector space. The dual space $E^{\ast }$ will be again $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$.
\[VIthm\]**(Existence of DUE)** Let assumption *(A1’), (A2)* hold. In addition, assume that the effective delay operator $\Psi: \Lambda\rightarrow \big(\mathcal{L}^2[t_0,\,t_f]\big)^{|\mathcal{P}|}$ is continuous. Then the dynamic user equilibrium problem as in Definition \[duedef\] has a solution.
The proof is divided into four parts.
**Part 1.** Our strategy for demonstrating existence is to adapt Theorem \[mainthm\] to the locally convex topological vector space $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$, and its subset $\Lambda $. By assumption, the map $h\mapsto \Psi (\cdot ,\,h)$ is continuous from $\Lambda $ to the space of $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$. If $\Lambda $ were compact and convex, we would immediately demonstrate the desired result. However, $\Lambda $ is bounded, closed and convex, but not compact in $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$.
**Part 2.** We will instead employ finite-dimensional approximations of $\Lambda $. In order to proceed, consider for each $n\geq 1$ the uniform partition of interval $[t_{0},\,t_{f}]$ with $2^{n}$ sub-intervals $$\begin{aligned}
t_{0}~& =~t^{0}~<~t^{1}~<~t^{2}~\ldots ~<~t^{2^{n}}~=~t_{f} \\
t^{i}-t^{i-1}~& =~{\frac{t_{f}-t_{0}}{2^{n}}}\qquad i=1,\ldots ,2^{n}\end{aligned}$$Then consider the following sequence of finite-dimensional subsets $$\Lambda _{n}~\doteq ~\left\{ h\in \Lambda :\quad h_{p}(\cdot )\hbox{ is
constant on }\lbrack t^{i-1},\,t^{i}),\quad p\in \mathcal{P}\right\}
~\subset ~\Lambda \label{subspace}$$We claim that for each $n\geq 1$, $\Lambda _{n}$ is compact and convex in $\big(\mathcal{L}^{2}([t_{0},\,t_{f}])\big)^{|\mathcal{P}|}$. Indeed, given any $h^{n,1},\,h^{n,2}\in \Lambda _{n}$, and $\alpha \in \lbrack 0,\,1]$, $\alpha \,h^{n,1}+(1-\alpha )\,h^{n,2}$ is clearly nonnegative and constant on each $[t^{i-1},\,t^{i}]$ for$\,i=1,\ldots ,2^{n}$. In addition, for any origin-destination pair $(i,\,j)\in \mathcal{W}$, by definition ([feasible]{}), $$\sum_{p\in \mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}\alpha
\,h^{n,1}(t)+(1-\alpha )\,h^{n,2}(t)\,dt~=~\alpha \,Q_{ij}+(1-\alpha
)\,Q_{ij}~=~Q_{ij}$$This verifies that $\Lambda _{n}$ is convex. To see compactness, we define the map $\mu :\Lambda _{n}\rightarrow \mathcal{\Re }_{+}^{2^{n}\times |\mathcal{P}|},\,h\mapsto (a_{i,\,p}:\,1\leq i\leq 2^{n},\,\,p\in \mathcal{P})
$ where each vector $(a_{1,p},\ldots ,a_{2^{n},p})$ is the coordinate of $h_{p}$ under the natural basis $\{e_{n}^{i}\}_{i=1}^{2^{n}}$, where $$e_{n}^{i}(t)~=~\begin{cases}
1\qquad & t\in \lbrack t^{i-1},\,t^{i}) \\
0\qquad & \hbox{else}\end{cases}$$Clear, the map $\mu $ is one-to-one. Now consider any sequence $\{h^{n \left(\nu\right)}\}_{\nu \geq 1}\subset \Lambda _{n}$, and the sequence of their images $\{\mu \big(h^{n\left( v\right) }\big)\big\}_{\nu \geq 1}\subset
\mathcal{\Re }_{+}^{2^{n}\times |\mathcal{P}|}$. By (\[feasible\]), $\big\{\mu (h^{n(\nu )}\big\}_{\nu \geq 1}$ is uniformly bounded by the following quantity $$\max_{i,j}{\frac{2^{n}\,Q_{ij}}{t_{f}-t_{0}}}$$By the Bolzano-Weierstrass theorem, there exists a convergent subsequence $\big\{\mu (h^{n(\nu \,^{\prime })})\big\}_{\nu \,^{\prime }\geq 1}$. By construction, the subsequence $\big\{h^{n(\nu \,^{\prime })}\big\}_{\nu
\,^{\prime }\geq 1}$ must converge uniformly to some $\hat{h}^{n}$. In view of the compact interval $[t_{0},\,t_{f}]$, we conclude that this convergence is also in the $\Vert \cdot \Vert _{\mathcal{L}^{2}}$ norm. It now remains to show $\hat{h}^{n}\in \Lambda _{n}$, then our claim follows from sequential compactness of $\Lambda _{n}$. Clearly $\hat{h}^{n}\geq 0$ and is constant on the sub-intervals $[t^{i-1},\,t^{i}),\,i=1,\ldots ,2^{n}$. Moreover, since $h^{n(\nu \,^{\prime })}$ are uniformly bounded, by the Dominated Convergence Theorem, $$Q_{ij}~=~\lim_{\nu \,^{\prime }\rightarrow \infty }\sum_{p\in \mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}h_{p}^{n(\nu \,^{\prime })}(t)\,dt~=~\sum_{p\in
\mathcal{P}_{ij}}\int_{t_{0}}^{t_{f}}\hat{h}_{p}^{n}(t)\,dt$$This implies $\hat{h}^{n}\in \Lambda _{n}$. Thereby, the claim is substantiated.
**Part 3.** For each $n\geq 1$, apply Theorem \[mainthm\] to $\Lambda
_{n}$ and obtain $h^{n,\ast }\in \Lambda _{n}$ such that $$\Big<\Psi \big(\cdot ,\,h^{n,\ast }\big),\,h^{n,\ast }(\cdot )-h^{n}(\cdot )\Big>~\geq ~0,\qquad \forall ~h^{n}\in \Lambda _{n} \label{finite}$$Where the $\big<\,,\,\big>$ is the inner product (duality) defined in ([nnorm]{}). It is easy to observe that (\[finite\]) implies that if $h_{p}^{n,\ast }(t)>0,\,t\in \lbrack t^{j},\,t^{j+1}]$, then $$\int_{t^{j}}^{t^{j+1}}\Psi (t,\,h^{n,\ast })\,dt~=~\min_{0\leq k\leq
2^{n}}\int_{t^{k}}^{t^{k+1}}\Psi (t,\,h^{n,\ast })\,dt \label{viimply}$$Recall (\[psimindef\]), (\[fmaxdef\]), choose any constant $\mathcal{M}$ such that $$\mathcal{M}~>~{\frac{3F^{max}}{\psi _{min}^{\prime }}} \label{upperbound}$$We then claim that for all $n\geq 1$, there must hold $h_{p}^{n,\ast
}(t)\leq \mathcal{M}$ for all $t\in \lbrack t_{0},\,t_{f}],\,p\in \mathcal{P}
$. Otherwise, assume there exists some $p\in \mathcal{P}$, some $\nu \geq 1$ and some $0\leq i\leq 2^{\nu }$ with $$h^{v,\ast }(t)~\equiv ~\mu ~>~\mathcal{M}\qquad t\in \lbrack
t^{i},\,t^{i+1}],$$By choosing $[t_{0},\,t_{f}]$ large enough, we can assume that $i\geq 1$. Now consider the interval $[t^{i-1},\,t^{i}]$, and the quantity $\displaystyle\sup_{t\in \lbrack t^{i-1},t^{i}]}\Psi (t,\,h^{v,\ast })$, by possibly modifying the value of the function $\Psi (\cdot ,\,h^{v,\ast })$ at one point, we can obtain $t^{\ast }\in \lbrack t^{i-1},\,t^{i}]$ such that $$\Psi (t^{\ast },\,h^{v,\ast })~=~\sup_{t\in \lbrack t^{i-1},t^{i}]}\Psi
(t,\,h^{v,\ast })$$Now, let $\tau _{p}(t,\,h^{v,\ast })=t+D_{p}(t,\,h^{v,\ast })$ be the arrival time function. According to FIFO and assumption *(A2)*, we deduce that for all $t\in \lbrack t^{i},\,t^{i+1}]$, $$\tau _{p}(t,\,h^{v,\ast })-\tau _{p}(t^{\ast },\,h^{v,\ast })~\geq ~{\frac{(t-t^{i})\,\mu }{F^{max}}} \label{boundedeqn1}$$this implies, together with (\[psimindef\]) that $$\psi _{ij}\big(\tau _{p}(t,\,h^{v,\ast })\big)-\psi _{ij}\big(\tau
_{p}(t,\,h^{v,\ast })\big)~\geq ~\psi _{min}^{\prime }\big(\tau
_{p}(t,\,h^{v,\ast })-\tau _{p}(t^{\ast },\,h^{v,\ast })\big)~\geq ~\psi
_{min}^{\prime }\cdot {\frac{(t-t^{i})\,\mu }{F^{max}}} \label{boundedeqn2}$$(\[boundedeqn2\]) and (\[vppsidef\]) imply $$\Psi _{p}(t,\,h^{v,\ast })-\Psi _{p}(t^{\ast },\,h^{v,\ast })~\geq ~{\frac{\psi _{min}^{\prime }\,\mu }{F^{max}}}(t-t^{i})-(t-t^{\ast }),\qquad \forall
~t\in \lbrack t^{i},\,t^{i+1}] \label{boundedeqn3}$$Integrating (\[boundedeqn3\]) from $t^{i}$ to $t^{i+1}$ and a simple calculation yield $$\int_{t^{i}}^{t^{i+1}}\Psi (t,\,h^{v,\ast })\,dt-(t^{i+1}-t^{i})\Psi
(t^{\ast },\,h^{v,\ast })~\geq ~{\frac{(t^{i+1}-t^{i})^{2}}{2}}\cdot {\frac{\psi _{min}^{\prime }\,\mu }{F^{max}}}+{\frac{(t^{i+1}-t^{i})}{2}}\cdot
\left( t^{\ast }-{\frac{t^{i}+t^{i+1}}{2}}\right) \label{boundedeqn4}$$Notice since $t^{\ast }\in \lbrack t^{i-1},\,t^{i}]$, we deduce $t^{\ast
}-(t^{i}+t^{i+1})/2\geq -3/2(t^{i+1}-t^{i})$. (\[boundedeqn4\]) then becomes $$\int_{t^{i}}^{t^{i+1}}\Psi (t,\,h^{v,\ast })\,dt-(t^{i+1}-t^{i})\Psi
(t^{\ast },\,h^{v,\ast })~\geq ~{\frac{(t^{i+1}-t^{i})^{2}}{2}}\left( {\frac{\psi _{min}^{\prime }\,\mu }{F^{\max }}}-3\right) ~>~0 \label{boundedeqn5}$$This yields the following contradiction to (\[viimply\]) and hence ([finite]{}) $$\int_{t^{i}}^{t^{i+1}}\Psi (t,\,h^{v,\ast
})\,dt~>~\int_{t^{i-1}}^{t^{i}}\Psi _{p}(t,\,h^{v,\ast })\,dt$$
**Part 4.** By previous steps, we have obtained a uniformly bounded sequence of vector-valued functions $\big\{h_{n}^{\ast }\big\}_{n\geq 1}$ satisfying (\[finite\]). By taking a subsequence, we can assume the weak convergence in both $\mathcal{L}^{2}$-space and $\mathcal{L}^{1}$-space $$h^{n,\ast }~\longrightarrow ~h^{\ast },\qquad n~\rightarrow ~\infty$$for some $h^{\ast }\in \Lambda $. We claim that for all $h\in \Lambda $, $$\Big<\Psi \big(\cdot ,\,h^{\ast }\big),\,h^{\ast }(\cdot )-h(\cdot )\Big>~\geq ~0$$Indeed, given any $h\in \Lambda $, there exists piecewise-constant approximation $\{h_{n}\in \Lambda _{n}\}_{n\geq 1}$, which converges to $h$ both point-wise and in the $\Vert \cdot \Vert _{\mathcal{L}^{2}}$ norm. According to (\[finite\]), we have $$\Big<\Psi _{p}\big(\cdot ,\,h^{n,\ast }\big),\,h^{n,\ast }(\cdot
)-h_{n}(\cdot )\Big>~\geq ~0 \label{eqn1}$$Notice that the map $h\mapsto \Psi _{p}(\cdot ,\,h)$ is continuous with respect to the $\Vert \cdot \Vert _{\mathcal{L}^{2}}$ norm, by the continuity of the inner product, we pass (\[eqn1\]) to the limit $$\Big<\Psi _{p}\big(\cdot ,\,h^{\ast }\big),\,h^{\ast }(\cdot )-h(\cdot )\Big>~\geq ~0$$
The next theorem establishes the existence of SRDC DUE with generalized Vickrey model, which is an immediate consequence of Theorem \[contthm\] and Theorem \[VIthm\].
\[gvmthm\] Let assumption *(A1’)* hold, then the dynamic user equilibrium as in Definition \[duedef\] with generalized Vickrey model has a solution.
Notice that the assumption A2 is satisfied by the Vickrey’s model and hence the GVM. Thus in the statement of the theorem, A2 is omitted.
Conclusion
==========
We have established the existence of the continuous-time simultaneous route-and-departure choice DUE for the generalized Vickrey model (GVM), a generalization of Vickrey’s model, and plausible regularity conditions that are easy to check and rather weak. It is significant that ours is the first DUE existence result without the a priori bounding of departure rates (path flows) and the most general constraint relating path flows to the trip table. In fact, our method of proof successfully overcomes two major hurdles that have stymied other researchers:
1. the set of feasible flows $\Lambda $ is intrinsically non-compact in $\mathcal{L}^{2}$-space as well as $\mathcal{L}^{1}$-space; and
2. a direct topological argument requires *a priori* bounds for the path flows, where those bounds do not arise from any behavioral argument or theory.
Theorem \[VIthm\] is a general result that subsumes all SRDC DUE models regardless of the arc dynamics, flow propagation and arc delay function employed, as long as the effective delay operator is continuous.
[99]{} Bressan, A., Han, K., 2011. Optima and equilibria for a model of traffic flow. SIAM Journal on Mathematical Analysis 43 (5), 2384-2417.
Bressan, A., Han, K., 2012. Nash equilibria for a model of traffic flow with several groups of drivers. ESAIM: Control, Optimization and Calculus of Variations.
Browder, F.E., 1968. The fixed point theory of multi-valued mappings in topological vector spaces. Math. Annalen 177, 283-301.
Daganzo, C.F., 1994. The cell transmission model: A simple dynamic representation of highway traffic. Transportation Research Part B 28B (4), 269-287.
Daganzo, C.F., 2005. A variational formulation of kinematic waves: basic theory and complex boundary conditions. Transportation Research Part B 39, 187-196.
Evans, L.C., 2010. Partial Differential Equations. Second edition. American Mathematical Society, Providence, RI.
Friesz, T.L., 2010. Dynamic Optimization and Differential Games, Springer, New York.
Friesz, T.L., Bernstein, D., Suo, Z., Tobin, R., 2001. Dynamic network user equilibrium with state-dependent time lags. Network and Spatial Economics 1 (3/4), 319-347.
Friesz, T.L., Bernstein, D., Smith, T., Tobin, R., Wie, B., 1993. A variational inequality formulation of the dynamic network user equilibrium problem. Operations Research 41 (1), 80-91.
Friesz, T.L., Bernstein, D., Suo, Z., Tobin, R.L., 2001. Dynamic network user equilibrium with state-dependent time lags. Networks and Spatial Economics 1, 319-347.
Friesz, T.L., Luque, F.J., Tobin, R.L., Wie, B.W., (1989). Dynamic network traffic assignment considered as a continuous time optimal control problem. Operations Research 6, 893-901.
Friesz, T.L., Mookherjee, R., 2006. Solving the dynamic network user equilibrium problem with state-dependent time shits. Transportation Research Part B 40, 207-229.
Friesz, T.L., Han, K., Meimand, A., Neto, P., Yao, T., 2012. Dynamic user equilibrium based on a hydrodynamic model. Submitted to Transportation Research Part B.
Friesz, T.L., Luque, J., Tobin, R.L., Wie, B.W., 1989. Dynamic network traffic assignment considered as a continuous time optimal control problem. Operations Research 37 (6), 893-901.
Friesz, T.L., Kim, T., Kwon, C., Rigdon, M.A., 2010. Approximate network loading and dual-time-scale dynamic user equilibrium. Transportation Research Part B 45, 176-207.
Han, K., Friesz, T.L., Yao, T., 2012a. A partial differential equation formulation of the continuous-time Vickrey’s model: methodology and theoretical analysis. To appear in Transportation Research Part B.
Han, K., Friesz, T.L., Yao, T., 2012b. A partial differential equation formulation of the continuous-time Vickrey’s model: computational method and application. To appear in Transportation Research Part B.
Han, L., Ukkusuri, S., Doan, K., 2011. Complementarity formulation for the cell transmission model based dynamic user equilibrium with departure time choice, elastic demand and user heterogeneity. Transportation Research Part B 45, 1749-1767.
Lax, P.D., 1957. Hyperbolic systems of conservation laws II. Comm. Pure and Applied Mathematics 10, 537- 566.
Lighthill, M., Whitham, G., 1955. On kinematic waves. II. A theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London: Series A 229, 317- 345.
Lo, H., Szeto, W., 2002. A cell-based variational inequality formulation of the dynamic user optimal assignment problem. Transportation Research Part B 36 (5), 421-443.
Merchant, D.K., Nemhauser, G.L., 1978a. A model and an algorithm for the dynamic traffic assignment problem. Transportation Science 12, 183-199.
Merchant, D.K., Nemhauser, G.L., 1978b. Optimality conditions for a dynamic traffic assignment model. Transportation Science 12 (3), 200–207.
Mounce, R., 2006. Convergence in a continuous dynamic queuing model for traffic networks. Transportation Research Part B 40, 779-791.
Nie, X., Zhang, H.M., 2005. A comparative Study of Some Macroscopic Link Models Used in Dynamic Traffic Assignment. Networks and Spatial Economics 5, 89-115.
Pang, J., Han, L., Ramadurai, G., Ukkusuri, S., in press. A continuous-time dynamic equilibrium model for multi-user class single bottleneck traffic flows. Mathematical Programming.
Ran, B., Hall, R.W., Boyce, D.E., (1996). A link-based variational inequality model for dynamic departure time/route choice. Transportation Research Part B 30B, 31-46.
Richards, P.I., 1956. Shockwaves on the highway. Operations Research 4, 42-51.
Smith, M.J., Wisten, M.B., 1994. Lyapunov methods for dynamic equilibrium traffic assignment. In proceedings of the second Meeting of the EURO working group on Urban Traffic and Transportations, 223-245, INRETS, Paris.
Smith, M.J., Wisten, M.B., 1995. A continuous day-to-day traffic assignment model and the existence of a continuous dynamic user equilibrium. Annals of Operations Research 60, 59-79.
Vickrey, W.S., 1969. Congestion Theory and Transport Investment. The American Economic Review 59 (2), 251-261.
Wie, B.W., Tobin, R.L., Carey, M., 2002. The existence, uniqueness and computation of an arc-based dynamic network user equilibrium formulation. Transportation Research Part B 36, 897-918.
Zhu, D.L., Marcotte, P., 2000. On the existence of solutions to the dynamic user equilibrium problem. Transportation Science 34 (4), 402–414.
[^1]: Note that, by refering to the network loading procedure, we are neither employing nor suggesting a sequential approach to the study and computation of DUE. Rather a subset of the equations and inequalites comprising a complete DUE model may be grouped in a way that identifies a traffic assignment subproblem and a network loading subproblem. Such a grouping and choice of names is merely a matter of convenient language that avoids repetitive reference to the same mathematical expressions. Use of such language does not alter the need to solve both the assignment and loading problems consistently and, thus, simultaneously. A careful reading of the mathematical presentation made in subsequent sections makes this quite clear.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate two examples of stochastic processes whose lifts to geometric rough paths require a renormalisation procedure to obtain convergence in rough path topologies. Our first example involves a physical Brownian motion subject to a magnetic force which dominates over the friction forces in the small mass limit. Our second example involves a lead-lag process of discretised fractional Brownian motion with Hurst parameter $H \in (1/4,1/2)$, in which the stochastic area captures the quadratic variation of the process. In both examples, a renormalisation of the second iterated integral is needed to ensure convergence of the processes, and we comment on how this procedure mimics negative renormalisation arising in the study of singular SPDEs and regularity structures.'
author:
- 'Y. Bruned, I. Chevyrev, and P. K. Friz'
bibliography:
- 'AllRefs.bib'
title: Examples of renormalized SDEs
---
Introduction
============
In recent years, the theory of regularity structures [@Hairer14] has been proposed to give meaning to a wide class of singular SPDEs. A central feature of the theory is the notion of renormalisation, specifically “negative renormalisation” [@BHZ16], which is required to obtain convergence of random models to a meaningful limit. It is well-known that this procedure is inherit to the problem since naive approximations of such equations typically fail to converge (with a number of notable exceptions, including a special variant of gKPZ [@Hairer16]). The same feature thus naturally appears in other solution theories which have been proposed to solve such equations, including the theories of paracontrolled distributions [@GIP15] and the Wilsonian renormalisation group [@Kupiainen16].
Viewing regularity structures as a multidimensional generalisation of the theory of rough paths [@Lyons98], it is natural to ask how renormalisation manifests itself in the latter. As solution theories to singular S(P)DEs, both share the common feature that one must give meaning to often analytically ill-posed higher order terms (iterated integrals) of distributions, which is typically done through some stochastic means. However, a key difference in applications of rough paths to SDEs and rough SPDEs [@FrizVictoir10] is that one can usually give meaning to such terms as limits of iterated integrals of mollifications of the irregular noise without the need of renormalisation.
The purpose of this note is to demonstrate situations in rough paths theory which fall outside this usual setting and for which renormalisation is a necessary feature. Specifically, we construct two stochastic processes whose lifts to geometric rough paths fail to converge without a renormalisation procedure akin to the one encountered in the theory of regularity structures.
The first is a physical Brownian motion subject to a magnetic force which dominates over the friction forces in the small mass limit. This example builds on the work [@FrizGassiat15], where a similar situation was considering with a constant magnetic field.
The second is a lead-lag process of a discretized path, which we take to be fractional Brownian motion with Hurst parameter $H \in (1/4,1/2]$. The stochastic area of this lead-lag process captures the quadratic variation of the discretized path, and thus, as one can expect, the second iterated integral fails to converge as the mesh of the discretisation goes to zero (unless $H=1/2$). This example is motivated from a similar Hoff process considered for semi-martinagles in [@Flint16].
In both examples we demonstrate an explicit renormalisation procedure of the second iterated integral under which the processes converge in rough path topologies. These (diverging) counter-terms serve precisely the same re-centring role encountered in regularity structures (for a direct comparison, consider the renormalisation of PAM [@Hairer14; @GIP15] where only one diverging term needs to be considered). In turn, rough differential equations driven by the renormalised and unrenormalised rough paths are related to one another by the addition of diverging terms, which again mimics the situation encountered in singular SPDEs. We refer to the upcoming work [@Bruned16] for a much more detailed study of this relation.
[**Acknowledgements.**]{} P.K.F. is partially supported by the European Research Council through CoG-683166 and DFG research unit FOR2402. I.C., affiliated to TU Berlin when this project was commenced, was supported by DFG research unit FOR2402.
Magnetic field blow-up {#subsec:magnetic}
======================
Consider a physical Brownian motion in a magnetic field with dynamics given by $$m \ddot{x} = -A\dot{x} + B\dot{x} + \xi, \; \; x(t) \in \R^d,$$ where $A$ is a symmetric matrix with strictly positive spectrum (representing friction), $B$ is an anti-symmetric matrix (representing the Lorentz force due to a magnetic field), and $\xi$ is an $\R^d$-valued white noise in time. We shall consider the situation that $A$ is constant whereas $B$ is a function of the mass $m$.
We rewrite these dynamics as $$\begin{aligned}
dX_t &= \frac{1}{m} P_t dt, \; \; X_0 = 0, \\
dP_t &= -\frac{1}{m} M P_t dt + dW_t, \; \; P_0 = 0,\end{aligned}$$ where $M = A - B$, and we have chosen the starting point as zero simply for convenience. We furthermore introduce the parameter $\varepsilon^2 = m$ and write $X^\varepsilon_t, P^\varepsilon_t$, and $M^\varepsilon = A-B^\varepsilon$ to denote the dependence on $\varepsilon$.
We are interested in the convergence of the processes $P^\varepsilon$ and $M^\varepsilon X^\varepsilon$ in rough path topologies. Let $G^2(\R^d)$ and $\g^2(\R^d)$ denote the step-$2$ free nilpotent Lie group and Lie algebra respectively. Let us also write $\g^2(\R^d) = \R^d \oplus \g^{(2)}(\R^d)$ for the decomposition of $\g^2(\R^d)$ into the first and second levels, where we identify $\g^{(2)}(\R^d)$ with the space of anti-symmetric $d\times d$ matrices.
For every $\varepsilon > 0$, define the matrix $$C^\varepsilon = \int_0^\infty e^{-M^\varepsilon s}e^{-(M^{\varepsilon})^* s}ds,$$ and the element $$v^\varepsilon = -\frac{1}{2}(M^\varepsilon C^\varepsilon - C^\varepsilon(M^\varepsilon)^*) \in \g^{(2)}(\R^d).$$
For any $v \in \g^{(2)}(\R^d)$, $p \in [1,3)$, and $p$-rough path $(Z_{s,t},\Z_{s,t}) \in G^2(\R^d)$ (where we ignore zeroth component $1$), we define the translated rough path $T_v(Z_{s,t}, \Z_{s,t})$ by $$\label{eq:TvDef}
T_v(Z_{s,t}, \Z_{s,t}) = (Z_{s,t}, \Z_{s,t} + (t-s)v^\varepsilon).$$
Consider the $G^2(\R^d)$-valued processes $$\begin{aligned}
(P^\varepsilon_{s,t}, \Pbb^\varepsilon_{s,t}) &= \left(P^\varepsilon_{s,t}, \int_s^t P^\varepsilon_{s,r} \otimes \circ dP^\varepsilon_r\right), \\
(Z^\varepsilon_{s,t}, \Z^\varepsilon_{s,t}) &= \left(M^\varepsilon X^\varepsilon_{s,t}, \int_s^t M^\varepsilon X_{s,r} \otimes d(M^\varepsilon X^\varepsilon)_r \right),\end{aligned}$$ and the canonical lift of the Brownian motion $W$ $$(W_{s,t}, \W_{s,t}) = \left(W_{s,t}, \int_s^t W_{s,r} \otimes \circ dW_r\right),$$ where the integrals in the definition of $\Pbb^\varepsilon_{s,t}$ and $\W_{s,t}$ are in the Stratonovich sense.
The following proposition establishes the convergence of the “renormalised” paths $T_{v^\varepsilon}(P^\varepsilon_{s,t}, \Pbb^\varepsilon_{s,t})$ and $T_{v^\varepsilon}(Z^\varepsilon_{s,t}, \Z^\varepsilon_{s,t})$.
\[thm:magneticConv\] Suppose that $$\label{eq:MBound}
\lim_{\varepsilon \rightarrow 0} |M^\varepsilon|\varepsilon^\kappa = 0 \; \textnormal{ for some } \kappa \in [0,1].$$ Then for any $\alpha \in [0,1/2-\kappa/4)$ and $q < \infty$, it holds that $T_{v^\varepsilon}(P^\varepsilon, \Pbb^\varepsilon) \rightarrow (0,0)$ and $T_{v^\varepsilon}(Z^\varepsilon, \Z^\varepsilon) \rightarrow (W,\W)$ in $L^q$ and $\alpha$-H[ö]{}lder topology as $\varepsilon \rightarrow 0$. More precisely, as $\varepsilon \rightarrow 0$, in $L^q$ $$\sup_{s,t \in [0,T]} \frac{|P^\varepsilon_{s,t}|}{|t-s|^\alpha} + \sup_{s,t \in [0,T]} \frac{|\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon|}{|t-s|^{2\alpha}} \rightarrow 0.$$ and $$\sup_{s,t \in [0,T]} \frac{|Z^\varepsilon_{s,t} - W_{s,t}|}{|t-s|^\alpha} + \sup_{s,t \in [0,T]} \frac{|\Z^\varepsilon_{s,t} + (t-s)v^\varepsilon - \W_{s,t}|}{|t-s|^{2\alpha}} \rightarrow 0.$$
The rest of the section is devoted to the proof of Theorem \[thm:magneticConv\] which builds on the proof of [@FrizGassiat15] Theorem 1.
We set $Y^\varepsilon = P^\varepsilon/\varepsilon$ and obtain that $$\begin{aligned}
dY^\varepsilon_t &= -\varepsilon^2 M^\varepsilon Y^\varepsilon_t dt + \varepsilon^{-1}dW_t \\
dX^\varepsilon_t &= \varepsilon^{-1}Y^\varepsilon_t dt.\end{aligned}$$ For fixed $\varepsilon$, we introduce the Brownian motion $\tilde W^\varepsilon_\cdot = \varepsilon W_{\varepsilon^{-2} \cdot}$ and consider $$\begin{aligned}
d\tilde Y^\varepsilon_t &= -M^\varepsilon\tilde Y^\varepsilon_t dt + d\tilde W^\varepsilon_t.
%\\
%d\tilde X^\varepsilon_t &= \tilde Y^\varepsilon_t dt.\end{aligned}$$ Observe that we have the pathwise equalities $$\label{eq:pathwiseEq}
Y^\varepsilon_\cdot = \tilde Y^\varepsilon_{\varepsilon^{-2}\cdot},
%(Y^\varepsilon_\cdot, \varepsilon^{-1} X^\varepsilon_{\cdot}) = (\tilde Y^\varepsilon_{\varepsilon^{-2}\cdot}, \tilde X^\varepsilon_{\varepsilon^{-2}\cdot}),$$ and since $Y^\varepsilon_0 = 0$, we have $$\label{eq:YSol}
\tilde Y^\varepsilon_t = \int_0^t e^{-M^\varepsilon (t-s)} d\tilde W^\varepsilon_s.$$
The dependence of $M^\varepsilon$ on $\varepsilon$ is, by construction, only though $B^\varepsilon$, the anti-symmetric part of $M^\varepsilon$. In particular, since the symmetric part $A$ stays constant and has strictly positive spectrum, it follows that for some $\lambda > 0$, $\Real(\sigma(M^\varepsilon)) \subset (\lambda,\infty)$ for all $\varepsilon > 0$. In particular, $$\label{eq:expBound}
\sup_{\tau > 0} \sup_{\varepsilon > 0} \frac{|e^{-\tau M^\varepsilon}|}{e^{-\lambda \tau}} < \infty.$$ We see then that $$\sup_{\varepsilon > 0} |C^\varepsilon| < \infty$$ and $$\label{YtildeL2}
\sup_{\varepsilon > 0} \sup_{0 \leq t < \infty} \EEE{|\tilde Y^\varepsilon_t|^2} < \infty.$$
\[lem:YBounds\] There exists $C_1 > 0$ such that for all $\varepsilon \in (0,1]$ and $s,t \in [0,T]$ $$\EEE{\left| Y^{\varepsilon}_{s,t} \right|^2}^{1/2} \leq C_1 \min\{\varepsilon^{-1}|t-s|^{1/2}, 1\}$$ and $$\EEE{\left| \int_s^t Y^{\varepsilon}_r \otimes Y^{\varepsilon}_r dr - (t-s)C^\varepsilon \right|^2}^{1/2} \leq C_1 \min\{\varepsilon |t-s|^{1/2}, |t-s|\}.$$
The first inequality is clear from , and . For the second, from , we see that for every $r > 0$, $\tilde Y^{\varepsilon}_r$ has distribution $\NN(0, C^\varepsilon_r)$ where $$C^\varepsilon_r = \int_0^r e^{-M^\varepsilon (r-u)} e^{-(M^\varepsilon)^*(r-u)} du = \int_0^r e^{-M^\varepsilon u} e^{-(M^\varepsilon)^* u} du.$$ Hence $Y^\varepsilon_r = \tilde Y_{\varepsilon^{-2}r}$ has distribution $\NN(0, C^\varepsilon_{\varepsilon^{-2} r})$. Thus $$\EEE{\int_s^t Y^{\varepsilon}_r \otimes Y^{\varepsilon}_r dr} = \int_s^t C^\varepsilon_{\varepsilon^{-2} r} dr = \int_s^t \int_0^{\varepsilon^{-2} r} e^{-M^\varepsilon u} e^{-(M^\varepsilon)^* u} du dr =: \mu_{s,t}^\varepsilon.$$ Observe that from $$\begin{aligned}
|\mu_{s,t}^\varepsilon - (t-s) C^\varepsilon| &\leq \int_s^t \int_{\varepsilon^{-2} r}^\infty |e^{-M^\varepsilon u} e^{-(M^\varepsilon)^* u}| du dr \\
&\leq C_2\int_s^t \int_{\varepsilon^{-2}r}^\infty e^{-2\lambda u} dudr \\
&\leq C_3 \int_s^t e^{-2\lambda \varepsilon^{-2}r}dr \\
&\leq C_4 \min\{\varepsilon^2, |t-s|\} \\
&\leq C_4 \min\{\varepsilon |t-s|^{1/2}, |t-s|\}.\end{aligned}$$ We now claim that $$\EEE{\left| \int_s^t Y^{\varepsilon}_r \otimes Y^{\varepsilon}_r dr - \mu_{s,t}^\varepsilon \right|^2} \leq C_5 \min\{\varepsilon^2 |t-s|, |t-s|^2\},$$ from which the conclusion follows. Indeed, by Fubini and Wick’s formula $$\begin{aligned}
\EEE{\left(\int_s^t Y^{\varepsilon,i}_r Y^{\varepsilon,j}_r dr\right)^2}
&= \int_{[s,t]^2} \EEE{Y^{\varepsilon,i}_r Y^{\varepsilon,j}_r Y^{\varepsilon,i}_u Y^{\varepsilon,j}_u} dr du \\
&= \int_{[s,t]^2} \EEE{Y^{\varepsilon,i}_r Y^{\varepsilon,j}_r}\EEE{Y^{\varepsilon,i}_u Y^{\varepsilon,j}_u} dr du \\
&+ \int_{[s,t]^2} \EEE{Y^{\varepsilon,i}_r Y^{\varepsilon,i}_u}\EEE{Y^{\varepsilon,j}_r Y^{\varepsilon,j}_u} dr du \\
&+ \int_{[s,t]^2} \EEE{Y^{\varepsilon,i}_r Y^{\varepsilon,j}_u}\EEE{Y^{\varepsilon,j}_r Y^{\varepsilon,i}_u} dr du \\
%&\leq (\mu_{i,j}^\varepsilon)_{s,t}^2 + \int_{[s,t]^2} \EEE{Y^{\varepsilon,i}_u Y^{\varepsilon,j}_r}^2 + \EEE{Y^{\varepsilon,i}_r Y^{\varepsilon,j}_u}^2 dr du \\
&\leq (\mu_{i,j}^\varepsilon)_{s,t}^2 + 4\int_{[s,t]^2} \left|\EEE{Y^{\varepsilon}_u \otimes Y^{\varepsilon}_r}\right|^2 \mathbf{1}\{r \leq u\} dr du.\end{aligned}$$
Observe that for $r \leq u$ $$\EEE{Y^\varepsilon_u \mid Y^\varepsilon_r} = e^{-\varepsilon^{-2} M^\varepsilon(u-r)} Y_r^\varepsilon.$$ and so $$|\EEE{Y^\varepsilon_u\otimes Y^\varepsilon_r}|^2 \mathbf{1}\{r \leq u\} \leq C_6 e^{-\varepsilon^{-2}2\lambda(u-r)} |C^\varepsilon_{\varepsilon^{-2}r}| \leq C_7 e^{-\varepsilon^{-2}2\lambda(u-r)}.$$ Thus $$\begin{aligned}
\EEE{\left(\int_s^t Y^{\varepsilon,i}_r Y^{\varepsilon,j}_r dr - (\mu_{i,j}^\varepsilon)_{s,t}\right)^2}
&= \EEE{\left(\int_s^t Y^{\varepsilon,i}_r Y^{\varepsilon,j}_r dr\right)^2} - (\mu_{i,j}^\varepsilon)_{s,t}^2 \\
&\leq C_8 \int_s^t \int_r^t e^{-\varepsilon^{-2}2\lambda(u-r)} du dr \\
&\leq C_9 \min\{\varepsilon^2 |t-s|, |t-s|^2\}\end{aligned}$$ as claimed.
\[lem:PBounds\] There exists $C_{10} > 0$ such that for all $\varepsilon \in (0,1]$ and $s,t \in [0,T]$ $$\norm{P^\varepsilon_{s,t}}_{L^2} \leq C_{10} \min\{\varepsilon, |t-s|^{1/2}\}$$ and $$\norm{\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon}_{L^2} \leq C_{10} |M^\varepsilon| \min\{\varepsilon |t-s|^{1/2}, |t-s|\}$$
The first inequality is immediate from Lemma \[lem:YBounds\]. For the second, we have $$\begin{aligned}
\Pbb^\varepsilon_{s,t} &= \varepsilon^2 \int_s^t Y^\varepsilon_{s,r} \otimes \circ dY^\varepsilon_r \\
&= -\int_s^t Y^\varepsilon_{s,r} \otimes M^\varepsilon Y^\varepsilon_r dr + \varepsilon\int_s^t Y^\varepsilon_{s,r}\otimes dW_r + \frac{1}{2}(t-s)I.\end{aligned}$$ Since $Y^\varepsilon_{s,r} \otimes M^\varepsilon Y^\varepsilon_r = (Y^\varepsilon_{s,r} \otimes Y^\varepsilon_r)(M^\varepsilon)^*$ and we can directly verify that $v^\varepsilon = C^\varepsilon(M^\varepsilon)^* - \frac{1}{2}I$, we have $$\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon = -\left(\int_s^t Y^\varepsilon_{s,r} \otimes Y^\varepsilon_r dr - (t-s)C^\varepsilon\right) (M^\varepsilon)^* + \varepsilon \int_s^t Y^\varepsilon_{s,r} \otimes dW_r.$$ From Lemma \[lem:YBounds\], we see that $$\norm{\varepsilon \int_s^t Y^\varepsilon_{s,r} \otimes dW_r}_{L^2} \leq C_{11}\min\{\varepsilon|t-s|^{1/2},|t-s|\}.$$ Furthermore, by Fubini and Wick’s formula, we can readily show $$\norm{\int_s^t Y^\varepsilon_s \otimes Y^\varepsilon_r dr}_{L^2} \leq C_{12} \min\{\varepsilon |t-s|^{1/2}, |t-s|\}.$$ It now follows from Lemma \[lem:YBounds\] that $$\norm{\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon}_{L^2} \leq C_{13} |M^\varepsilon| \min\{\varepsilon |t-s|^{1/2}, |t-s|\}.$$
\[lem:ZBounds\] There exists $C_{14} > 0$ such that for all $\varepsilon \in (0,1]$ and $s,t \in [0,T]$ $$\EEE{|Z^\varepsilon_{s,t} - W_{s,t}|^2}^{1/2} \leq C_{14} \min\{\varepsilon, |t-s|^{1/2}\}$$ and $$\EEE{\left|\Z^\varepsilon_{s,t} + (t-s)v^\varepsilon - \W_{s,t}\right|^2}^{1/2} \leq C_{14} |M^\varepsilon| \min\{\varepsilon |t-s|^{1/2}, |t-s|\}.$$
The first inequality follows from $Z^\varepsilon_{s,t} = W_{s,t} - \varepsilon Y^\varepsilon_{s,t}$ and Lemma \[lem:YBounds\]. For the second, we have $$\begin{aligned}
\int_s^t Z^\varepsilon_{s,r} \otimes dZ^\varepsilon_r
&= \int_s^t Z^\varepsilon_{s,r} \otimes dW_r - \varepsilon\left( \int_s^t Z^\varepsilon_{r} \otimes dY^\varepsilon_r - Z^\varepsilon_s\otimes Y^\varepsilon_{s,t}\right) \\
%&= \int_s^t Z^\varepsilon_{s,r} \otimes dW_r - \varepsilon \left( Z^\varepsilon_t\otimes Y^\varepsilon_t - Z^\varepsilon_s\otimes Y^\varepsilon_s - \int_s^t dZ_r \otimes Y^\varepsilon_r - Z^\varepsilon_s\otimes (Y^\varepsilon_t - Y^\varepsilon_s)\right) \\
&= \int_s^t Z^\varepsilon_{s,r} \otimes dW_r - \varepsilon \left( Z^\varepsilon_t\otimes Y^\varepsilon_t - \int_s^t dZ_r \otimes Y^\varepsilon_r - Z^\varepsilon_s\otimes Y^\varepsilon_t\right) \\
&= \int_s^t Z^\varepsilon_{s,r} \otimes dW_r - \varepsilon Z_{s,t}^\varepsilon\otimes Y_t^\varepsilon + \int_s^t M^\varepsilon Y^\varepsilon_r \otimes Y^\varepsilon_r dr.\end{aligned}$$ We see that $$\norm{\int_s^t Z^\varepsilon_{s,r} \otimes dW_r - \int_s^t W_{s,r} \otimes dW_r}^2_{L^2} \leq C_{15}\min\{\varepsilon^2|t-s|, |t-s|^2\}.$$ Furthermore, by Fubini and Wick’s formula, we can readily show $$\norm{\varepsilon Z^\varepsilon_{s,t}\otimes Y^\varepsilon_t}^2_{L^2} = \norm{\int_s^t M^\varepsilon Y^\varepsilon_{r}\otimes Y^\varepsilon_t dr}^2_{L^2} \leq C_{16}|M^\varepsilon| \min\{\varepsilon^2 |t-s|, |t-s|^2\}.$$ Finally, by Lemma \[lem:YBounds\] $$\norm{\int_s^t M^\varepsilon Y^\varepsilon_r \otimes Y^\varepsilon_r dr - (t-s)M^\varepsilon C^\varepsilon|}_{L^2} \leq C_{17}|M^\varepsilon| \min \{\varepsilon |t-s|^{1/2}, |t-s|\}.$$ It follows that $$\norm{\Z^\varepsilon_{s,r} - \W_{s,t} - (t-s)(M^\varepsilon C^\varepsilon - \frac{1}{2}I)}_{L^2} \leq C_{18}|M^\varepsilon| \min \{\varepsilon |t-s|^{1/2}, |t-s|\}.$$ We can directly verify $v^\varepsilon = -M^\varepsilon C^\varepsilon + \frac{1}{2}I$, from which the conclusion follows.
Observe that condition implies that $$\lim_{\varepsilon \rightarrow 0}|M^\varepsilon|\varepsilon = 0.$$ From Lemmas \[lem:PBounds\] and \[lem:ZBounds\], along with Gaussian chaos, we thus obtain the pointwise convergence as $\varepsilon \rightarrow 0$ for any $q < \infty$ and $s,t \in [0,T]$ in $L^q$ $$|P^\varepsilon_{s,t}| + |\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon|^{1/2} \rightarrow 0$$ and $$|Z^\varepsilon_{s,t} - W_{s,t}| + |\Z^\varepsilon_{s,t} + (t-s)v^\varepsilon - \W_{s,t}|^{1/2} \rightarrow 0.$$ Furthermore, since $\min\{\varepsilon |t-s|^{1/2}, |t-s|\} \leq \varepsilon^{\kappa}|t-s|^{1-\kappa/2}$ for all $\kappa \in [0,1]$, condition , Lemmas \[lem:PBounds\] and \[lem:ZBounds\], and Gaussian chaos imply that for any $q < \infty$ there exists $C_q > 0$ such that for all $s,t \in [0,T]$ and $$\begin{aligned}
\sup_{\varepsilon \in (0,1]} \EEE{|P^\varepsilon_{s,t}|^q} &\leq C_q|t-s|^{q/2}, \\
\sup_{\varepsilon \in (0,1]} \EEE{|Z^\varepsilon_{s,t} - W_{s,t}|^q} &\leq C_q|t-s|^{q/2}\end{aligned}$$ and $$\begin{aligned}
\sup_{\varepsilon \in (0,1]} \EEE{|\Pbb^\varepsilon_{s,t} + (t-s)v^\varepsilon|^q} &\leq C_q|t-s|^{q(1-\kappa/2)}, \\
\sup_{\varepsilon \in (0,1]} \EEE{|\Z^\varepsilon_{s,t} + (t-s)v^\varepsilon - \W_{s,t}|^q} &\leq C_q|t-s|^{q(1-\kappa/2)}.\end{aligned}$$ Applying Theorem A.13 of [@FrizVictoir10] completes the proof.
Rough lead-lag process {#subsec:Hoff}
======================
Consider a path $X : [0,1] \mapsto \R^d$. Let $n \geq 1$ be an integer and write for brevity $X^n_i = X_{i/n}$. Consider the piecewise linear path $\tilde X^n : [0,1] \mapsto \R^{2d}$ defined by $$\begin{aligned}
\tilde X^n_{2i/2n} &= (X^n_i, X^n_i), \\
\tilde X^n_{(2i+1)/2n} &= (X^n_i, X^n_{i+1}),\end{aligned}$$ and linear on the intervals $\left[\frac{2i}{2n}, \frac{2i+1}{2n}\right]$ and $\left[\frac{2i+1}{2n}, \frac{2i+2}{2n}\right]$ for all $i = 0,\ldots, n-1$. Note that this is a variant of the Hoff process considered in [@Flint16].
Denote by $\tilde \Xbf^n_{s,t} = \exp(\tilde X^n_{s,t} + \Abb^n_{s,t})$ the level-$2$ lift of $\tilde X^n$, where $\Abb^n_{s,t}$ is the $(2d) \times (2d)$ anti-symmetric L[é]{}vy area matrix given by $$\Abb^n_{s,t}
%= \left( \begin{array}{cc} \Abb^{b,b}_{s,t} & \Abb^{b,f}_{s,t} \\ \Abb^{f,b}_{s,t} & \Abb^{f,f}_{s,t} \end{array} \right)
= \frac{1}{2}\left(\int_s^t \tilde X^n_{s,r} \otimes d\tilde X^n_r - \int_s^t \tilde X^n_{s,r} \otimes d\tilde X^n_r\right).$$
Let $H \in (0,1)$ and consider a fractional Brownian motion $B^H$ with covariance $R(s,t) = \frac{1}{2}(t^{2H} + s^{2H} - |t-s|^{2H})$. Let $X : [0,1] \mapsto \R^d$ be $d$ independent copies of $B^H$.
Recall the definition of $T_v$ from . We are interested in the convergence in rough path topologies of $T_{\tilde v^n}(\tilde \Xbf^n)$ where $\tilde v^n \in \g^{(2)}(\R^{2d})$ is appropriately chosen. Define the (diagonal) $d\times d$ matrix $$v^n = \frac{1}{2}\EEE{\sum_{k=0}^{n-1} (X^n_{k+1}-X^n_k) \otimes (X^n_{k+1}-X^n_k)} = \frac{n^{1-2H}}{2} I,$$ and the anti-symmetric $(2d) \times (2d)$ matrix $$\tilde v^n = \left( \begin{array}{cc} 0 & -v^n \\ v^n & 0 \end{array} \right) \in \g^{(2)}(\R^{2d}).$$ Finally, consider the path $\tilde X = (X,X) : [0,1] \mapsto \R^{2d}$, its canonically defined L[é]{}vy area $\Abb$ (which exists for $1/4 < H \leq 1$), and its level-$2$ lift $\tilde \Xbf = \exp(\tilde X + \Abb)$. The following is the main result of this subsection.
\[thm:HoffConv\] Suppose $1/4 < H \leq 1/2$. Then for all $\alpha \in [0, H)$ and $q < \infty$, it holds that $T_{\tilde v^n}(\tilde \Xbf^n) \rightarrow \tilde \Xbf$ in $L^q$ and $\alpha$-H[ö]{}lder topology. More precisely, as $n \rightarrow \infty$, in $L^q$ $$\sup_{s,t \in [0,T]} \frac{|\tilde X^n_{s,t} - \tilde X_{s,t}|}{|t-s|^\alpha} + \sup_{s,t \in [0,T]} \frac{|\Abb^n_{s,t} + (t-s)\tilde v^n - \Abb_{s,t}|}{|t-s|^{2\alpha}} \rightarrow 0.$$
The rest of the section is devoted to the proof of Theorem \[thm:HoffConv\]. We first state two lemmas which are purely deterministic.
Let $Y^n : [0,1] \mapsto \R^{d}$ be the piecewise linear interpolation of $X$ over the partition $\left(0, \frac{1}{n}, \ldots, \frac{n-1}{n}, 1\right)$, let $\tilde Y^n = (Y^n,Y^n) : [0,1] \mapsto \R^{2d}$, and let $\Ybb^n$ be the L[é]{}vy area of $\tilde Y^n$.
\[lem:XYDiff\] Let $s \in [\frac{m}{n},\frac{m+1}{n}]$ and $t \in [\frac{k}{n}, \frac{k+1}{n}]$ with $s < t$, and define $$\begin{aligned}
\Delta_1 &= n\left(\frac{m+1}{n} \wedge t - s \right) |X^n_{m+1} - X^n_m|, \\
\Delta_2 &= |X^n_k - X^n_{m+1}| \; \textnormal{ if $k > m$}, \; \; 0 \textnormal{ if $k= m$} \\
\Delta_3 &= n\left(t-\frac{k}{n}\vee s\right)|X^n_{k+1} - X^n_k|.\end{aligned}$$ There exists a constant $C_1 > 0$ such that for all $n \geq 1$ and $0 \leq s < t \leq 1$, it holds that $$|\tilde X^n_{s,t} - \tilde Y^n_{s,t}| \leq C_1\left( \Delta_1 + \Delta_3 \right).$$ and, if $k > m$, $|\Abb^n_{s,t} - \Abb^n_{\frac{m+1}{n}, \frac{k}{n}}|$ and $|\Ybb^n_{s,t} - \Ybb^n_{\frac{m+1}{n}, \frac{k}{n}}|$ are bounded above by $$C_1 \left(\Delta_1^2 + (\Delta_1+ \Delta_2 + \Delta_3)\Delta_3 + \Delta_1 \Delta_2 \right).$$
Direct calculation and triangle inequality.
The second part of the above lemma essentially allows us to work over the partition $\left(0, \frac{1}{n}, \ldots, \frac{n-1}{n}, 1\right)$, on which computations are easier.
\[lem:partitionPoints\] Suppose $0 \leq m \leq k \leq n$.
1\) For all pairs $1 \leq i,j \leq d$ and $d+1 \leq i,j \leq 2d$ $$\left(\Abb^n_{\frac{m}{n}, \frac{k}{n}}\right)^{i,j} = \left(\Ybb^n_{\frac{m}{n}, \frac{k}{n}}\right)^{i,j}.$$
2\) For all $1 \leq i \leq d < j \leq 2d$ $$\left(\Abb^n_{\frac{m}{n}, \frac{k}{n}}\right)^{i,j} = \left(\Ybb^n_{\frac{m}{n}, \frac{k}{n}}\right)^{i,j} - \frac{1}{2} \sum_{r=m}^{k-1} (X^{n,i}_{r+1}-X^{n,i}_r)(X^{n,j}_{r+1}-X^{n,j}_r)$$
Denote $\tilde X^n = (M^n, N^n)$, so that $M^n$ is the lag component, and $N^n$ is the lead. The first equality is clear since $M^n$ and $N^n$ are simply reparametrisations of $Y^n$ over the interval $[\frac{m}{n},\frac{k}{n}]$. For the second, observe that $$\int_{m/n}^{k/n} M^{n,i}_{m/n, r} dN^{n,j}_r = \sum_{r=m}^{k-1} (X^{n,i}_{r}-X^{n,i}_{m}) (X^{n,j}_{r+1} - X^{n,j}_{r})
%= \sum_{r=m}^{k-1} X^{n,i}_{r}(X^{n,j}_{r+1} - X^{n,j}_{r}) - X^{n,i}_{m} (X^{m,j}_{k} - X^{n,j}_{m})$$ and $$\int_{m/n}^{k/n} N^{n,j}_{m/n, r} dM^{n,i}_r = \sum_{r=m}^{k-1} (X^{n,j}_{r+1}-X^{n,j}_{m}) (X^{n,i}_{r+1} - X^{n,i}_{r}).
%= \sum_{r=m}^{k-1} X^j_{r+1}(X^i_{r+1} - X^i_{r}) - X^j_{m} (X^i_{k} - X^i_{m}).$$ Remark now that the signature of $Y^n$ over $[\frac{m}{n},\frac{k}{n}]$ is $$e^{X_{m+1} - X_{m}} \ldots e^{X_k - X_{k-1}},$$ so that a calculation with the CBH formula gives $$\begin{aligned}
\left(\Ybb^n_{m/n,k/n}\right)^{i,j}
%&= \frac{1}{2}\left( \int_{m/n}^{k/n} M^i_{m/n, r} dM^j_r - \int_{m/n}^{k/n} M^j_{m/n, r} dM^i_r \right) \\
&= \frac{1}{2}\sum_{r=m}^{k-1} (X^{n,i}_r-X^{n,i}_m)(X^{n,j}_{r+1}-X^{n,j}_r) - (X^{n,j}_r-X^{n,j}_m)(X^{n,i}_{r+1}-X^{n,i}_r).
%&= \frac{1}{2}\sum_{r=1}^k X^i_r X^j_{r+1} - X^j_{r}X^i_{r+1} - X^i_m(X^j_k-X^j_m) + X^j_m(X^i_k-X^i_m).\end{aligned}$$ Using the fact that $$\begin{aligned}
\left(\Abb^n_{m/n,k/n}\right)^{i,j} &= \frac{1}{2}\left( \int_{m/n}^{k/n} M^{n,i}_{m/n, r} dN^{n,j}_r - \int_{m/n}^{k/n} N^{n,j}_{m/n, r} dM^{n,i}_r \right),
%&= \frac{1}{2}\sum_{r=m}^{k-1} X^i_{r}(X^j_{r+1} - X^j_{r}) - X^j_{r+1}(X^i_{r+1} - X^i_{r}) -X^i_{m} (X^j_{k} - X^j_{m}) + X^j_{m} (X^i_{k} - X^i_{m}) \\
%&= \Ybb^{i,j}_{m/n,k/n} - \frac{1}{2} \sum_{r=m}^{k-1} (X^i_{r+1}-X^i_r)(X^j_{r+1}-X^j_r).\end{aligned}$$ the conclusion readily follows.
We now return to the specific case that $X : [0,1] \mapsto \R^d$ is given by $d$-independent copies of a fractional Brownian motion with Hurst parameter $H \in (0,1)$. In particular, this implies that for all $s,t,a,b \in [0,1]$ $$\label{eq:covariance}
\EEE{(X^i_t-X^i_s)(X^j_b-X^j_a)} = \delta_{i,j} \frac{1}{2}(|t-a|^{2H} + |s-b|^{2H} - |t-b|^{2H} - |s-a|^{2H}).$$ Consider the $(2d) \times (2d)$ anti-symmetric matrix $$\Pbb^n_{s,t} = \Abb^n_{s,t} - \Ybb^n_{s,t} + (t-s)\tilde v^n.$$
\[lem:PControlPartition\] There exists $C_2 > 0$ such that for all $H \leq 1/2$, $n \geq 1$, and $0 \leq m \leq k \leq n$ $$\norm{\Pbb^n_{m/n,k/n}}_{L^2} \leq C_2 \frac{(k-m)^{1/2}}{n^{2H}}.$$
Denote $K = k-m$. By part (2) of Lemma \[lem:partitionPoints\], we have $$\norms{\Pbb^n_{m/n,k/n}} \leq \sum_{i,j=1}^d \norms{\sum_{r=m}^{k-1} (X^{n,i}_{r+1} - X^{n,i}_r)(X^{n,j}_{r+1}- X^{n,j}_r) - \frac{K}{n}v^n_{i,j}}.$$ Observe moreover that $$\frac{K}{n}v^n_{i,j} = \EEE{\sum_{r=m}^{k-1}(X^{n,i}_{r+1} - X^{n,i}_r)(X^{n,j}_{r+1}- X^{n,j}_r)} = \delta_{i,j} K n^{-2H},$$ and that for all $r,\ell \in \{0,\ldots, n-1\}$ $$\EEE{(X^{n,i}_{r+1} - X^{n,i}_r)(X^{n,i}_{\ell+1} - X^{n,i}_\ell)} = \frac{n^{-2H}}{2}(|r-\ell+1|^{2H} + |r-\ell-1)|^{2H} - 2|r-\ell|^{2H}).$$ Then for all $i \neq j$, by independence of the components of $X$, $$\begin{aligned}
& \EEE{\left(\sum_{r=m}^{k-1} (X^{n,i}_{r+1} - X^{n,i}_r)(X^{n,j}_{r+1}- X^{n,j}_r) \right)^2}
\\ &= \EEE{\sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} X^i_{r,r+1}X^i_{\ell,\ell+1} X^j_{r,r+1} X^j_{\ell,\ell+1}} \\
&= \sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} \EEE{(X^{n,i}_{r+1} - X^{n,i}_r)(X^{n,i}_{\ell+1} - X^{n,i}_{\ell})}^2 \\
&= \sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} \frac{n^{-4H}}{4}\left(|r-\ell+1|^{2H} + |r-\ell-1|^{2H} - 2|r-\ell|^{2H} \right)^2 \\
&= \frac{n^{-4H}}{4} \sum_{x=-K+1}^{K-1} (K-|x|)(|x+1|^{2H} + |x-1|^{2H} - 2x^{2H})^2 \\
&=: \psi(n,K).\end{aligned}$$ Likewise for $i=j$, by Wick’s formula, $$\begin{aligned}
\EEE{\left(\sum_{r=m}^{k-1} (X^{n,i}_{r+1} - X^{n,i}_r)^2 \right)^2}
%&= \EEE{\sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} (X^i_{r,r+1}X^i_{\ell,\ell+1})^2} \\
%&= \sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} \EEE{(X^i_{r,r+1}X^i_{\ell,\ell+1})^2} \\
%&= \sum_{r=m}^{k-1} \EEE{(X^i_{r,r+1})^4} + 2\sum_{m \leq r < \ell \leq k-1} \\
&= \sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} \EEE{(X^{n,i}_{r+1} - X^{n,i}_{r})^2} \EEE{(X^{n,i}_{\ell+1} - X^{n,i}_{\ell})^2} \\
&+ 2\EEE{(X^{n,i}_{r+1} - X^{n,i}_{r})(X^{n,i}_{\ell+1} - X^{n,i}_{\ell})}^2 \\
%&= \sum_{r=m}^{k-1} \sum_{\ell=m}^{k-1} n^{-4H} + 2\frac{n^{-4H}}{4}\left(|r-\ell+1|^{2H} + |r-\ell-1|^{2H} + 2|r-\ell|^{2H} \right)^2 \\
&= K^2 n^{-4H} + 2\psi(n,K) \\
&= \left(\frac{K}{n}v^n_{i,i}\right)^2 + 2\psi(n,K).\end{aligned}$$ It hence follows that $$\norm{\Pbb^n_{m/n, k/n}}_{L^2}^2 \leq C_3\psi(n,K).$$ The conclusion now follows since one can readily show that there exists $C_4 > 0$ such that for all $H \leq 1/2$, $n \geq 1$, and $0 \leq K \leq n$ $$\psi(n,K) \leq C_4K n^{-4H}$$ (in fact the inequality holds for all $H < 3/4$, though with a constant in general depending on $H$).
\[lem:YPBounds\] There exists $C_5 > 0$ such that for all $H \in \left(\frac{1}{4},\frac{1}{2}\right]$, $n \geq 1$ and $0 \leq s < t \leq 1$ $$%\norm{\Ybb^n_{s,t}}_{L^2},
\norm{\Pbb^n_{s,t}}_{L^2} \leq C_5|t-s|^{2H}.$$
Suppose $s \in [\frac{m}{n},\frac{m+1}{n}]$ and $t \in [\frac{k}{n}, \frac{k+1}{n}]$. If $m=k$, then $\Abb^n_{s,t} = \Ybb^n_{s,t} = 0$ and $|t-s| < n^{-1}$, so that $$|\Pbb^n_{s,t}| = \left(t- s\right)|\tilde v^n| \leq |t-s|^{2H}.$$ For the case $k > m$, following the notation of Lemma \[lem:XYDiff\], note that $\EEE{\Delta_1^2}$ and $\EEE{\Delta_3^2}$ are bounded above by $$n^{-2H} \min\{n^2|t-s|^2,1\}.$$ It readily follows that for all $\ell \in \{1,2,3\}$ $$\EEE{\Delta^2_\ell} \leq |t-s|^{2H}.$$ Moreover, we have $$\left(t-\frac{k}{n} + \frac{m+1}{n} - s\right)|\tilde v^n| \leq \min \{|t-s|, n^{-1} \} n^{1-2H} \leq |t-s|^{2H}.$$ Hence Lemma \[lem:XYDiff\] implies that $$\begin{gathered}
|\Pbb^n_{s,t} - \Pbb^n_{(m+1)/n, k/n}| \leq 2C_1\left(\Delta_1^2 + (\Delta_1+ \Delta_2 + \Delta_3)\Delta_3 + \Delta_1 \Delta_2 \right) \\ + \left(t-\frac{k}{n} + \frac{m+1}{n} - s\right)|\tilde v^n|,\end{gathered}$$ and so by Gaussian chaos $$\norm{\Pbb^n_{s,t} - \Pbb^n_{(m+1)/n, k/n}}_{L^2} \leq C_6|t-s|^{2H}.$$ The conclusion now follows from Lemma \[lem:PControlPartition\] since $(k-m-1)^{1/2}n^{-2H} \leq |t-s|^{2H}$ for all $H \geq 1/4$.
Let $0 \leq s < t \leq 1$. We observe that as $n \rightarrow \infty$, it readily follows from Lemmas \[lem:XYDiff\] and \[lem:PControlPartition\] that in $L^q$ $$|\tilde X^n_{s,t} - \tilde Y^n_{s,t}| \rightarrow 0$$ and $$|\Abb^n_{s,t} + (t-s)\tilde v^n - \Ybb^n_{s,t}| \rightarrow 0.$$ Furthermore, by Gaussian chaos, Lemma \[lem:XYDiff\] implies $$\sup_{n \geq 1} \EEE{|\tilde X^n_{s,t} - \tilde Y^n_{s,t}|^q} \leq C_q |t-s|^{qH},$$ while Lemma \[lem:YPBounds\] implies $$\sup_{n \geq 1} \EEE{|\Abb^n_{s,t} + (t-s)\tilde v^n - \Ybb^n_{s,t}|^q} \leq C_q |t-s|^{2qH}.$$ Applying Theorem A.13 of [@FrizVictoir10], and the fact that $\exp(\tilde Y^n + \Ybb^n) \rightarrow \tilde \Xbf$ in $\alpha$-H[ö]{}lder topology in $L^q$ ([@FrizVictoir10] Theorem 15.42), completes the proof.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report the first coherent excitation of intramolecular vibrational modes via the nonlinear interaction of a TeraHertz (THz) light field with molecular liquids. A TeraHertz-TeraHertz-Raman pulse sequence prepares the coherences with a broadband, high-energy, (sub)picosecond TeraHertz pulse, that are then measured in a TeraHertz Kerr effect spectrometer via phase-sensitive, heterodyne detection with an optical pulse. The spectrometer reported here has broader TeraHertz frequency coverage, and an increased sensitivity relative to previously reported TeraHertz Kerr effect experiments. Vibrational coherences are observed in liquid diiodomethane at 3.66 THz (122 cm$^{-1}$), and in carbon tetrachloride at 6.50 THz (217 cm$^{-1}$), in exact agreement with literature values of those intramolecular modes. This work opens the door to 2D spectroscopies, nonlinear in TeraHertz field, that can study the dynamics of condensed-phase molecular systems, as well as coherent control at TeraHertz frequencies.
Keywords: TeraHertz, Nonlinear and 2D Spectroscopy, Molecular Dynamics
author:
- 'Marco A. Allodi'
- 'Ian A. Finneran'
- 'Geoffrey A. Blake'
title: Nonlinear TeraHertz Coherent Excitation of Vibrational Modes of Liquids
---
Introduction
============
The microscopic structure and dynamics of condensed-phase matter determine its physical behavior. In liquids, the different types of interactions between a molecule and its surroundings fundamentally shape those dynamics and determine properties as diverse as vapor pressure, electrical conductivity, and a liquid’s ability to solvate different compounds. Gaining new insight into the relationship between molecular dynamics and molecular interactions requires a continued expansion of the frontiers of light-matter interactions.
The TeraHertz (THz, 3-330 cm$^{-1}$, 0.1-10 THz), or far-infrared, region of the electromagnetic spectrum is one such frontier that offers novel probes of a variety of condensed-matter systems such as charge transport in semiconductors,[@Rice2013; @Zaks2012] the biophysics of vision,[@Groma2008] and the coherent control of molecular reactions.[@LaRue2015] Of particular interest to physical chemistry, the THz-active degrees of freedom correspond to soft, large-amplitude modes of a liquid that participate directly in the molecular dynamics. At 298 K, $k_BT$ is 207 cm$^{-1}$. Thus, given the energy of THz modes relative to $k_BT$, these soft, predominantly intermolecular, modes are easily populated near room temperature, and require further study because measuring their dynamics may be key to understanding the properties of liquids such as water.[@SAHamm2013]
A central, open question about such molecular interactions involves the nature of molecular vibrations in liquids and how the vibrational dynamics on a femtosecond-to-picosecond (fs-to-ps, or ultrafast) timescale affect the physical properties of the fluid. The class of nonlinear, ultrafast, spectroscopic measurements, in the perturbative regime, that scale as the applied electric field cubed, generally known as third-order, or $\chi^{(3)}$, nonlinear spectroscopies, have a proven ability to interrogate (sub)ps dynamics. For isotropic systems, such as amorphous glasses or molecular liquids, the $\chi^{(3)}$ terms are the first nonlinear contributions to the perturbative expansion of the polarization due to the symmetry of the system.[@Boyd; @Mukamel] Specifically, two-dimensional (2D) optical and infrared techniques as well as Optical Kerr effect (OKE) spectroscopy are robust third-order methods that continue to enrich our understanding of liquids.[@Fayer2001] The third-order polarization measured in these experiments reports upon the beating of coherences that are generated and perturbed, in a controllable fashion, by the laser pulses delivered to the sample.
However, for 2D measurements in the mid-IR, the modes studied are not thermally populated and thus only indirectly report on the dynamics of the THz-active degrees of freedom. Given the considerable challenges of simply generating and detecting intense THz pulses, the first attempt to investigate these soft THz modes of liquids via sub-ps 2D spectroscopy was with a 2D Raman-based approach.[@Tokmakoff1997] This technique never became broadly implemented because the measured signal is rigorously fifth-order in electric field ($\chi^{(5)}$), which leads to complicated alignment and weak signals that are overwhelmed by cascaded third-order processes.[@Blank1999; @Jansen2000]
Developing a third-order 2D THz spectroscopy would enable new experiments to explore the structure and dynamics of condensed-phase systems. Indeed, pioneering 2D THz work has been done on a variety of systems such as semiconductor quantum wells,[@Kuehn2009; @Kuehn2011] graphene,[@Woerner2013] and nonlinear crystals.[@Somma2014] These techniques demonstrate the breadth of science that can be addressed with THz techniques.
Another approach to third-order THz spectroscopy of molecular systems has made use of [*hybrid*]{} pulse sequences, most successfully 2D Raman-THz.[@SAHamm2013] Such studies have shown that Raman-THz-THz, THz-Raman-THz, and THz-THz-Raman pulse sequences can be quite sensitive to intermolecular modes because they combine both dipolar and polarizability interactions. The combination of interactions specifically provides insight into the anharmonicity of both the modes (mechanical) and the dipole moment surface(s).[@Ikeda2015; @Hamm2012] These hybrid techniques are distinct from 2D Raman spectroscopy, a $\chi^{(5)}$ approach where the field interactions occur via the polarizability, and from OKE spectroscopy, which can measure linear Raman spectra and also employs only polarizability interactions.[@McMorrow1991; @Righini1993] While the pulse sequences used in 2D Raman-THz spectroscopy (Raman-THz-THz, and THz-Raman-THz) enable investigations of molecular dynamics, the measured signals are linear in THz field.[@SAHamm2013]
Nonlinear THz approaches would provide a means of manipulating coherences and exploring the dynamics of molecular systems directly via interactions on the thermally-populated potential energy surface – offering the possibility of measuring and controlling the dynamics of a liquid by interacting with the modes that directly result from the intermolecular interactions. As discussed almost 20 years ago by Okumura and Tanimura, tools such as 2D THz spectroscopy would complement Raman-based approaches that required the involvement of high-lying virtual states, and provide new routes to the observation of dynamics that may not be visible as perturbations to a mode thousands of wavenumbers above the energy of the intermolecular modes (e.g. an intramolecular OH stretch around 2500 cm$^{-1}$ to 3500 cm$^{-1}$).[@Okumura1998]
A THz-THz-Raman (TTR) pulse sequence, for example, would augment those already used in 2D Raman-THz spectroscopy[@Ikeda2015] and generate a $\chi^{(3)}$ signal, which is nonlinear in THz field. The TTR pulse sequence is identical to THz Kerr Effect (TKE) spectroscopy, where large THz field strengths (in excess of 100 kV/cm) create a transient birefringence that is detected by the optical pulse, making TKE spectroscopy the first technique to measure a signal that is nonlinear in THz field on molecular systems.[@Hoffmann2009] Additionally, in a centrosymetric material, this pulse sequence yields a backgroundless measurement of a $\chi^{(3)}$ process. It is the backgroundless/low-background nature of many 2D IR measurements that has been a major part of the success of that technique.
Here, we report the first nonlinear coherent excitation of intramolecular vibrational modes of liquids using THz Kerr effect spectroscopy. We employ a TTR pulse sequence to initiate intramolecular vibrational coherences via the nonlinear interaction of a THz light field with the liquid. With this experimental setup, the excitation proceeds via a resonant-two-photon-like mechanism, which is a distinct from a Raman excitation pathway. While coherent excitation of THz molecular modes was first observed over 25 years ago in the gas phase,[@vanExter1989] the measurement of nonlinear coherent excitation in room temperature liquids opens the door to 2D spectroscopies, which are nonlinear in THz field, and that could ultimately allow for the coherent control of low-frequency vibrational modes in the condensed phase.
Experimental Methods
====================
![ The heterodyne-detected THz Kerr effect spectrometer used for THz-THz-Raman experiments. Note that the 800 nm probe beam is [*s*]{}-polarized and that the THz polarizer is set at 45$^{\circ}$ with respect to that polarization. \[spectrometer\]](figure1.eps){width="8.8cm"}
A diagram of the spectrometer used in these experiments can be seen in Figure \[spectrometer\]. The heart of the laser system is a Legend Elite USP Ti:Sapphire regenerative amplifier (Coherent, Inc.) that produces 35 fs, 4.5 mJ pulses, centered around 800 nm, at a 1 kHz repetition rate. These pulse are used to drive an ultrafast optical parametric amplifier (OPA, Light Conversion Ltd.) that down-converts the pulses further into the near-IR. Approximately 450 mW of 1450 nm light from the signal beam of the OPA passes through a 4-N,N-dimethylamino-4$^{\prime}$-N$^{\prime}$-methyl-stilbazolium 2,4,6-trimethylbenzenesulfonate (DSTMS) crystal (Rainbow Photonics AG) set on a 3 mm diameter aperture to produce $\sim$160 nJ THz pulses. After magnifying the beam waist by a factor of 7.5, the THz beam passes through a wire-grid polarizer set at 45$^{\circ}$ relative to the polarization of the 800 nm probe beam and is then focused onto the sample to generate a THz electric field in excess of 300 kV/cm. To maximize the magnitude of the THz electric field at the sample, the chirp of the 800 nm regenerative amplifier pulses is adjusted while monitoring the THz generation efficiency via electro-optic sampling.[@Wu1995; @Lu1997]
Roughly 2 mW of the regenerative amplifier 800 nm light is split off before the OPA and sent down a delay line (Newport ILS150HA with ESP300 controller) fitted with a retroreflector for precise control of the arrival time of the 800 nm probe pulse at the sample. The probe beam is focused through a hole in the back of the THz-focusing off-axis parabolic mirror in a co-linear geometry with the THz pulse. Both beams hits the sample that is held in a Suprasil quartz (fused silica) cuvette of 1 mm path length.
The 10$^5$:1 polarizer (Throlabs LPVIS050-MP2), $\lambda$/4 plate, and a Wollaston prism placed after the sample – that serves to split the orthogonal polarizations of the probe beam – are critical to the heterodyne-detection scheme. By adjusting the 800 nm polarizer such that almost all the light is in the [*s*]{}-polarization, the small fraction of light that is [*p*]{}-polarized acts as a 90$^{\circ}$ out-of-phase local oscillator that is heterodyned with the molecular Kerr effect signal on a pair of photodiodes. This scheme ensures that only the backgroundless birefringence signal is measured for centrosymmetric media.[@McMorrow1991; @Mukamel] An iris in front of the photodiode is used to attenuate the [*s*]{}-polarized beam so that it is of a comparable magnitude to the [*p*]{}-polarized beam.[@Brunner2014; @Johnson2014] The 1450 nm beam is modulated with an optical chopper at 500 Hz and photodiode signals are sent to a digital lock-in amplifier (SRS SR830), where they are subtracted, and the TTR signal is measured. We verified that the detected TTR signal scales as the applied THz electric field squared ($E_{THz}^2$) by inserting high-resistivity Si plates into the THz path before the wire-grid polarizer and measuring the peak TTR signal of carbon disulfide (Sigma-Aldrich, $>$99%). A plot of this data can be seen in the supplemental information, Figure S1.[^1] More details regarding the instrumentation and optics can be found in our earlier work on linear THz spectroscopy and references therein.[@Allodi2014; @Ioppolo2014]
The liquids studied in this experiment were $>$99% purity and were used as received without further purification. All measurements were taken in a nitrogen () purged atmosphere to remove the components of ambient air, especially water, that are strong THz absorbers. The optical quality of the fused silica cuvette and liquid samples were such that minimal scattered 800 nm light was observed. This significantly increases the sensitivity of the TKE spectrometer.
Results and Discussion
======================
The third-order electric field measured, $E^{(3)}_{Sig}(t)$, is directly proportional to the third-order molecular polarization $P_{i}^{(3)}(t)$ after a phase shift. The polarization is given by: $$P_{i}^{(3)}(t,\tau)=E^{pr}_j (t-\tau) \int^{\infty}_0 d\tau ^{\prime} \, R^{(3)}_{ijkl}(\tau^{\prime})E^{THz^*}_k (t-\tau^{\prime})E^{THz}_l (t-\tau^{\prime}), \:\:\:$$ where $R^{(3)}_{ijkl}(\tau)$ is the 3$^{rd}$-order response function of the liquid, $\tau$ is the time between the THz and 800 nm pulses, and $E^{pr}(t-\tau)$ is the probe pulse electric field.[@Zanni; @Mukamel] The response function, $R^{(3)}_{ijkl}(\tau)$, where the indices correspond to polarization, can be separated into electronic and nuclear components. Since the THz field is non-resonant with any electronic transitions, the electronic component will only contribute to the signal when the applied THz electric field is present, and only the nuclear components of the response function contain information about the molecular dynamics. Given that $R^{(3)}_{ijkl}(\tau)$ is directly related to the emitted polarization, we can relate the measured signal, $S(\tau)$, to the molecular dynamics initiated with the THz pulse.
![Measured third-order HD-TKE signals, generated by a TeraHertz-TeraHertz-Raman pulse sequence, of the liquids in this study, labeled with the point groups of the isolated molecule. Vibrational coherences are clearly visible as oscillations in the diiodomethane and carbon tetrachloride signal. Carbon disulfide yielded the largest response. The bottom trace is the square of the THz electric field applied in these measurements. At time zero, the 800 nm pulse is probing the peak TKE response of the liquid. Positive times on the abscissa denote an interaction with the THz pulse applied first followed by the 800 nm detection pulse. The blue trace below the tetrahydrofuran data shows the empty cuvette background response. Note that all molecular data are normalized and vertically offset for clarity.\[TD\_data\]](figure2.eps){width="8.6cm"}
The heterodyne detection scheme employed allows for the maximum recovery of information about $E^{(3)}_{Sig}(\tau)$. Since the photodiodes that detect the 800 nm probe light function as square-law detectors, the intensity of the light ($I(\tau)$) is measured and the combined signal is integrated over the time variable $t$ because the detectors are not fast enough to measure the oscillating signal in time directly. Without a local oscillator provided by another light field, the homodyne signal can be written as: $$\begin{centering}
S_{hom}(\tau)=\int^{\infty}_{0}dt \:|E^{(3)}_{Sig}(t,\tau)|^2 = I_{sig}(\tau).
\end{centering}$$ By taking the square of the signal and integrating over time, information about the phase of the light field is lost. This phase information is especially important if the signal contains any oscillatory components going as $e^{i\omega t}$, since the integral over the modulus squared would cancel these oscillatory terms.[@Zanni]
![ Modeling of the HD-TKE signal. The experimental data for are plotted at the top in black. The complete model is plotted directly below in red. The individual components of the fit, cuvette (blue), orientational (green), and electronic (magenta), are plotted together on the final line.\[model\]](figure3.eps){width="9cm"}
In contrast, the heterodyne-detected signal can be written as, $$\begin{aligned}
S_{het}(\tau)=\int^{\infty}_{0} dt \: |E_{LO}(t-\tau) + E^{(3)}_{Sig}(t,\tau)|^2\: \: \:\:\: \\
= I_{LO}(\tau) + I_{sig}(\tau) + 2 Re(\int^{\infty}_{0} dt E_{LO}(t-\tau) E^{(3)}_{Sig}(t,\tau)),\:\: \:\:\:\end{aligned}$$ where $Re$ denotes the real part of a complex number that can be measured by physical detectors. The contribution from $I_{LO}(\tau)$ can be removed by a number of common-mode-rejection techniques. In this work, we remove $I_{LO}(\tau)$ by using balanced photodiodes as discussed above. The $I_{sig}(\tau)$ term can be neglected since is small compared to the LO and signal interference term. Thus, heterodyne detection increases the instrumental sensitivity since the measurement scales both with the magnitude of the local oscillator electric field, created using the polarizer, and that of the molecular signal’s electric field. We confirmed the linearity of the heterodyne detection by adjusting the power of the 800 nm beam and observing that the measured signal scaled linearly with the applied 800 nm average power.
If the duration of the LO is shorter than the timescale of the dynamics of interest in $R^{(3)}_{ijkl}(\tau)$, then the envelope of LO pulse can be thought of as a sampling function and approximated as a delta function.[@Torre2007] As such we can write: $$S_{het}(\tau) = 2 Re(\int^{\infty}_{0} dt \, \delta(t-\tau) E^{(3)}_{Sig}(t,\tau)) = 2 Re( E^{(3)}_{Sig}(\tau)).$$ This makes heterodyne detection sensitive to the phase of the emitted signal field, as opposed to homodyne measurements that only capture the signal-field squared. In reality, it is the phase difference that survives the integral, but since the phase of the LO is known, the phase of the signal is also known. The relationship described in eq. (3) ensures that the heterodyne signal is linearly proportional to the response function while the homodyne signal is proportional to the response function squared.[@Torre2007]
The signals generated by a TTR pulse sequence in a variety of room temperature liquids, along with the trace of the THz electric field squared ($E^*_{THz}E_{THz}$), measured via electro-optic sampling in a GaP(110) crystal [@Wu1995; @Lu1997], are presented in Figure \[TD\_data\]. We refer to these equivalently as heterodyne-detected TKE (HD-TKE) or TTR signals. The zero point on the time axis corresponds to the point where the peak of the 800 nm pulse is overlapped with the peak of the THz pulse, yielding maximum signal. At positive times, the 800 nm pulse interacts with the sample after the peak of the THz pulse, while at negative times, the 800 nm pulse interacts with the sample first. This time delay is controlled to high accuracy and precision with an opto-mechanical delay line.
The measured nonlinear TTR response near zero time follows the electric field squared of the THz pulse (E$^2_{THz}$).[@Zhong2008] The nuclear part of the signal remains zero until $\tau=0$ and contains all of the potentially retrievable information about the molecular dynamics of the liquid.[@Righini1993] It is well established in the literature that rotational diffusion of individual monomers returning to an isotropic distribution of molecular orientations is the dominant contribution to the OKE and TKE signals starting several picoseconds after the peak, and that this post-input pulse response can be fit with a decaying exponential. For example, such behavior is clearly visible in the carbon disulfide (CS$_2$) HD-TKE data presented in Figure \[TD\_data\]. The large molecular polarizability of CS$_2$ leads to a large electronic response followed by an orientational response that quickly dominates the signal.
The recovery of a nonlinear TTR signal in tetrahydrofuran (THF, EMD Millipore, $>$99.5%) demonstrates the significant sensitivity improvements that result from heterodyne detection. Our data show a clear response from THF, which was not seen in previous work. The pre-pulse seen in the data before 0 ps is the TTR signature of the fused silica sample cuvette. It is present in all of the data in Figure \[TD\_data\], and provides a reference for the relative molecular responses. The blue trace in Figure \[TD\_data\] shows the cuvette response, scaled to match the pre-pulse seen in the the THF trace and slightly offset for clarity. The cuvette signal lasts for just over 4 ps. It is quite similar in size to that of the tetrahydrofuran signal, but stable from experiment to experiment. As a result, the cuvette response can be removed from the full trace to isolate the nonlinear TTR signal of strongly (THz) absorbing liquids. Straight-forward improvements to the apparatus, such as the inclusion of a liquid jet, will improve the sensitivity in future experiments. Finally, we note that the cuvette signal is so weak relative to the TKE response of the other liquids studied that its presence in the data does not affect the conclusions drawn from these experiments.
To further understand the interaction of the THz pulse with diiodomethane (Sigma-Aldrich, $>$99%), we used a three component model of the data, including electronic, orientational, and cuvette responses, the components and results of which are seen in Figure \[model\]. For the electronic response we applied a minimal moving average to the E$^2_{THz}$ pulse measured by electro-optic sampling. The orientational response was modeled based on the previous work of Hoffman et al.[@Hoffmann2009] by stepping over the measured E$^2$ pulse with the following finite-difference solution to the differential equation: $$S(t+ \Delta t)= S(t) + [E^2(t+ \Delta t) - S(t)/\tau_o]*\Delta t ~~,$$ where $S(t)$ is the measured birefringence signal, $\Delta t$ is the time step, $E^2$ is the measured THz pulse, and $\tau_o$ is the orientational decay time (13 ps for ). The cuvette response was measured with an empty cuvette, and the full modeled response is a linear combination of the three components that are each scaled to match the data. The qualitative agreement with the results is quite good. When all three components are considered, a fourth oscillatory signal is clearly visible in the CH$_2$I$_2$ and CCl$_4$ traces.
{width="\textwidth"}
In diiodomethane, these previously unmeasured oscillations, highlighted in Figure \[beats\]a, correspond to a vibrational coherence that lasts for several ps, excited by the high-field-strength THz pulse. While the THz excitation pulse extends to roughly 1 ps after $\tau=0$ (as seen in the E$^2_{THz}$ trace at the bottom of Figure \[TD\_data\]), the oscillating signal lasts for *at least* another 5 ps. Both the electrons and nuclei in the molecule will respond to the THz field, but the Born-Oppenheimer approximation tells us that the electron motion will cease only a few fs after the THz pulse has propagated through the sample.[@Born1927] As a result, once the THz pulse no longer interacts with the sample, the only contribution to the THz Kerr effect signal arises from the nuclear portion of the response function. This separation of nuclear and electronic response has been demonstrated in the THz Kerr effect work of Hoffmann et al.[@Hoffmann2009]; is well established in the OKE literature [@Zhong2008]; and can be seen in the model of the data employed in Figure \[model\].
Given that the coherence lasts much longer than the THz pulse, the observed oscillations cannot result from the electronic response to the applied THz field, but rather must arise from the coherently-excited molecular motions in the liquid. While the orientational component is also part of the nuclear response, the coherences remain after biexponential fits are used to remove both the initial electronic response, and the orientational response, as Figure \[beats\]b shows.Two distinct time constants in the orientational decay require a biexponential model for a good fit, with the shorter timescale resulting from molecular dynamics akin to those revealed in the so-called intermediate response in OKE spectroscopy. The use of the biexponential model and the intermediate response will be discussed in more detail in an upcoming publication.
These coherences were likely not observed in the previous TKE measurements on molecular systems because of the limited frequency content available via room-temperature LiNbO$_3$ tilted-pulse-front THz generation and because of the phase information, high sensitivity, and dynamic range provided by heterodyne detection. Finally, we checked the signal scaling of these oscillations and they match the scaling of the orientational response to the applied fields.
The nature of the chemical bonds in CH$_2$I$_2$ leads to a fundamental normal vibrational mode, $\nu_4$, at 3.66 THz (122 cm$^{-1}$). [@Johnson2006] We confirm that the coherence corresponds to this mode by taking a numerical Fourier transform of the data starting at time zero and using an asymmetric Hann window for apodization.[@Ioppolo2014; @Galvao2007] The magnitude of the complex-valued Fourier transform (also known as the absolute value of the Fourier transform) is plotted in the right panel of Figure \[beats\]. Similar coherences are also observed in carbon tetrachloride (CCl$_4$, Sigma-Aldrich, $>$99.5%). Here, the two interactions with the THz field are able to excite a coherence in the intramolecular $\nu_2$ mode at 6.5 THz (217 cm$^{-1}$)[@Gabelnick1976] that is also visible in the right panel of Figure \[beats\].
These results are in excellent agreement with the literature values reported for the pure liquids[@Johnson2006; @Gabelnick1976], and well within the Fourier transform-limited spectral resolution of $\sim$5 cm$^{-1}$ produced by the 6 ps traces. Indeed, the difference between the literature value for the peak center and our measurement for diiodomethane is only 1 cm$^{-1}$. We also note that while the linewidth of these two features is produced by the observed dephasing time of the coherences, conclusive statements as to whether the homogeneous linewidth is commensurate with that observed in linear spectroscopy of the liquid(s) will require higher signal-to-noise data.
![A Jablonski diagram showing the relative energy of the interactions and a representative double-sided Feynman diagram showing one possible pathway that leads to the vibrational coherence of the $\nu _4$ vibrational mode in diiodomethane. The green arrows correspond to THz absorption processes. The red arrows indicate an anti-Stokes Raman process. Note that the two green arrows in the Feynman diagram are slightly separated by time $\Delta {\rm t _ {12}}$ for diagramatic purposes, however both interactions come from the same THz pulse as indicated by $\Delta {\rm t _ {12}}=0$. The orange trace on the right of the Jablonski diagram corresponds to the power spectrum of the THz electric field and shows the bandwidth covered by the pulse. The green trace corresponds to the power spectrum of the THz electric field squared. The spacing between THz and optical pulses is labeled $\tau$. These figures are drawn to show that the system starts with some number of vibrational quanta, $\nu$, and is excited through some intermediate state, $i$, to a coherence between a state of $\nu + 1$ and $\nu$, which generates the observed TKE signal. \[Jablonski\]](figure5.eps){width="8.8cm"}
The results from diiodomethane and carbon tetrachloride are the first measurements of vibrational coherences that result from the *nonlinear* coherent excitation with THz radiation ($E^2_{THz}$) in a condensed-phase, molecular system. The measured signal must be third-order in electric field (a $\chi^{(3)}$ process) given the isotropic nature of the samples. It is also backgroundless because we do not spectroscopically detect the linear THz response. Our signal is detected on Si photodiodes at 800 nm, which do not respond to incident THz photons, and we do not spectroscopically resolve the 800 nm photons we detect. Thus, we are not measuring a linear optical spectrum with our birefringence measurements. Finally, the second-order response is zero since $\chi^{(2)}$ is zero by symmetry in any isotropic system.[@Boyd] TKE measurements then detect the transient birefringence that results from the nonlinear coherent excitation of THz active modes.
One can see that this is a birefringence measurement by considering the polarization of the THz excitation and optical probe beams. The third-order response function, $R^{(3)}_{ijkl}(\tau)$, is a rank 4 tensor, where the indices correspond to the polarization of the electric fields of the light applied. All $\chi^{(3)}$ nonlinear spectroscopies follow the same polarization rules. In the dipole approximation, the different polarizations must be in pairs, e.g. YYXX, since an odd number of terms, e.g. YYYX, would result in an odd number of cosine terms in the integral over the angular dependence of the response function, which would integrate to zero.[@Zanni] Thus, the allowed polarizations in the dipole approximation are XXXX, YYXX, XYXY, XYYX, (and further substitutions of X and Y), where, for example, XXYY would be pump with Y-polarization twice, pump on X polarization once, then detect emission from the X-polarization. Both OKE and TKE signals are a linear combination of XXXX and YYXX, since the probe pulse is tilted at 45 degrees with respect to the pump. Specifically, the pump is oriented in the X+Y polarization (+45 degrees), the 800 nm probe is aligned to be initially s-polarized, Y, and we detect a signal that is p-polarized, X. As shown by Mukamel, the fact that the signal is 90$^{\circ}$ out-of-phase from the polarization of the probe ensures that this samples the birefringence, as opposed to dichroism.[@Mukamel] In addition, the overall signal is XXXX-YYXX, equal to the XYXY response by symmetry,[@Torre2007] so R$^{(3)}_{ijkl}$ = R$_{XYXY}$ in eq. (1), which is the same polarization response as OKE spectroscopy. That is, the transient birefringence, which is analyzed with a crossed polarizer to yield the signal, is measured from the differences in polarized emission from the sample.
Thus, the coherences observed in these measurements are generated via two, simultaneous, THz field interactions analogous to resonant two-photon absorption. Figure \[Jablonski\] contains both a Jablonski diagram and a double-sided Feynman diagram illustrating this process. One can think of this as an absorption mechanism in a three level system, where two photons are absorbed during the temporal duration of the THz pulse. The system starts out with some number of vibrational quanta, $\nu$, on a given vibrational manifold. In one possible pathway, the first photon (corresponding to the first green arrow in either the Jablonski or Feynman diagram in Figure \[Jablonski\]) takes the system into a coherence between the $\nu$ state and some intermediate state, while the second photon excites the system further to generate the coherence between the $\nu$ and the $\nu +1$ state. Other pathways starting from excited vibrational states likely contribute, given that $k_BT$ at 300 K is 6.2 THz. However, in this work, the time between the two THz field interactions, denoted by $\Delta {\rm t _ {12}}$ in Figure \[Jablonski\], cannot be independently controlled, and without this experimental handle, these measurements cannot determine the exact pathways. A mechanism involving a three-level system certainly would generate the oscillatory features observed in the TKE traces, and the Feynman diagram in Figure \[Jablonski\] represents one such pathway that would create such a coherence.
Several different groups have used a three-level system to explain the nonlinear spectra of low-frequency modes in molecular systems. This is similar to the mechanism used by Fleischer et al. to understand the two-quantum coherences generated by intense THz pulses in a gas-phase molecular sample,[@Fleischer2012] by Tokmakoff et al. to describe the pathways that contribute to 2D Raman signals,[@Tokmakoff1997] and by Hamm and Savolainen to describe hybrid Raman-THz spectroscopy.[@Hamm2012] Also, several authors have discussed how, for this type of mechanism, the signal should be sensitive to the anharmonicity of the vibrational modes, and that the effect of the anharmonicity should be quite clear in a 2D TTR experiment.[@Hamm2012; @Ikeda2015]
The orange trace on the right side of the Jablonski diagram in Figure \[Jablonski\] depicts the bandwidth of the THz pulse, which extends to roughly 2 THz. This is insufficient to directly excite the 3.66 THz mode of CH$_2$I$_2$. Further, we note that the Suprasil quartz cuvette used here has a strong cutoff starting around 3 THz, where $\alpha$ is around 15 cm$^{-1}$, increasing to $\sim$20 cm$^{-1}$ at 3.6 THz, and rising rapidly to $\alpha = 50~$cm$^{-1}$ by 5.5 THz.[@CunninghamPhD2010] As such, few 3.66 THz photons access the liquid, likely precluding a resonant dipolar mechanism involving any frequencies at or above the $\nu _4$ fundamental.
The THz pulse energies available here place the experiment squarely in the perturbative regime. As such, only third-order terms contribute to the observed signal, as can be shown by doing a simple estimation of the Rabi frequency for a 0.1 D dipole (a likely overestimation of the transition dipole of the 122 cm$^{-1}$ mode in CH$_2$I$_2$) in a 300 kV/cm electric field. When multiplied by the 200 fs FWHM of the THz field, a pulse strength of $\pi$/165 results. Excitation squarely in the strong-field regime requires a $\pi$/2 pulse, that would need THz-pulse field strengths of $\sim$30 MV/cm (or roughly 100 times the current field strength).
A strong-field excitation mechanism would lead to higher-order polarization terms contributing to the signal. If this were the case, one might expect to observe coherences between higher-lying vibrational modes as well, given our detection bandwidth. As discussed above, the coherences are detected in the time domain by adjusting the timing ($\tau$) between the arrival of the THz pulse and the $\sim$35 fs 800 nm pulse by sweeping a delay line. The delay line runs at a constant velocity, collecting points that amount to a 5 fs spacing. The convolution of the 35 fs pulse with this oversampling should yield an effective spacing of independent points in the time domain of 10-20 fs. The duration of the 800 nm pulse then limits the bandwidth of the measurement to 25-50 THz. As a result, if our generation mechanism did indeed generate coherences of higher-lying modes, up to $\sim$50 THz, they would be present in the data. We have confirmed that there are no such features in the data and have added a plot to the supplementary information showing the extended frequency-domain response.$^{27}$
Consequently, the only way to put more than 3 THz of energy into the vibrational mode of CH$_2$I$_2$ is via a process like resonant two-photon absorption involving a three-level system (shown in Figure \[Jablonski\]), which is clearly distinct from a Raman process. Given the 2 THz bandwidth contained within the pulse, two THz field interactions can potentially access a mode 4 THz above the initial energy of the molecule(s). However, the measured intramolecular mode of carbon tetrachloride sits at 6.5 THz. As such, the exact mechanism of excitation in CCl$_4$ is unclear. It may be a result of excitation from thermally excited states, and future experiments with a temperature controlled cell would help support or refute this hypothesis.
Conclusion
==========
In this work, we have measured the coherent excitation of intramolecular vibrational modes initiated by a nonlinear TeraHertz interaction with a molecular liquid. While the details of the excitation mechanism are still unclear, two quanta of THz light are needed to generate the observed response, and the intermolecular modes below 3 THz may well be involved in enhancing the signal strength.[@Mukamel] One means of testing this conjecture is with studies of the temperature dependence of the TTR response. In liquids such as CH$_2$I$_2$, and CCl$_4$, THz frequency modes are thermally populated, and it is precisely these modes that determine the molecular dynamics. Thus, they are distinct from modes in the mid- or near-IR, where $h\nu \gg k_B T$.
As a result, this heterodyne-detected TTR technique opens up new areas in the study of molecular interactions in liquids, especially with the further development of approaches that can manipulate and detect vibrational coherences in a variety of systems – including highly polar species. In particular, the stage is set for a two-THz-pulse TTR experiment, where a coherence is initiated with one THz pulse and subsequently perturbed with another THz pulse, or where the time spacing between the two pulses is varied to maximize the nonlinear THz coherent excitation. This will ultimately result in a 2D TTR spectroscopy. Beyond pure liquids, there are many different condensed-phase systems such as molecular ices[@Allodi2014], or polypeptides[@Yamaguchi2005; @Yamamoto2005] that have complex THz spectra with distinct features. The dynamics of these modes, or even the specific intra- versus intermolecular nature of the modes themselves, remain poorly understood, and 2D TTR experiments would be able to address many open questions.
The multi-cycle nature of the pulses used in this experiment would fundamentally limit the temporal resolution that could be achieved with a two-dimensional approach, but switching to a plasma-based gas photonic THz source, or thinner crystal emitters, would yield sub-cycle THz pulses – thus opening the possibility of $<$100 fs resolution in nonlinear THz measurements.[@Kim2008] Such temporal resolution may be required to understand the anharmonicity of intra- and intermolecular vibrations and would provide novel data that probe how the molecular dynamics on ultrafast timescales affects the physical properties in the condensed phase.
Extensions of this work will eventually enable coherent control of vibrational modes in liquids. Since THz pulses can directly access the soft modes that participate in the room-temperature dynamics, varying the time between pulses in a two-pulse scheme will provide one handle for the coherent excitation, as has already been demonstrated in the gas phase.[@Fleischer2012] In addition, controlling the shape of THz field[@Chen2011] can provide another handle that can be used for THz coherent control. Given the straight-forward improvements that are possible, these TTR results offer the first glimpse into unexplored realms of condensed-phase dynamics and control provided by THz radiation.
Acknowledgments
===============
The authors gratefully acknowledge financial support from the National Science Foundation (Grants No. CHE-1214123, CHE-0722330, and the NSF Graduate Research Fellowship Program). MAA acknowledges current support from a Yen Postdoctoral Fellowship from The Institute for Biophysical Dynamics at The University of Chicago.
[42]{}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.110.137404) @noop [****, ()]{} [****, ()](\doibase
10.1073/pnas.0706336105) [****, ()](\doibase 10.1103/PhysRevLett.115.036103) [****, ()](\doibase 10.1073/pnas.1317459110) @noop [**]{} (, ) @noop [**]{} (, ) , ed., @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRevLett.79.2702) [****, ()](\doibase http://dx.doi.org/10.1063/1.479591) [****, ()](\doibase http://dx.doi.org/10.1063/1.481795) [****, ()](\doibase http://dx.doi.org/10.1063/1.3120766) [****, ()](\doibase
10.1021/jp1099046) [****, ()](http://stacks.iop.org/1367-2630/15/i=2/a=025039) [****, ()](\doibase 10.1103/PhysRevLett.112.146602) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
http://dx.doi.org/10.1016/S0009-2614(98)00968-3) [****, ()](\doibase http://dx.doi.org/10.1063/1.3271520) [****, ()](\doibase 10.1364/OL.14.001128) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1364/JOSAB.31.000904) [****, ()](\doibase
10.1364/JOSAB.31.001035) @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) , ed., [**](https://books.google.com/books?id=zTLY-ypXUBYC) (, ) @noop [****, ()]{} [****, ()](\doibase 10.1002/andp.19273892002) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.1840409) [****, ()](\doibase 10.1103/PhysRevLett.109.123603) **, @noop [Ph.D. thesis]{}, () [****, ()](\doibase http://dx.doi.org/10.1063/1.1857080) [****, ()](\doibase {10.1529/biophysj.105.067447}) @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.3624919)
[^1]: See supplemental material at \[URL will be inserted by AIP\] for the plot of TTR signal scaling as a function of applied THz field, Figure S1, a calculation of THz field strength, and the frequency-domain data out to the Nyquist bandwidth of the measurement, Figure S2.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Anisotropic pairing between fermion species with different fermi momenta opens two-dimensional areas of gapless excitations, thus producing a spatially homogeneous state with coexisting superfluid and normal fluids. This breached pairing state is stable and robust for arbitrarily small mismatch and weak p-wave coupling.'
author:
- 'E. Gubankova'
- 'E.G. Mishchenko'
- 'F. Wilczek'
title: 'Breached superfluidity via p-wave coupling'
---
[*Introduction.*]{} Recently there has been considerable interest in the possibility of new forms of superfluidity that could arise when one has attractive interactions between species with different fermi surfaces. This is stimulated by experimental developments in cold atom systems [@coldatoms] and by considerations in high-density QCD [@ARW]. Possible coexistence of superfluidity with gapless excitations is an especially important qualitative issue. Spatially homogeneous superfluid states that coexist with gapless modes at isolated points or lines in momentum space arise in a straightforward way when BCS theory is generalized to higher partial waves [@Mineev]. Gapless states also are well known to occur in the presence of magnetic impurities [@Abrikosov] and, theoretically, in states with spontaneous breaking of translation symmetry [@LOFF], where the gapless states span a two-dimensional fermi surface. Strong coupling between different bands also may lead to zeros in quasiparticle excitations and gapless states [@Volovik]. For spherically symmetric (s-wave) interactions a spherically symmetric [*ansatz*]{} of this type naturally suggests itself when one attempts to pair fermions of two different species with distinct fermi surfaces, and a pairing solution can be found [@Sarma; @LW; @GLW; @ABR; @SH]. The stability of the resulting state against phase separation [@Bedaque] or appearance of a tachyon in the gauge field (negative squared Meissner mass) [@WuYip] is delicate, however [@FGLW; @AKR]. It appears to require some combination of unequal masses, momentum-dependent pairing interactions, and long-range neutrality constraints. Here we demonstrate another possibility: direction-dependent interactions, specifically p-wave interactions. In this case, stability appears to be quite robust. It seems quite reasonable, intuitively, that expanding an existing (lower-dimensional) locus of zeros into a two-dimensional zone should be significantly easier than producing a sphere of gapless excitations “from scratch”. We shall show that it even occurs for arbitrarily small coupling and small Fermi surface mismatch.
Interactions relevant to pairing can be dominated by p-wave (or higher) harmonics under several circumstances. If the s-wave interaction is repulsive, it will not be subject to the Cooper instability, and will not induce pairing. The Cooper instability can be regarded as an enhancement of the effective interaction for attractive channels as one integrates out high-energy modes near the Fermi surface. Thus the effective Hamiltonian will come to resemble the form we assume if the interspecies interactions are repulsive in s-wave but attractive in p-wave. Fermi statistics forbids diagonal (intraspecies) s-wave interactions; if the higher partial waves are repulsive, or weakly attractive, the model discussed here will apply. In the context of cold atom systems, tuning to an appropriate p-wave Feshbach resonance, as recently reported in [@Zhang], can insure interspecies p-wave dominance.
A crucial difference between the model we consider and the conventional p-wave superfluid system, $^{3}$He, lies in the distinguishability of the paired species. So although there are two components, there is no approximate quasispin symmetry, and no analogue of the fully gapped B phase [@BW]. In the absence of a magnetic field $^{3}$He has an approximate $SO(3)_L\times SO(3)_S\times U(1)$ symmetry under separate spatial, rotations, spin rotations, and number, which is spontaneously broken to the diagonal $SO(3)_{L+S}$ in the B phase. The residual symmetry enforces a gap of uniform magnitude in all directions in momentum space. The systems we consider have quite different symmetry and breaking patterns, for example $U(1)_{L_z}\times U(1)_A\times U(1)_B \rightarrow U(1)_{L_z + A +B}$ for two spin-polarized species A, B in a magnetic field, or $SO(3)_L\times U(1)_A\times U(1)_B \rightarrow U(1)_{L_z + A +B}$ if the magnetic field can be neglected. The reduced residual symmetry allows for interesting direction-dependent structure in momentum space. (In the A phases $^{3}$He pairs effectively as two separate single-species systems, which again is quite different from our set-up.)
Experimental realizations of $p$-wave interaction in cold atom systems have been reported recently in Ref. [@TRJB]. Feshbach resonance in $p$-wave occurs between $^{40}$K atoms in $f=9/2$, $m_f=-7/2$ hyperfine states. This is in contrast to the $s$-wave resonance, which occurs between non-identical $f=9/2$, $m_f=-9/2$ and $f=9/2$, $m_f=-7/2$ states [@LRTB]. A promising system for possible observation of the $p$-wave breached pairing superconductivity is a mixture of $f=9/2$, $m_f=-9/2$ and $f=9/2$, $m_f=-7/2$ atoms $^{40}$K tuned into the repulsive side of the $s$-wave Feshbach resonance. Different densities (Fermi momenta) of $m_f=-9/2$ and $m_f=-7/2$ particles can be prepared using different magnitudes of an initial additional magnetic field, which is then removed. Large atomic relaxation times ensure that the created (metastable) states will exist long enough to allow formation of a superfluid phase.
[*Model:*]{} Having in mind cold atoms in a magnetic trap with atomic spins fully polarized by a magnetic field, we consider a model system with the two species of fermions having the same Fermi velocity $v_F$, but different Fermi momenta $p_F\pm I/v_F$. The effective Hamiltonian is $$\label{ham} H=\sum_{\bf p} [ \epsilon^A_{ p} a_{\bf p}^\dagger
a_{\bf p}+ \epsilon^B_{p} b_{-\bf p}^\dagger b_{\bf -p} -
\Delta_{\bf p}^{*} a_{\bf p}^\dagger b^\dagger _{-\bf p} -
\Delta_{\bf p} b_{-\bf p} a_{\bf p}]$$ with $\epsilon^A_{p}=\xi_p +I,~\epsilon^B_{p}=\xi_p -I$, $\xi_p=v(p-p_F)$, $\Delta_{\bf p}=\sum_{\bf k} V_{\bf p-k}\langle
a_{\bf k}^\dagger b^\dagger _{-\bf k}\rangle$. Here the attractive inter-species interaction is $-V_{\bf p-k}$ within the “Debye” energy $2\omega_D$ around the Fermi surface ($\omega_D\gg I$), and the intra-species interaction is assumed to be either repulsive or negligibly small. Excitations of the Hamiltonian (\[ham\]), $E_{\bf p}^{\pm}=\pm \sqrt{\xi^2_p+\Delta_{\bf p}^2}
+I$, are gapless provided that there are areas on the Fermi surface where $I>\Delta_{\bf p}$. The gap equation at zero temperature, $$\Delta_{\bf p}=\frac{1}{2}\sum_{\bf k}V_{\bf p-k}
\frac{\Delta_{\bf k}}{\sqrt{\xi^2_k+\Delta_{\bf k}^2}}~
\theta\left(\sqrt{\xi^2_k+\Delta_{\bf k}^2}-I\right),$$ can be simplified by taking the integral over $d\xi_p$, $$\begin{aligned}
\Delta_{\bf n} &=&\nu \int \frac{d \o_{\bf n'} }{4\pi}
V({\bf n,n'}) \Delta_{\bf n'} \left(\ln{\frac{1}{|\Delta_{\bf
n'}|}} \right. \nonumber\\
&+& \left.\Theta (I-|\Delta_{\bf n'}|)\ln{\frac{|\Delta_{\bf
n'}|}{I+\sqrt{I^2-\Delta^2_{\bf n'}}} } \right).
\label{gapeq}\end{aligned}$$ where $\nu=p_F^2/(2\pi^2 v_F)$ is the density of states. In the last expression it is assumed that $I$ and $\Delta_{\bf n}$ are dimensionless, and scaled with the “Debye” energy: $I \to 2\omega_D I$, $\Delta_{\bf n} \to 2\omega_D \Delta_{\bf n}$. In deriving Eq. (\[gapeq\]) we neglect dependence of $V_{\bf p-k}$ on the absolute values of ${\bf p}$ and ${\bf k}$; this is valid for $\omega_D \ll E_F$. At weak coupling we may linearize in the partial wave expansion, $V({\bf n,n'})=\sum_{l,m} V_lY_{lm}({\bf n})Y^{\star}_{lm}({\bf n'})$. Assuming p-wave dominance, we parameterize $V({\bf n,n'})=g ({\bf n}\cdot {\bf n'})$ with $g>0$. P-wave gap parameters can arise in the forms $Y_{10}$ and $Y_{1\pm 1}$, describing polar and planar phases respectively.
[*Polar phase:*]{} $\Delta_{\bf n} \sim Y_{10}({\bf n})$. We look for a solution in the form $\Delta_{\bf n}=\Delta
\cos({\bf n,z})$ where ${\bf z}$ is a fixed but arbitrary direction (rotational symmetry is broken). The gap equation becomes, $$\begin{aligned}
&-&\frac{1}{\nu g} = \int\limits_0^{\pi/2} d\theta
\sin{\theta}\cos^2{\theta} \ln\left(\Delta\cos{\theta}\right) \\
&+& \int\limits_{\theta^*}^{\pi/2} d\theta \sin{\theta}
\cos^2{\theta} \ln\left(\frac{z+\sqrt{z^2-\cos^2{\theta} }}
{\cos{\theta} } \right) \nonumber \label{gap_polar}\end{aligned}$$ where $\theta^*=\arccos{z}$, for $z=I/\Delta <1$, and $\theta^*=0$ for $z>1$. Performing the integrations (detailed calculations will be given elsewhere [@long]) we obtain the algebraic gap equation, $$\begin{aligned}
&& \ln\left(1/y\right)=z^3\frac{\pi}{4} ~,\hspace{1.5cm}(z<1)
\nonumber\\
&& \ln\left(\frac{1/y}{z+\sqrt{z^2-1}}\right)
=-\frac{z}{2}\sqrt{z^2-1} +\frac{z^3}{2}\arcsin[{{z}^{-1}}]~,
\nonumber\\
&& \hspace{4cm} (z>1)
\label{curve_polar}\end{aligned}$$ where $y=\Delta/\Delta_0$ is the relative magnitude of the gap compared to its value at $I=0$, $$\Delta_0^{pol}={\rm exp}\left(-\frac{3}{\nu g}+\frac{1}{3}\right)
\approx 1.40~{\rm exp}\left(-\frac{3}{\nu g}\right).$$ for weak coupling. There is a factor $3$ in the exponent with anisotropic interaction instead of $1$ as in the s-wave BCS. For small values of $z$ the solution to the gap equation is $y=1-\frac{\pi x^3}{4}$ with $x=I/\Delta_0$. We depict the solution of the polar phase gap equation in Fig. 1, with the following numerical values of the characteristic points: $x_A=(4/3\pi e)^{1/3}=0.538$, $y_A=e^{-1/3}=0.717$ (at the point A $y'(x)\rightarrow\infty$), $y_C=e^{-1/3}/2=0.358$ (at the point C $y(x)=0$).
\[fig2\]
[*Planar phase:*]{} $\Delta_{\bf n} \sim Y_{11}({\bf n})$, $\Delta_{\bf n} \sim Y_{1-1}({\bf n})$. We now look for a solution in the form $\Delta_{\bf n}
=\Delta\sin{({\bf n},{\bf z})e^{i\phi}}$, where $\phi$ is the polar angle in the plane perpendicular to ${\bf z}$. The gap equation becomes, $$\begin{aligned}
-\frac{2}{\nu g} &=& \int\limits_0^{\pi/2} d\theta
\sin^3{\theta} \ln{\left(\Delta\sin{\theta}\right)}
\nonumber\\
&&+ \int\limits_0^{\theta^*} d\theta \sin^3{\theta}
\ln\left(\frac{z+\sqrt{z^2-\sin^2{\theta} }}{ \sin{\theta} }
\right) \label{gap_planar}\end{aligned}$$ where $\theta^*=\arcsin{z}$, for $z<1$, and $\theta_0=\pi/2$, for $z>1$. Performing the integration we obtain the algebraic gap equation, $$\begin{aligned}
\ln{(1/y)}=-\frac{z^2}{4}+\frac{z}{8}(3+z^2)
\ln{\left|\frac{1+z}{1-z} \right|}+\frac{1}{2} \ln{|1-z^2|},
\label{curve_planar}\end{aligned}$$ where again $y$ is the relative magnitude of the gap compared to its value for a zero mismatch. For the planar phase, $$\Delta_0^{pl}=\frac{1}{2}~ {\rm exp}\left(-\frac{3}{\nu
g}+\frac{5}{6}\right)\approx 1.15~{\rm exp}\left(-\frac{3}{\nu
g}\right).$$ For small values of $z$ the solution to the gap equation has the form $y=1-\frac{3x^4}{4}$ and $x$ is defined as before. Note that the planar phase gap is more robust than the polar phase, being perturbed by the fourth power instead of the third. Solution of the gap equation for the planar phase is depicted in Fig. 1, with the following numerical values of the characteristic points: $x_A=0.674$, $y_A=0.787$, $z_A=x_A/y_A=0.856$; $x_C=e^{-5/6}= 0.435$.
[*Stability:*]{} The condensation energy is given by (at $T=0$), $$\begin{aligned}
\Omega_s-\Omega_n&=&\nu\int\frac{d \o_{\bf n} }{4\pi}
\left(-\frac{|\Delta_{\bf n}|^2}{2}+I^2\right.\nonumber\\
&-&\left.I\sqrt{I^2-|\Delta_{\bf
n}|^2}~\Theta(I-|\Delta_{\bf n}|) \right).\end{aligned}$$ Evaluating this expression for $z=I/\Delta<1$, we obtain for polar phase $$\Omega_s-\Omega_n=\nu \Delta^2 \left(-\frac{1}{6}
-\frac{\pi z^3}{4}+z^2\right),$$ which is negative for $z<0.537$, and for planar phase $$\Omega_s-\Omega_n=\nu \Delta^2 \left(-\frac{1}{3}
+\frac{z^2}{2}+\frac{z(1-z^2)}{4} \ln\frac{1+z}{1-z}\right),$$ which is negative for $z<0.623$. For our specific model Hamiltonian, at weak coupling, the planar phase is more stable. For $I>\Delta$ the condensation energy is always positive, indicating that the lower branches are unstable.
Following the standard methods in the theory of superconductivity [@AGD] we calculate the super-currents in our system under the influence of homogeneous in space vector potential ${\bf
A}$. The super-current is anisotropic, $j_i=\frac{e^2
N}{m}\kappa_{ik}A_k$ with the components given by ($\kappa_{xx}=\kappa_{yy}$), $$\left[\begin{array}{l} \kappa_{zz}\\ \kappa_{xx}
\end{array}\right] = 1- \frac{3I}{2} \int\ \frac{d \o_{\bf n} }{4\pi}
\left[\begin{array}{l}\cos^2{\theta} \\
\sin^2{\theta}
\end{array} \right] \frac{\Theta(I-|\Delta_{\bf n}|)}{\sqrt{I^2-|\Delta_{\bf
n}|^2}}.$$ For the polar phase, assuming $z>1$ we find, $$\left[ \begin{array}{l} \kappa_{zz}\\ \kappa_{xx}
\end{array}\right] =\begin{array}{l}
1-3z^3/4\pi, \\
1-3\pi z/4 +3\pi z^3/8,
\end{array}$$ The coefficient $\kappa_{xx}$ becomes negative at $z\geq 0.480$ ($\kappa_{zz}$ at higher values of $z\geq 0.752$) indicating an instability with respect to a transition into some inhomogeneous state (probably similar to a LOFF state). For the planar phase, $$\begin{aligned}
\left[ \begin{array}{l} \kappa_{zz}\\ \kappa_{xx}
\end{array}\right] =
1\mp \frac{3z^2}{4} -\frac{3z}{8}(1\mp z^2)\ln\left(
\frac{1+z}{1-z}\right).\end{aligned}$$ The coefficients $\kappa_{xx}$ and $\kappa_{zz}$ remain always positive for the whole range of $z<0.623$ where the gap equation (\[curve\_planar\]) has stable solutions. Thus, we find that the planar phase has lower energy and higher density of Cooper pairs than the polar phase and is therefore more stable.
[*Specific Heat:*]{} The important manifestation of the BCS states with gapless excitations is the appearance of the term linear in temperature in the specific heat, which is characteristic for a normal Fermi liquid. The specific heat is given by $$\label{spec_heat}
C=\sum_{{\bf p}}\left(E_{\bf p}^{+}\frac{\partial
n(E_{\bf p}^+)}{\partial T} + E_{\bf p}^{-}\frac{\partial
n(E_{\bf p}^-)}{\partial T}\right),$$ where $E^{\pm}_{\bf p}=\pm\sqrt{\xi^2_p+\Delta_{\bf n}^2}+I$. At low temperatures $T\ll I$ the first term in Eq. (\[spec\_heat\]) gives an exponentially small contribution. The second term, with $E^{-}$, in Eq. (\[spec\_heat\]) is, $$C=\frac{\nu}{4T^2}\int\limits_{-\infty}^\infty d\xi
\int\frac{d \o_{\bf n} }{4\pi}
\frac{\left(\sqrt{\xi^2+|\Delta_{\bf n}|^2}-I
\right)^2}{\cosh^2{\left[\frac{\sqrt{\xi^2+|\Delta_{\bf n}|^2}-I}{2T}\right]}}.$$ Performing the integration, we calculate the contribution of the gapless modes to the specific heat at $T\ll I$ to be $$C=\nu T\frac{\pi}{6} \left\{
\begin{array}{ll}\pi z,& \text{polar phase},
\\ 4z^2,& \text{planar phase}.
\end{array} \right.$$ As expected, the “normal” contribution to the specific heat, is proportional to the area occupied by the gapless modes, i.e. the $I/\Delta$ strip around the equator for the polar phase and the $I^2/\Delta^2$ islands around the poles for the planar phase.
[*Conclusion and Comments:*]{} We have presented substantial evidence that our simple model supports the planar phase gapless superfluidity in the ground state. For $I\ll \Delta$ the gapless modes contribute high powers in terms of mismatch, $\sim I^4$ for the solution and $\sim I^2$ for the heat capacity, i.e. they represent small perturbations. The residual continuous symmetry of this state, and its favorable energy relative to plausible competitors (normal state, polar phase) suggest that it is a true ground state in this model. The planar phase is symmetric under simultaneous axial rotation and gauge (i.e., phase) transformation. Also, we obtain a positive density of superconducting electrons, suggesting that inhomogeneous LOFF phases are disfavored at small $I$.
In some respects the same qualitative behavior we find here in the p-wave resembles what arose in s-wave [@FGLW]. Namely, isotropic s-wave superconductivity has two branches of solution: the upper BCS which is stable and – for simple interactions – fully gapped, and the lower branch which has gapless modes but is unstable. The striking difference is that in p-wave the upper branch retains stability while developing a full two-dimensional fermi surface of gapless modes. Thus the anisotropic p-wave breached pair phase, with coexisting superfluid and normal components, is stable already for a wide range of parameters at weak coupling using the simplest (momentum-independent) interaction. This bodes well for its future experimental realization.
In our model, which has no explicit spin degree of freedom, gapless modes occur for either choice of order parameter with residual continuous symmetry. By contrast, for $^{3}$He in the B phase the p-wave spin-triplet order parameter is a $2\times 2$ spin matrix, containing both polar and planar phases components, there are no zeros in the quasiparticle energies, and the phenomenology broadly resembles that of a conventional s-wave state [@BW]; in the A phase (which arises only at $T\neq 0$ [@Leggett]) the separate up and down spin components pair with themselves, in an orbital P-wave, and no possibility of a mismatch arises.
Experimentally, the microscopic nature of the pairing state can be revealed most directly by probing the momentum distribution of the fermions, including angular dependence. Time of flight images, obtained when trapped atoms are released from the trap and propagate freely, reflect this distribution.
It is possible that the emergent fermi gas of gapless excitations develops, as a result of residual interactions, secondary condensations. Also, one may consider analogous possibilities for particle-hole, as opposed to particle-particle, pairing. In that context, deviations from nesting play the role that fermi surface mismatch plays in the particle-particle case. We are actively investigating these issues.
The authors thank E. Demler, M. Forbes, O. Jahn, R. Jaffe, B. Halperin, G. Nardulli, A. Scardicchio, O. Schroeder, A. Shytov, I. Shovkovy, D. Son, V. Liu for useful discussions. This work is supported in part by funds provided by the U. S. Department of Energy (D.O.E.) under cooperative research agreement DF-FC02-94ER40818, and by NSF grant DMR-02-33773.
[99]{}
For a review, see Nature 416, 205 (2002).
M. Alford, K. Rajagopal, and F. Wilczek, Nucl. Phys. [**B537**]{}, 443 (1999).
V. P. Mineev and K. V. Samokhin, [*Introduction to Unconventional Superconductivity*]{} (Gordon and Breach Science Publishers, 1999).
A. A. Abrikosov, [*Fundamentals of the Theory of Metals*]{} (Elsevier, 1988).
A. I. Larkin and Yu. N. Ovchinnikov, Sov. Phys. JETP [**20**]{}, 762 (1965); P. Fulde and R. A. Ferrell, Phys. Rev. [**135**]{}, A550 (1964).
G. E. Volovik, Phys. Lett. A [**142**]{}, 282 (1989).
G. Sarma, Phys. Chem. Solid [**24**]{}, 1029 (1963); S. Takada and T. Izuyama, Prog. Theor. Phys. [**41**]{}, 635 (1969).
V. Liu and F. Wilczek, Phys. Rev. Lett. [**90**]{}, 047002 (2003).
E. Gubankova, V. Liu, and F. Wilczek, Phys. Rev. Lett. [**91**]{}, 032001 (2003).
M. Alford, J. Berges, and K. Rajagopal, Phys.Rev.Lett. [**84**]{}, 598 (2000).
I. Shovkovy and M. Huang, Phys.Lett. B [**564**]{} (2003) 205.
P. F. Bedaque, H. Caldas, and G. Rupak, Phys. Rev. Lett. [**91**]{}, 247002 (2003).
S. T. Wu and S. Yip, Phys. Rev. A [**67**]{}, 053603 (2003).
M. Forbes, E. Gubankova, V. Liu, and F. Wilczek, Phys. Rev. Lett. [**94**]{}, 017001 (2005).
M. Alford, C. Kouvaris, and K. Rajagopal, hep-ph/0407257, to appear in [*Strong and Electroweak Matter 2004*]{}.
J. Zhang, et al., quant-ph/0406085, to appear in Phys. Rev. A.
R. Balian and N. R. Werthamer, Phys. Rev. [**131**]{}, 1553 (1963).
C. Ticknor, C.A. Regal, D.S. Jin, and J.L. Bohn, Phys. Rev. A [**69**]{}, 42712 (2004).
T. Loftus, C.A. Regal, C. Ticknor, J.L. Bohn and D.S. Jin, Phys. Rev. Lett. [**88**]{}, 173201 (2002).
E. Gubankova, E. G. Mishchenko, and F. Wilczek, cond-mat/0411328.
A. A. Abrikosov, L. P. Gorkov, and I. E. Dzyaloshinskii, [*Methods of Quantum Field Theory in Statistical Physics*]{}, (Dover, New York, 1975).
A. J. Leggett, Rev. Mod. Phys. [**47**]{}, 331 (1975).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an efficient technique based on low-rank separated approximations for the computation of three-dimensional integrals in the computer code DEPOSIT that describes ion-atomic collision processes. Implementation of this technique decreases the total computational time by a factor of $\sim 10^3$. The general concept can be applied to more complicated models.'
address:
- 'Skolkovo Institute of Science and Technology, Novaya St. 100, Skolkovo, Odintsovsky district, 143025 Moscow Region, Russia'
- 'Institute of Numerical Mathematics, Gubkina St. 8, 119333 Moscow, Russia'
author:
- 'Mikhail S. Litsarev'
- 'Ivan V. Oseledets'
bibliography:
- 'bibtex/our.bib'
- 'bibtex/tensor.bib'
- 'refs.bib'
- 'misha.bib'
title: 'Low rank approximations for the DEPOSIT computer code.'
---
Low rank approximation ,2D cross ,Separated representation ,Exponential sums ,3D Integration ,Slater wave function ,Ion-atom collisions ,Electron loss
Introduction. {#IntroSect}
=============
The computer code DEPOSIT [@litsarev-cpc-2013] is intended to describe ion-atomic collision processes. It allows to calculate total and multiple electron loss cross sections $\sigma$ and $\sigma_m$ ($m$ is the number of ionized electrons), the deposited energies $T(b)$, ($b$ is the impact parameter of the projectile ion) and ionization probabilities $P_m(b)$. It is based on the energy deposition model introduced by N. Bohr [@bohr1915] and developed further by A. Russek and J. Meli [@RussekMeli1970], C.L. Cocke [@cocke_pra1979], and V.P. Shevelko *at al.* [@litsarev-jpb-2008]. Theoretical development of the DEPOSIT is presented in [@litsarev-jpb-2008; @litsarev-nimb-2009; @litsarev-jpb-2009; @litsarev-jpb-2010]. Examples of calculations are reported in [@litsarev-springer-2012; @litsarev-hci-2012; @uspekhi2013; @litsarev-jpb-2014]. Detailed description of the code and user guide are given in [@litsarev-cpc-2013].
The cross sections and ionization probabilities needed for estimation of the losses and lifetimes of fast ion beams, background pressures and pumping requirements in accelerators and storage rings are, in fact, functionals of the deposited energy $T(b)$, which in turn is a three-dimensional integral over the coordinate space. To calculate any of these parameters one has to compute $T(b)$ in all points of the $b$-mesh.
The integral $T(b)$ is a bottleneck of the program, and it is required to be done as fast as possible. In the previous work [@litsarev-cpc-2013] an advanced quadrature technique was used, and the computational time has appeared to be much faster in comparison with direct usage of uniform meshes. It takes several seconds to compute one point $T(b)$ for one atomic shell at fixed $b$. For complex ions, the total computation takes few hours on one processor core and is not enough fast. To overcome this issue a fully scalable parallel variant of the algorithm was proposed. Nevertheless, the computational time is still large.
In this work, we present an entirely different approach for computing $T(b)$ in many points of the $b$-mesh, based on low rank approximations of matrices and tensors. The main idea is to approximate the functions to be integrated by a sum of products of univariate functions, effectively decreasing the dimensionality of the problem. This involves active usage of numerical and analytical tools.
The definition of $T(b)$ involves a function of two variables (the energy gain $\Delta E$ during an ion-atomic collision) and a function of three variables (electron density in Slater-type approximation). Details and definitions are given in Section \[TbDefSect\]. The integral is computed in Cartesian coordinates, which are better suited for the construction of separable representation than spherical coordinates used in the original DEPOSIT code.
In Section \[LowRankSect\] for a function of two variables we use the pseudo-skeleton decomposition of matrices [@tee-mosaic-1996; @gtz-psa-1997; @gtz-maxvol-1997] computed via a variant of the incomplete cross approximation algorithm [@tee-cross-2000]. We show numerically that the function in question can be well-approximated by a separable function in Section \[NumericSect\]. Thus, the approximation can be computed in $\mathcal{O}(n)$ time, where $n$ is the number of grid points in one dimension.
In Section \[RhoExpSect\] the Slater-type function of three variables is decomposed by the exponential sums approach [@beylkin-expsum-2005; @beylkin-expsumrev-2010]. The integral is immediately reduced to a two-dimensional one of a simpler structure.
Combining these two representations we obtain in Section \[FastTb2DSect\] an efficient algorithm with $\mathcal{O}(n)$ scaling, in comparison with $\mathcal{O}(n^3)$ complexity for direct integration over a three-dimensional mesh. The computation of $T(b)$ on the whole $b$-mesh takes less then one minute and total speedup of the program is about $\sim 10^3$ times. Illustrative examples are given in Section \[NumericSect\].
All the equations related to the physical model are written in atomic units.
Numerical procedure {#sectTb}
===================
Statement of the problem {#TbDefSect}
------------------------
The deposited energy $T(b)$ is defined as a three-dimensional integral over coordinate space centered in the projectile ion. $$\label{Tb3DIntegral}
T(b)=\sum_{\gamma} \int \Delta E_{\gamma}(p)
\, \rho_{\gamma}(r)
\, d^{3} \mathbf{r}.$$ The sum here is over all atomic shells denoted by $\gamma=nl$, $n$ is the *principle quantum number* and $l$ is the *orbital quantum number*. The electron density $\rho_{\gamma}(r)$ is taken in a Slater-type approximation $$\label{rhoDef}
\rho_{\gamma}(r)=C_{\gamma} r^{\alpha_{\gamma}} e^{-2\beta_\gamma r}$$ with integer $\alpha_{\gamma}$, real positive $\beta_{\gamma}$ and normalization condition $$\int_{0}^{\infty} \rho_{\gamma}(r) dr=N_{\gamma},$$ where $N_{\gamma}$ is the number of electrons in a $\gamma$-shell. The gain of kinetic energy $\Delta E_{\gamma}$ is a smooth finite function of parameter $|\mathbf{p}|$ without any singularities. The impact parameter $\mathbf{p}$ of the ion’s electron is a function of $\mathbf{b}$ and $\mathbf{r}$. In frame of the moving projectile the following equality holds $$\label{pbrrelation}
p^2=(b-r\cos\theta)^2 + (r\cos\varphi \sin \theta)^2.$$ For details we refer the reader to the paper [@litsarev-cpc-2013]. In Cartesian coordinates $\Delta E_{\gamma}(p)$ as a function of parameter $p$ depends only on $x$ and $z$ as it follows from the equation (\[pbrrelation\]) $$\label{TbDef}
T_{\gamma}(b)=\iiint \! \Delta E_{\gamma}(x,z - b) \rho_{\gamma}(x,y,z) dx dy dz.$$ Thus, we need to compute the integral . From here and bellow index $\gamma$ will be skipped for the sake of simplicity and only one shell will be considered in the following equations.
System $\gamma$-Shell $r$ $T_{\mbox{\footnotesize{cross}}}$ (sec) $\varepsilon$ $N$ $r$ $T_{\mbox{\footnotesize{cross}}}$ (sec) $\varepsilon$ $N$ $r$ $T_{\mbox{\footnotesize{cross}}}$ (sec) $\varepsilon$ $N$
--------------- ---------------- ------ ----------------------------------------- --------------- -------- ------ ----------------------------------------- --------------- -------- ----- ----------------------------------------- --------------- ------
$Au^{26+}+ O$ $4df^{17}$ $13$ $0.21$ $10^{-6}$ $1024$ $21$ $0.42$ $10^{-9}$ $1024$ 24 2.41 $10^{-9}$ 4096
$4sp^{8}$ $13$ $0.21$ $21$ $0.33$ 24 2.40
$3d^{10}$ $14$ $0.19$ $22$ $0.42$ 26 2.58
$3sp^{8}$ $16$ $0.25$ $24$ $0.54$ 29 2.64
$2sp^{8}$ $17$ $0.28$ $25$ $0.56$ 30 2.71
$1sp^{2}$ $17$ $0.27$ $25$ $0.56$ 30 2.70
$U^{28+}+ Xe$ $5sp^{4}$ $14$ $0.20$ $10^{-6}$ $1024$ $22$ $0.50$ $10^{-9}$ $1024$ 26 2.17 $10^{-9}$ 4096
$4df^{24}$ $15$ $0.23$ $24$ $0.52$ 27 2.58
$4sp^{8}$ $17$ $0.28$ $25$ $0.55$ 30 2.69
$3d^{10}$ $17$ $0.27$ $25$ $0.54$ 30 2.77
$3sp^{8}$ $17$ $0.27$ $25$ $0.55$ 30 2.75
$2sp^{8}$ $17$ $0.26$ $25$ $0.54$ 30 2.71
$1sp^{2}$ $17$ $0.26$ $25$ $0.55$ 30 2.69
Low rank approximation. {#LowRankSect}
-----------------------
Let $F(x, y)$ be a function of two variables $x, y$ where point $(x, y)$ is in a certain rectangle $[a_x,b_x] \otimes [a_y,b_y]$. The function is said to be in the *separated form* if it can be represented as a sum of products of univariate functions: $$\label{FxyCanonical}
F(x,y)=\sum_{\alpha=1}^r \sigma_{\!\alpha} \,u_{\alpha}(x)g_{\alpha}(y).$$ The minimal number $r$ such that exists will be called *separation rank*. Direct generalization of to multivariate functions is referred to as canonical polyadic (CP, also known as CANDECOMP/PARAFAC) [@kolda-review-2009].
If the function is in the separated form, the integration is simplified a lot. Indeed, $$\label{lrint:2dex}
\iint F(x,y) dx dy = \sum_{\alpha=1}^r \sigma_{\alpha}
\! \int_{a_x}^{b_x} \!\!\! u_{\alpha}(x) dx \! \int_{a_y}^{b_y} \!\!\! g_{\alpha}(y) dy,$$ and the problem is reduced to the computation of one-dimensional integrals, which can be computed using fewer quadrature points than the original integral.
The discretization of one-dimensional integrals in by some quadrature formula with nodes $x_i \in [a_x, b_x]$, $i = 1,\ldots, n$, $y_j \in [a_y, b_y]$, $j = 1,\ldots,m$ and weights $w^{(x)}_i$, $w^{(y)}_j$, leads to the approximation $$\label{IntegralS1D1D}
\iint F(x,y) dx dy \approx \sum_{\alpha=1}^r \sigma_{\alpha}
\sum_{i=1}^n w^{(x)}_i u_{\alpha}(x_i)
\sum_{j=1}^m w^{(y)}_j g_{\alpha}(y_j).$$ On the other hand, direct two-dimensional quadrature with separated weights in $x$ and $y$ can be used for the original integral: $$\label{integral2D}
\iint F(x,y) dx dy \approx
\sum_{i=1}^n w^{(x)}_i
\sum_{j=1}^m w^{(y)}_j
F(x_i,y_j).$$ Comparison of two representations and leads to the following discrete approximation problem $$\label{lrint:2ddiscr}
F(x_i, y_j) \approx
\sum_{\alpha=1}^r \sigma_{\alpha} u_{\alpha}(x_i) g_{\alpha}(y_j),$$ which is a discrete analogue of . Equation can be written in the matrix form: $$A \approx U \Sigma G^{\top},$$ where $A$ is an $n \times m$ matrix with elements $A_{ij}=F(x_i, y_j)$, $U$ is an $n \times r$ matrix with elements $U_{i\alpha}=u_{\alpha}(x_i)$, $G$ is an $m \times r$ matrix with elements $G_{\! j\alpha}=g_{\alpha}(y_j)$ and $\Sigma$ is an $r \times r$ diagonal matrix with elements $\sigma_{\alpha}$ on the diagonal. This is a standard *low-rank approximation problem* for a given matrix. Provided that a good low-rank approximation exists, there are very efficient *cross approximation algorithms* [@tee-cross-2000; @bebe-2000] that need only $\mathcal{O}((n + m)r)$ elements of a matrix to be computed.
By using of our implementation of the cross approximation algorithm we decompose the energy gain $\Delta E(x,\tilde z)$ in the form . In Table \[Table1Ranks\] the ranks $r$ and other numerical parameters are given for particular systems. Description of these parameters can be found in Section \[NumericSect\].
Exponential sums. {#RhoExpSect}
-----------------
For a function $\rho(x, y, z)$ defined in the separation of variables can be done analytically [@beylkin-expsum-2005; @hackbra-expsum-2005; @GHK-ten_inverse_ellipt-2005; @beylkin-expsumrev-2010]. The main idea is to approximate the Slater density function by a sum of Gaussians $$\label{gaussianExpand}
\rho(r) \approx \sum_{k=0}^{K} \lambda_k e^{-\eta_k r^2},
\qquad r=\sqrt{x^2 + y^2 + z^2}.$$ Once the approximation is computed, the separation of variables in Cartesian coordinates comes for free $$\label{rhoSeparated}
\rho(x,y,z) \approx \sum_{k=0}^{K} \lambda_k \,
e^{-\eta_k x^2} \, e^{-\eta_k y^2} e^{-\eta_k z^2}.$$ The technique for the computation of the nodes $\lambda_k$ and the weights $\eta_k$ is based on the computation of the inverse Laplace transform.
Let us consider a function $f_{\alpha \beta}(t)$ such that its Laplace transform is function $F_{\alpha \beta}(s)$ $$\label{FabsqS}
F_{\alpha \beta}(s)=
\int^{\infty}_{0} e^{-st} f_{\alpha \beta}(t) \,dt,$$ of the following form: $$\label{FabsqS1}
F_{\alpha \beta}(s)\equiv \frac{\rho(\!\sqrt{s}\,)}{C}
={\left(\!\sqrt{s}\right)}^{\alpha} e^{-2\beta \!\sqrt{s}}
$$ where $\alpha$ and $\beta$ are parameters of the Slater density . The inverse Laplace transform $f_{\alpha \beta}(x)$ can be computed analytically for the known $F_{\alpha \beta}(s)$. In \[App:Laplace\] we present explicit expressions for the functions $f_{\alpha \beta}(t)$ corresponding to the functions for integer $\alpha$ and real positive $\beta$.
Once is given and the function $f_{\alpha \beta}(t)$ is known, the integral is approximated by a quadrature formula $$\label{lr:quadexp}
\rho(r) \approx C \sum_{k=0}^K
w_k e^{t_k} f_{\alpha \beta}(e^{t_k}) e^{-r^2 e^{t_k}},$$ where $w_k$ and $t_k$ are quadrature weights and nodes, respectively. The procedure to compute the weights and the nodes was taken from the paper [@beylkin-expsumrev-2010]. For the reader’s convenience we give the formula and its derivation in \[App:rhoIntegral\].
According to equation $$\lambda_k = C \, w_k e^{t_k} f_{\alpha \beta}(e^{t_k}),
\qquad \eta_k = e^{t_k}.$$ It appears that only several quadrature points (at fixed $r$) are required to achieve the accuracy of the expansion of order $10^{-7}$.
Fast computation of $T(b)$. {#FastTb2DSect}
---------------------------
The three-dimensional integral $T(b)$ defined in can be reduced to a two-dimensional integral by means of the decomposition $$T(b)=\sum_{k=0}^{K} \lambda_k \iint \! \Delta E(x,z - b) \,
e^{-\eta_k x^2} e^{-\eta_k y^2} e^{-\eta_k z^2}
dx dy dz$$ and analytical evaluation of the one-dimensional Gaussian integral $$\int^{\infty}_{-\infty} e^{-\eta\, y^2}dy=\sqrt{\frac{\pi}{\eta}},$$ $$\label{Tb2DGauss}
T(b)=\sqrt{\pi}
\sum_{k=0}^{K} \frac{\lambda_k}{\sqrt{\eta_k}}
\iint \! \Delta E(x,z - b) \, e^{-\eta_k x^2} e^{-\eta_k z^2} dx dz.$$ Suppose that $\Delta E(x,z-b)$ has been decomposed as follows $$\label{DeltaEDecomp}
\Delta E(x, z - b) \approx
\sum_{\alpha=1}^r \sigma_{\alpha} u_{\alpha}(x) g_{\alpha}(z-b).$$ Then the integration can be reduced to a sequence of one-dimensional integrations. $$\label{Tb1d1dsum}
T(b)=\sqrt{\pi}
\sum_{k=0}^{K} \frac{\lambda_k}{\sqrt{\eta_k}}
\sum_{\alpha=1}^r \sigma_{\alpha}
I_{\alpha k} J_{\alpha k}(b),$$ $$\label{If1D}
I_{\alpha k}=\int_{a_x}^{b_x} \! u_{\alpha}(x) e^{-\eta_k x^2} dx,$$ $$\label{Ig1D}
J_{\alpha k}(b)=
\int_{a_y}^{b_y} \! g_{\alpha}(z-b) e^{-\eta_k z^2} dz.$$ For the numerical approximation of the integrals and we use the quadrature formula with uniform quadrature nodes (although any suitable quadrature can be used) $$\label{NumIka}
I_{\alpha k}=
\sum_{i} w^{(x)}_{i} u_{\alpha}(x_i) e^{-\eta_k x_i^2},$$ $$\label{xiPoints}
x_i=-a_x + i \, h_x, \qquad
0 \le i \le 2N_x, \qquad
h_x = a_x/N_x,$$ $$\label{Jzmb}
J_{\alpha k}(b)=
\sum_{j}
w^{(z)}_j \,g_{\alpha}(z_j-b) e^{-\eta_k z_j^2},$$ $$\label{zjPoints}
z_j=-a_z + j \, h_z, \qquad
0 \le j \le 2N_z, \qquad
h_z = a_z/N_z.$$ We sample the impact parameter $b$ (which can take only positive values) with the *same step* $h_z$ $$\label{bPoints}
b_l= l \, h_z, \qquad
0 \le l \le N_z.$$ This allows us to introduce a new variable $\tilde z = z - b$ discretized as $$\label{ztildakPoints}
\tilde z_k=-2a_z + k \, h_z, \qquad
0 \le k \le 3N_z,$$ and such that for the boundary conditions , , $$z_j - b_l = \tilde z_{j-l+N_z}.$$ The approximation problem reduces to a low-rank approximation of the extended $(2N_x + 1) \times (3N_z + 1)$ matrix $$\label{Eextand}
\Delta E(x_i, \tilde z_j)\approx
\sum_{\alpha=1}^r \sigma_{\alpha} u_{\alpha}(x_i) g_{\alpha}(\tilde z_j).$$ This should be done only once (using the cross approximation algorithm), and the final approximation of the integral reads $$\label{Jakblfast}
J_{\alpha k}(b_l) \approx
\sum_{j}
w^{(z)}_j \,g_{\alpha}(\tilde z_{j-l+N_z}) e^{-\eta_k \tilde z_j^2}.$$ The calculation of $T(b)$ can be summarized in the following algorithm.
compute the decomposition for $\rho(r)$ compute the cross approximation for the matrix $\Delta E(x_i,\tilde z_j)$ defined in compute the integral $I_{\alpha k}$ defined in compute the integral $J_{\alpha k}(b_l)$ defined in compute $T_{\gamma}(b_l)$, equation
System $\gamma$-Shell $N_w$ $T_s$ ($\times10^{-3}$ sec) $T_{D}$ (sec) $T_{D}/T_s$
--------------- ---------------- ------- ----------------------------- --------------- -------------
$Au^{26+}+ O$ $4df^{17}$ $74$ $7.94$ $3.89$ $490$
$4sp^{8}$ $69$ $4.92$ $3.83$ $778$
$3d^{10}$ $73$ $3.59$ $3.88$ $1080$
$3sp^{8}$ $72$ $3.81$ $3.82$ $1003$
$2sp^{8}$ $107$ $2.42$ $3.86$ $1592$
$1sp^{2}$ $209$ $1.24$ $3.88$ $3120$
$U^{28+}+ Xe$ $5sp^{4}$ $62$ $10.1$ $3.94$ $390$
$4df^{24}$ $70$ $6.05$ $3.90$ $644$
$4sp^{8}$ $67$ $5.00$ $3.94$ $788$
$3d^{10}$ $71$ $3.88$ $3.92$ $1011$
$3sp^{8}$ $70$ $3.52$ $3.90$ $1106$
$2sp^{8}$ $105$ $1.99$ $3.87$ $1945$
$1sp^{2}$ $207$ $1.04$ $3.88$ $3723$
: Timings to compute $T(b)$ at fixed $b$ are presented for two cases: the DEPOSIT code (old) $T_D$ and the code based on the separated representations $T_s$. Collision systems are the same as in Table \[Table1Ranks\]. Number of terms in the expansion is labeled by $N_w$. The calculations were carried out for accuracy $\varepsilon=10^{-7}$ and $[-8,8]\otimes[-16,8]$ mesh with $4097 \times 6145$ points. The last column shows the speedup of the program. []{data-label="Table2Times"}
Numerical experiments. {#NumericSect}
----------------------
The most important parameter in is the rank $r$. It determines the complexity of the algorithm (the smaller $r$, the better). In Table \[Table1Ranks\] we present the ranks (and other numerical parameters) calculated for the energy gain $\Delta E(x, \tilde z)$ corresponding to different ion-atomic collisions. As it follows from the numerical experiments, the ranks are small. It means that the cross decomposition allows to decrease the size of the problem from $O(n^2)$ elements to $O(r\cdot n)$ elements where $r \ll n$.
In Table \[Table2Times\] we present the program speedup for every atomic shell. Details are given in the caption of the table. In sums and the terms less then $\epsilon = 10^{-20}$ were thrown out for every $x_i$ and $\tilde z_j$. It is readily seen, that the use of the technique based on the separated representations allows to decrease the total time to compute $T(b)$ by a factor of $\sim 10^3$ compared to the previous version. In practice the computational time is reduced from several hours to one minute or less on the same hardware.
Conclusions and future work {#ConclSect}
===========================
We proposed a new technique for the computation of three-dimensional integrals based on low-rank and separated representation, that significantly reduces the computational time. The general concept can be applied to more complicated models (like ion-molecular collisions with electron loss and charge-changing processes) that lead to multidimensional integrals. For the multidimensional case we plan to use the fast approximation techniques based on the tensor train (TT) format [@osel-tt-2011].
Acknowledgements {#AcknowledgeSect .unnumbered}
================
This research was partially supported by RFBR grants 12-01-00546-a, 14-01-00804-a, 13-01-12061-ofi-m.
Inverse Laplace transform sources {#App:Laplace}
=================================
For integer $\alpha$ and real positive $\beta$ the inverse Laplace transform $f_{\alpha \beta}(t)$ of $F_{\alpha \beta}(s)$ from equation may be calculated analytically and expressed via *the Kummer’s confluent hypergeometric function* $M(a,b;z)$ ([@abramowitz-stegun], chapter 13) as follows $$f_{\alpha\beta}(t)=
\frac{M\left(1+\frac{\alpha}{2},\frac{1}{2},-\frac{\beta^2}{t}\right)}
{t^{1+\frac{\alpha}{2}} \Gamma\left(-\frac{\alpha}{2} \right)}
-2\,\beta\,\frac{M\left(\frac{3+\alpha}{2},\frac{3}{2},-\frac{\beta^2}{t}\right)}
{t^{\frac{3+\alpha}{2}} \Gamma\left(-\frac{1+\alpha}{2} \right)},$$ where $$M(a,b;z)=1+\frac{a}{b}\frac{z}{1!}+
\frac{a(a+1)}{b(b+1)}\frac{z^2}{2!}+\ldots$$ and $\Gamma(x)$ is *the Gamma function*.
Below we present the most interesting $f_{\alpha \beta}(t)$ explicitly. Due to the difference of the normalization conditions in spherical and Cartesian coordinates for the Slater density $$\rho(r)= N_{\gamma}\frac{(2\beta)^{2\mu+1}}{\Gamma({2\mu+1})} r^{2 \mu} e^{-2\beta},$$ the parameter $\alpha$ is related to the parameter $\mu$ as follows $$\alpha = 2 \mu - 2.$$ The number of electrons in the shell $\gamma$ is labeled as $N_{\gamma}$. The parameter $\mu$ is greater or equal to unity. It is an integer or half-integer depending on *the principal quantum number* $n$ and *the orbital quantum number* $l$ of the atomic shell. Details can be found in [@slater1960; @shevelko1993]. For example, $\mu_{1s^2}=1$, $\alpha=0$; $\mu_{2sp^8}=2$, $\alpha = 2$; $\mu_{4d^{10}}=3.5$, $\alpha = 5$. Finally, $$f_{0 \beta}(t)= \frac{g_0\left(t/\beta^2 \right)}{ \sqrt{\pi} \, \beta^2},
\quad g_0(t)= \frac{e^{-\frac{1}{t}}}{t^{3/2}}$$ $$f_{1 \beta}(t)= \frac{g_1\left(t/\beta^2\right)}{ 2\!\sqrt{\pi} \,\beta^3},
\quad
g_1(t) = -\frac{e^{-\frac{1}{t}}}{t^{3/2}}\left(1- \frac{2}{t} \right)$$ $$f_{2 \beta}(t)=\frac{3\, g_2\left(t/ \beta^2\right)} {2\!\sqrt{\pi} \, \beta^4},
\quad
g_2(t)=-\frac{e^{-\frac{1}{t}}} {t^{5/2}} \left(1- \frac{2}{3t} \right)$$ $$f_{3 \beta}(t)= \frac{3 \, g_3\left(t/\beta^2\right)}{4\!\sqrt{\pi}\,\beta^5},
\quad
g_3(t)=\frac{e^{-\frac{1}{t}}}{t^{5/2}}
\left(1 - \frac{4}{t} + \frac{4}{3t^2} \right)$$ $$f_{4 \beta}(t)=\frac{15 \,g_4\left(t/\beta^2\right)}{4 \! \sqrt{\pi} \, \beta^6},
\quad
g_{4}(t)=\frac{e^{-\frac{1}{t}}}{t^{7/2}}
\left(1 - \frac{4}{3t} + \frac{4}{15t^2} \right)$$ $$f_{5 \beta}(t)=\frac{15\, g_5\left(t/\beta^2 \right)}{8 \! \sqrt{\pi} \, \beta^{7}},
\quad
g_5(t)=-\frac{e^{-\frac{1}{t}}}{t^{7/2}} \left(1- \frac{6}{t} +\frac{4}{t^2} - \frac{8}{15t^3} \right)$$ $$f_{6 \beta}(t)=\frac{105\, g_6\left(t/\beta^2\right)}{8\! \sqrt{\pi}\, \beta^8},
\quad
g_6(t) = -\frac{e^{-\frac{1}{t}}}{ t^{9/2}}
\left(1- \frac{2}{t} +\frac{4}{5t^2}- \frac{8}{105t^3} \right)$$
Quadrature formula for the Laplace integral {#App:rhoIntegral}
===========================================
To obtain the decomposition for given $\alpha$ and $\beta$ we make a substitution $s \rightarrow s^2$ into the equation $$F_{\alpha \beta}(s^2)=s^{\alpha} e^{-2\beta s}=
\int^{\infty}_{0} e^{-s^2 x} f_{\alpha \beta}(x) \,dx,$$ then introduce another variable $x=e^{t}$ $$\label{IntegralExpt}
F_{\alpha \beta}(s^2)=s^{\alpha} e^{-2\beta s}=
\int^{\infty}_{-\infty} e^{-s^2 e^{t}+t} f_{\alpha \beta}(e^{t}) dt.$$ Good news is that the function under the integral has exponential decay both in the spatial and frequency domains, therefore the truncated trapezoidal (or more advanced) rule gives the optimal convergence rate. The final approximation has the form $$\label{IntegralGaussWeights}
F_{\alpha \beta}(s^2) \approx \sum_{k=0}^K
w_k e^{t_k} f_{\alpha \beta}(e^{t_k}) e^{-s^2 e^{t_k}},$$ where parameters of the formula $$t_k = a_t + k h_t, \quad
h_t = (b_t - a_t)/K$$ have to be selected in such a way that the resulting quadrature formula approximates the integral for a wide range of parameter $s$. Typically, the choice $a_t \gtrsim -3$, $b_t \lesssim 45$, and $K\sim 250$ gives good accuracy ($\le 10^{-7}$). As an example, in Table \[Table2Times\] the required number of terms in sum is presented. Accurate error analysis can be found in [@beylkin-expsumrev-2010].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss a certain class of two-dimensional quantum systems which exhibit conventional order and topological order, as well as two-dimensional quantum critical points separating these phases. All of the ground-state equal-time correlators of these theories are equal to correlation functions of a [*local*]{} two-dimensional classical model. The critical points therefore exhibit a time-independent form of conformal invariance. These theories characterize the universality classes of two-dimensional quantum dimer models and of quantum generalizations of the eight-vertex model, as well as ${\mathbb Z}_2$ and non-abelian gauge theories. The conformal quantum critical points are relatives of the Lifshitz points of three-dimensional anisotropic classical systems such as smectic liquid crystals. In particular, the ground-state wave functional of these [*quantum Lifshitz points*]{} is just the statistical (Gibbs) weight of the ordinary 2D free boson, the 2D Gaussian model. The full phase diagram for the quantum eight-vertex model exhibits quantum critical lines with continuously-varying critical exponents separating phases with long-range order from a ${\mathbb Z}_2$ deconfined topologically-ordered liquid phase. We show how similar ideas also apply to a well-known field theory with non-Abelian symmetry, the strong-coupling limit of $2+1$-dimensional Yang-Mills gauge theory with a Chern-Simons term. The ground state of this theory is relevant for recent theories of topological quantum computation.'
author:
- Eddy Ardonne
- Paul Fendley
- Eduardo Fradkin
title: Topological Order and Conformal Quantum Critical Points
---
Introduction
============
During the past decade and a half there has been an intense search for new kinds of theories describing quantum condensed-matter systems. Many experimental results have implied that strongly-correlated fermionic systems exhibit qualitatively new types of physical behavior. The now-classic example of this is the fractional quantum Hall effect, where one of the striking consequences of strong correlations is that the Laughlin quasiparticles have fractional charge and fractional statistics, even though though the microscopic degrees of freedom are electrons with integer charge and fermionic statistics \[\].
Traditionally one classifies different phases in terms of order parameters which give a global characterization of the physical state. In turn, the local fluctuations of this order parameter field drive the phase transitions between ordered and disordered states of these systems. This viewpoint, pioneered by Landau and his school, has been extremely successful in condensed matter physics and in other areas of physics, such as particle physics, through the powerful underlying concept of spontaneous symmetry breaking. Much of the structure of modern theory of critical phenomena is based on this point of view \[\].
However, there are many different experimentally-realizable phases (and even more realizable theoretically!) in the fractional quantum Hall effect, but no [*local*]{} order parameter distinguishes between them. These phases are incompressible liquid states which have a fully gapped spectrum and do not break any symmetries of the Hamiltonian. The lack of a local order parameter led to many interesting discussions of the off-diagonal long-range order in the Hall effect \[\]. One particularly elegant way of characterizing the order in the fractional quantum Hall effect is as [*topological order*]{} \[\]. The topological order parameters are non-local; they are expectation values of operators which are lines or loops. Because of this, they can (and do) depend on topology: their value depends on the genus of the two-dimensional surface on which the electrons live. One interesting characteristic of a [*topological phase*]{} is that the correlation functions in the ground state do not depend on the locations of the operators, but only on how the loops braid through each other. In addition, the degeneracies of these non-symmetry breaking ground states on topologically non-trivial manifolds are determined by the topology of these manifolds \[\].
Although so far the only unambiguous experimental realizations of topological phases are in the fractional quantum Hall effect, there has been considerable effort to find, both theoretically and experimentally, condensed matter systems whose phase diagrams may exhibit topological ground states. Much of the current work involves studying fractionalized phases in time-reversal invariant systems (see e.g. \[\]). One reason is that the “normal state" of high-temperature superconductors lacks an electron-like quasiparticle state in its spectrum. There are reasons to believe that frustrated magnets may also exhibit fractionalized behavior as well.
A particularly well-known and simple model with a topological phase is the [*quantum dimer model*]{}, which was invented as a way of modeling the short-range resonating-valence-bond theory of superconductivity \[\]. The degrees of freedom of this two-dimensional model are classical dimers living on a two-dimensional lattice. With a special choice of Hamiltonian (called the RK point), the exact ground-state wave function can be found \[\]. When the dimers are on the square lattice, the result is a critical point. If one deforms this special Hamiltonian, one generically obtains ordered phases. However, a topological phase occurs in the quantum dimer model on the triangular lattice \[\]. When the quantum dimer model is in a topological phase, an effect analogous to fractionalization occurs \[\]. This is called spin-charge separation. One can view the dimers as being created by nearest-neighbor pairs of lattice electrons in a spin-singlet state. Even though the fundamental degrees of freedom (the electrons) have both spin and charge, one finds that the basic excitations have either charge (holons) or spin (spinons), but not both. To prove this occurs, one must show that if one breaks apart an electron pair (dimer) into two holons or two spinons, they are deconfined. For the triangular-lattice quantum dimer model, this was shown in Ref.; the analogous statement in terms of holon-holon correlators was proven in Ref.. At a quantum critical phase transition between an ordered/confining phase and a disordered/deconfining phase (or between different confining states), confinement is lost: the RK point is deconfining \[\].
The notion of spin-charge separation is one of the basic assumptions behind the RVB theories of high-temperature superconductivity \[\], which effectively can be regarded as strongly-coupled lattice gauge theories. In $2+1$-dimensional systems spin-charge separation can only take place if these gauge theories are in a deconfined phase \[\]. In $2+1$-dimensions this is only possible for discrete gauge symmetries. For a continuous gauge group, say $U(1)$ or $SU(2)$, $2+1$-dimensional gauge theories are always in a confining phase, unless the matter fields carry a charge higher than the fundamental charge so that the gauge symmetry is broken to a discrete subgroup \[\]. Thus, the only consistent scenarios for spin-charge separation necessarily involve an effective discrete gauge symmetry, which in practice reduces to the simplest case ${\mathbb Z}_2$. Of particular interest is the fact that the low-energy sector of the deconfined phases of discrete gauge theories are the simplest topological field theories \[\].
Many of these ideas have their origin in the conceptual description of confined phases of gauge theories as monopole condensates, and of their deconfined states as “string condensates" \[\]. In gauge theories it has long been known that their phases cannot characterized by a local order parameter, since local symmetries cannot be spontaneously broken. The phases of gauge theories are understood instead in terms of the behavior of generally non-local operators such as Wilson loops and disorder operators \[\], a concept borrowed from the theory of the two-dimensional Ising magnet \[\].
Interesting as they are, the applicability of these ideas to the problem of high-temperature superconductivity and other strongly correlated systems is still very much an open problem. Topological fractionalized ground states are not the only possible explanation of the unusual physics of the cuprates. In fact, when constructing local microscopic models of strongly-correlated systems which are suspected to have fractionalized phases, many theorists have found that instead these models have a strong tendency to exhibit spatially-ordered states, [*a.k.a.*]{} “valence bond crystals", which appear to compete with possible deconfined states. It is now clear that the regimes of strongly-correlated systems which may favor fractionalized phases also favor, and perhaps more strongly, non-magnetic spatially ordered states of different types, including staggered flux states \[\], or $d$-density wave states \[\], and electronic analogs of liquid crystalline phases \[\]. By now there are a number of examples of models with short-range interactions whose phase diagrams contain both fractionalized and spatially-ordered phases \[\]. It has recently been proposed that deconfined critical points may describe the quantum phase transitions between ordered Néel states and valence bond crystals \[\].
For several reasons, most of the studies of topological order and quantum critical points have focused on examples with two spatial dimensions. The experimental reason is that the Hall effect is two-dimensional, and typical strongly-correlated systems, such as the cuprate high-temperature superconductors, are often effectively two-dimensional. Theoretically, it is because in two dimensions particles can have exotic statistics interpolating between bosonic and fermionic \[\]. A common characteristic of topological phases in two dimensions is the presence of exotic statistics, which occur in the fractional quantum Hall effect \[\]. The statistics can even be non-abelian: in some cases, the change in the wave function depends on the order in which particles are exchanged \[\]. Systems with non-abelian statistics are particularly interesting because they are useful for error correction in quantum computers \[\].
In this paper we will discuss models with topological phases and ordered phases, as well as quantum phase transitions separating them. There has also been a great deal of interest in quantum critical points in and of themselves \[\]. At a quantum critical point, the physics is of course scale invariant, but it need not be Lorentz invariant. The quantum critical points discussed in this paper have dynamical critical exponent $z=2$, instead of the usual $z=1$ of a Lorentz-invariant theory. This allows for some striking new physics. The action of these $z=2$ quantum critical points is invariant under time-independent conformal transformations of the two-dimensional space. A remarkable consequence is that the [*ground-state wave functionals*]{} of the field theories discussed here are conformally invariant in space. This means that the ground state wave functional is invariant under any angle-preserving coordinate transformations of space. For two-dimensional space, there is an infinite set of such transformations, as is familiar from studies of two-dimensional conformal field theory\[\]. This sort of behavior is not common at all: the action of a field theory at a critical point is often scale invariant (and also conformally invariant), but the ground-state wave functional itself in general is not. We dub critical points with this behavior [*conformal quantum critical points*]{}.
One of the consequences of the conformal [*invariance*]{} of the ground state wave function is that all the equal-time correlators of the quantum theory are equal to suitable correlation functions of observables of a two-dimensional Euclidean conformal field theory. We will exploit this connection in this paper quite extensively. However, just as important, conformal invariance of the wave function implies that the ground state of this $2+1$-dimensional theory at a conformal quantum critical point must have zero resistance to shear stress in the two-dimensional plane.
This can be seen as follows. Consider an infinitesimal local distortion of the geometry of the two-dimensional plane represented by an infinitesimal change $\delta g_{ij}(x)$of the two-dimensional metric, as is conventional in the theory of elasticity\[\][^1]. Let $\ket{\Psi}$ be the ground state wave function for the undistorted plane and $\ket{\Psi(g)}$ be the ground state wave function in the distorted plane with two-dimensional metric $g_{ij}(x)=\delta_{ij}+\delta g_{ij}(x)$. Under this distortion the Hamiltonian of the system changes by an amount $$\delta H(g)=\int d^2 x \; \frac{\delta H}{\delta g_{ij}(x)} \;
\delta g_{ij}(x)+\ldots
\label{deltaH}$$ To first order in perturbation theory in $\delta H$, the change of the ground state wave energy is $$\delta E_0=\frac{\me{\Psi}{\; \delta H(g)\; }{\Psi}}{\ipr{\Psi}{\Psi}}
\equiv \langle \delta H(g) \rangle= \int d^2x \;
\displaystyle{\Big\langle {\frac{\delta H }{\delta g_{ij}(x)}\Big\rangle}}
\delta g_{ij}(x)+\ldots
\label{change}$$ where $E_0$ is the exact ground state energy of the distorted system. On the other hand, the change of the norm of the ground state wave function $\|\Psi\|$ is, to all orders in perturbation theory, given by\[\] $$\|\Psi\|^2=\frac{\partial E_0}{\partial\varepsilon_0}$$ where $\varepsilon_0$ is the ground state energy of the undistorted system. Thus, the change of the norm $\|\Psi\|$ is determined by the ($ 2 \times 2$) [*stress tensor*]{} $T_{ij}$ of the $2+1$-dimensional theory $$T_{ij}(x)=\Big\langle \frac{\delta H\; }{\delta g_{ij}(x)}\Big\rangle\ .
\label{stress}$$ On the other hand, we can regard $\| \Psi\|^2$ as the partition function $Z$ of a two-dimensional Euclidean conformal field theory. This theory has an Euclidean [*stress-energy*]{} tensor, $T^{\rm cft}_{ij}$, defined by\[\] $$T^{\rm cft}_{ij}=-\Big\langle \frac{\delta S_{\rm cft}}{\delta g_{ij}(x)}\Big\rangle=\frac{\delta \ln Z}{\delta g_{ij}(x)}\ ,
\label{2Dcft}$$ which essentially coincides with the [*stress*]{} tensor of the $2+1$-dimensional quantum field theory defined above. Scale invariance, rotational invariance and conservation require that $T_{ij}$ be a conserved (divergence free) symmetric traceless tensor. Consequently, the effective Hamiltonian (as well as the action) at this quantum critical point can depend on the spatial gradients of the field only through the “spatial curvature", [*e.g.*]{} $(\nabla^2\varphi)^2$ in a scalar field theory. In other words, at a conformal quantum critical point for a scalar theory, the stiffness vanishes: the usual $(\nabla\varphi)^2$ term is not possible. This means that the dynamical critical exponent of this quantum critical theory must be $z=2$. We call such theories quantum Lifshitz theories; we will discuss such critical points in detail.
In this paper we discuss lattice models which exhibit both ordered/confined phases and disordered/deconfined phases. We will also discuss the field-theory description of these phases and of the phase transitions. To simplify matters, and to be able to obtain exact results, we will introduce models whose ground-state wave function will be known exactly and whose properties we will be able to determine quite explicitly. In this sense, these models are a generalization of the quantum dimer model at the RK point. The basis of the Hilbert space of these models is the configuration space of a two-dimensional classical statistical-mechanical system or Euclidean field theory. Each of these basis states is defined to be orthogonal with respect to the others. An arbitrary state in this Hilbert space can therefore be described as some linear combination of these basis elements. Describing the Hilbert space in such a fashion is not particularly novel. The unusual feature of the models we will discuss is that the ground-state wave function can be expressed as in terms of the action or Boltzmann weights of a [*local*]{} two-dimensional classical theory. The normalization of the wave function will then be the partition function or functional integral of the classical two-dimensional model. This special property is why the wave functionals at the critical points are have a time-independent conformal invariance at their critical points. The field theory of these conformal quantum critical points can be extended to describe nearby ordered and disordered phases, including their confinement properties. We will study this quite explicitly in a quantum generalization of the eight-vertex model. However, much of the physics we discuss should apply to topological phases and ($z=2$) quantum critical points in general.
We will also study theories with a continuous non-abelian symmetry. We show that, interestingly enough, it is very difficult to construct a non-trivial conformal quantum critical point with such a symmetry. We do find a Hamiltonian whose ground state is the doubled Chern-Simons theory of Ref. \[\]. This is a time-reversal invariant theory of interest in topological quantum computation and in (ordinary) supercondutivity; it is in a gapped topological phase.
In section \[sec:scale-wf\] we discuss the simplest model with a scale-invariant critical wave function, the quantum dimer model at the RK point. Here we also introduce the quantum Lifshitz model, the effective field theory of these new quantum critical points. In section \[sec:2dwavefn\], we generalize the relation between the quantum dimer model and the scalar field theory discussed in section \[sec:scale-wf\] to include perturbations which drive the system in to a quantum disordered/deconfined phase or to a ordered/confined phase. In section \[sec:q8v\], we define the quantum eight-vertex model by finding a Hamiltonian whose ground-state wave function is related to the classical eight-vertex model. This will allow us to find quantum critical lines with variable critical exponents separating a ${\mathbb Z}_2$-ordered phase from a topologically-ordered phase. It will also allow us to place a number of previously-known models, in particular that of Ref. \[\], in a more general setting. We show in detail how to use the known results from the Baxter solution of the classical model to map out the critical behavior of the quantum theory. In particular we analyze in detail the confinement and deconfinement properties of the different phases and at criticality. In section \[sec:YMCS\], we study the non-abelian case, and see that the strongly-coupled limit of Yang-Mills theory with a Chern-Simons term has a wave functional local in two-dimensional classical fields \[\]. This theory is in a phase with topological order. In three appendices we give details of the correlators of the quantum Lifshitz field theory (Appendix \[app:gaussian\]), and of the gauge-theory construction of the quantum six-vertex (Appendix \[app:gauge\]) and eight-vertex (Appendix \[app:z2gauge\]) models.
Scale-Invariant Wave Functions and Quantum Criticality {#sec:scale-wf}
======================================================
The simplest lattice model we discuss is the quantum dimer model; the simplest field theory we dub the quantum Lifshitz theory. They both provide very nice illustrations of the properties discussed in the introduction. In the quantum dimer model, the space of states consists of close-packed hard-core dimers on a two-dimensional lattice. A quantum Hamiltonian therefore is an operator acting on this space of dimers, taking any dimer configuration to some linear combination of configurations. In every configuration exactly one dimer must touch every site, so any off-diagonal term in the Hamiltonian must necessarily move more than one dimer. The simplest such operator is called a “plaquette flip”: if one has two dimers on opposite sites of one plaquette, one can rotate the dimers around the plaquette without effecting any other dimers. For example, for the $i$th plaquette on the square lattice one has $$\begin{picture}(350,30)
\thicklines
\put(-40,6){$\hat{F}_i:$}
\put(0,0){\line(0,1){20}}
\put(0.3,0){\circle*{4}}
\put(0.3,20){\circle*{4}}
\put(20,0){\line(0,1){20}}
\put(20.3,20){\circle*{4}}
\put(20.3,0){\circle*{4}}
\put(40,8){$\longrightarrow$}
\put(70,0){\line(1,0){20}}
\put(70,20){\line(1,0){20}}
\put(70.3,0){\circle*{4}}
\put(70.3,20){\circle*{4}}
\put(90.3,0){\circle*{4}}
\put(90.3,20){\circle*{4}}
\put(170,6){and}
\put(240,0){\line(1,0){20}}
\put(240,20){\line(1,0){20}}
\put(240.3,0){\circle*{4}}
\put(240.3,20){\circle*{4}}
\put(260.3,0){\circle*{4}}
\put(260.3,20){\circle*{4}}
\put(280,8){$\longrightarrow$}
\put(310.3,0){\circle*{4}}
\put(310.3,20){\circle*{4}}
\put(310,0){\line(0,1){20}}
\put(330,0){\line(0,1){20}}
\put(330.3,0){\circle*{4}}
\put(330.3,20){\circle*{4}}
\end{picture}$$ The operator $\hat{F}_i$ is defined as zero on any other dimer configuration around a plaquette (i.e. if the $i$th plaquette is not flippable). We define the operator $\hat{V}_i$ as the identity if the plaquette is flippable, and zero otherwise.
The Rokhsar-Kivelson Hamiltonian for the quantum dimer model \[\] $$H_{RK} = \sum_i (\hat{V}_i - \hat{F}_i)
\label{HRK}$$ has the remarkable property that one can find its ground states exactly. They have energy zero, and every state (in a given sub-sector labeled by global conserved quantities) appears with equal amplitude in its ground-state wave function. These properties follow from the facts that $H_{RK}$ is self-adjoint, and $(\hat{V}_i - \hat{F}_i)^2 =
2(\hat{V}_i - \hat{F}_i)$. Hamiltonians of the form $H = \sum_i
Q^\dagger_i Q_i$ necessarily have eigenvalues $E$ obeying $E\ge
0$. Moreover, if one can find a state annihilated by all the $Q_i$, then it is necessarily a ground state. The equal-amplitude sum over all states is indeed such a state. In the Schrödinger picture, the wave function for this state is easy to write down. Define $Z$ as the number of all dimer configurations in some finite volume. This is precisely the classical partition function of two-dimensional dimers with all configurations weighted equally. Then the properly-normalized ground-state wave function for any basis state $|C\rangle$ in the Hilbert space (any classical dimer configuration $C$) is $$|\Psi_0\rangle=\frac{1}{\sqrt{Z}} \sum_C |C\rangle \Rightarrow
\Psi_0(C) = \frac{1}{\sqrt{Z}}$$ The wave function of the quantum system is indeed related to the classical system.
One can extend this sort of analysis to compute equal-time correlators in the ground state. One finds simply that these correlators are given by the correlation functions of the two-dimensional classical theory. Thus for $H_{RK}$ for dimers on the square lattice, one finds algebraic decay of the correlation functions \[\]. This model is then interpreted as a critical point between two ordered phases of the dimers \[\]. However, for the analogous Hamiltonian on the triangular lattice, the classical two-dimensional correlators are exponentially decaying \[\]. One can also show that spinon-type excitations (sites without a dimer) are deconfined on the triangular lattice \[\]. This means the quantum dimer model with $H_{RK}$ on the triangular lattice is interpreted as being in a “liquid” phase, which has a mass gap and exponential decay of interactions, but which has no non-zero local order parameter.
Such a relation between two-dimensional quantum theories is not limited to lattice models, nor are the ground-state wave functions required to be equal-amplitude sums over all configurations. We will construct now a simple (non-Lorentz invariant) two-dimensional quantum critical field theory a theory whose ground state wave function represents a two-dimensional conformal theory.
Consider a free boson $\varphi(x,t)$ in two spatial dimensions and one time dimension. Instead of the usual Hamiltonian quadratic in derivatives, we use one which has been conjectured by Henley \[\] to belong to the same universality class as the square-lattice quantum-dimer model. It is $$H = \int d^2x \; \left[\frac{\Pi^2}{2} + \frac{\kappa^2}{2}(\nabla^2\varphi)^2\right]
\label{eq:Hboson}$$ where $\Pi=\dot\varphi$ as usual. The associated Euclidean action for the field $\varphi$ is $$S=\int d^3x \left[\frac{1}{2} \left(\partial_\tau \varphi\right)^2+\frac{\kappa^2}{2}\left(\nabla^2 \varphi\right)^2\right]
\label{3D-Euclidean}$$ This system, Eq. (\[3D-Euclidean\]), also arises in three-dimensional classical statistical mechanics in the field-theory description of Lifshitz points \[\], for example in (smectic) liquid crystals. For this reason we will call the system with Hamiltonian (\[eq:Hboson\]) the [*quantum Lifshitz model*]{}. Of particular relevance to our discussion is the long-ago observation by Grinstein \[\] that this system is analogous to the two-dimensional Euclidean free boson in that it represents a line of fixed points parametrized by $\kappa$.
Let us rederive this result by quantizing the Hamiltonian (\[eq:Hboson\]). We impose the canonical commutation relations $$[\varphi(\vec x),\Pi(\vec x^{\prime})]=i\delta(\vec x -\vec x^{\prime})
\label{eq:ccr}$$ so in the Schr[ö]{}dinger picture the canonical momentum is the functional derivative $\Pi(\vec x)= -i{\delta}/{\delta\varphi(\vec x)} $. The Schr[ö]{}dinger equation for the wave functional $\Psi[\varphi]$ is then $$\int d^2 x\,
\left[-\12 \left(\frac{\delta}{\delta\varphi}\right)^2 + \frac{\kappa^2}{2}(\nabla^2\varphi)^2
\right]\Psi[\varphi] = E\Psi[\varphi] .
\label{eq:SE}$$ We can find the ground-state wave function in the same fashion as we did for $H_{RK}$. Indeed, if we define $$Q(x)\equiv \frac{1}{\sqrt{2}} \left(\frac{\delta}{\delta\varphi} +
\kappa \; \nabla^2\varphi \right)\ , \qquad\quad
Q^\dagger(x)\equiv \frac{1}{\sqrt{2}} \left(-\frac{\delta}{\delta\varphi}
+ \kappa\; \nabla^2\varphi \right) ,
\label{eq:QAbelian}$$ the (normal-ordered) quantum Hamiltonian is then $$H= \12 \int d^2x \; \left\{Q^\dagger(\vec x), Q(\vec x)\right\}
-\varepsilon_{\rm vac} V
\equiv \int d^2x \; Q^\dagger(\vec x) Q(\vec x)
\label{eq:HQQ}$$ which is Hermitian and positive. Here $V$ is the spatial volume (area) of the system, and we have normal-ordered the Hamiltonian by subtracting off the (UV divergent) zero-point energy density $$\varepsilon_{\rm vac}=-\frac{\kappa}{2} \lim_{\vec y \to \vec
x} \nabla_x^2 \delta(\vec x-\vec y)>0 \ .$$ Any state annihilated by $Q(x)$ for all $x$ must be a zero-energy ground state. The corresponding ground-state wave functional $\ipr{[\varphi]}{{\rm vac}}=\Psi_0[\varphi]$ satisfies $Q\Psi_0 [\varphi] =0 $, where $Q$ is defined in Eq. (\[eq:QAbelian\]). This is simply a first-order functional differential equation, and is easily solved, giving $$\Psi_0[\varphi] = \frac{1}{\sqrt{{\mathcal Z}}}e^{\displaystyle{-\frac{\kappa}{2} \int d^2x \left(\nabla \varphi(x)\right)^2}}
\label{eq:Psi0}$$ where ${\mathcal Z}$ is the normalization $${\mathcal Z}=\int [{\mathcal D} \varphi] \;
e^{\displaystyle{-\kappa \int d^2x \; \left(\nabla \varphi\right)^2}}\ .
\label{eq:Zboson}$$ The probability of finding the ground state in the configuration $\ket{[\varphi]}$ is therefore $$\big\vert \Psi_0[\varphi]\big\vert^2=\frac{1}{{\mathcal Z}} e^{\displaystyle{-\kappa \int d^2x \; \left(\nabla \varphi\right)^2}}\ .
\label{eq:probab}$$ Consequently, the ground state expectation value of products of Hermitian local operators ${\mathcal O}[\varphi(\vec x)]$ reduces to expressions of the form $$\me{{\rm vac}}{ {\mathcal O} [\varphi( {\vec x}_1)] \ldots {\mathcal O}[\varphi( {\vec x}_n)]}{{\rm vac}}
=\frac{1}{\mathcal Z} \int [{\mathcal D} \varphi] \; {\mathcal O}[\varphi(\vec x_1)] \ldots {\mathcal O}[\varphi(\vec x_n)] \;
e^{\displaystyle{-\kappa \int d^2x \; \left(\nabla \varphi\right)^2}} \ .
\label{eq:mapping}$$
This two-dimensional quantum theory has a deep relation with a two-dimensional classical theory: the ground-state expectation value of all local observables are mapped one-to-one to correlators of a two-dimensional massless Euclidean free boson. The latter is a well-known conformal field theory, and its correlation functions are easily determined (for convenience we give them explicitly in Appendix \[app:gaussian\]). This two-dimensional critical field theory is conformally-invariant, so the equal-time correlators of the quantum theory must reflect this. This scalar field theory is therefore not only quantum critical but it also has a time-independent conformal invariance.
The equal-time expectation values of the “charge operators" ${\mathcal O}[\varphi]$, as well as the correlation functions of the dual vortex (or “magnetic") operators discussed in Appendix \[app:gaussian\], exhibit a power-law behavior as a function of distance, as expected at a quantum critical point. As shown also in Appendix \[app:gaussian\], their autocorrelation functions also exhibit scale invariance albeit with a dynamic critical exponent $z=2$. This behavior of the equal-time correlator was shown earlier to occur in the quantum dimer model on the square lattice at the RK point: there is a massless “resonon” excitation, and the equal-time correlation functions for two static holons has a power law behavior equal to that of the monomer correlation function in the classical 2D dimer model on the square lattice.
However, not all theories whose ground state can be found in this fashion need be critical with power-law behavior. As noted above, the quantum dimer model on the triangular lattice is not. As we will discuss in section \[sec:2dwavefn\], adding say a mass-like term to $Q(x)$ in the scalar field theory gives a theory with exponentially-decaying correlations in the ground state. Thus one can understand phase transitions in such theories as well, and we will explore several of these in this paper. Nevertheless, we expect that when the quantum phase transition is continuous, the quantum critical points of generic theories of this type will have the basic structure of the quantum Lifshitz model. As discussed in the introduction, only quantum Lifshitz points can be conformal quantum critical points.[^2]
In the next sections we will show that generalizations of this $z=2$ quantum Lifshitz Hamiltonian also describe the quantum phase transitions between generalizations of the valence bond crystal states and quantum disordered states which describe deconfined topological fluid phases (provided these quantum phase transitions are continuous). Notice that phase transitions from deconfined to confined, uniform and translationally-invariant states are described by the standard Lorentz invariant $z=1$ critical point of gauge theories \[\].
These examples show that one can obtain precise information about some 2d quantum systems in terms of known properties of 2d classical systems. The trick of doing so is in finding a set of $Q_i$ (or $Q(x)$ in the continuum) which annihilates the equal-amplitude state (or some other desired state), and then defining $H = \sum_i Q^\dagger_i Q_i$ \[\]. This seems like it should be possible to do for any classical 2d theory, and indeed, there are many known examples of this sort. However, it is not clear for a given 2d classical theory one can always find a Hamiltonian which is both local and ergodic (in this context, ergodic means that the Hamiltonian will eventually take the system through all of phase space with a given set of conserved quantum numbers). It is also not clear that even if such a Hamiltonian exists, whether it will have any physical relevance.
Moreover, this simple relation of the ground-state wave function of a 2d quantum system to a 2d classical system is not at all generic: the quantum dimer model with $H_{RK}$ and this quantum Lifshitz field theory are quite special. To illustrate this, let us discuss briefly the ground-state wave functions of standard quantum field theories at a (quantum) critical point. Consider first the most common case, the Lorentz invariant $\varphi^4$ field theory at criticality. Below $D=4$ space-time dimensions this critical theory is controlled by its non-trivial Wilson-Fisher fixed point. The resulting theory is massless and in general it has an anomalous dimension $\eta \neq
0$. Scale and Lorentz invariance fully dictate the behavior of [ *all*]{} the correlation functions at this fixed point. General fixed point theories are scale invariant and, in addition, they exhibit an enhanced, generally finite-dimensional, conformal symmetry. It is a very special feature of $D=1+1$-dimensional Lorentz-invariant fixed point theories that they exhibit a much larger, infinite-dimensional, conformal invariance. This enhanced symmetry leads to a plethora of critical behaviors in $1+1$ dimensions. In contrast, there are relatively few known distinct critical points in higher dimensions for Lorentz-invariant field theories.
It is well known that the knowledge of all the equal-time correlation functions determines completely the form of the ground state wave function, [*i.e.*]{} in the Schrödinger representation of the field theory \[\]. For a general theory, the ground-state wave function is a non-local and non-analytic functional of the field configuration. Thus, at the Wilson-Fisher fixed point, which describes theories with only a global conformal (scale) invariance, the structure of the ground state wave function is quite complicated. For instance, the probability of a [constant]{} field configuration $\varphi(\vec
x)=\varphi$ at a critical point has the universal form \[\] $$\left\vert \Psi_{\rm vac}(\varphi)\right\vert^2=A \;
e^{\displaystyle{-B \; \vert \varphi \vert^{1+\delta}}}
\label{eq:constant}$$ For a Lorentz-invariant $\varphi^4$ theory the universal critical exponent is given by $\delta=(d+2-\eta)/(d-2+\eta)$, and $A$ and $B$ are two non-universal constants. Therefore, at criticality the wave function in general is a non-analytic non-local functional of the field configuration. In contrast, the ground-state wave function of a $1+1$-dimensional relativistic interacting fermions (a Luttinger liquid), which is a conformal field theory, has a universal non-local non-analytic Jastrow-like power law factorized form \[\] consistent with the form found by the Bethe-ansatz solution of the Calogero-Sutherland model \[\]. This structure is a consequence of the (local) conformal invariance of the $1+1$-dimensional theory. Even in $1+1$ dimensions, the wave function is generically non-local.
Dimers, fermions and the quantum Lifshitz field theory {#sec:2dwavefn}
======================================================
In Section \[sec:scale-wf\], we showed how to find the exact ground-state wave functions of the quantum dimer model with Rokhsar-Kivelson Hamiltonian $H_{RK}$, and the quantum Lifshitz scalar field theory. In this section, we will describe their properties in more detail, and some simple generalizations.
From the square to the triangular lattice {#sec:from}
-----------------------------------------
The ground states of $H_{RK}$ are the sum over all classical dimer configurations in a sector with equal amplitudes \[\]. There are a number of useful generalizations to models where the ground state is still a sum over all states in a sector, but not necessarily with equal amplitudes. One interesting case is a quantum dimer model which interpolates between the square and triangular lattices.
A triangular lattice can be made from a square lattice by adding bonds across all the diagonals in one direction. The classical dimer model on the square lattice can be deformed continuously into the triangular-lattice model by assigning a variable weight $w$ for allowing dimers along these diagonals: for $w=0$ we have the original square-lattice model, while $w=1$ gives the triangular-lattice one \[\]. There is a two-dimensional quantum Hamiltonian which has the $w$-dependent classical dimer model as its ground state. For dimers on opposite sides of a plaquette of the original square lattice, the Hamiltonian remains $H_{RK}$. In addition, however, parallel dimers on adjacent diagonals can also be flipped \[\], see figure \[triflips\].
Like the flip on the square lattice, this flip can be done without violating the close-packing and hard-core constraints.
When the off-diagonal terms in the Hamiltonian $H_w$ consist solely of these flips, $H_w$ breaks up into two-by-two blocks like $H_{RK}$ does. In the ground state dimers along diagonals should get a weight $w$, so the Hamiltonian $H_{w}$ must explicitly depend on $w$. To construct a Hamiltonian with the desired ground state, we find a set of operators $Q_i$, each of which annihilates this state. For example, denote a configuration with dimers on adjacent diagonals as $|1\rangle$, and $|2\rangle$ as the configuration to which it is flipped, as shown in fig. \[triflips\]. The $Q_i$ acting on these two configurations is then $$Q_i = \frac{1}{w^2+w^{-2}}
\begin{pmatrix}
w^{-2}&-1\\
-1&w^2
\end{pmatrix}
\label{Qtriang}$$ where the first row and column correspond to state $|1\rangle$, while the second corresponds to state $|2\rangle$. There are three types of $Q_i$: $Q^{(\hbox{square})}_i$ acts on dimers on opposite sides of a plaquette of the original square lattice, while $Q^{(\hbox{horiz})}_i$ and $Q^{(\hbox{vert})}_i$ are associated to flips involving dimers on the ‘diagonal’ links; see figure \[triflips\]. The operator $Q^{(\hbox{square})}_i$ is given by (\[Qtriang\]) with $w=1$. We then take $$H_w = \sum_i \left[Q^{(\hbox{square})}_i +
w Q^{(\hbox{vert})}_i + wQ^{(\hbox{horiz})}_i \right]$$ where the sum is over all plaquettes $i$. Each $Q_i$ is a projection operator, so $Q_i=Q^\dagger_i =(Q_i)^2$.
We have defined the operator $Q_i$ so that it annihilates the state $w^2|1\rangle\ +\ |2\rangle$. The ground state of $H_w$ is then the sum over classical dimer model states with each state weighted by $w^{\mathcal D}$, where ${\mathcal
D}$ is the number of dimers along diagonals in that state. More precisely, one must find the conserved quantities for a given value of $w$ and boundary conditions; the sum over all states with weight $w^{\mathcal D}$ in that sector is an eigenstate of $H_{RK}$ with zero energy. The ground-state wave function for a configuration with ${\mathcal D}$ diagonal dimers is then $$\Psi_0[{\mathcal D}] = \frac{w^{\mathcal D}}{\sqrt{Z(w^2)}}
\label{psiD}$$ where $Z(w)$ is the partition function for the classical dimer model with diagonal dimers receiving weight $w$. Note the $w^2$ in the argument of $Z$ in the denominator: this is because in quantum mechanics probabilities are given by $|\Psi_0|^2$. For the square or triangular lattice, this is unimportant, because $w=0$ or $w=1$, both of which have $w^2=w$. For $w=1$, we have the equal-amplitude sum over all dimer states of the triangular lattice, the model discussed in Ref. \[\]. For $w = 0$, we recover a slight generalization of the original square-lattice quantum-dimer model of Rokhsar and Kivelson \[\]. In this limit, this Hamiltonian reduces to $H_{RK}$ plus a potential term forbidding dimers on adjacent diagonal links. Isolated diagonal dimers are still allowed for $w\to 0$, but since none of them can be flipped, the Hamiltonian does not affect them at all. Thus they can be viewed as fixed zero-energy defects in the square-lattice quantum-dimer model. The ground-state wave function for a given set of defects is the equal-amplitude over all configurations of dimers on the sites without defects.
Since we know the exact ground-state wave function for any $w$, one would like to compute the correlation functions in this model. In most two-dimensional lattice models, even those solvable by the Bethe ansatz, this is extraordinarily difficult or impossible. However, the classical dimer model is special in that one can do such computations, because like the two-dimensional Ising model, it is essentially free-fermionic. Precisely, its partition function and correlators can be written in terms of the Pfaffian (the square root of the determinant) of known matrices \[\]. One can rewrite the Pfaffians in terms of a functional integral over Grassmann variables at every site on the lattice \[\]. The action in the case of equal Boltzmann weights is quadratic in the Grassmann variables, so one can compute easily any ground-state correlation function using the dimers, because the dimers can be written in terms of the fermions. This was discussed for the triangular lattice in \[\]. The correlators of spinon-like or holon-like excitations are much more complicated, but the computation was done for $w=0$ in \[\], and for arbitrary $w$ in \[\]. On the lattice, the holon is a defect or monomer, a site without a dimer. The holon-creation operators are not local in terms of the fermionic variables, in a manner reminiscent of how the spin and fermion operators are non-local with respect to each other in the two-dimensional Ising model. The holon two-point function is valuable in that it gives an order parameter for the phase with topological order: if it is non-vanishing as two holons are taken far apart, the holons are deconfined and we are indeed in a topological phase. The existence of topological order was previously established for the triangular lattice $w=1$ \[\]; in \[\] the explicit correlator was computed, and indicates the topologically-ordered phase exists for any non-zero $w$.
The ground state of the quantum dimer model with $H_{w}$ is therefore well understood for any $w$. There are no exact results for the excited states, however. In fact, since $H_w$ does not have any action upon empty sites, one can give the holon any gap desired without changing the ground states. It is therefore useful to find a continuum limit and study the field theory describing this model. In other words, we would like to understand a field theory with partition function equal to the continuum limit of partition function $Z(w)$ in Eq. (\[psiD\]). Since ground-state correlators for the square lattice are algebraically decaying, the $w=0$ model is critical and should have a sensible continuum limit. Indeed, when $w=0$, the Grassmann variables turn into a single free massless Dirac fermion field; its action is the usual rotationally-invariant kinetic term. The dimer correlation length $\xi$ was computed exactly as a function of $w$; for the triangular lattice it is about one lattice spacing, while $\xi$ diverges as $1/w$ for $w$ small \[\]. Thus there is a field-theory description of the continuum limit of the classical dimer model valid as long as $w$ is scaled to zero with the lattice spacing $a$ such that $w/a$ remains finite. Since the action is still quadratic in the Grassmann variables, the resulting fermionic field theory remains free. However, the Dirac fermion receives a mass proportional to $w/a$ \[\].
There are several useful aspects of taking the continuum limit, apart from finding the excited-state spectrum. Correlators are easier to compute: for $w=0$ one can use conformal field theory \[\], while in the scaling limit $w\to 0$ one can use form-factor techniques \[\]. Another useful fact is that (ignoring boundary conditions), a Dirac fermion can be described in terms of two decoupled Ising field theories. In the continuum, the holon can be written in terms of the product of the spin field in one Ising model with the disorder field in the other Ising model. Taking $w$ away from zero amounts to giving one Ising order field an expectation value, and the other disorder field an expectation value. \[\]. Thus one can see directly in continuum that the holon order parameter is non-vanishing for $w\ne 0$. We will see in the next section that one can also understand the physics of the quantum Lifshitz critical line for all $\kappa$ in terms of two (coupled) Ising models.
The critical field theory {#critft}
-------------------------
We would therefore like to find a natural-looking quantum field-theory Hamiltonian which has as its ground-state wave functional $$\Psi_0[\psi] = \frac{e^{-S_{Dirac}[\psi]}}{\sqrt{Z_{Dirac}}}
\label{psidirac}$$ where $S_{Dirac}$ is the usual action for a rotationally-invariant action for a free Dirac fermion in two Euclidean dimensions. Since this wave functional involves Grassman numbers, the easiest way to think of $|\Psi_0|^2$ as a weight in the path integral defining all correlators. While it is possible to find a Hamiltonian acting on this fermionic basis, it is more convenient and more intuitive to instead use bosonic variables. It is more convenient because correlators in a massless Dirac fermion theory (including those involving the product of spin fields) can be bosonized, meaning that they can be written in terms of correlators of free scalar fields \[\]. It is more intuitive because the classical dimer model on the square lattice has a simple description in terms of a “height” variable \[\]. A height is an integer-valued variable, which typically in the continuum limit turns into a scalar field. This description will allow us also to make contact with the quantum Lifshitz model.
Recently, Moessner [*et al.*]{} \[\] generalized Henley’s argument of Ref. \[\], and used the connection between quantum dimer models and their dual quantum roughening (height) models \[\] to argue that the Hamiltonian of Eq. (\[eq:Hboson\]) actually defines the universality class of quantum critical points between valence bond crystal phases. The nature of the phase transition between valence bond crystal states is the focus of much current research. Quite recent results by Vishwanath [*et al.*]{} \[\] [^3], and by Fradkin [*et al.*]{} \[\], show that the transition between valence bond crystals is generically first order, as expected from a simple Landau argument. Nevertheless, when the transition is continuous, it is described by the Hamiltonian of Eq. (\[eq:Hboson\]) which must be regarded as a (rather rich) multicritical point.
To see how the height description arises, let us go back to the square-lattice quantum dimer model, where $H_w$ reduces to $H_{RK}$. To map the square-lattice classical dimer model onto a height model, one first assigns a height variable to each plaquette. In going around a vertex on the even sub-lattice clockwise, the height changes by $+3$ if a dimer is present on the link between the plaquettes, and by $-1$ if no dimer is present on that link. On the odd sub-lattice, the heights change by $-3$ and $+1$ respectively. The flip operator $\hat{F}_i$ on a plaquette $i$ changes the height on that plaquette by either $\pm 4$. To take the continuum limit, is convenient to turn this into a model with heights on the sites. We define $h$ on each site to be the average value of the four plaquette heights around that site[^4]; see figure \[heights\].
(80,20)(0,-8) (0,5)(25,0)[4]{}[(1,0)[10]{}]{} (5,0)(25,0)[4]{}[(0,1)[10]{}]{} (5,5)[(0,1)[5]{}]{} (30,5)[(1,0)[5]{}]{} (55,5)[(0,-1)[5]{}]{} (80,5)[(-1,0)[5]{}]{} (2,7)(25,0)[4]{}[$0$]{} (7,7)[$3$]{} (32,7)[$-1$]{} (57,7)[$-1$]{} (82,7)[$-1$]{} (7,1)[$2$]{} (32,1)[$2$]{} (57,1)[$-2$]{} (82,1)[$-2$]{} (2,1)[$1$]{} (27,1)[$1$]{} (52,1)[$1$]{} (74,1)[$-3$]{} (1,-8)[$h=3/2$]{} (26,-8)[$h=1/2$]{} (50,-8)[$h=-1/2$]{} (76,-8)[$h=-3/2$]{}
To avoid overcounting configurations, we identify the height $h$ with $h+4$. The flip operator $\hat F$ corresponds to changing $h\to h \pm 1$ on all four sites around a plaquette (the $\pm$ depending on the sub-lattice). Columnar order for dimers corresponds to an expectation value for $h$, while staggered order for dimers corresponds to an expectation value for $\partial
h$. One can obtain a Hamiltonian with ordered ground states by allowing the coefficients of the two terms in Eq. (\[HRK\]) to be different. If the coefficient of $\hat F_i$ is larger, this favors columnar order; if the coefficient of $\hat V_i$ is larger, staggered order is favored. It is widely assumed that $H_{RK}$, where the coefficients are equal, describes a phase transition between the two kinds of order \[\]. However, this has never been proven, and there exists the possibility of an intermediate “plaquette" phase\[\].
We would now like to take the continuum limit of the quantum dimer model in its height description. In this limit, we identify the height $h$ with a scalar field $4\varphi(x)$. Like the height, the scalar field must be periodic, so we identify $\varphi$ with $\varphi +1$. We have already noted that the continuum correlators of the square-lattice quantum dimer model are those of a massless Dirac fermion. The correlators in the ground state of the quantum Hamiltonian in terms of $\varphi$ must be identical. The quantum Lifshitz Hamiltonian Eq. (\[eq:Hboson\]) and Eq. (\[eq:HQQ\]) has ground-state correlators of the form given in Eq. (\[eq:mapping\]). When $\kappa^{-1}=2\pi$, these correlators are precisely those of a Dirac fermion; this is a result of the widely-known procedure known as bosonization \[\]. In fact, correlators of many two-dimensional critical classical statistical mechanical systems, not just free fermions, can be written in terms of exponentials of a free boson \[\]; we collect some of these results in Appendix \[app:gaussian\]. In the next section we will display quantum lattice models whose continuum limit corresponds to all values of $\kappa$.
The Hamiltonian $H_{RK}$ on the square lattice in the continuum limit is therefore identified with the quantum Lifshitz Hamiltonian with $\kappa^{-1}=2\pi$. This also allows a qualitative understanding of the physics away from the RK point \[\]. A phase with staggered order should have an expectation value of $\partial_x
\varphi$ or $\partial_y \varphi$ in the continuum theory. Adding a term $(\nabla\varphi)^2$ to the Hamiltonian with negative coefficient will drive the system into such an ordered phase. Adding this term with positive coefficient will favor a constant value of $\varphi$, driving the system into columnar order. Adding terms like $\cos(2\pi\varphi)$ will also drive the system into a phase with columnar order. Thus one expects that at a critical point like that described by $H_{RK}$ on the square lattice, the coefficients of $(\nabla\varphi)^2$ and $\cos(2n\pi\varphi)$ will vanish. This leaves Eq. (\[eq:Hboson\]) as the simplest non-trivial Hamiltonian with the desired properties. The requirement that the ground state be equivalent to a free fermion then fixes the coefficient $\kappa$; note that if desired $\kappa$ can scaled out of the Hamiltonian by redefining the compactification relation to be $\varphi\sim \varphi +
\sqrt{\kappa}$.
It is not at all clear whether critical $2+1$-dimensional field theory and the continuum limit of $H_{RK}$ are identical for excited states, although the above heuristic argument is very suggestive. The Hamiltonian for the scalar field theory Eq. (\[eq:HQQ\]) is purely quadratic in the field $\varphi$, so one can obtain essentially any desired information exactly. $H_{RK}$ is not so simple on the lattice, but one may hope the two models are in the same universality class. In any event, we can easily extract all the excited-state energies for the quantum Lifshitz field theory. The operators $Q(x)$ and $Q^\dagger(x)$ are essentially harmonic-oscillator creation and annihilation operators: the equal-time commutation relation Eq. (\[eq:ccr\]) implies that $$\left[ Q(\vec x), Q^\dagger(\vec y) \right] = \kappa \;
\nabla^2\delta^{(2)} (\vec{x} - \vec{y})
\label{QQccr}$$ The ground state is indeed annihilated by all $Q(x)$, so we can create excited states by acting with $Q^\dagger(x)$. The commutation relations Eq. (\[QQccr\]) mean that the dispersion relation is $E =
\kappa p^2$. This theory is gapless but not Lorentz-invariant: the dynamical critical exponent is $z=2$.
The exponent $z=2$ can also be seen by looking at the classical action associated with this Hamiltonian. This will also allow us to make contact with the earlier three-dimensional statistical-mechanical results. The action consistent with the Hamiltonian of Eq. (\[eq:Hboson\]) and the canonical commutation relations of Eq. (\[eq:ccr\]) is $${\mathcal S}=\int d^3x \left[\12 \left(\partial_t
\varphi\right)^2-\frac{\kappa^2}{2}(\nabla^2\varphi)^2\right]
\label{eq:Sboson}$$ Clearly, this action is not Lorentz invariant and has $z=2$. It is rotationally invariant only in the XY plane. Defining the imaginary time $\tau=it$ gives the Euclidean action (\[3D-Euclidean\]). The imaginary time axis $\tau$ can be regarded as the $z$-coordinate of a three-dimensional classical system in which $\varphi(\vec x,\tau)$ is an angle-like variable and the action represents the spin-wave approximation of an anisotropic classical $XY$ model. In general one would have expected a term proportional to the operator $\left(\nabla
\varphi\right)^2$ with a finite positive stiffness in the plane. This is so in the $XY$ ferromagnetic phase. On the other hand, if the stiffness becomes negative, there is an instability to a modulated helical phase. The action of Eq. (\[3D-Euclidean\]) represents the Lifshitz point, the critical point of this phase transition \[\] where the stiffness vanishes. This effective action also plays a central role in the smectic A-C transition \[\] and in other classical liquid crystal phase transitions associated with the spontaneous partial breaking of translation and/or rotational invariance \[\]. The compactified version of the problem (the identification $\varphi\sim \varphi + 1$) has also been considered in this context, for example in Ref. \[\]. The choice of period in general depends on the physical context of the problem.
The square-lattice quantum dimer model and the scalar field theory with Hamiltonian Eq. (\[eq:HQQ\]) are both at critical points, in that the correlators in the ground state are algebraically decaying. The quantum dynamics implied by this Hamiltonian must be [*compatible*]{} with the 2D time-independent conformal invariance. In particular, the spectrum of the quantum theory must be gapless and, as we learned from this example, the dynamic critical exponent must be $z=2$. Notice however, that $z=2$ alone does not guarantee a gapless (or even critical) theory. Indeed, instructive counter-examples to this statement are well known in the theory of (the absence of) quantum roughening \[\] where quantum fluctuations destroy the critical behavior and lead to an ordered state through an order-from-disorder mechanism.
The off-critical field theory {#offcritft}
-----------------------------
We have thus shown that the continuum limit of the square-lattice quantum dimer model is described by the quantum Lifshitz model at a special point $\kappa^{-1}=2\pi$. We also argued that at least some deformations of the two models result in ordered phases. However, we saw at the beginning of this section that not all deformations of the square-lattice quantum dimer model result in an ordered phase. Allowing dimers across the diagonals with Hamiltonian $H_w$ results in a topologically-ordered phase, where the order parameter is not local.
In this subsection, we find a bosonic field theory describing the topological phase in the continuum limit. We showed above that in the scaling limit $w\to 0$ with $w/a$ finite, the ground-state wave function Eq. (\[psidirac\]) can be written in terms of a free massive Dirac fermion of mass proportional to $w/a$. In two dimensions, the bosonic version of a massive Dirac fermion is the sine-Gordon model at a particular coupling. Precisely, the two-dimensional fermion action is equivalent to $$S_{2d} = \int d^2 x \left[ {\kappa}(\nabla\varphi)^2 - \lambda
\cos(2\pi\varphi) \right] .
\label{sgaction}$$ For the free-fermion case, we have $\kappa=1/2\pi$; we will discuss the more general case in section \[sec:q8v\]. In the fermion language, different values of $\kappa$ correspond to adding a four-fermion coupling to $S_{2d}$. To find a Hamiltonian with this two-dimensional action describing the ground state, we again find an operator $Q(x)$ and define the Hamiltonian via Eq. (\[eq:HQQ\]). The operator $$Q (x) \equiv \frac{1}{\sqrt{2}} \left(\frac{\delta}{\delta\varphi} +
\kappa \; \nabla^2\varphi + \frac{\lambda}{2\pi} \sin(2\pi\varphi)\right)
\label{Qsg}$$ annihilates the wave functional $\Psi \propto e^{-S_{2d}}$. Because of the extra term in $Q$, the commutator $\left[ Q(\vec x), Q^\dagger(\vec y) \right]$ is not a simple c-number, but in fact depends on the field configuration $\varphi$, $$\left[ Q(\vec x), Q^\dagger(\vec y) \right] = \kappa
\nabla^2\left(\delta^{(2)} (\vec{x} - \vec{y}) \right)
+ \lambda \sin(2\pi \varphi(\vec{x})) \delta^{(2)}(\vec{x}-\vec{y}) \ .
\label{QQccrpot}$$ Thus, normal-ordering the Hamiltonian is not just an innocent ground state energy shift: the two parts of (\[eq:HQQ\]) are not the same here. To obtain the desired ground-state wave functional, we must define $H$ of the form $\int Q^{\dagger} Q$. The three-dimensional version of this model was discussed in Ref. \[\].
This Hamiltonian is not quadratic in the field $\varphi$ except at the critical point $m=0$, so that even with this fine tuning this model cannot be solved simply. Since $Q$ and $Q^\dagger$ do not have simple commutation relations, we cannot simply find the spectrum of this theory. Of course, one can compute properties in the fermionic picture, but as noted before, computations involving spin fields are non-trivial in this basis as well. However, since there are dimensionful parameters in the Hamiltonian and no spontaneous breaking of a continuous symmetry, it seems likely that the Hamiltonian is gapped. Moreover, in the limit with $\lambda/\kappa$ finite, the action Eq. (\[sgaction\]) reduces to that of a free massive boson. Then one can solve the model explicitly, and the quantum Hamiltonian indeed has a gap. When $\lambda$ is reduced to a finite value, the gap should remain[^5].
There are many terms in the Hamiltonian of this field theory, and their coefficients must be fine-tuned to enable us to compute the ground-state wave function explicitly. There are terms like $\cos(4\pi\varphi)$ and $(\nabla\varphi)^2$ which, as noted above, tend to order the system. However, the exact lattice results for $H_w$ from \[\] show that the ground-state correlators for $w$ small are those of the bosonic Hamiltonian with these special couplings. Thus this model is not ordered but rather topologically ordered. Moreover, since the model is gapped, we expect that its physical properties are robust and persist even when the coefficients are tuned away from this special point. Thus there must exist a topological phase, not just an isolated point. An interesting open problem is to understand how large (in coupling constant space) the topological phase is, as compared to the ordered phases.
A quantum eight-vertex model {#sec:q8v}
============================
In the last section, we discussed a $2+1$-dimensional theory whose ground-state wave function is simply described in terms of a classical two-dimensional bosonic field. With vanishing potential and a particular value of the coupling $\kappa$, it is believed to describe the continuum limit of the quantum dimer model on the square lattice. In this section, we will study lattice models which in the continuum limit allow arbitrary values of $\kappa$. These models also have a quantum critical line separating an ordered phase from a topologically-ordered phase.
The degrees of freedom in our model are those of the classical two-dimensional eight-vertex model. These are arrows placed on the links of a square lattice, with the restriction that the number of arrows pointing in at each vertex is even. This means that there are eight possible configurations at each vertex, which we display in figure (\[fig8v\]).
The classical Boltzmann weights for a given vertex in the zero-field eight-vertex model are usually denoted by $a$, $b$, $c$ and $d$, as shown in the figure. Since we are interested in rotationally-invariant theories, we set $a=b$ in the following; moreover, since we can rescale all the weights by a constant, we set $a=b=1$. A typical configuration is displayed in figure (\[fig8vtypical\]);
the Boltzmann weight of such a configuration is given by the product of Boltzmann weights of the vertices.
The classical eight-vertex model is integrable, and many of its properties can be derived exactly \[\]. For $a=b=1$, it has ordered phases for $c>d+2$ and $d>c+2$. In these phases the ${\mathbb Z}_2$ symmetry of flipping all the arrows is spontaneously broken. Critical lines with continuously varying exponents at $c=d+2$ and $d=c+2$ separate the ordered phases from the disordered one $|c-d|<2$. The correlation length diverges as \[\] $$\xi \sim \big| |c-d| -2 \big|^{-\pi/(2\mu)}\ , \qquad
\mu \equiv 2 \tan^{-1}(\sqrt{cd})\ ,
\label{critexp}$$ near these critical lines. (For $\pi/\mu$ an even integer this is multiplied by $\log | |c-d| -2 |$ .) When $c=0,d\leq 2$ or $d=0,c\leq 2$, the model is also critical; in fact the partition function on this line can be mapped onto that for the order-disorder critical line. For $d=0$, the exponent in Eq. (\[critexp\]) diverges: there is a Kosterlitz-Thouless transition as one brings $c$ through $2$. Another useful result for the classical correlation length is that it is zero on the line $c=d$; this is the state of maximal disorder.
An order parameter which will be useful later comes by rewriting the model in terms of an Ising spin at the center of each plaquette. This description is best thought of as two Ising models, with spins $\tau(A)$ on one sublattice, and $\tau(B)$ on the other. Then the Boltzmann weights can be written in terms of two Ising couplings between nearest sites on the $A$ lattice and on the $B$ lattice, and a four-spin coupling between the two $A$ and two $B$ spins around a site of the original lattice. The polarization operator of the eight-vertex model becomes $\tau(A)\tau(B)$. One finds that its expectation value is non-vanishing in the ordered phase, and vanishes in the disordered phase $|c-d|<2$ \[\]. One also can define Neél-like staggered order parameters, in terms of $\tau(A)$ and $\tau(B)$ individually, which do not vanish in the ordered phase. Along the line $cd=1$, the four-spin coupling vanishes, so the eight-vertex model turns into two decoupled Ising models; the model here in this case be solved by using Pfaffian techniques \[\]. Along the line $c=d$, the two Ising couplings vanish, leaving only the four-spin coupling. Thus this line in the classical model has an extra ${\mathbb Z}_2$ gauge symmetry.
The classical eight-vertex model has a number of useful dualities \[\]. They can be described by defining the combinations $W_1=(a+b)/2$, $W_2=(a-b)/2$, $W_3=(c+d)/2$ and $W_4=(c-d)/2$. The partition function is invariant under the exchange of any two of the $W_j$ and under the $W_j\to -W_j$ for any $j$. These dualities, for example, map the critical line $c=d+2$ to the critical line $d=0,c\le 2$ by exchanging $W_1$ with $W_4$. In Ising language this amounts to performing Kramers-Wannier duality on one of the two types of Ising spins. The line $c=d+2$ is invariant under the exchange $W_1\leftrightarrow W_3$. In Ising language, this duality amounts to taking the Kramers-Wannier dual of both types of Ising spins. Denoting the dual spins as $\mu(A)$ and $\mu(B)$, duality means therefore that in the disordered phases, the expectation value $\langle \tau(A) \tau(B)\rangle$ is non-vanishing.
Construction of the quantum eight-vertex Hamiltonian {#sec:8vqh}
----------------------------------------------------
We now define a quantum Hamiltonian acting on a Hilbert space whose basis elements are the states of this classical eight-vertex model. To define such a Hamiltonian, we first need the analog of the flip operator in the quantum dimer model. A flip operator needs to be ergodic: by flips on various plaquettes one should be able to reach all the states with the same global conserved quantities. The simplest such operator for the eight-vertex model is the operator which reverses all the arrows around a given plaquette. We write this flip operator $\hat{\mathcal F}_i$ explicitly in gauge-theory language in appendix \[app:z2gauge\]. Note that as opposed to the quantum dimer model on the square or triangular lattice, all configurations in the quantum eight-vertex model are flippable[^6]: $\hat{\mathcal F}_i$ preserves the restriction that an even number of arrows be pointing in or out at each vertex.
The simplest Hamiltonian has no potential energy, just a flip term. It is convenient to write this in terms of a projection operator: for $I$ the identity matrix, we have $(I-\hat{\mathcal F}_i)^2
=2(I-\hat{\mathcal F}_i)$. Then the Hamiltonian $$H_{c=d=1}= \sum_i (I-\hat{\mathcal F}_i)
\label{hkitaev}$$ has a ground state corresponding to the equal-amplitude sum over all eight-vertex model states. In terms of the Boltzmann weights introduced above, this is the state with $a=b=c=d=1$. A Hamiltonian with the same ground state was introduced by Kitaev \[\]. There the eight-vertex-model restriction of having an even number of arrows in and out at each vertex was not required a priori, but instead a term was introduced giving a positive energy to vertices not obeying the restriction. The zero-energy ground state can therefore include no such vertices, so the ground state for the model of \[\] is indeed the sum over the eight-vertex-model configurations with equal weights. Because every plaquette is flippable and all configurations have equal weights, different $\hat{\mathcal F}_i$ commute. Therefore all the terms in Eq. (\[hkitaev\]) commute with each other, so they can be simultaneously diagonalized and their eigenstates can easily be found, as for the quantum Lifshitz field theory. The model is gapped, and is in a topologically-ordered phase \[\]. This follows as well from the results for the classical eight-vertex model discussed above: for $c=d=1$ all Ising couplings vanish resulting in two decoupled Ising models at infinite temperature. At this point the order parameter vanishes, but the non-local order parameter does not. We can thus interpret the product of dual Ising variables $\mu(A)\mu(B)$ as a topological order parameter.
This ground state can also be mapped onto the ground state of a ${\mathbb Z}_2$ gauge theory deep in its deconfined phase. In Appendix \[app:z2gauge\], we give a detailed derivation of the $ {\mathbb
Z}_2$ gauge theory of the full quantum eight-vertex model we define below, as well as the its dual theory which we will use to characterize some of the phases. In Appendix \[app:gauge\] we discuss the $U(1)$ gauge theory description of the quantum six-vertex model, which describes the limit $d=0$, and by duality, the lines $c=0$, $c^2=d^2+2$ and $d^2=c^2+2$.
We now find a two-parameter quantum Hamiltonian whose ground state is a sum over the states of the eight-vertex model with amplitudes given by the classical Boltzmann weights with arbitrary $c$ and $d$. We are still keeping $a=b=1$ to preserve two-dimensional rotational invariance, but this restriction can be relaxed if desired. Our model is neither the simplest nor the most natural extension of the classical eight-vertex model: a simpler Hamiltonian was proposed by Chakravarty \[\] in the context of $d$-density waves. Although we believe that our Hamiltonian and Chakravarty’s describe the same physics, we are not aware of any simple mapping between these models. Another lattice model related to the quantum eight-vertex model discussed here was introduced in Ref.. The main virtue of the construction that we use here is the structure of the ground-state wave function.
Finding a quantum Hamiltonian with a known ground state is straightforward to do by using the trick discussed above (and in Ref.). Namely, we find a Hamiltonian of the form $$\label{q8vham}
H_{q8v} = \sum_i w_i {\mathcal Q}_i$$ where ${\mathcal Q}_i={\mathcal Q}_i^\dagger \propto {\mathcal Q}_i^2$. To yield the desired ground state, each operator ${\mathcal Q}_i$ must annihilate the sum over states with each state weighted by $c^{N_c}
d^{N_d}$, where $N_c$ and $N_d$ are the number of $c$ and $d$ type vertices in that state. In particular, we look for a ${\mathcal Q}_i$ of the form $${\mathcal Q}_i = \sum_i \left[\hat{\mathcal V}_i - \hat{\mathcal F}_i\right]
\label{q8v}$$ where $\hat{\mathcal V}_i$ is diagonal and depends on the Boltzmann weights for the four vertices at the corners of the plaquette $i$. Since $(\hat{\mathcal F}_i)^2=I$, this Hamiltonian breaks into $2$ by $2$ blocks like $H_{RK}$ for the quantum dimer model. If we choose the potential ${\mathcal V}_i$ so that the blocks are of the form $$\begin{pmatrix}
v & -1\\
-1 & v^{-1}
\end{pmatrix} \ ,$$ ${\mathcal Q}_i$ will have the desired properties.
To find $\hat{\mathcal V}_i$, let $n_c$ be the number of $c$-vertices at the corners of the plaquette $i$, and let $\widetilde{n}_c$ be the number of $c$-vertices around the plaquette after it is flipped by $\hat{\mathcal F}_i$. Likewise, let $n_d$ be the number of $d$-vertices around the plaquette, while $\widetilde{n}_d$ is the number of $d$ vertices in the flipped configuration. Note that $\hat{\mathcal F}_i$ always flips a $c$ or $d$ vertex to an $a$ or $b$ vertex, and vice versa. Consequently, we take the operators $\hat
{\mathcal V}_i$ to be of the form $$\hat{\mathcal V}_i =
c^{\widetilde{n}_c - n_c} d^{\widetilde{n}_d - n_d}
\label{Vpot} \ .$$ Explicitly, we can write the projectors as follows $$\mathcal{Q}_i = \begin{pmatrix}
c^{\widetilde{n}_c - n_c} d^{\widetilde{n}_d - n_d} & -1 \\
-1 & c^{\widetilde{n}_c - n_c} d^{\widetilde{n}_d - n_d} \\
\end{pmatrix} \ .$$ This ${\cal Q}_i$ is indeed proportional to a projection operator: the Hamiltonian (\[q8vham\]) has the classical eight-vertex model as an eigenstate. This holds for any choice of the $w_i$, but in the ${\mathbb Z}_2$ gauge-theory language of appendix \[app:z2gauge\], it is natural to set all $w_i=1$. The amplitude of the ground-state wave function of this $H_{q8v}$ on a state with $N_c$ $c$-vertices and $N_d$ $d$-vertices is $$\Psi_0[N_c,N_d] = \frac{c^{N_c} d^{N_d}}{\sqrt{Z(c^2,d^2)}} \ ,$$ where $Z(c,d)$ is the partition function of the classical two-dimensional eight-vertex model with weights $a=1$, $b=1$, $c$ and $d$. The arguments in the denominator are $c^2$ and $d^2$ because averages in the quantum model are calculated with respect to $|\Psi_0|^2$.
In section \[sec:2dwavefn\], we discussed how by taking the limit $w\to 0$ in $H_w$, one can recover a slightly-generalized square-lattice quantum dimer model which allows defects with no dynamics. Similarly, here one can find a quantum six-vertex model with some defects allowed by taking the limit $d \rightarrow
0$.[^7] Configurations with $n_d>\widetilde{n}_d$ on any plaquette will receive infinite potential energy and so are disallowed. Some useful facts are that $n_c+n_d +
\widetilde{n}_c +\widetilde{n}_d =4$, and $n_c-\widetilde{n}_c =0$ mod 2 and $n_d-\widetilde{n}_d =0$ mod 2. When $d\to 0$, configurations with $n_d=\widetilde{n}_d=0$ are the flippable plaquettes of the six-vertex model, and are obviously included in the ground-state wave function. Configurations with $n_d=\widetilde{n}_d=1$ or $n_d=\widetilde{n}_d=2$ are not suppressed, but the flip therefore preserves the number of $d$ vertices around each plaquette. Thus in the $d\to 0$ limit we can view these $d$ vertices as defects in the quantum six-vertex model. The ground-state wave function is a sum over all allowed states with a given set of plaquettes with defects. The amplitude of each configuration in this sum is proportional to $c^{N_c}$. We discuss the relation between the quantum six- and eight-vertex models in more detail in the Appendices \[app:gauge\] and \[app:z2gauge\].
The phase diagram {#sec:8vphases}
-----------------
Now that we have the exact ground-state wave function for the quantum eight-vertex Hamiltonian of , we can use it to determine the phase diagram. Given the fact that the probability of a configuration of arrows in the ground state wave function is equal to the Boltzmann weight of a classical eight-vertex model, we can deduce much of the physics of the quantum theory (at least its equal-time properties) directly from the Baxter solution of the classical 2D eight-vertex model \[\], as well as from Kadanoff’s classic work on its critical behavior \[\]. The only change is that the weights must be squared here, since in quantum mechanics we weigh configurations with $|\Psi|^2$. The phase diagram is displayed in figure \[eight\]; note that the axes are labeled by $c^2$ and $d^2$.
We will now use this knowledge, as well a simple perturbative arguments in the quantum theory, to determine the phase diagram, the behavior of physical observables in the different phases, and (much of) their critical behavior. A useful fact is that for $cd=1$, the classical model with partition function $Z(c^2,1/c^2)$ is equivalent to two decoupled Ising models. This decoupling is a property only of the wave function, not of the Hamiltonian of the full $2+1$-dimensional quantum theory. Since this line $cd=1$ goes through both ordered and disordered phases, much of the physics of the quantum eight-vertex model can be described at least qualitatively in terms of decoupled Ising models. In particular, for any values of $c$ and $d$ except those on the critical line, correlators decay exponentially fast with distance with a correlation length $\xi$ which diverges as the phase boundary is approached, in a manner given in Eq.. This exponential decay occurs in general, not just on the decoupling curve.
As with the dimer models discussed in section \[sec:2dwavefn\], the partition function for $cd=1$ can be expressed in terms of Grassmann variables with only quadratic terms, i.e. free fermions. Duality means that $Z(c^2,d^2)$ is free-fermionic for $c^4+d^4=2$ as well. The correlators in the continuum limit of the (critical) square-lattice quantum dimer model are therefore identical to those obtained for $c=\sqrt{2}$, $d=0$. The special point $c=d=1$ discussed above and in Ref. \[\] (labeled in fig. \[eight\] as “Kitaev”) is also free fermionic. At this point one does not need the Pfaffian techniques to compute correlators exactly, and one finds that the model is in a disordered phase in the Ising-spin language. However, the expectation value $\langle\mu(A)\mu(B)\rangle$ is non-vanishing, so there is topological order at this point.
Let us now discuss the different phases of this system.
1. [*The Ordered (Confined) Phase*]{}:\
From the known phase diagram of the classical eight-vertex model \[\], we conclude that the ground state of the [*quantum*]{} model with the Hamiltonian of Eq. has an [*ordered*]{} phase for $c^2>d^2+2$ (and also for $d^2>c^2+2$). That this is an ordered phase can be seen easily by considering the limit $c \to \infty$ (with $d$ fixed). In this limit the ground state is dominated by just two configurations, related to each other by a lattice translation of one lattice spacing, which have a $c$ vertex on every site, as shown in fig. \[ddw\]. In this phase the staggered polarization operator $\langle \tau(A)\tau(B)\rangle$ has a non-vanishing expectation value. This result can also be obtained directly from the Hamiltonian of the ${\mathbb Z}_2$ gauge theory, Eq. , since for $c$ large the potential energy term $H_V$ dominates and in it the piece associated with the $c$ projection operators.
The ordered phase is also [*confining*]{}. Below we will discuss the behavior of the Wilson loop operator and show that in this regime it obeys an area law, the hallmark of confinement \[\]. We will also show that the energy of a state with two static sources grows linearly with their separation. We should also note here that the equal-time fermion correlation function has an exponential decay in this phase, suggesting that this phase may support massive fermionic excitations.
2. [*The Disordered (Deconfined) Phase*]{}:\
From the exact solution of the classical model, we know that there is a [*disordered*]{} phase for $c^2<d^2+2$ ($d^2<c^2+2$) (see Fig. \[eight\]). This is also a [*deconfined*]{} phase. This is most easily seen by taking a point deep in this phase, such as the Kitaev point $a=b=c=d=1$ \[\]. In fact, all along the line $a=b=1$ and $c=d$, the correlation length of the eight-vertex model is zero \[\]. All points in the disordered phase have this line as their RG stable fixed point (all points on the line are equivalent). This is the infinite temperature limit of the classical 2D eight-vertex model. In this regime the expectation value of the polarization operator $\langle \tau(A)\tau(B)\rangle$ is zero (and its correlation vanishes at a length scale of the lattice spacing.). Thus this state does not exhibit long-range order, but since the dual variable $\langle \mu(A)\mu(B)\rangle$ is non-vanishing, it exhibits topological order. Quantum mechanically, the wave function is the equal-amplitude superposition of all configurations of arrows consistent with the eight-vertex restrictions. This state is deconfined since in this limit the Hamiltonian reduces to the flip term (plaquette). It is well known \[\] that in this regime the Wilson loop operator has a perimeter law behavior and that the energy of two static sources is finite and independent of their separation. The low energy ([*i. e.*]{} long-distance) sector of this phase describes a [*topological*]{} ${\mathbb Z}_2$ deconfined theory, equivalent to the Kitaev point.
3. [*Critical Behavior*]{}:\
The 2D [*quantum*]{} theory has lines of critical points $c^2=d^2+2$ (and $d^2=c^2+2$). There are also lines of critical points for $d=0$ and $0\leq c^2 \leq 2$ (and $c=0$ and $0\leq d^2 \leq 2$). All of these critical lines are (up to duality transformations) equivalent to the six-vertex model. Notice that, since the energy of this quantum state is exactly zero (by construction), the ground-state energy does not have singularities at the phase transitions. This is of course a peculiarity (or rather a pathology) of this model and it certainly non-generic: any perturbation leading to a non-vanishing energy will lead to singularities in the energy of the ground state.
In section \[critft\], we discussed how the ground-state correlators of the (multi) critical square-lattice quantum dimer model were the same as those of the quantum Lifshitz theory, Eq. , at $\kappa^{-1}=2\pi$ \[\]. The same equivalence holds for the quantum eight-vertex model for all values of $\kappa$. Moving along the critical line results in changing $\kappa$ as given in (\[eq:kappa8v\],\[eq:kappa8v2\]). Kadanoff \[\] showed that the critical behavior of the classical eight-vertex model can be mapped exactly to the critical behavior of the two-dimensional Gaussian model, the free boson described by Eq. and Eq. . The equal-time correlators in the quantum Lifshitz model summarized in Appendix \[app:gaussian\] are therefore identical to those of the quantum eight-vertex model on its critical lines. Moreover, when $d=0$ and $c^2\le 2$, the eight-vertex model becomes the six-vertex model, which has a height description. Thus the same heuristic arguments applied in the last section can be applied here, again implying that the quantum six-vertex model and the quantum Lifshitz theory are in the same universality class. In particular, the dynamics of the quantum eight-vertex model along its critical lines will obey scaling with a dynamic exponent $z=2$.
To compete the identification, we need to express the coupling constant $\kappa$ of the quantum Lifshitz model in terms of the eight-vertex parameters $c$ and $d$. This is easiest to do by computing the dimension $x$ of the “energy” operator, which when added to the action of (classical) critical theory, moves it away from criticality. Combining universality with Baxter’s exact result (\[critexp\]) for the correlation-length exponent gives $x=2(1-\mu/\pi)$. We have normalized the boson $\varphi$ in Appendix \[app:gaussian\] so that the energy operator is given by $\cos(2\varphi)$, which has dimension $x=1/(2\pi\kappa)$. Note that at the free-fermion point $\kappa^{-1}=2\pi$, the energy operator has dimension 1, the dimension of a 2D fermion mass term. Combining the two and using a simple trigonometric identity yields $$\kappa^{-1}=8 \cot^{-1}\left(cd\right).\qquad\qquad \hbox{for }|c^2-d^2|=2
\label{eq:kappa8v}$$ and by duality $$\kappa^{-1}=8 \cot^{-1}\left(\sqrt{\frac{4}{c^4} -1}\right)\qquad\qquad
\hbox{for } c^2\le 2, d=0.
\label{eq:kappa8v2}$$ At the Kosterlitz-Thouless transition point $c^2=2,d=0$ where the two critical lines meet, the dimension $x=0$ as expected, and both formulas give $\kappa^{-1}=4\pi$ here.
This equivalence to a scalar field theory can be extended to the scaling region near the critical lines, by repeating the arguments of section \[offcritft\]. The perturbing operator is $\cos(2\varphi)$; symmetry forbids $\cos(\varphi)$ from being added to the action of the classical theory. Thus near to the critical lines the effective field theory will be of the form (\[Qsg\]) and Eq. (\[eq:HQQ\]); these field theories will have the correct critical exponents.
Operators {#sec:operators}
---------
Equal-time correlation functions of operators which are diagonal in the arrow representation of the eight-vertex model are identical to correlation functions of the classical eight-vertex model. Unfortunately, these correlation functions can be computed exactly only along the free-fermion lines, by using Pfaffian techniques. Likewise correlation functions of off-diagonal operators ([*e.g.*]{} flip operators) can only be determined at some special points deep in the ordered and disordered phases by using directly the quantum eight-vertex Hamiltonian. However, at or near the critical lines, the asymptotic behavior of the correlators can be found by using universality. The operators of interest in the eight-vertex model can all be expressed in terms of the charge and vortex operators of the Gaussian model \[\].
The operators of interest in the quantum eight-vertex model can be readily identified in the quantum Lifshitz model. Thus we have a full identification of a $2+1$-dimensional quantum critical theory. As a direct consequence of being able to identify these operators we can also study how these operators perturb the critical theory, at least within a renormalization group argument along the lines of Ref. \[\]; the analog in the quantum dimer model is our analysis of $H_w$ for $w$ small at the end of section 2.
To identify the eight-vertex operators in the bosonic language as in Ref. \[\], we first study the six-vertex line, where there is a height description. There the height is local in terms of spin variables, so we expect the boson $\varphi$ to be local as well. Standard bosonization techniques \[\] show that only products of spin operators from both Ising models can be written in terms of the boson. One finds that $\cos(\varphi)\sim \tau(A)\tau(B)$. Note that this is consistent with the identification of $\cos(2\varphi)=2\cos^2(\varphi)-1$ with the perturbing operator; the usual Ising operator product gives the fusion rule $\tau(A)\tau(A)\sim
1 + {\cal E}(A)$, where ${\cal E}(A)$ is the energy operator.
To study the operators in more depth, we utilize a $2+1$-dimensional ${\mathbb Z}_2$ gauge theory of the quantum eight-vertex model derived in Appendix \[app:z2gt\]. The degrees of freedom of this gauge theory reside on the links of the square lattice and have a one-to-one correspondence to arrow configurations of the eight-vertex model. The Hamiltonian of the ${\mathbb Z}_2$ gauge theory is $$H_{\rm q8v} = H_{\rm V} + H_{\rm flip}
\label{q8vham2}$$ where $H_{\rm flip}$ has the form of a sum over plaquettes of ${\mathbb Z}_2$ flip operators[^8]: $$\label{sflip}
H_{\rm flip} = - \sum_{\vec x} \sigma^3_1(\vec x) \; \sigma^3_2(\vec x+\vec e_1) \; \sigma^3_2(\vec x) \; \sigma^3_1(\vec x+\vec e_2)$$ The potential energy terms are combinations of operators which project onto the allowed select $a$, $b$, $c$ and $d$ vertices, and assign the weights $a$, $b$, $c$ and $d$ to different contributions to the wave function. The explicit form of the potential energy terms $H_{\rm
V} $ is given in Eq. . As in all ${\mathbb Z}_2$ gauge theories \[\], the physical states of this theory satisfy the constraint of gauge invariance (“Gauss Law") $$\label{scon8v}
G(\vec x)=\sigma^1_1(\vec x) \; \sigma^1_2(\vec x) \; \sigma^1_1(\vec x-\vec e_1) \; \sigma^1_2(\vec x-\vec e_2) =
1 \qquad \forall \vec x $$ which in this context simply expresses the restrictions on the configurations allowed in the eight-vertex model. The operator $G(\vec x)$ is the generator of local time-independent gauge transformations \[\].
In Appendix \[app:duality\] we derive the dual theory of the quantum eight-vertex Hamiltonian of Eq. . The degrees of freedom of the dual theory are defined on the sites of the dual square lattice. The Hamiltonian is $$H_{\rm q8v, dual} = H_{\rm V,dual} - \sum_{\vec r} \tau^1(\vec r) \ ,
\label{q8v-dual2}$$ where $H_{\rm V,dual}$ is given in Eq. and Eq. . We will use the both the gauge theory and its dual to investigate these phases.
Let us discuss briefly the observables of this model, how they behave in the ordered (confined) and disordered (deconfined) phases and what critical behavior they exhibit at the phase transition lines. As we note in Appendix \[app:duality\], the quantum eight-vertex model retains the two-sublattice structure of the classical model. The ${\mathbb Z}_2 \times {\mathbb Z}_2$ symmetry of the classical model also survives in the quantum theory.
1. [*Order and disorder operators*]{}: We denote by $\tau(A)$ and $\tau(B)$ the $\tau^3$ (order) operator for sublattice $A$ and $B$ respectively. In the gauge theory this operator is a disorder or kink operator \[\]; in this 2D gauge theory this is just the flux (monopole) operator of Refs., or the vison operator of Refs.. The order operators are just the order parameters of the sublattice Ising models. As such they acquire an expectation value in the ordered phase (where the staggered polarization has an expectation value as well). In the disordered phase their equal-time correlation functions decay exponentially with distance (as do their connected equal-time correlators in the ordered phase). The precise behavior of this correlation function is only known on the decoupling curves where they reduce to the correlation functions of the 2D classical Ising model, and on the critical lines! The scaling dimension of this spin field (which is known as the “twist” field in conformal field theory \[\]), is $1/8$ not only at the decoupling point (as is well known from the 2D Ising model) but along the entire phase boundary. We note however, that the spin fields can not be represented directly in terms of the boson, but only in its orbifold \[\].
Similarly, we will denote by $\mu(A)$ and $\mu(B)$ the “frustration" (or “fractional domain wall") or [*disorder operator*]{} \[\] for sublattices $A$ and $B$ respectively. In the 2D classical theory, this is the dual of the order operators and in conformal field theory it is also represented by a twist field. It has the same scaling dimension as the spin field. On the other hand, in the gauge theory, the disorder operators correspond to a local violation of the Gauss-Law constraint, Eq. , by imposing that $G(\vec x)=-1$ just at site $\vec x$, [*i.e.*]{} a static source ([*i.e.*]{} an “electrode", or Polyakov loop) at site $\vec x$. This operator is analogous to the holon (or “monomer") operator in the Rokhsar-Kivelson quantum dimer model. This expectation value of products of two operators of this type serves as a test for [*confinement*]{}. The disorder operators have a non-zero expectation value in the quantum disordered phase. A simple minded calculation shows that, in [*in the ordered phase*]{}, the ground state with two such defects (or sources) separated at a distance $R$, has a non-zero energy $U(R)$, which grows [*linearly*]{} with $R$, $U(R) \sim \sigma R$, consistent with the fact that the defects create a fractional domain wall in an ordered state. We expect that the string tension $\sigma$ vanishes as the critical lines are approached, with a behavior dictated by that of the gap scale, [ *i.e.*]{}, $\xi^{-z}$. Thus, the excitations created by the disordered operator are [*confined*]{} in the ordered phase[^9]. In contrast, in the quantum disordered phase, the energy of a pair of defects saturates to a finite value if $R \gg \xi$. Hence, the disordered phase is deconfined.
It is also interesting to ask what is the [*energy cost*]{} of a set of vortex configurations on the quantum critical lines. This can be done in a number of ways. We note here that using path-integrals the change of the ground state energy can be found by computing the path integral in a background of vortices, normalized by the path-integral without vortices. Consider the simple (an general) case of two vortices of magnetic charges $\pm m$ separated a distance $R$. Let is denote this amplitude by $W(R,T)$, where $T$ is the (infinite) time-span of the system. In the dual gauge theory this is the same as the computation of two static Wilson (or Polyakov) loops, corresponding to two static sources with electric charges $\pm m$ at a distance $R$ from each other. The energy cost of these defects is $$U(R)=\lim_{T \to \infty} \left(-\frac{1}{T} \ln W(R,T)\right)
\label{energy-cost}$$ We can compute $U(R)$ by using the path-integral for the quantum Lifshitz theory of Eq. in a background of vortices, which amounts to modifying the action by the minimal coupling shift $$(\nabla^2 \phi)^2 \to ( \vec \nabla \cdot (\vec \nabla \phi-\vec A))^2$$ By explicit calculation we find that $$U(R)=0
\label{eq:zero}$$ In other terms, the interaction energy of vortices is zero! A simple way to understand this results is to recall that, in imaginary time, the quantum Lifshitz model is a theory of a smectic liquid crystal. The vortex state is just a configuration of screw dislocations running along the $z$ axis (imaginary time). Since the smectic is isotropic and has no resistance to shear in the $xy$ plane, these defects do not cost any energy. From a quantum mechanical point of view this result is consistent with the known fact that at the RK point of the quantum dimer model, the interaction energy of a pair of monomers is zero[^10]. In gauge theory language this means that at this critical point vortices are completely free.[^11] In contrast, at quantum critical points of Lorentz invariant gauge theories, which have $z=1$, it is known \[\] that at their critical point the effective interaction has the universal law $U(R)\sim1/R$, in [*all dimensions*]{}.
2. [*Polarization*]{}: The polarization operator, the local arrow configuration on a link, is the natural order parameter. In terms of the order operators it reads $P=\tau(A) \tau(B)$. As we saw above in the classical “antiferroelectric" ordered phase, $P$ is a natural order parameter, although appropriately-staggered expectation values of $\tau(A)$ and $\tau(B)$ also do not vanish in this ordered phase. The behavior of the polarization operator on the critical lines of the classical model was studied by Kadanoff and Brown \[\], who found that its scaling dimension is $\Delta_P=1/8\pi \kappa$, with $\kappa$ given by Eq.. From the results of Appendix \[app:gaussian\] we see that we can identify the polarization operator with the boson operator $P \sim \cos(\varphi)$ in the quantum Lifshitz theory, which has the same scaling dimension.
3. [*Mass term*]{}: In the classical theory the mass term, or “energy density", is the product of two (dual) spin variables $\tau$ on nearest neighboring sites on the same sublattice. We will denote them by ${\mathcal E}(A)$ and ${\mathcal E}(B)$ respectively. Although in the quantum theory these operators no longer correspond to an energy density, we will define ${\mathcal E} =\left( {\mathcal
E}(A)+{\mathcal E}(B)\right)/\sqrt{2}$, as usual, and refer to this operator as to the [*mass term*]{} since it drives the theory away from criticality. In Ref. \[\] it is found that the scaling dimension of this operator is $\Delta_{\mathcal
E}=1/2\pi \kappa=4 \Delta_P$, which led to the identification ${\mathcal E} \sim \cos(2\varphi)$. If we regard this operator as acting on the critical line along the $d=0$ axis, [*i.e.*]{} the $6$-vertex model, we see that this is a relevant operator for the quantum Lifshitz model as well. Grinstein’s RG arguments \[\] show that this perturbation drives the system into a massive phase with a residual unbroken ${\mathbb Z}_2$ symmetry. This is just the disordered phase of the quantum eight-vertex model, and this perturbation drives the system towards the Kitaev point. A similar analysis works on the $c^2=d^2+2$ critical line.
4. [*Wilson loops*]{}: As it is well known gauge theories, the natural test for confinement is the behavior of the Wilson loop operator $W_\gamma=\prod_{\ell \in \gamma} \sigma^3(\ell)$, where $\{ \ell \}$ is a set of links that belong to a closed path $\gamma$ on the (direct) lattice (see Ref. \[\] and references therein). Since these operators are off-diagonal in the arrow representation, [*i.e.*]{} in the representation in which $\sigma^1$ is diagonal, they do not have an analog in the classical 2D eight-vertex model. A simple argument \[\] shows that in the ordered state, which for $c \to \infty$ is essentially an eigenstate of $\sigma^1$, the expectation value of the Wilson loop is zero and that first non-zero contribution comes from acting $N(\gamma)$ times with the flip operator, where $N(\gamma)$ is the number of plaquettes enclosed inside the path $\gamma$, and that the behavior thus found has the form $\exp(-{\rm constant}\; N(\gamma))$, [*i.e.*]{} an [*area law*]{}. Thus the ordered state is a confining phase. Conversely, deep in the disordered phase, the ground state is essentially an eigenstate of the flip operator and we get a perimeter law, [*i.e.*]{} deconfinement.
5. [*Spinors*]{}: The classical 2D eight-vertex model has two Majorana fermion operators $\psi(A)$ and $\psi(B)$, constructed as usual as products of order and disorder operators \[\], $\psi\sim \tau \mu$. We can also define an analog of the spinors in the quantum theory (see the discussion in Appendix \[app:gaussian\] on charge and vortex operators in the $2+1$-dimensional quantum Lifshitz model). The scaling dimension of the spinors is \[\] $\Delta_\psi=\left(2\pi \kappa+1/2\pi\kappa\right)/4$, a result familiar from the Luttinger model.
6. [*Other operators*]{}. The classical 2D eight-vertex model contains a marginal operator $E={\mathcal E}(A) {\mathcal E}(B)$, with scaling dimension $\Delta_E=2$, which is responsible for the lines of fixed points with varying critical exponents. The quantum eight-vertex model discussed here has a similar behavior. In the quantum Lifshitz model the marginal operator is[^12] $E \sim \left(\nabla^2
\varphi\right)^2$. Another operator of interest is the crossover operator $\left( {\mathcal E}(A)-{\mathcal E}(B)\right)/\sqrt{2}$ which breaks the sublattice symmetry and has scaling dimension $2\pi
\kappa$, The operator product expansion of two crossover operators generates an operator symmetric under sublattice exchange, and is identified with a charge-two vortex operator. The scaling dimension of this operator is $8\pi \kappa$ and it is becomes marginal at $\kappa=1/4\pi$, driving the Kosterlitz-Thouless transition along the $d=0$ line.[^13] Other interesting operators are the “mixed" operators ${\tilde
P}=\tau(A)\mu(B)$ and ${\tilde P}^*=\mu(A) \tau(B)$, both with scaling dimension $\Delta_{\tilde P}=\pi \kappa/2$. Hence ${\tilde P}$ and ${\tilde P}^*$ are vortex operators as well; these are non-local with respect to $P$. They have been identified as the holon creation operators in the context of the quantum dimer model \[\].
Another quantum eight-vertex model {#another}
----------------------------------
Let us end this section by noting that there is a simpler quantum eight-vertex model that one can write down whose behavior is quite different from the one we have discussed here. Basically, this model will give weights to all the vertices, according to their type. In addition, there will be a flip term. In its dual form the model is $$\label{nonrk}
H = \sum_{\vec r} \bigl(
\varepsilon_a \mathcal{P}_a(\vec r) + \varepsilon_b \mathcal{P}_b(\vec r) +
\varepsilon_c \mathcal{P}_c(\vec r) + \varepsilon_d \mathcal{P}_d (\vec r)
\bigr) - \sum_{\vec r} \tau^1(\vec r) \ ,$$ where the projection operators are defined in Eq. . The conventional Baxter weights of the classical model ([*i.e.*]{} when the flip term is absent) are $a=e^{-\frac{\varepsilon_a}{k_B T}}$, etc.
We first note that this model is not of the Rokhsar-Kivelson type (except when $\varepsilon_a=\varepsilon_b=\varepsilon_c=\varepsilon_d$, which is a dual version of the Kitaev model). The ground state wave function does not look like the statistical weights of the classical two-dimensional model, [*i.e.*]{} it is not a linear superposition of states weighted according to the vertices present. Thus, strictly speaking we can not use the machinery of this paper to study the ground state of this model. We do expect, however, that this model will have the same ordered phases, just like the ferro-electric and anti-ferro-electric phases of the eight-vertex model, as well as the same confinement-deconfinement properties discussed above. In addition, we expect that these ordered phases have the same spectrum and general properties as the ones we found in the Rokhsar-Kivelson type model we studied above. Similarly, in the regime where the flip term dominates, the ground state of this model is a uniform quantum disordered state, with the same properties as those of the topological phase of the model studied here.
However, the quantum critical properties of Eq. (\[nonrk\]) are completely different from the quantum eight-vertex model discussed in the rest of this section. They instead have a number of features in common with well-known three-dimensional classical models. By setting $\varepsilon_a+\varepsilon_b=\varepsilon_c+\varepsilon_d$, this model reduces to two interpenetrating two-dimensional Ising models in a transverse field, as the four-body interactions cancel. Thus along these curves this model has second-order phase transitions in the universality class of the classical three-dimensional Ising model, as opposed to the quantum eight-vertex model discussed above. The latter has $z=2$, while the model of Eq. therefore has dynamical exponent $z=1$ and is Lorentz invariant in the critical regime. It is easy to see that this behavior extends beyond the decoupling point. Indeed, the universality class of the 3D classical Ising model is controlled by the Wilson-Fisher fixed point and is accessible by the $4-\epsilon$ expansion. Away from the decoupling point we have then two 3D Ising models which are coupled through their energy density. This theory is in the universality class of a two-component real field $(\varphi_1,\varphi_2)$, whose $U(1)$ symmetry is explicitly broken down to a discrete ${\mathbb Z}_2 \times
{\mathbb Z}_2$ symmetry by quartic operators of the form $\varphi_1^2
\varphi_2^2$. This is a cubic symmetry-breaking perturbation and it is well known to be perturbatively relevant at the Wilson-Fisher fixed point of the two decoupled Ising models, but irrelevant at the Wilson-Fisher fixed point for the classical $XY$ model[^14]. It is well known from perturbative $4-\epsilon$ RG studies that the transition near the decoupling point that the transition may also become a fluctuation-induced first order transition depending on the sign of the effective coupling (see for instance Refs.).
Non-Abelian Topological States {#sec:YMCS}
==============================
In this section we broaden our scope and discuss field theories with continuous non-abelian symmetries. One can of course obtain a quantum critical point with non-abelian symmetries by taking copies of the quantum Lifshitz Hamiltonian. By choosing the charge lattice of the vertex operators carefully one can describe a theory with a non-abelian symmetry, say $SU(N)$. However the equivalent conformal field theories thus obtained always have integer central charge and are at level $1$. Although critical, such theories only support excitations with abelian statistics. One may try to attempt to generalize this result to models by writing down a theory with more structure, such as one related to a Wess-Zumino-Witten model. However, we will find that, contrary to naive expectation, such theories will turn out not to be critical but, instead, in a phase with a finite gap, a topological phase, just like the fractional quantum Hall effect. We will see that it is not at all clear how to find a conformal quantum critical point with a non-trivial non-abelian symmetry.
Nevertheless it is possible to use the approach of the previous sections to discuss topological non-abelian phases, even if any associated non-trivial quantum critical points are not yet known or do not exist. By analogy with the theory of fractional quantum Hall states, as well as from general results in Chern-Simons gauge theory, one expects that a time-reversal-breaking topological phase in $2+1$-dimensions should have an effective field theory description in terms of a Chern-Simons gauge theory at some level $k$. On closed manifolds this is a topological field theory in the sense that the partition function is independent of the metric of $2+1$-dimensional space-time, and that the expectation values of its gauge-invariant observables, the Wilson loops, depend only on the topology of the loops such as the knot invariants \[\].
By analogy with the approach pursued in the previous sections, we will discuss the properties of the wave functionals of this theory. This will lead us to consider first the strong-coupling limit of Yang-Mills theory with a Chern-Simons term, and later the topological sector of this theory. We will find the ground-state wave functional of this theory, and show how it is related to a two-dimensional Wess-Zumino-Witten (WZW) model for a specific choice of polarization ([*i.e.*]{}, the choice of canonically conjugate variables) which is natural in the strong coupling limit of the Yang-Mills Chern-Simons theory. Even though the WZW model is a conformal field theory with algebraically-decaying correlators, the gauge-invariant ground-state correlators in the two-dimensional quantum theory are exponentially decaying. Topological invariance will only be attained by correlators of gauge-invariant operators, Wilson and Polyakov loops, and only at long distances. The excitations obey non-abelian statistics: when bringing particles around one another, not only does one pick up phases, but the order in which they are exchanged affects the final state.
We will also study a doubled Chern-Simons theory which, contrary to the fractional quantum Hall effect, is time-reversal invariant. Using by now standard methods, we will construct the ground-state wave functionals of this doubled theory for a gauge group $G$ at level $k$, and use its connection with a gauged WZW model to understand its properties. We note that the doubled theory is often used to circumvent technical problems involved in the use of the formally anomalous wave functional of the undoubled theory: see the discussion in the subsections \[sec:ymcswzw\] and \[sec:doubled\]. The topological properties of the doubled theory have been recently discussed in depth by Freedman [*et al.*]{} \[\].
Field theories with continuous non-abelian symmetries {#sec:ftnon-abelian}
-----------------------------------------------------
Let us attempt first to construct a non-abelian version of the quantum Lifshitz theory, which one might hope would still be a quantum critical theory. We will follow the same procedure as in the abelian case. Since there are many two-dimensional Euclidean critical points with non-abelian symmetry, the first thing to try is the above procedure for making a two-dimensional quantum theory from a two-dimensional classical theory. This procedure is indeed fairly general. One can easily construct a set of a appropriate projectors $Q_i$ in a lattice model without local constraints. However, the caveat “without constraints” is quite important: many interesting classical lattice models have local constraints like in the dimer model and in the six- and eight-vertex models. This means that one must find a flip operator which respects the constraints. There is no guarantee that there is a set of flips which are both local and ergodic. This is particularly apparent in models where the two-dimensional degrees of freedom are closed loops. At a critical point, arbitrarily-long loops are an important part of the configuration space, so one must construct some sort of flip which still acts non-trivially but locally on these loops. In the continuum limit, it is easy to imagine such flips, but it is not always obvious how to make them act consistently on the lattice. Nevertheless, the Rokhsar-Kivelson Hamiltonian and the quantum eight-vertex model provide examples of models where one can solve this problem.
For field theories, the issue is similar. If a theory has an action $S_{2d}$ and one ignores constraints such as gauge invariance, one can easily find projection operators $Q(x)$ which annihilate the state weighted by $e^{-S_{2d}}$. For a single field $\varphi$, we have $$Q= \frac{\delta}{\delta\varphi} + \frac{\delta S_{2d}}{\delta{\varphi}}
\label{Qdef}$$ and the Hamiltonian is defined via Eq. (\[eq:HQQ\]). This $Q$ is hardly unique: one can multiply it on the left by any operator without changing the ground state. It is important to note that this construction is not always easy to implement. Many interesting two-dimensional field theories do not have a simple-to-describe action: they are usually defined instead in terms of representations of an infinite-dimensional symmetry algebra such as the Virasoro algebra, arising from conformal invariance \[\]. For example, the action of a $G/H$ coset model ([*e.g.*]{} a conformal minimal model) is written as a gauged WZW model such as we will discuss below; one must then worry about the constraints of gauge invariance.
A point we wish to emphasize, however, is that even if one succeeds in constructing such a Hamiltonian, the $2+1$-dimensional theory may be qualitatively very different from the classical two-dimensional theory. For example, the classical theory may be critical and have algebraically-decaying correlators, but the quantum theory may have exponentially-decaying correlators in the ground state. The reason is that quantum-ground-state correlators are weighted by $|\Psi_0|^2$, not $\Psi_0$. For the models discussed above, this has effects which are easy to take account of: it is the reason that the axes in figure \[eight\] are labeled by $c^2$ and $d^2$ instead of $c$ and $d$. If however $\Psi_0$ has a complex part, the effects can be much more dramatic.
Let us first discuss the simplest two-dimensional classical field theories with a non-trivial non-abelian symmetry. The $G$ principal chiral model is written in terms of a unitary matrix field $g$ taking values in some simple Lie group $G$, with action $$S_{PCM} = -\frac{1}{8\pi u}\int d^2 x \, \hbox{Tr}\left[
g^{-1}\partial^a g \, g^{-1}\partial_a g\right] \ .
\label{spcm}$$ This action has a global $G_L \times G_R$ symmetry under $g\to L g
R^\dagger$, where $L$ and $R$ are elements of separate groups dubbed $G_L$ and $G_R$. This model looks critical: the coupling constant $u$ is naively dimensionless. However, in two dimensions, the one-loop beta function is proportional to the curvature of the target-space manifold $G$. Simple Lie groups $G$ are curved: for example, $SU(2)$ as a manifold is isomorphic to a three-sphere. Moreover, one finds that the trivial fixed point at $u=0$ is unstable, and the model has a finite correlation length proportional to an exponential in $1/u$. Correlators decay exponentially, not algebraically. The phenomenon where a naively dimensionless coupling becomes dimensionful due to loop corrections is known in the particle physics literature as dimensional transmutation; it is familiar in condensed-matter physics in the Kondo problem and the Hubbard model at half-filling. Another reason a gap should appear is that the Mermin-Wagner-Coleman theorem does not allow Goldstone bosons in two Euclidean dimensions. One would obtain $S_{PCM}$ in a theory as the low-energy limit of a fermionic theory where one attempts to spontaneously break the chiral symmetry $G_L \times G_R$ to its diagonal $G_D$ subgroup. The would-be Goldstone bosons would take values on the manifold $G_L \times G_R
/G_D \cong G$. This satisfies the theorem by having the low-energy excitations get a mass and restoring the symmetry to the full $G_L
\times G_R$.
A critical theory with non-abelian symmetry $G$ does occur when an extra term, the Wess-Zumino term, is added to $S_{PCM}$. This term is easiest to write in a three-dimensional space $M$ which has the two-dimensional space of interest as a boundary; we assume the two-dimensional space has no boundary. It is $$\Gamma(g) = \frac{1}{12\pi} \int_M d^3 x\ \epsilon^{\alpha\mu\nu}
\hbox{Tr}\left[g^{-1}\partial_\alpha g\, g^{-1}\partial_\mu g\,
g^{-1}\partial_\nu g\right]
\label{swz}$$ One finds that different ways of extending space to three dimensions result only in changing $\Gamma(g)$ by $2\pi$ times an integer. Thus we can add $ik\Gamma(g)$ to $S_{PCM}$ for any integer $k$. This model has a critical point at $u = 1/|k|$ \[\]. This theory, with action $$I(g) = S_{PCM}(g,u=1/|k|) - ik\Gamma(g) \ ,
\label{swzw}$$ has conformal invariance and is known as the $G_k$ Wess-Zumino-Witten (WZW) model. A great deal of information is known about the WZW model; the most important for our current purposes is that the correlators decay algebraically, as they must at a critical point.
Since the WZW model has an explicit action, it is easy to use the trick in Eq. (\[Qdef\]) to find a two-dimensional Hamiltonian which has a ground-state wave functional with amplitudes given by the WZW action, $$\Psi_0[g] = e^{-I(g)}.
\label{psiwzw}$$ The WZW model is critical, so one might guess that the quantum version with this wave function would also be critical, in the fashion of the theories in the previous sections. However, it is not. The reason is that the Wess-Zumino term has an $i$ in front of it, so $$|\Psi_0[g]|^2 = e^{-2S_{PCM}(g,u=1/|k|)}$$ Thus equal-time correlators in the ground state are weighted with the action of a theory with a finite correlation length, the principal chiral model. These correlators decay exponentially.
We do note that when the level $k=1$, one can use an alternate method to find the 2d WZW correlators in the ground-state of a quantum theory. All the fields of $SU(N)_1$ can be written in terms of $N-1$ free bosons \[\]. The combination of the corresponding quantum Lifshitz theories will then have the $SU(N)_1$ correlators in its ground state. Likewise, $O(N)_1$ and $SU(2)_2$ can be written in terms of free Majorana fermions. Since the different bosons or fermions here do not interact, the physics is the same as that discussed earlier; in particular the model is essentially free, and the excitations are abelian.
Yang-Mills, Chern-Simons, and Wess-Zumino-Witten {#sec:ymcswzw}
------------------------------------------------
Since it is not possible to have a wave functional with $|\Psi[g]|^2
=e^{-I[g]}$, we will need to work harder to find how the physics of the WZW model can arise in $2+1$ dimensions. A connection between the WZW model and the wave functions of $2+1$-dimensional Chern-Simons gauge theory has long been known \[\]. Chern-Simons theory without any matter fields is a topological field theory: the physical states are Wilson and Polyakov loops, and their correlation functions do not depend on distance but only on topological properties of the loops. These correlators can be expressed in terms of the Verlinde numbers, which also describe the (chiral) fusion rules of the WZW model \[\]. These topological field theories have been discussed in detail in many places. One discussion of issues closely related to those of interest here is the recent work of Freedman [*et al.*]{} \[\]; we will discuss the relation of our results to this paper in the next subsection.
In this subsection we will describe a wave functional which after squaring and averaging over fields indeed yields the WZW partition function. We will show that this wave functional is the ground state of a 2+1-dimensional gauge theory. This all sounds like exactly what we want for a quantum critical point, but in a gauge theory, physical observables must be gauge invariant. Gauge-invariant states will turn out to have a gap, and gauge-invariant correlators in the ground state will be exponentially decaying. Introducing a gauge field therefore will not yield a quantum critical point with non-abelian symmetry. It will yield, however, something just as interesting: an explicit Hamiltonian for a theory in a topological phase.
Before discussing construction of these wave functionals, which was done in great detail by Witten \[\] and whose construction we will follow here, it is worthwhile to discuss first its physical meaning and to emphasize some well-known facts of Chern-Simons gauge theories \[\]. The ground state wave function of any field theory can be viewed as the quantum mechanical amplitude of some arbitrary state into the vacuum (or ground state), [*i.e.*]{} what we normally call the vacuum state in a given representation. As such this amplitude is the functional integral of the quantum field theory on an open manifold bounded by the initial time surface. However, the partition function of a topological gauge theory such as Chern-Simons is only gauge invariant on a manifold without boundary. Thus, the ground state wave function has a gauge anomaly. Nevertheless, if the wave function is used to compute expectation values of [*gauge-invariant observables*]{}, the result is gauge invariant. The reason is that the computation of expectation values (as well as all inner products) involves the conjugate wave function, [*i.e.*]{} the amplitude to evolve from the vacuum state in the remote past into the chosen state at the fixed time surface. Consequently, the gauge anomaly cancels in the computation of the expectation values of gauge-invariant operators. In this way, at a formal level, the computation of expectation values leads to a formally “doubled" theory even though the number of degrees of freedom has not changed. In subsection \[sec:doubled\] we will discuss a physically doubled theory which has a formal relation with what we do in this Subsection. There is an extensive literature on the technical issues involved in this problem: the relevant discussion for the analysis done in this paper can be found in Refs. \[\].
We should note here that the gauge anomaly of the wave function is formally (mathematically) analogous to the gauge anomaly of Chern-Simons theory on manifold with a spatial edge. In the latter, the anomaly is physical: to cancel the anomaly the gauge-invariant theory must include physical degrees of freedom residing at this $1+1$-dimensional boundary. In the incompressible fractional quantum Hall fluid, this results in physical edge states \[\].
Another important fact is that, since Chern-Simons gauge theory is a topological field theory, the partition function on a closed space-time manifold is independent of the metric. However, the boundary of the manifold and the choice of polarization break the general coordinate invariance of Chern-Simons theory. In particular, the choice of holomorphic polarization induces a conformal structure in the wave functional \[\] which is absent in other polarizations \[\]. We will see that the conformal wave function arises naturally in a specific theory, the strong-coupling limit of the Yang-Mills Chern-Simons theory.
The wave functional of interest was discussed in depth by Witten \[\]. It involves the WZW field $g$ coupled to a gauge field $A_i$ taking values in the Lie algebra of $G$. It is most convenient to write the gauge field in complex coordinates $A_z\equiv (A_1 -i A_2)/2$ and $A_{\overline z} \equiv (A_1+ iA_2)/2$. When the symmetry $G_R$ of the WZW model is gauged, the closest thing to a gauge-invariant action is $$I(g,A) = I(g) + \frac{k}{4\pi} \int d^2z\, \hbox{Tr}\left[2A_{\overline z}
g^{-1} \partial_z g\ -\ A_{\overline z} A_z\right]
\label{iga}$$ This action is not gauge invariant; the subgroup $G_R$ of $G_L\times G_R$ is anomalous \[\]. Indeed, the fields under gauge transformations $U(z,\overline{z})$ in $G_R$ as $$g\to gU\ ,\qquad A_i \to U^{-1} A_i U + U^{-1}\partial_i U$$ so that the action $I(g,A)$ transforms as $$I(g,A)\to I(g,A) + \frac{k}{4\pi} \int d^2 z\, \hbox{Tr}\left[
A_{z} U\partial_{\overline z} U^{-1}
- A_{\overline z} U\partial_z U^{-1}\right]
- i k \Gamma(U).
\label{igagauge}$$ Note, however, that the variation of $I(g,A)$ under gauge transformations is independent of $g$. If we were to simply square a wave functional proportional to $e^{-I(g,A)}$, the WZW term in the action would cancel and we would be back a gauged principal chiral model. Instead, we define a wave functional depending only on $A$ by doing the path integral over $g$: $$\Psi[A] \equiv \int [Dg] e^{-I(g,A)}.
\label{psiA}$$
This wave functional does yield the full WZW partition function after squaring and integrating over $A$ \[\]. We define this integrated $|\Psi|^2$ as $$|\Psi|^2\equiv
\frac{1}{\hbox{vol }\hat{G}} \int [DA] \left|\Psi[A]\right|^2 ,$$ where we have divided the measure by the volume of the gauge group $\hat G$ because it follows from Eq. (\[igagauge\]) that even though $\Psi[A]$ is not gauge-invariant, $\overline{\Psi[A]}\Psi[A]$ is. Substituting the definition Eq. (\[psiA\]) gives $$\begin{aligned}
|\Psi|^2 &=& \frac{1}{\hbox{vol }\hat{G}} \int [DA][Dg][Dh^{-1}] \;
e^{\displaystyle{-I(g,A) - (I(h^{-1},A))^*}}\ ,\\
I(g,A) + (I(h^{-1},A))^*&=&
I(g)+I(h)+\frac{k}{2\pi}\int d^2 z\, \hbox{Tr}
\left[A_{\overline z} g^{-1}\partial_z g -
A_z \partial_{\overline z} h\cdot h^{-1} - A_{\overline z} A_z\right] \end{aligned}$$ Note that we have defined $A$ so that $(A_{z})^\dagger = -A_{\overline z}$, [*i.e.*]{} covariant derivatives have no $i$ in them. We have also used the easily-proven fact that $I(h)=I^*(h^{-1})$. Since the integral over $A$ is Gaussian, it can easily be done, giving $$|\Psi|^2 = \frac{1}{\hbox{vol }\hat{G}} \int
[Dg][Dh^{-1}] \exp\left(-I(g)-I(h)+
\frac{k}{2\pi}\int d^2 z \, \hbox{Tr}\left[g^{-1}\partial_z g\,
\partial_{\overline{z}} h\cdot h^{-1}\right]\right) \ .$$ This can be simplified by using the Polyakov-Wiegmann formula \[\] $$I(gh) = I(g) + I(h) - \frac{k}{2\pi}\int d^2 z \,
\hbox{Tr}\left[g^{-1}\partial_z g\,
\partial_{\overline{z}} h\cdot h^{-1}\right].
\label{pw}$$ Thus the integrand depends only on the product $f\equiv gh$, so we can change variables to $f$, which cancels the volume of the gauge group. This yields the final expression for the normalization \[\] $$|\Psi|^2= \int [Df] e^{-I(f)} \ .
\label{psisquared}$$ Thus, remarkably, the integrated square of the wave functional $\Psi[A]$ ends up giving the full WZW path integral, including the imaginary piece: because of the path integral, the right-hand-side of Eq. (\[psisquared\]) is positive and real as it must be.
To reemphasize a point made earlier, (\[psisquared\]) does not mean we have now found a $2+1$-dimensional quantum critical point whose equal-time correlators are those of the WZW model. We must first find a Hamiltonian which has $\Psi[A]$ in Eq. (\[psiA\]) as its ground state. Once having found that, the physically-relevant operators are only those satisfying the proper gauge-invariance properties and regularization. After having done so, we will end up seeing that despite sharing some key properties with the WZW model, the $2+1$-dimensional model is gapped and has exponentially-decaying correlators.
To find such a Hamiltonian, let us go back to the definitions Eq. (\[iga\]) and Eq. (\[psiA\]) of the wave functional $\Psi[A]$. Because the only dependence on $A_z$ is through the quadratic term, we have $$\left(\frac{\delta}{\delta A_z} - \frac{k}{4\pi} A_{\overline z}\right)
\Psi[A]=0.
\label{wit1}$$ With a little more work \[\], one also can show that $$\left(D_{\overline z} \frac{\delta}{\delta A_{\overline z}} +
\frac{k}{4\pi} D_{\overline z} A_{z}
-\frac{k}{2\pi} F_{\overline{z} z}\right)\Psi[A] = 0,
\label{wit2}$$ where we have define the covariant derivatives via $D_ig = \partial_i g
- gA_i$ and $D_iA_j = \partial_i A_j + [A_i,A_j]$. The field strength is defined as $F_{\overline{z}z} = \partial_{\overline z} A_z -
\partial_z A_{\overline z} + [A_{\overline z}, A_z]$.
We would therefore like to find a Hamiltonian with a ground-state wave function satisfying Eq. (\[wit1\]) and Eq. (\[wit2\]). As noted in Ref. \[\], these equations arise in the canonical quantization of Chern-Simons theory \[\], so this suggests we look there. The precise Hamiltonian turns out to be given by the strong-coupling limit of Yang-Mills theory with a Chern-Simons term \[\]. In the action, we have a gauge field $A_\mu$ taking values in the Lie algebra of $G$; here $\mu=0,1,2$ and $A_\mu$ depends on space and time. The action for the strong-coupling limit of Yang-Mills theory on a three-manifold $M$ includes only the electric-field term, namely $$S_{SC} = \frac{1}{2e^2}\int_M\,\hbox{Tr}\left[F_{0i}F^{0i}\right]\ .
\label{ssc}$$ This term is not Lorentz-invariant, but does preserve two-dimensional rotational symmetry. The Chern-Simons term is $$S_{CS} = \frac{k}{4\pi} \int_M \epsilon^{\mu\nu\alpha}
\hbox{Tr} \left[A_\mu\partial_{\nu} A_\alpha + \frac{2}{3}
A_\mu A_\nu A_{\alpha} \right]\ .
\label{scs}$$ Under gauge transformations $U(x)$ belonging to $G$, the integrand in $S_{SC}$ is invariant, but the integrand in $S_{CS}$ is not: $$S_{CS} \to S_{CS} + \frac{k}{4\pi} \int_M
\,\epsilon^{\mu\nu\alpha}\, \hbox{Tr} \left[
\partial_{\mu}\left(A_\alpha + U
\partial_\nu U^{-1}\right)\right] + k \Gamma(U).
\label{csgauge}$$ For the manifold $M$ a three-sphere, the latter term turns out to be the winding number of the gauge transformation $U(x)$, and is an integer times $2\pi k$ \[\]. The Chern-Simons term is gauge-invariant if $M$ has no boundary, and $k$ is an integer. If $M$ has a spatial boundary, one must include massless chiral fermions on the edge to restore gauge invariance, giving for example the famous edge modes in the fractional quantum Hall effect \[\].
The Hamiltonian with (\[psiA\]) as a ground-state wave functional comes from canonically quantizing the theory with action $$S= S_{CS} + S_{SC},
\label{scssc}$$ following, for instance, Refs. \[\]. Since this is a fairly standard computation, we will be brief here. The gauge invariance allows us to fix temporal gauge $A_0=0$, so that the degrees of freedom are the gauge fields $A_1$ and $A_2$. Their canonical momenta are $$\Pi_i = \frac{1}{e^2}F_{0i} + \frac{k}{8\pi} \epsilon_{ij}A_j \ .
\label{canmon}$$ The Chern-Simons term contributes nothing to the Hamiltonian, because all the terms are first-order in time derivatives. The classical Hamiltonian in the strongly-coupled limit is therefore $$H= \frac{1}{e^2}\int d^2 x\, \hbox{Tr}\left[(F_{0i})^2\right]
={e^2}\int d^2 x\, \hbox{Tr}\left[\left(\Pi_i - \frac{k}{8\pi}
\epsilon_{ij}A_j\right)^2\right].$$ Expanding $A$ and $F$ in terms of generators $T^a$ of the Lie algebra of $G$, we impose the canonical commutation relations $$[A_i^a(\vec{x}),\Pi_j^b(\vec{y})] = i \delta_{ij} \delta^{ab}
\delta^{(2)}(\vec{x}-\vec{y})\ .
\label{ccrA}$$ One of the effects of the Chern-Simons term is that $A_1$ and $A_2$ do not commute. In the Schrödinger picture, $\Pi_j$ is given by $$\Pi_j = -i \frac{\delta}{\delta A_j}
\label{pij}$$ operating on the wave functionals.
The Hamiltonian of this theory in the Schrödinger picture has a very simple form. Because $F_{0i}$ includes $\Pi_i$, one has the usual ordering ambiguity in the quantum Hamiltonian. We define $$E = \frac{i}{e^2}\left(F_{01} + i F_{02}\right) =
\frac{\delta}{\delta A_z} - \frac{k}{4\pi} A_{\overline{z}}
\label{Edef}$$ so that $$[E^a(\vec{x}), (E^b(\vec{y}))^\dagger] = \frac{k}{2\pi}\delta^{ab}
\delta^{(2)}(\vec{x}-\vec{y}).
\label{ccrE}$$ The operators $E$ and $E^\dagger$ are like annihilation and creation operators. We then normal-order the Hamiltonian as in section \[sec:scale-wf\], subtracting the vacuum energy to give $$H=e^2 \int d^2 x\, \hbox{Tr}\left[ E^\dagger(\vec{x}) E(\vec{x})\right] \ .
\label{HEE}$$ Thus the Hamiltonian of strongly-coupled Yang-Mills theory with a Chern-Simons term is precisely of the form Eq. (\[eq:HQQ\]), like all the others discussed in this paper. Moreover, the annihilation relation $E\Psi[A] =0$ here is identical to the relation Eq.(\[wit1\]). We thus have found an explicit Hamiltonian whose ground state obeys the first of the two relations satisfied by $\Psi[A]$ above.
Finding the Hamiltonian alone does not complete the canonical quantization of gauge theories with a Chern-Simons term. Fixing the gauge $A_0=0$ still allows time-independent gauge transformations. Moreover, there is no time derivative $\dot{A}_0$ in Eq. (\[scssc\]), so $A_0$ should be viewed as a Lagrange multiplier which results in a constraint. When we are canonically quantizing the theory in $A_0=0$ gauge, this constraint is implemented on the wave functions. Specifically, one has $$\left( D_{\overline z} \frac{\delta}{\delta A_{\overline z}} +
D_{z} \frac{\delta}{\delta A_{z}}
+ \frac{k}{4\pi} \partial_z A_{\overline z}
- \frac{k}{4\pi} \partial_{\overline z} A_{z}\right) \Psi[A]=0.
\label{gausslaw}$$ This operator is the generator of time-independent gauge transformations on $A_z$ and $A_{\overline z}$, so this condition amounts to requiring that the wave function be invariant under such transformations. Another way of viewing this constraint is as requiring that the wave functionals obey the non-abelian version of Gauss’ Law. Requiring Gauss’ Law along with Eq. (\[wit1\]) yields Eq. (\[wit2\]), the other desired relation: adding Eq. (\[wit1\]) and Eq. (\[wit2\]) together yields Eq. (\[gausslaw\]). Moreover, the fact (\[igagauge\]) that $\Psi[A]$ is not gauge invariant is precisely the effect of the fact (\[csgauge\]) that the Chern-Simons action is not gauge invariant when there are boundaries \[\]; one can think of the constant-time slice required to define a wave functional as a boundary in space time.
We have seen that the wave functional Eq. (\[psiA\]) indeed describes a zero-energy ground state of the $2+1$-dimensional theory with action Eq. (\[scssc\]). We now need to understand the correlators in the ground state and the excited states. Luckily, the former issue has been studied in Refs. \[\], and the latter in Refs. \[\]. We will show in the next subsection that the ground-state are those of a topological field theory, and that one can prove there is a gap in the spectrum.
The doubled theory {#sec:doubled}
------------------
It is both convenient and physically important to study the wave function of the “doubled” theory, where we have two gauge fields $A$ and $B$. The three-dimensional action is the sum of an action of the form of Eq. (\[scssc\]) for both fields, where the Chern-Simons term for the field $A$ has coefficient $k$, while that for $B$ has coefficient $-k$. As opposed to Chern-Simons theory with a single field, the doubled theory is also invariant under time reversal and parity, if the fields $A$ and $B$ are exchanged under these transformations[^15]. The two fields are not coupled in the action, so formally the wave functional of the doubled theory factorizes: $$\chi[A,B] \propto
\Psi[A]\overline{\Psi[B]}$$ There is also a compelling technical reason to study the doubled theory. We will show that the doubled theory amounts to gauging a non-anomalous symmetry. In particular, after integrating out one of the gauge fields (a simple Gaussian integration), the wave functional is gauge invariant in the remaining field. The theory can be quantized consistently; effectively the regularization (the measure of the path integral) couples the two copies. Thus in spite of this factorization of the wave function, the operators that create the physical states are made of operators acting on each sector, carefully glued together to satisfy the requirements of gauge invariance.
Canonically quantizing the doubled theory is straightforward, since the action splits into decoupled pieces. The ground-state wave functional of the doubled theory can be written as \[\] $$\chi[A,B] \equiv \int [Dg] e^{-I(g,A,B)}
\label{eq:chiAB}$$ where $$I(g,A,B) \equiv I(g)
+ \frac{k}{4\pi} \int d^2z \hbox{Tr}\left[2A_{\overline z}
g^{-1} \partial_z g + 2B_{z}
g \partial_{\overline z} g^{-1} - A_{\overline z} A_z
- B_{\overline z} B_z + 2 B_z g A_{\overline z} g^{-1}
\right]
\label{igab}$$ One can prove that $\chi[A,B]\propto \Psi[A]\overline{\Psi [B]}$ indirectly by showing that $\chi[A,B]$ satisfies both (\[wit1\]) and (\[wit2\]) for $A$, and the conjugate equations for $B$. Directly, we prove this by first noting that the Polyakov-Wiegmann identity (\[pw\]) yields $$I(g,A) + (I(h,B))^* = I(h^{-1}g,A,B) - \frac{k}{2\pi}
\int d^2 z\, \hbox{Tr}\left[
(gD_{\overline z} g^{-1}) (hD_z h^{-1})\right]\ .$$ where the covariant derivatives are $D_{\overline z} g^{-1} = \partial_{\overline z} g^{-1} +
A_{\overline z} g^{-1}$ and $D_{z} h^{-1} = \partial_{z} h^{-1} +
B_{z} h^{-1}$. In the path integral we can split apart the last term by introducing an auxiliary gauge field $C$: $$\begin{aligned}
\Psi[A]\overline{\Psi[B]} &=& \int [Dg][Dh^{-1}] e^{-I(g,A) -
(I(h,B))^*}\\
&=& \int [Dg][Dh^{-1}][DC] \exp\Big(-I(h^{-1}g,A,B)\\
&&\qquad\qquad -
\frac{k}{2\pi} \int d^2 z\, \hbox{Tr}\left[
\alpha C_{ z} (gD_{\overline z} g^{-1}) + \alpha C_{\overline z}
(hD_z h^{-1}) + {\alpha}^2 C_z C_{\overline z} \right]\Big)\ , \end{aligned}$$ where $\alpha$ is a (small) coupling constant. The terms linear in $C$ can be removed by redefining $g$ and $h$. Namely, under $g\to g +
\delta g$ for small $\delta g$, the action $$I(g,A) \to I(g,A) + \frac{k}{2\pi} \int d^2 z\, \hbox{Tr}\left[
gD_{\overline z} g^{-1} \partial_{z}(\delta g\, g^{-1})\right]\ .$$ Thus if we redefine the fields $g$ and $h$ so that $\delta g$ and $\delta h^{-1}$ obey $\alpha C_z = \partial_z(\delta g\, g^{-1})$ and $\alpha C_{\overline z} = -\partial_{\overline z}(h\, \delta h^{-1})$, the mixed terms cancel and the field $C$ decouples. The integrand then only depends on the combination $h^{-1}g$, so changing variables to $f=h^{-1}g$ yields $\chi[A,B]$, up to a factor of the volume of the gauge group: $$\Psi[A]\overline{\Psi[B]} = \int[Dg][Dh^{-1}] e^{-I(h^{-1}g,A,B)} =
\hbox{vol }\hat{G}\ \chi[A,B].$$
To understand this wave functional $\chi[A,B]$, we compute its integrated norm. One can easily do one of the two functional integrals over the gauge fields in the same manner as before, yielding \[\] $$\int [DB] \left|\chi[A,B]\right|^2 = \int[Dg] e^{-I_{G/G}(g,A)}
\label{chisquared}$$ where $I_{G/G}$ is the action of the $G/G$ gauged WZW model, namely $$I_{G/G}(g,A) =
I(g) + \frac{k}{2\pi} \int d^2z\, \hbox{Tr}\left[A_{\overline z}
g^{-1} \partial_z g + A_{z}
g \partial_{\overline z} g^{-1} -
A_{\overline z} A_z + A_z g A_{\overline z} g^{-1}
\right]\ .
\label{igg}$$ As opposed to $I(g,A)$ in Eq. (\[iga\]), this action is gauge invariant, because one is gauging the full $G_L\times
G_R$ symmetry of the WZW model. This means the path integral of the doubled theory is well defined and free of anomalies, as well as parity-invariant.
Let us compare the doubled theory to the undoubled one. For the undoubled theory, we found that the effective partition function $|\Psi|^2$, the ground-state wave functional squared and integrated over the (one) gauge field, was that of the WZW model. However, constructing excited states and operators directly in the undoubled theory appears to be problematic due to the gauge anomaly discussed above. Although it would seem that [*a priori*]{} one could not require that operators be gauge invariant if the ground-state wave functional itself is not, as we emphasized in Subsection \[sec:ymcswzw\], from general considerations of Chern-Simons theory we know that the physical observables [*are*]{} gauge invariant, that [*only*]{} gauge invariant observables must be considered, and that their expectation values are free of any anomalies. However, while this is apparent in the path-integral construction of the quantum theory, it is not so apparent if one is to use the wave function, [*i.e.*]{} in terms of a chiral Euclidean WZW model. As we discussed above, at this level one is led to introduce a formal “doubled" theory even for the undoubled theory. Thus, although the wave function itself factorizes the physical states cannot be constructed in terms of arbitrary factors from each sector of the doubled theory. Hence, the requirement of gauge invariance can spoil the apparent factorization suggested by the wave function $\chi[A,B]$ of Eq. (\[eq:chiAB\]).
The formally correct way to define the correlators in the undoubled theory is in the doubled theory: one can always introduce another field $B$ and then integrate it out. The reason for doing this is that the doubled theory can be properly regulated, because (\[chisquared\]) is gauge invariant. All the formal manipulations done above are well founded, because the path integral can be defined properly. In contrast in a truly doubled theory, the gauge-invariant physical observables couple to both fields which must then be regarded as genuine degrees of freedom. In contrast, in the undoubled theory, the additional degrees of freedom are a formal device used to regulate the theory[^16]. This means that correlators in both theories are those of the $G/G$ gauged WZW model, which is anomaly-free. These correlators have been studied in great detail \[\]. In particular, careful discussions of the proper regularization of this theory can be found in these papers. As common in non-abelian gauge theories, to properly do these path integrals, one needs to introduce fermionic ghosts. One then finds a BRST charge $Q_{BRST}$ obeying $(Q_{BRST})^2=0$; physical states are annihilated by it. We can thus consistently demand that states and physical operators be gauge invariant.
A relation between the $G/G$ gauged WZW model and topological field theory was conjectured in \[\] and proven in \[\]. In particular, it was shown that once we integrate over $A$ as well, the resulting partition function $|\chi|^2$ is independent of the metric of two-dimensional space. The correlators in the ground state therefore are independent of distance, and given by a topological field theory. A topological field theory is obtained by studying the states which are annihilated by $Q_{BRST}$ but not given by $Q_{BRST}$ acting on something else. In mathematical language, the physical states of the topological theory are given by the cohomology of $Q_{BRST}$. It was derived directly in \[\] that the correlators of the $G/G$ topological field theory are given in terms of the Verlinde numbers \[\], which give the dimensions of conformal blocks in the ordinary WZW model.
Going to the doubled theory has therefore allowed us to not only avoid technical problems, but also to prove that the ground state correlators are those of a topological field theory. We should stress once again that the doubled theory used here has twice as many degrees of freedom as the undoubled theory. It is a [*physically distinct*]{} time-reversal invariant theory, unlike the undoubled theory which breaks time-reversal symmetry. This physical difference is apparent from the construction of its observables: it has twice as many “anyons", which come in time-reversed pairs. On a torus, the Wilson loops of the corresponding topological field theory wind around both cycles \[\]. The weights of the loops can be defined locally, and one can see how the non-abelian statistics arise \[\]. The results of this section show how this precise topological field theory arises as the ground state of a specific Hamiltonian. We note in addition that if one specializes these results to an abelian gauge field, the topological field theory obtained is known as a (two-dimensional) BF theory \[,\]. Recently, Freedman, Nayak and Shtengel \[\] have discussed (reasonably local) lattice models of interacting fermions and bosons, which they argue have topological ground states with some of the topological algebraic structure of the doubled non-abelian Chern-Simons theories. It is reasonable to expect that the universal long-distance structure of the wave functions of these topological ground states, at least deep in the topologically-ordered phase, has the same structure of the wave functions we discussed in this section.
Since we have an explicit Hamiltonian Eq. (\[HEE\]) (plus the analogous term for the $B$ gauge field), we can go beyond the topological field theory describing the ground state. Because of the commutation relations (\[ccrE\]), $E^\dagger$ acts like a creation operator. However, in a non-abelian gauge theory, $E$ and $E^\dagger$ are not gauge invariant, so the appropriate states are slightly more complicated than just $E^\dagger \Psi[A]$. The simplest candidate is \[\] $$\psi^a(x,A,B) = E^{\dagger b}(x)
\int[Dg]\,\hbox{Tr}\left[T^a g(x) T^b g^{-1}(x)\right]
e^{-I(g,A,B)} \ .$$ Once this amplitude is squared and $B$ is integrated over, one obtains a gauge-invariant probability. In the $2+1$-dimensional picture, one can think of one of these states as a Polyakov loop, which intersects two-dimensional space at a single point $x$. The commutation relation Eq. (\[ccrE\]) shows that this is indeed an eigenstate with a gap proportional to $k$. A more thorough treatment, taking into account the measure of the path integral, shows that the gap is shifted to $e^2(k+2c_A)/(4\pi)$, where $c_A$ is the quadratic Casimir of the adjoint representation of the Lie algebra of $G$ \[\]. For our purposes, the important point is that there is indeed a gap.
We have thus seen that correlators of Wilson loops in the ground state of strongly-coupled Yang-Mills theory are given by a topological field theory. We have also seen that the theory has a gap, so it is indeed in a topological phase. The last thing we would like to discuss is the ground-state correlators for operators not in the topological theory. In other words, we wish to consider the full set of physical operators (i.e. those annihilated by $Q_{BRST}$), including those which are given by $Q_{BRST}$ acting on something else. Because our Hamiltonian involves only the electric and not the magnetic field, it is not Lorentz invariant. Thus the existence of a gap does not immediately require that correlators decay exponentially. However, in this theory, they do, as implied by the results of \[\]. We noted above that proper quantization of gauged WZW models requires introducing fermionic ghosts, and a fermionic operator $Q_{BRST}$. A fermionic symmetry suggests the appearance of supersymmetry, and indeed the $G/G$ topological field theory can also be obtained from a supersymmetric field theory where the supersymmetry charge is “twisted” into $Q_{BRST}$ \[\]. The supersymmetric field theory also describes the correlators of the full theory, not just the topological subsector. For the case where $G=SU(N)$ at level $k$, the appropriate two-dimensional field theory is a (twisted) supersymmetric sigma model, where the bosonic fields take values on the “Grassmannian” manifold $U(N+k)/(U(N)\times
U(k))$ \[\]. It is likely that the other simple Lie groups end up giving supersymmetric sigma models on the analogous Grassmannians.
This means we have now come full circle! We started without gauge fields, and found that the equal-time correlators are those of the principal chiral model, a two-dimensional sigma model with a curved target space. These correlators are exponentially decaying. We then introduced gauge fields, in the hope of finding a quantum critical point. After this lengthy discussion, we have ended up showing the ground state is described by a supersymmetric two-dimensional sigma model with a curved target space, the Grassmannian. This means that the correlators here are exponentially decaying as well, and the theory is in a topological phase whose properties are encoded in the wave function $\chi[A,B]$.
We are grateful to M.P.A. Fisher, M. Freedman, S. Kivelson, C. Nayak, K. Shtengel, T. Senthil, S. Sondhi, M. Stone, A. Vishwanath and X.G. Wen for many illuminating conversations on quantum dimer models, topological phases and beyond. We are grateful to M. Stone for motivating us to keep track of factors of $i$. We also thank M. Freedman, C. Nayak and Z. Wang for organizing a very stimulating conference on topological order at the American Institute of Mathematics, during which part of this work was done. This work was supported in part by the National Science Foundation through the grants NSF-DMR-01-32990 at the University of Illinois and NSF-DMR-0104799 at the University of Virginia. The work of P.F. was also supported by the DOE under grant DEFG02-97ER41027.
Operators of the quantum Lifshitz field theory {#app:gaussian}
==============================================
In addition to the products of field operators $\varphi(\vec x)$, in what follows we will be interested in two types of local operators: charge and vortex operators. The charge operators are $${\mathcal O}_n(\vec x)=e^{\displaystyle{-i n\; \varphi(\vec x)}}
\label{eq:charge}$$ where $n \in {\mathbb Z}$. This operator creates a boson coherent-state which we will refer to as a [*charge*]{} $n$ excitation. The vortex operators are $${\tilde{\mathcal O}}_m(\vec x)=e^{\displaystyle{i \int d^2z \; \varphi(\vec z)\; \Pi(\vec z)}}
\label{eq:vortex}$$ where $$\varphi(\vec z)=m \arg(\vec z-\vec x)$$ where $0\leq \arg(\vec z-\vec x) \leq 2\pi$ is the argument of the vector $\vec z-\vec x$ (with a branch cut defined arbitrarily along the negative $x$ axis). The action of the operator ${\tilde{\mathcal O}}_m(\vec x)$ on an eigenstate of the field operator $\ket{[\varphi]}$ is simply a shift $$e^{\displaystyle{i \int d^2z \; \varphi(\vec z)\; \Pi(\vec z)}} \ket{[\varphi]}=
\ket{[\varphi(\vec x)-\varphi(\vec x)]}
\label{eq:gauge}$$ In other words, it amounts to s singular gauge transformation. Therefore, its action is equivalent to coupling the field $\varphi$ to a vector potential whose space components $\vec A$ satisfy $$\oint_\gamma d \vec z \cdot \vec A[\vec z]=2\pi m
\label{eq:circulation}$$ for all closed paths $\gamma$ which have the point $\vec x$ in their interior, and zero otherwise. In particular, the wave function of the state resulting from the action of the vortex operator on the ground state is: $$\Psi_m[\vec x]=\me{[\varphi]}{{\tilde{\mathcal O}}_m(\vec x)}{{\rm vac}}
=\frac{1}{\sqrt{\mathcal Z}}
e^{\displaystyle{-\frac{\kappa}{2} \int d^2z \; \left(\vec \nabla \varphi-\vec A\right)^2}}
\label{eq:vortex-wf}$$ where $\vec A$ is any vector field which satisfies Eq. (\[eq:circulation\]). The (equal-time) ground state expectation value of a product of vortex operators with magnetic charges $\{m_l\}$, [*i.e.*]{} the [*overlap*]{} of the state with k vortices at locations $\vec x_l$ and magnetic charge $m_l$ with the vortex-free ground state wave function, is therefore $$\me{{\rm vac}}{{\tilde{\mathcal O}}_{m_1}(\vec x_1)\ldots {\tilde{\mathcal O}}_{m_k}(\vec x_k)}{{\rm vac}}=
\frac{1}{{\mathcal Z}} \int {\mathcal D} \varphi \;
e^{\displaystyle{-\kappa \int d^2z \; \left(\vec \nabla \varphi-\vec A\right)^2}}
\label{eq:multivortex}$$ where ${\mathcal Z}$ is given by Eq. (\[eq:Zboson\]). The vector potential in Eq. (\[eq:multivortex\]) satisfies $$\varepsilon_{ij} \nabla_i A_j=2\pi \sum_{l=1}^k m_l \delta^2 (\vec z-\vec x_l)
\label{eq:vortex-config}$$ This result is equivalent to the expectation value of the vortex operator in the 2D classical $c=1$ compactified free bose field discussed extensively by Kadanoff \[\] (see also \[\]).
The boson propagator of this theory, in imaginary time $t$, is $$G(\vec x- {\vec x}^\prime, t-t^\prime)=\langle \varphi(\vec x, t) \; \varphi({\vec x}^\prime, t^\prime)\rangle
= \int \frac{d\omega}{2\pi}
\int \frac{d^2q}{(2\pi)^2}
\;
\frac{
e^{\displaystyle{i \omega (t-t^\prime)-i\vec q \cdot (\vec x -{\vec x}^\prime)}}
}{\omega^2+\kappa^2 \left({\vec q}^{\;2}\right)^2}
\label{eq:phiprop}$$ which has a short-distance logarithmic divergence. From now on we will use instead the regularized (subtracted) propagator $$G_{\rm reg}(\vec x,t) \equiv G(\vec x,t)-G(a,0)
=-\frac{1}{8\pi \kappa}\left[ \ln\left(\frac{\vert \vec x \vert^2}{a^2}\right)
+\Gamma\left(0,\frac{\vert \vec x \vert^2}
{4\kappa \vert t \vert}\right)\right]
\label{eq:regprop}$$ where $a$ is a short-distance cutoff and $\Gamma(0,z)$ is the incomplete Gamma function $$\Gamma(0,z)=\int_z^\infty \frac{ds}{s} \; e^{\displaystyle{-s}}
\label{eq:incomplete}$$ The regularized propagator has the asymptotic behaviors $$G_{\rm reg}(\vec x,t)=
\begin{cases}
-\displaystyle{\frac{1}{4\pi \kappa}} \ln\left(\displaystyle{\frac{\vert \vec x\vert}{a}}\right), & \textrm{for} \; \vert t \vert \to 0 \\
-\displaystyle{\frac{1}{8\pi \kappa}} \ln\left(\displaystyle{\frac{4 \kappa \vert t \vert}{a^2\gamma}}\right) , & \textrm{for} \; \vert \vec x \vert \to a
\end{cases}
\label{eq: asymptotic}$$ where $\ln \gamma=\mathbf{C}=0.577\ldots$ is the Euler constant.
The correlation functions of the charge operators are $$\langle {\mathcal O}_n(\vec x,t)^\dagger {\mathcal O}_n({\vec x}^{\; \prime},t^\prime)\rangle=
e^{\displaystyle{\; n^2 \; G_{\rm reg}(\vec x-{\vec x}^{\; \prime},t-t^\prime)}}
\label{eq:charge-corr}$$ At equal (imaginary) times, $|t-t^\prime|\to 0$, it behaves like $$\langle {\mathcal O}_n(\vec x,0)^\dagger {\mathcal O}_n({\vec x}^{\;\prime},0)\rangle=
\left(\displaystyle{\frac{a}{\vert \vec x-{\vec x}^{\; \prime} \vert}}\right)^{\displaystyle{\frac{ n^2}{4 \pi\kappa}}}
\label{eq:charge-eqtime}$$ which implies that the operator ${\mathcal O}_n$ has (spacial) scaling dimension $$\Delta_n=\frac{ n^2}{8 \pi \kappa}
\label{eq:dimension-electric}$$ For , its asymptotic behavior is instead given by $$\langle {\mathcal O}_n(\vec 0,t)^\dagger {\mathcal O}_n({\vec 0},t^\prime)\rangle=
\left(\displaystyle{\frac{a^2\gamma}{4\kappa \vert t-t^\prime\vert}}\right)^{\displaystyle{\frac{ n^2}{8\pi \kappa}}}
\label{eq:charge-auto}$$ This behavior is manifestly consistent with a dynamical critical exponent $z=2$.
It is straightforward to show by an explicit calculation of the overlap of Eq. , which is completely analogous to the classical vortex correlation functions of the 2D Gaussian model \[\], that the (spacial) scaling dimension of the vortex of magnetic charge $m$ is, $$\Delta_m=2\pi \kappa m^2
\label{eq:dimension-magnetic}$$
$U(1)$ Gauge-theory for the quantum six-vertex model {#app:gauge}
====================================================
In this Appendix, we describe the quantum six-vertex model in the language of gauge theory. The quantum eight-vertex model will be described in appendix \[app:z2gauge\]. We will follow closely the gauge theory description of the quantum dimer model, which is described in detail in \[\]. We note that this gauge theory is [*not*]{} simply an abelian version of the gauge theory discussed in section \[sec:YMCS\].
We define link variables $E_i (\vec x)$, where $\vec x$ labels the vertices and $i=1,2$ indicates the direction; the unit vector in direction $i$ is denoted by $\vec e_i$. The link variables are integer valued, and can be viewed as the eigenvalues of angular momentum operators which we will also denote by $E_i (\vec x)$. We assign the values $E_1=1$ and $E_2=1$ to the right and up going arrows respectively, while for the left and down going arrows we have $E_1=-1$ and $E_2=-1$.
(20,20)(0,0) (0,10)[(1,0)[20]{}]{} (10,0)[(0,1)[20]{}]{} (10,10)[(0,1)[6]{}]{} (10,10)[(0,-1)[6]{}]{} (0,10)[(1,0)[6]{}]{} (20,10)[(-1,0)[6]{}]{} (7,7)[$\vec x$]{} (-19,2)[$E_2(\vec x -\vec e_2)=-1$]{} (15,6)[$E_1(\vec x) = -1$]{} (11,15)[$E_2(\vec x) = 1$]{} (-17,12)[$E_1(\vec x - \vec e_1)=1$]{}
An example is given in figure \[econfig\]. In the Hamiltonian, we need a term which will lead to a restriction to those states in which all the $E_i(\vec x)$’s have values $\pm 1$. Such a term is $$H_{\rm E} = \frac{1}{k_1} \sum_{\vec x, i}
\left(
E_i^2 (\vec x)-1
\right)^2 \ ,$$ in the limit $k_1 \rightarrow 0$.
The six-vertex model also requires that the same number of arrows point in and out at a vertex. This is precisely a lattice version of Gauss’ law. Defining the lattice divergence as ${\Delta^{\!\! -}}_i E_i^{\vphantom{-}} (\vec x) \equiv
E_1(\vec x) - E_1(\vec x -\vec e_1) + E_2(\vec x) - E_2(\vec x -\vec e_2)$, one easily finds that only for the six vertices of type $a,b$ and $c$, we have $$\label{con6v}
{\Delta^{\!\! -}}_i E_i^{\vphantom{-}} (\vec x) = 0 \ .$$ The lattice differentiation ${\Delta^{\!\! -}}_i$ is defined by ${\Delta^{\!\! -}}_i f (\vec x) \equiv f(\vec x) - f(\vec x -\vec e_i)$. In the following, when dealing with sums over the plaquettes of the lattice, we will frequently use $\Delta_i f (\vec x) \equiv {\Delta^{\!\! +}}_i f (\vec x) \equiv
f(\vec x+ \vec e_i) - f(\vec x)$. Of course, the constraint has to commute with all the terms in the Hamiltonian. In the following, we will find that this is indeed the case.
The two main ingredients in the gauge theory description are the flip term which flips the flippable plaquettes (see below) and a potential term, which give a finite weight to (only) the flippable plaquettes. Flippable plaquettes are those which have both $n_d=\widetilde{n}_d =0$. Pictorially, a flippable plaquette here is one where the arrows around the plaquette point either all clockwise or all counterclockwise: $$\setlength{\unitlength}{.5 mm}
\begin{picture}(20,20)(0,0)
\put(5,0){\line(0,1){20}}
\put(15,0){\line(0,1){20}}
\put(0,5){\line(1,0){20}}
\put(0,15){\line(1,0){20}}
\put(5,5){\vector(1,0){6}}
\put(15,5){\vector(0,1){6}}
\put(15,15){\vector(-1,0){6}}
\put(5,15){\vector(0,-1){6}}
\end{picture}
\raisebox{4 mm}{$\Longleftrightarrow$}
\begin{picture}(20,20)(0,0)
\put(5,0){\line(0,1){20}}
\put(15,0){\line(0,1){20}}
\put(0,5){\line(1,0){20}}
\put(0,15){\line(1,0){20}}
\put(15,5){\vector(-1,0){6}}
\put(15,15){\vector(0,-1){6}}
\put(5,15){\vector(1,0){6}}
\put(5,5){\vector(0,1){6}}
\end{picture} \ .$$ In terms of the electric field, this can be written as $$\label{hpot6v}
H_{\rm V} = \frac{V}{64} \sum_{{
\setlength{\unitlength}{1.5 mm}
\begin{picture}(1,1)(0,0)
\put(0,0){\line(1,0){1}}
\put(0,0){\line(0,1){1}}
\put(1,1){\line(-1,0){1}}
\put(1,1){\line(0,-1){1}}
\end{picture}
}}
(\Delta_2 E_1)^2 (\Delta_1 E_2)^2(E_1-E_2)^2 \ ,$$ where the sum is over all the plaquettes of the lattice. The factors $(\Delta_2 E_1)^2$ and $(\Delta_1 E_2)^2$ make sure that the arrows on opposite links of the plaquette are anti-parallel. The factor $(E_1-E_2)^2$ checks if two arrows on one vertex are both pointing clockwise or both counterclockwise.
As the $E_i(\vec x)$’s have the integers as their eigenvalues, the canonically conjugate operators, $a_i(\vec x)$ are phases, i.e. $0\leq a_i(\vec x)<2\pi$. Using the commutation relations $$\left[a_j(\vec x),E_{j'}(\vec x ')\right] =
i \delta_{jj'}\delta_{\vec x,\vec x'} \ ,$$ it is easy to show that the operators $e^{\pm i a_j (\vec x)}$ act as raising and lowering operators on the $E$’s, and thus e.g. the operator $e^{-2 i a_1 (\vec x)}$ will flip the arrow pointing outward from $\vec x$ to the arrow pointing inward. We can use these raising and lowering operators to write the term in the Hamiltonian which flips the flippable plaquettes $$\label{hflip6v}
H_{\rm t} =
-2 t \sum_{{
\setlength{\unitlength}{1.5 mm}
\begin{picture}(1,1)(0,0)
\put(0,0){\line(1,0){1}}
\put(0,0){\line(0,1){1}}
\put(1,1){\line(-1,0){1}}
\put(1,1){\line(0,-1){1}}
\end{picture}
}}
\cos (2\textstyle\sum\nolimits_{{
\setlength{\unitlength}{2 mm}
\begin{picture}(1,1)(0,0)
\put(0,0){\line(1,0){1}}
\put(0,0){\line(0,1){1}}
\put(1,1){\line(-1,0){1}}
\put(1,1){\line(0,-1){1}}
\put(1,0){\vector(0,1){1}}
\end{picture}
}} a_j(\vec x)) \ ,$$ where $\sum_{{
\setlength{\unitlength}{2 mm}
\begin{picture}(1,1)(0,0)
\put(0,0){\line(1,0){1}}
\put(0,0){\line(0,1){1}}
\put(1,1){\line(-1,0){1}}
\put(1,1){\line(0,-1){1}}
\put(1,0){\vector(0,1){1}}
\end{picture}
}} a_j(\vec x)=\Delta_1 a_2 (\vec x) -\Delta_2 a_1 (\vec x)$ is the oriented sum of the $a$’s around a plaquette. The total Hamiltonian of the gauge theory version of the 6-vertex model is therefore $$H_{\rm 6v} = H_{\rm E} + H_{\rm t} + H_{\rm V} \ .$$ As usual, the Rokhsar-Kivelson point is located at $t=V$.
We will proceed by going to the dual formulation of the theory, and show that the theory is equivalent to a height model (which is well know). Doing the duality basically amounts to solving the (electrostatic) constraint. In the process, it gets replaced by a magnetic constraint. To solve the constraint, we introduce the new variables $S(\vec r)$, which live on the sites of the dual lattice (or plaquettes of the direct lattice); these operators have the integers as their spectrum. In addition, we need the fields $B_i (\vec r)$, which live on the links of the dual lattice. We can now write the “electric" fields $E_i$ as follows: $$\label{dual}
E_i (\vec x) =
\epsilon_{ij} \left({\Delta^{\!\! -}}_j S (\vec r) + B_j (\vec r)\right) \ .$$ Substituting this in the constraint ${\Delta^{\!\! -}}_i E_i^{\vphantom{-}}=0$ gives the ‘magnetic’ constraint $\epsilon_{ij}{\Delta^{\!\! -}}_j B_k^{\vphantom{-}} (\vec r) = 0$, so $B_k$ is curl free, and can be written as a gradient. But, as the there are no sources, we can do even better, as becomes clear when we interpret $S (\vec r)$ as a height variable which lives on the plaquettes of the direct lattice.
We will first recall the known fact that the configurations of the 6-vertex model can be mapped onto height configurations. The rules are as follows. First, pick a reference site, and give it a reference height, say $S (\vec 0)=0$. Then, if one crosses an outgoing arrow clockwise (both seen from the vertex), the height decreases by one, while crossing an incoming arrow (again in a clockwise manner), the height increases by one. As all the vertices have two incoming and two outgoing arrows, this indeed gives a consistent height configuration. So, as an example, around an $a$ vertex we have
(10,10)(0,0) (5,0)[(0,1)[10]{}]{} (0,5)[(1,0)[10]{}]{} (5,0)(0,5)[2]{}[(0,1)[3.5]{}]{} (0,5)(5,0)[2]{}[(1,0)[3.5]{}]{} (7,7)[0]{} (6,1)[-1]{} (1,1)[0]{} (1,7)[1]{}
.
We now assume that the $S(\vec r)$’s appearing in Eq. (\[dual\]) can in fact be interpreted as the heights of the plaquettes. Because of the duality Eq. (\[dual\]), the $B_i$ are now completely determined by the $E_i$, because they determine $S$ via the height rule. So we found that interpreting the $S (\vec r)$ as heights amounts to picking a gauge for the $B_i (\vec r)$. Combining the height rule of the previous paragraph with Eq. (\[dual\]), we easily find that $B_j (\vec r) \equiv 0$.
To complete the duality transformation, we need to transform the flip term. This will involve the canonically conjugate variable to $S(\vec
r)$. Let us call this the momentum $P (\vec r)$, which satisfies $\left[ P(\vec r),S(\vec r')\right] = i \delta_{\vec r,\vec r'}$. Again, acting with $e^{i P(\vec r)}$ on the plaquette at $\vec r$ will increase the eigenvalue of $S(\vec r)$ by one. Flipping a plaquette changes $S$ by $\pm 2$, so we find that we can write the flip term of the Hamiltonian as $H_{\rm t} = -2 t \sum_{\vec r} \cos (2 P(\vec r))$. In other words, we find that the circulation around the plaquette at $\vec r$ is given by $\sum_{{
\setlength{\unitlength}{2 mm}
\begin{picture}(1,1)(0,0)
\put(0,0){\line(1,0){1}}
\put(0,0){\line(0,1){1}}
\put(1,1){\line(-1,0){1}}
\put(1,1){\line(0,-1){1}}
\put(1,0){\vector(0,1){1}}
\end{picture}
}} a_j(\vec x) = P (\vec r)$.
In principle, we could go on to describe the eight-vertex model in a similar fashion. The main difference with the six-vertex model is that the constraint is no longer satisfied. One can introduce a matter field which has no-zero values on the $d$ vertices. In addition, a flip term which flips every plaquette has to be constructed. This flip term has to commute with the new constraint. It turns out that this is indeed possible, but the flip term will involve the conjugate of the matter field. Also, one would need a potential term which gives weights plaquettes according to the vertices present on the plaquette. However, as there is a more natural gauge description, which is based on the $\mathbb{Z}_2$ symmetry of the eight-vertex model, we will discuss and use that description of the quantum eight vertex model in the next appendix.
A $\mathbb{Z}_2$ gauge theory for the quantum eight vertex model {#app:z2gauge}
================================================================
In this appendix, we will discuss a $\mathbb{Z}_2$ gauge theory which can be viewed as an extension of the Kitaev model which incorporates vertex weights differing from unity. The model we will discuss is of the Rokhsar-Kivelson type, but it is not the simplest quantum generalization of the classical eight-vertex model, as we pointed out in section \[sec:q8v\].
The $\mathbb{Z}_2$ gauge theory {#app:z2gt}
-------------------------------
In this model, spins living on the bonds of the square lattice are the degrees of freedom. Thus, on every link of the square lattice, we define a Pauli algebra of $2 \times 2$ Hermitian matrices $\sigma^a_j(\vec x)$, where $a=1,2,3$ labels the three Pauli matrices, and for a lattice site $\vec x$, we denoted the orientation of the link by $j=1,2$ ($1=$horizontal and $2=$vertical). (Thus, the degrees of freedom live half-way between the lattice sites $\vec x$ and $\vec
x+\vec e_j$, where $\vec e_j$ is a unit vector along the direction $j$.) In what follows we will take the states $|\uparrow\; \rangle$ (an “up" spin) and $|\downarrow\;\rangle$ ( a “down" spin) as the states which diagonalize $\sigma^1$ (instead of $\sigma^3$, as it is customary). The relation with the eight-vertex model is simple. Up-spins correspond to arrows pointing up or to the right, while down spins correspond to arrows pointing down or to the left. Around each vertex, the number of up-spins has to be even. In this section, we will denote the vertex by $\vec x$ and the associated plaquette by its south-west corner $\vec x$.
For details, see Ref. \[\]. The constraint can written in a simple form[^17] $$\label{scon8v2}
\sigma^1_1(\vec x) \; \sigma^1_1(\vec x-\vec e_1) \; \sigma^1_2(\vec x) \; \sigma^1_2(\vec x-\vec e_2)=1, \qquad \forall \vec x$$ The term which flips all the arrows around a plaquette can be written in terms of $\sigma^{3}$’s $$\label{sflip2}
H_{\rm flip} =-\sum_{\vec x} \sigma^3_1(\vec x) \; \sigma^3_1(\vec x+\vec e_2)\; \sigma^3_2(\vec x)\; \sigma^3_2(\vec x+\vec e_1)$$ This flip operator commutes with the constraint, because a vertex and a plaquette have either $0$ or $2$ common bonds. Hence, both operators can be diagonalized simultaneously. In the model considered by Kitaev \[\], equal weight is given to all types of vertices, so no term in the Hamiltonian is required. To go beyond the point $a=b=c=d=1$, we need a term which will weight the plaquettes according to the types of vertices present. Thus, we will need to introduce operators which can discern among the various vertices. In addition, the “weight" of the vertex depends on its position in the plaquette under consideration.
The notation that we use is summarized in figure \[sigmaconfig\].
Let us start by giving the terms which give a non-zero contribution for $a,b,c$ and $d$ vertices respectively (but zero otherwise). Let us define the “vertex magnetizations" $$\begin{aligned}
S_a(\vec x)&=&\frac{1}{4}\left(\sigma^1_1(\vec x)+\sigma^1_2(\vec x)+\sigma^1_1(\vec x-\vec e_1)+\sigma^1_2(\vec x-\vec e_2)\right)
\nonumber \\
S_b(\vec x)&=&\frac{1}{4}\left(\sigma^1_1(\vec x)-\sigma^1_2(\vec x)+\sigma^1_1(\vec x-\vec e_1)-\sigma^1_2(\vec x-\vec e_2)\right)
\nonumber \\
S_c(\vec x)&=&\frac{1}{4}\left(\sigma^1_1(\vec x)-\sigma^1_2(\vec x)-\sigma^1_1(\vec x-\vec e_1)+\sigma^1_2(\vec x-\vec e_2)\right)
\nonumber \\
S_d(\vec x)&=&\frac{1}{4}\left(\sigma^1_1(\vec x)+\sigma^1_2(\vec x)-\sigma^1_1(\vec x-\vec e_1)-\sigma^1_2(\vec x-\vec e_2)\right)
\label{vms} \end{aligned}$$ With this notation, the projectors onto the vertices $a$, $b$, $c$ and $d$ are just the squares of the vertex magnetizations of Eq. (\[vms\]): $${\mathcal P}_a=S_a^{\;2}, \qquad {\mathcal P}_b=S_b^{\;2}, \qquad {\mathcal P}_c=S_c^{\;2}, \qquad {\mathcal P}_d=S_d^{\;2}
\label{P-S}$$ It is straightforward to show that these operators act as projection operators on vertices of type $a$, $b$, $c$ and $d$ respectively, [*i.e.*]{} they yield $1$ when acting on the corresponding vertex and zero otherwise. It is elementary to check that $${\mathcal P}_a+{\mathcal P}_b+{\mathcal P}_c+{\mathcal P}_d={\mathcal I}
\label{identity}$$ where ${\mathcal I}$ is the identity operator.
We can now write down the potential term which assigns weights to the plaquettes in the same way as done by the projectors of section \[sec:q8v\]. Flipping a plaquette will change a $d$ vertex to an $a$ vertex if the $d$ vertex is in the south-west or north-east “even” corner of the plaquette. A $d$ at the other “odd” corners will go to a $b$ under the flip. In short, $d_e \leftrightarrow a_e$, $d_o \leftrightarrow b_o$. For $c$’s this is opposite, namely $c_e \leftrightarrow b_e$, $c_o \leftrightarrow a_o$. The potential term will involve all four vertices around a plaquette, and thus we need to distinguish between the different position in the plaquette. This is achieved by the following, albeit rather cumbersome, term $$\label{q8vpotsig}
\begin{split}
H_{\rm V} &= \sum_{\vec x}
\left(
\frac{d}{a} {\mathcal P}_a (\vec x) +
\frac{c}{b} {\mathcal P}_b (\vec x) +
\frac{b}{c} {\mathcal P}_c (\vec x) +
\frac{a}{d} {\mathcal P}_d (\vec x)
\right) \times \\
&\left(
\frac{c}{a} {\mathcal P}_a (\vec x+e_1) +
\frac{d}{b} {\mathcal P}_b (\vec x+e_1) +
\frac{a}{c} {\mathcal P}_c (\vec x+e_1) +
\frac{b}{d} {\mathcal P}_d( \vec x+e_1)
\right) \times \\
&\left(
\frac{c}{a} {\mathcal P}_a (\vec x+e_2) +
\frac{d}{b} {\mathcal P}_b (\vec x+e_2) +
\frac{a}{c} {\mathcal P}_c (\vec x+e_2) +
\frac{b}{d} {\mathcal P}_d (\vec x+e_2)
\right) \times \\
&\left(
\frac{d}{a} {\mathcal P}_a (\vec x+e_1+e_2) +
\frac{c}{b} {\mathcal P}_b (\vec x+e_1+e_2) +
\frac{b}{c} {\mathcal P}_c (\vec x+e_1+e_2) +
\frac{a}{d} {\mathcal P}_d (\vec x+e_1+e_2)
\right) \ .
\end{split}$$ The potential term assigns potential energies to the plaquettes in the same way as is done by the projectors of section \[sec:q8v\]. That is, the potential is the product of vertex weights obtained by flipping the plaquette, divided by the product of the vertex weights of the plaquette itself. As mentioned before, the two-body terms only couple spins on the same sublattice. The most non-local terms in the potential energy term consist of eight-body interactions. Of course, a plaquette potential energy term of this sort is needed, if one assigns weights to plaquettes[^18]. Note that for $a=b=1$ and $c=d$ (or $c=\tfrac{1}{d}$), the eight-body terms cancel each other. However, there will remain four and six-body interactions, which will mix different sublattices.
The total Hamiltonian for the $\mathbb{Z}_2$ gauge theory of the quantum eight vertex model is $$H_{\rm q8v} = H_{\rm V} + H_{\rm flip}
\label{hq8v}$$ where the two terms are given by and . The states in this model have to satisfy the constraint . We can check these formulas in a few limits. For $a=b=c=d=1$, each factor in will be $1$ when acting on states satisfying the constraint. Hence, the potential term became the identity operator, which merely results in an energy shift, and thus we find back the Kitaev model, as we must.
Another interesting limit is $d \rightarrow 0$. As was discussed in section \[sec:8vqh\], this limit gives a slight generalization of the six-vertex model, in the sense that plaquettes which have $n_d=\widetilde{n}_d=1$ or $n_d=\widetilde{n}_d=2$ will have finite energy, and they can be interpreted as static defects. As can be seen from the potential term \[q8vpotsig\], plaquettes which have $n_d>\widetilde{n}_d$ will be suppressed as they receive infinite energy. The flip term has to be modified, because we need to have a flip term which commutes with the constraint, which has become a stronger statement in the six-vertex case, namely (c.f. Eq. ) $$\label{scon6v}
\sigma_1^1 (\vec x) - \sigma_{1}^1 (\vec x-\vec e_1) +
\sigma_2^1 (\vec x) - \sigma_{2}^1 (\vec x-\vec e_2) = 0 \qquad \forall \vec x \ .$$ The flip term which preserves the six-vertex constraints is $$\label{sflip6v}
H_{\rm flip,6v} = - \sum_{\vec x}
\bigl(\sigma^-_1 (\vec x) \; \sigma^-_2 (\vec x + e_1) \;
\sigma^+_1 (\vec x + e_2) \; \sigma^+_2 (\vec x) + {\rm h.c.} \bigr) \ ,$$ where the raising and lowering operators (in the representation we use) are given by $$\sigma^\pm = \tfrac{1}{2} (\sigma^3 \mp i\sigma^2)$$ To make contact with the flip term for the eight-vertex model, we rewrite Eq. as $$\begin{split}
H_{\rm flip,q8v} = & - \sum_{\vec x}
\bigl(\sigma^+_1 (\vec x) + \sigma^-_1 (\vec x)\bigr)
\bigl(\sigma^+_2 (\vec x + e_1) + \sigma^-_2 (\vec x + e_1)\bigr) \times \\
&\bigl(\sigma^+_1 (\vec x + e_2) + \sigma^-_1 (\vec x + e_2)\bigr)
\bigl(\sigma^+_2 (\vec x) + \sigma^-_2 (\vec x)\bigr) \ .
\end{split}$$ The flip term for the six-vertex model is therefore precisely the flip term for the eight-vertex model minus the terms which cause the six-vertex constraint to be violated. It is easily checked that the six-vertex flip term Eq. commutes with the constraint . We thus find that on the level of the wave function, the limit $d \rightarrow 0$ is smooth, as the amplitude of the configurations which contain $d$ vertices goes to zero. In addition, for $d \neq 0$, the flip term commutes with constraint , while for $d=0$, the flip term commutes with the $U(1)$ constraint . Thus, the symmetry is enhanced from $\mathbb{Z}_2$ for $d\neq 0$ to $U(1)$ for $d=0$, as was to be expected.
The dual of the gauge theory {#app:duality}
----------------------------
We now have a Rokhsar-Kivelson generalization of the eight-vertex model, in a gauge-theory language. We can use this representation of our model to study the various phases. However, this is more easily done in a dualized version, as the dual takes the form of an Ising model. In the dual picture, the spin degrees of freedom will live on the sites of dual square lattice, [*i.e.*]{} the centers of the plaquettes of the direct lattice. Thus, we will label by $\vec r$ the site of the dual lattice on the center of the plaquette labeled by $\vec x$ (its SW corner). Of course, the potential term in the dual language will still be quite formidable. We will denote the dual Pauli operators by $\tau^1$ and $\tau^3$. To start with the flip term, the product of $\sigma^3$’s around a plaquette becomes $\tau^1$ on the plaquette \[\] $$\label{tx}
\tau^1 (\vec r) =
\sigma^3_1(\vec x) \; \sigma^3_2(\vec x+\vec e_1) \;
\sigma^3_1(\vec x+ \vec e_2) \; \sigma^3_2(\vec x)$$ To see what happens with the constraint and the projector operators defined by Eq. (\[vms\]) and Eq. (\[P-S\]) we need the dual form of the $\sigma^1$’s living on the links. In term of the dual variables $\tau^3$, and using the notation of Fig. \[fig:dual\], the $\sigma^1$’s are given by $$\label{tz}
\sigma^1_1(\vec x)= \tau^3(\vec r) \; \tau^3(\vec r-\vec e_2), \qquad
\sigma^1_2(\vec x)= \tau^3(\vec r) \; \tau^3(\vec r-\vec e_1)$$
We thus easily find that the constraint is automatically satisfied (again, going to the dual picture amounts to solving the constraint). Also, it is trivial to show \[\] that the inverse relation, [*i.e.*]{} to express the dual lattice $\tau^3$ operators in terms of the $\sigma^1$ operators of the original lattice is $$\tau^3(\vec r)=\prod_{\ell \in \Gamma(\vec r)} \sigma^1(\ell)
\label{tau3}$$ where $\{\ell\}$ is a set of links of the direct lattice pierced by a path $\Gamma(\vec r)$ on the [*dual*]{} lattice ending at the dual site $\vec r$ (but which is otherwise arbitrary); see Fig. \[dual-path\].
Note that we need to choose the spin on one of the plaquettes; all the others are subsequently determined by the $\sigma^x$’s on the links. In terms of the dual variables, the projection operators for site $\vec x$, defined by Eq. , take the form $$\label{vms-dual}
\begin{split}
\mathcal{P}_a(\vec x) &= \frac{1}{4} \left(
1 + \mathcal{A}(\vec r) + \mathcal{B}(\vec r) + \mathcal{C}(\vec r)
\right) \\
\mathcal{P}_b(\vec x) &= \frac{1}{4} \left(
1 - \mathcal{A}(\vec r) - \mathcal{B}(\vec r) + \mathcal{C}(\vec r)
\right) \\
\mathcal{P}_c(\vec x) &= \frac{1}{4} \left(
1 - \mathcal{A}(\vec r) + \mathcal{B}(\vec r) - \mathcal{C}(\vec r)
\right) \\
\mathcal{P}_d(\vec x) &= \frac{1}{4} \left(
1 + \mathcal{A}(\vec r) - \mathcal{B}(\vec r) - \mathcal{C}(\vec r)
\right) \ , \\
\end{split}$$ where $\mathcal{A,B}$ and $\mathcal{C}$ are given by $$\begin{split}
\mathcal{A}(\vec r) &= \tau^3(\vec r-\vec e_1) \; \tau^3(\vec r-\vec e_2) \\
\mathcal{B}(\vec r) &= \tau^3(\vec r) \; \tau^3(\vec r-\vec e_1-\vec e_2) \\
\mathcal{C}(\vec r) &= \tau^3(\vec r) \; \tau^3(\vec r-\vec e_1) \;
\tau^3(\vec r-\vec e_1) \; \tau^3(\vec r - \vec e_1-\vec e_2) \ .
\end{split}$$
The interaction in terms of these projectors has the same structure as in the spin representation of the classical eight-vertex model: it consists of two-body terms on interpenetrating sublattices, and a four-body term coupling the two sublattices. All the two-body interaction terms will only couple spins on the same sublattice. The four- and six-body interaction terms (of which there are many!), couple the sublattices. The same holds for the eight-body term, naturally. The dual form of the theory is $$H_{\rm q8v, dual} = H_{\rm V,dual} - \sum_{\vec r} \tau^1(\vec r) \ ,
\label{q8v-dual}$$ where $H_{\rm V,dual}$ is given by , but now with the projectors given in Eq. . Thus, formally this theory takes the form of a (multi-spin) Ising model in a transverse field. However, the two-body interactions only couple spins on the same sublattices, together with the multi-spin terms conspire to change the quantum critical behavior from the conventional $z=1$ Lorentz-invariant criticality of the standard Ising model in a transverse field to the $z=2$ quantum critical behavior discussed in the rest of this paper.
Now that we found the dual version of our gauge theory, we would like to discuss the limits $a=b=c=d=1$ and $d=0$. Again, the first limit brings us back to the Kitaev point, because the potential term becomes the identity operator again, and we are left with the very simple spin flip term of Eq. , $H_{\rm f}=-\sum_{\vec r} \tau^1(\vec r)$. The limit $d \rightarrow 0$ is however more complicated in this dual gauge theory. First of all, we now do need a constraint, which was not present for $d \neq 0$. Moreover, the flip term now only can act, depending on the surrounding spins. Let us start by dualizing the constraint Eq. , which results in $$\label{tcon6v}
\bigl( \tau^3(\vec r) - \tau^3(\vec r-\vec e_1-\vec e_2) \bigr)
\bigl( \tau^3(\vec r-\vec e_1) + \tau^3(\vec r-\vec e_2) \bigr) = 0 \qquad
\forall \vec x \ .$$ Obviously, the eight-vertex flip term $\tau^1 (\vec r)$ does not commute with this constraint. To find a flip term which does commute with the constraint, we dualize the six-vertex flip term , which results in $$\label{tflip6v}
H_{\rm flip,q6v} = -\frac{1}{8}\sum_{\vec r} \tau^1 (\vec r)
\bigl( 1-\tau^3(\vec r+\vec e_1) \tau^3(\vec r+\vec e_2) \bigr)
\bigl( 1+\tau^3(\vec r-\vec e_1) \tau^3(\vec r+\vec e_2) \bigr)
\bigl( 1+\tau^3(\vec r+\vec e_1) \tau^3(\vec r-\vec e_2) \bigr) \ .$$ The factors in bracket can be seen to give a non zero result only on plaquettes which are flippable. Hence, this flip term commutes with the constraint . Of course, this can also be checked explicitly. Apart from a factor $\tau^2$, there are factors depending on the $\tau^3$’s coming form both the flip term and the constraint. The signs in this product conspire in such a way to render the commutator zero.
[128]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ()
, ** (, , ).
, ** (, , ).
, eds., ** (, ).
, ****, ().
, ****, ().
, ****, () .
, ****, () .
, , , ****, ().
, ****, ().
, , , ****, () .
, in **, edited by (, , ), .
, ** (, , ), .
, , , ****, () .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, (), .
, ****, ().
, , , ****, ().
, ****, ().
, ****, () .
, ****, ().
, ****, ().
, ****, ().
, ** (),
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, , , , ****, () .
, , , ****, ().
, ****, () .
, ****, () .
, , , ****, () .
, ****, () .
, , , , (), .
, eds., ** (, ).
, ****, ().
, ****, () .
(), .
, , , , .
, , , , ****, () .
, , , .
, ** (, , ).
, , , ** (, , ).
, ** (, , ).
, , , , (), .
, , (), .
, ****, ().
, ****, () .
, ****, ().
, ****, ().
, , , , in ** (, , ).
, ****, ().
(),
, ****, ().
, in **, edited by , , (, , ), p. , .
, ****, ().
, , , ****, () .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , (), .
, , , , (), .
, ****, ().
, ****, (), .
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, ** (, ),
, ****, ().
, , , ****, () .
() .
, ****, () .
, , , .
, , (), .
, ****, () .
, , (), .
, ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, () .
, in **, edited by (, ), .
, ****, ().
, ****, ().
, ****, ().
, ****, (),
, , , , ****, ().
, ** (, , ), .
, ****, ().
, , , , ****, ().
, ****, ().
, ****, (), .
, , , , ****, () .
, , , ****, () .
, ****, ().
, ****, () .
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, in ** () p. , .
, ****, () .
, , , ****, () .
, , , , ****, () .
(), .
, ****, () .
, , (), .
, ****, ().
, in **, edited by (), p. , .
, ****, ().
[^1]: Recall that the change in the metric, given by the strain tensor, is quadratic in the local deformation of the system.
[^2]: The role of conformal invariance in 2D classical dynamics with $z=2$ and anisotropic 3D classical Lifshitz points was considered recently in Refs. \[\].
[^3]: We became aware of Ref. \[\] as this paper was being completed.
[^4]: This construction has been employed extensively in statistical mechanics models of classical dimers and loops; see for instance \[\].
[^5]: The equal-time correlators of the vertex operators defined and computed in Appendix \[app:gaussian\] can also be computed away from the critical point by either by a naive semi-classical argument, which predicts the simple exponential decay we just discussed, or in a more sophisticated way by means of form factors \[\]. Apart from a subtle bound state structure in the spectral functions, the more sophisticated approach confirms the essence of the naive semi-classical result. On the other hand, the $z=2$ character of the critical theory suggests that time-dependent correlation functions must be consistent with this fact, and that the correlators must be functions of $x^2$ or $t$, and that the time dependent Euclidean auto-correlation functions may obtained from the equal-time correlator by replacing $x^2
\leftrightarrow |t|$. These arguments are consistent with the renormalization-group results of Ref. \[\]
[^6]: However, the quantum dimer models on the Kagome \[\] and Fisher \[\] lattices have Hamiltonians where all plaquettes are flippable.
[^7]: This Hamiltonian in the special case $c=1$, $d=0$ was discussed in Ref. \[\]. Moreover, the quantum six-vertex model should be in the same universality class as the“supersymmetric” $XY$ model introduced some time ago \[\]; this model defines a similar Hamiltonian acting on the 2D classical XY model, and presumably can also be understood in terms of a mapping to the quantum Lifshitz model. Another quantum six-vertex model has been proposed as a model of a planar pyrochlore lattice in Ref. \[\]. This six-vertex model is not of the Rokhsar-Kivelson type; it is the six-vertex limit of the simpler quantum eight-vertex model we discuss in section \[another\].
[^8]: Throughout this section and in Appendix \[app:z2gauge\] we use the standard notation used in lattice gauge theories \[\]. The $\sigma$’s are Pauli matrices; the superscript is a Pauli matrix label and the subscript indicates the spatial direction, $1=$ horizontal and $2=$ vertical. Notice that the notation used by Kitaev \[\] is somewhat different. Please see Appendix \[app:z2gauge\], and figures \[sigmaconfig\] and \[fig:dual\], for details of the notation that we use here.
[^9]: The holon operator of the quantum dimer model has the same behavior in both the columnar and staggered phases.
[^10]: We thank Shivaji Sondhi for this remark.
[^11]: This result, Eq. holds strictly only at the quantum Lifshitz fixed point (or rather line) fixed point, Eq. . However, marginally irrelevant operators such as $\left(\nabla \phi\right)^4$, discussed extensively in Refs.\[\], give rise to corrections to scaling which change the behavior of $U(R)$ from Eq. to a $(\ln R)/R^2$ law\[\]. We thank T. Senthil for pointing this out to us.
[^12]: As we noted above, on symmetry grounds a lattice model may generate additional operators which break rotational invariance, absent at this RK point, which may drive the transitions to be first order, see \[\].
[^13]: Notice that the charge-one vortex operator is not allowed in this case, which is the operator driving the KT transition in the 2D classical XY model \[\].
[^14]: This is a classic example of a dangerously-irrelevant operator, since it removes the would-be Goldstone boson from the broken-symmetry phase.
[^15]: For this reason, this theory has arisen for example in theories of quantum computation and in superconductivity \[,\], and in effective “coset" field theories of the non-abelian fractional quantum Hall states \[\].
[^16]: This approach is reminiscent of stone soup.
[^17]: Throughout we use the upper label to indicate the Pauli matrix and the lower label to indicate direction.
[^18]: In spite of the appearances, $H_V$ as given in Eq. respects rotational invariance; the apparent asymmetry is due to the use of the vertex weights as labels.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose using global orientation from inertial measurements, and the bias it induces on the shape of objects populating the scene, to inform visual 3D reconstruction. We test the effect of using the resulting prior in depth prediction from a single image, where the normal vectors to surfaces of objects of certain classes tend to align with gravity or be orthogonal to it. Adding such a prior to baseline methods for monocular depth prediction yields improvements beyond the state-of-the-art and illustrates the power of gravity as a supervisory signal.'
author:
- 'Xiaohan Fei, Alex Wong, and Stefano Soatto [^1] [^2]'
title: '**Geo-Supervised Visual Depth Prediction**'
---
<span style="font-variant:small-caps;">Introduction</span>
==========================================================
The visual world is heavily affected by gravity, including the shape of many artifacts such as buildings and roads, and even natural objects such as trees. Gravity provides a globally consistent orientation reference that can be reliably measured with low-cost inertial sensors present in mobile devices from phones to cars. We call a machine learning system able to exploit global orientation, [*geo-supervised*]{}. Gravity can be easily inferred from inertial sensors without the need for dead-reckoning, and the effect of biases is negligible in the context of our application.
To measure the influence of gravity as a supervisory signal, we choose the extreme example of predicting depth from a single image. This is, literally, an impossible task in the sense that there are infinitely many three-dimensional (3D) scenes that can generate the same image. So, any process that yields a point estimate has to rely heavily on priors. We call the resulting point estimate a [*hypothesis*]{}, or [*prediction*]{}, and use public benchmark datasets to quantitatively evaluate the improvement brought about by exploiting gravity. Of course, only certain objects have a shape that is influenced by gravity. Therefore, our prior has to be applied [*selectively*]{}, in a manner that is informed by the semantics of the scene.
Our approach to geo-supervised Visual Depth Prediction is based on training a system end-to-end to produce a map from a single image and an estimate of the orientation of gravity in the (calibrated) camera frame to an inverse depth (disparity) map. In one mode of operation, the training set uses calibrated and rectified stereo pairs, together with a semantic segmentation module, to evaluate a loss function differentially on the images where geo-referenced objects are present. In a second mode, we use monocular videos instead and minimize the reprojection (prediction) error. Optionally, we can leverage modern visual-inertial odometry (VIO) and mapping systems that are becoming ubiquitous from hand-held devices to cars.
The key to our approach is a prior, or regularizer, that selectively biases certain regions of the image that correspond to geo-referenced classes such as roads, buildings, vehicles, and trees. Specifically, points in space that lie on the surface of such objects should have normals that either align with, or are orthogonal to, gravity. This is in addition to standard regularizers used for depth prediction, such as left-right consistency and piecewise smoothness.
While at training time a semantic segmentation map is needed to apply our prior selectively, it is never passed as input to the network. Therefore, at test time it is not needed, and an image is simply mapped to the disparity.
The ultimate test for a prior is whether it helps improve end-performance. To test our prior, we first incorporated it into two top-performing methods, one binocular (Sect. \[sect-stereo\]) and one monocular (Sect. \[sect-mono\]), in the KITTI benchmark [@kitti], and showed consistent performance improvement in all metrics. To further challenge our prior, we took two other baselines which were not the top performers. We then added our prior and tested the results against the top performers in the latest benchmark. We also performed generalizability tests (Sect. \[sect-make3d\]), ablation studies (Sect. \[sect-ablation\]) and demonstrated our approach with VIO on hand-held devices (Sect. \[sect-visma\]).
<span style="font-variant:small-caps;">Related work</span> {#sect-related}
==========================================================
Early learning-based depth prediction approaches [@saxena2006learning; @saxena2009make3d; @konrad2013learning; @karsch2012depth] predict depth using local image patches and then refine it using Markov random fields (MRFs). Recent works [@eigen2014depth; @laina2016deeper] leverage deep networks to directly learn a representation for depth prediction where the networks are typically based on the multi-scale fully convolutional encoder-decoder structure. These methods are fully supervised and do not generalize well outside the datasets on which they are trained. Latest self-supervised methods [@garg2016unsupervised; @godard2017unsupervised; @zhou2017unsupervised] have shown better performance on benchmarks with better generalization.
There is a large body of work [@mahjourian2018unsupervised; @yin2018geonet; @wang2017learning; @zhan2018unsupervised] on self-supervised monocular depth prediction following Godard *et al*. [@godard2017unsupervised] and Zhou *et al*. [@zhou2017unsupervised], which simply use the reprojection error as a learning criterion, as has been customary in 3D reconstruction for decades. Generic priors such as piecewise smoothness and left-right consistency are also encoded into the network as additional loss terms. Our work is in-line with these self-supervised approaches, but we also exploit class-specific regularizers beyond the generic ones.
In terms of exploiting the relation of different geometric quantities in an end-to-end learning framework, closely related works include [@wang2016surge; @qi2018geonet; @liu2018planenet], where surface normals are explicitly computed by using either a network [@wang2016surge] or some heuristics [@qi2018geonet]. While the former is computation intensive, the latter relies on heuristics and thus is sub-optimal. In contrast, by using losses proposed in this paper, we directly regularize depth via the depth-gravity relation without a separate surface normal predictor. Besides, both [@wang2016surge] and [@liu2018planenet] are supervised, while ours is self-supervised with the photometric loss and guided by global orientation and the semantics of the scene. Earlier work on semantic segmentation [@shotton2008semantic] relied on local features, and have been improved by incorporating global context using various structured prediction techniques [@krahenbuhl2011efficient; @russell2009associative]. Starting from the work of Long *et al*. [@long2015fully], fully convolutional encoder-decoder networks have been a staple in semantic segmentation. Although we do not address semantic segmentation, we leverage per-pixel semantic labeling enabled by existing systems to aid depth prediction in the form of providing class-specific priors and an attention mechanism to selectively apply such priors, which is different from joint segmentation and depth prediction approaches [@jafari2017analyzing].
The idea of using class-specific priors to facilitate reconstruction is not new [@hane2013joint; @kundu2014joint]. In [@hane2013joint], class-specific shape priors in the form of spatially varying anisotropic smoothness terms are used in an energy minimization framework to reconstruct small objects. Though promising, this system does not scale well. An efficient inference framework [@krahenbuhl2011efficient] has been used with a CRF model over a voxel-grid to achieve real-time performance by [@kundu2014joint]. While all these methods explore class-specific priors in various ways, none has used them in an end-to-end learning framework. Also, all the methods above take range images as inputs, which are then fused with semantics during optimization, while ours exploits semantics at an earlier stage – when generating such range images which themselves can serve as priors for dense reconstruction and other inference tasks.
<span style="font-variant:small-caps;">Methodology</span>
=========================================================
In this section, we introduce our loss functions as regularizers added to existing models at training time, in addition to data terms (photometric loss) and generic regularizers (smoothness loss). We dub our loss semantically informed geometric loss (SIGL) because geometric constraints are selectively applied to certain image regions, where a semantic segmentation module informs the selection. Fig. \[fig-diagram\] illustrates part of our training diagram. In Sect. \[sect-baseline\], we review baseline models used in our experiments and show that the application of our losses on top of them improves performance (Sect. \[sect-exp\]).
Semantically informed geometric loss {#sect-sigl}
------------------------------------
During training, we assume to be given a partition of the image plane into semantic classes $c\in C$ that have a consistent geometric correlate. For instance, a pixel with image coordinates $(x, y) \in {\mathbb R}^2$ and class $c(x,y) =$ “road” is often associated to a normal plane oriented along the vertical direction (direction of gravity), whereas $c = $“building” has a normal vector orthogonal to it. We also assume we are given the calibration matrix $K$ of the camera capturing the images, so the pixel coordinates $(x,y)$ on the image plane back-project to points in space via $${\bf X} = \left[\begin{array}{c} X \\ Y \\ Z \end{array}\right] = K^{-1}
\begin{bmatrix}
x\\
y\\
1
\end{bmatrix}
Z(x, y)
\label{eq-back-project}$$ where $Z(x, y)$ is the depth $Z$ of the point along the projection ray determined by $(x, y)$.
Any subset $\Omega \subset {\mathbb R}^2$ of the image plane that is the image of a spatial plane with normal vector $N \in \mathbb R^3$, at distance $\| N \|$ from the center of projection, satisfies a constraint of the form ${\bf X}^T_i N = 1$ for all $i$, assuming the plane does not go through the optical center. Stacking all the points into matrix $\bar {\bf X}\doteq [{\bf X}_1, {\bf X}_2\cdots {\bf X}_M]^\top$, we have $\bar {\bf X} N = {\bf 1}$, where ${\bf 1}$ is a vector of $M$ ones, and $M=|\Omega|$ is the cardinality of the set $\Omega$. If the direction, but not the norm, of the vector $N$ is known, a scale-invariant constraint can be easily obtained by removing the mean of the points, so that (details in Sect. \[sect-derivation\]) $$({\bf I}-\frac{1}{M}{\bf 1}{\bf 1}^\top) \bar {\bf X} N = 0.
\label{eq-constraint}$$ The scale-invariant constraint above can be used to define a loss to penalize deviation from planarity: $$L_{HP}(\Omega_{HP}) = \frac{1}{|\Omega_{HP}|}\| ({\bf I}-\frac{1}{|\Omega_{HP}|}{\bf 1}{\bf 1}^\top)\bar {\bf X}\gamma\|_2^2
\label{eq-lhp}$$ where $N$ in Eq. is replaced by normalized gravity $\gamma$ due to the homogeneity of Eq. , and the squared norm is taken assuming the network predicts per-pixel depth $Z(x,y)$ up to additive zero-mean Gaussian noise. $\Omega_{HP} \subset \mathbb R^2$ is a subset of the image plane whose associated semantic classes have horizontal surfaces, such as “road”, “sidewalk”, “parking lot”, etc. We call this loss “horizontal plane” loss, where the direction of gravity $\gamma$ can be reliably and globally estimated.
Similarly, a “vertical plane” loss can be constructed to penalize deviation from a vertical plane whose normal $N$ has *both unknown direction and norm* but lives in the null space of $\gamma$, [*i.e.,*]{} $N \in {\mathcal N}(\gamma)$. Thus, the vertical plane loss reads $$L_{VP}(\Omega_{VP})=\min_{\substack{N \in \mathcal N(\gamma) \\ \|N\| =1}}
\frac{1}{|\Omega_{VP}|}\|({\bf I}-\frac{1}{|\Omega_{VP}|}{\bf 1}{\bf 1}^\top) \bar{\bf X}N)\|_2^2
\label{eq-lvp}$$ where the constraint $\|N\|=1$ avoids trivial solutions $N=0$ again due to the homogeneity of the objective; $\Omega_{VP}$ is a subset of the image plane whose associated semantic classes have vertical surfaces, such as “building”, “fence”, “billboard”, etc. The constrained minimization problem in the vertical plane loss $L_{VP}$ is due to the unknown direction of the surface normals and introduces some difficulties in training. We discuss approximations in Sect. \[sect-derivation\].
![*Illustration of geo-supervised visual depth prediction.* Our visual depth predictor is an encoder-decoder convolutional neural network with skip connections. At inference time, the network takes an RGB image as the only input and outputs an inverse depth map. At training time, gravity extracted from inertial measurements biases the depth prediction *selectively*, which is informed by semantic segmentation produced by PSPNet. The other identical stream of the network and the photometric losses used for training are omitted in this figure.[]{data-label="fig-diagram"}](figures/system_diagram.jpg){width="0.85\linewidth"}
Explanation of the objectives {#sect-derivation}
-----------------------------
Our idea is essentially to use priors about surface normals to regularize depth prediction. An intuitive way to achieve this is to compute the surface normals from the depth values first and then impose regularity, which will eventually bias the depth predictor via backpropagation. However, such a method involves normal estimation from depth, which can be problematic, especially with a simplistic but noisy normal estimator [@qi2018geonet].[^3] On the other hand, one could train a deep network to compute surface normals [@wang2016surge], which is costly. Therefore, *we do not compute surface normals but directly regularize the depth values* via the scale-invariant constraint Eq. which is a function of depth and the direction of gravity.
In what follows, we give an explanation of $L_{HP}$ Eq. from a statistical perspective. Let $M=|\Omega_{HP}|$ to avoid notation clutter and expand Eq. : $$\begin{aligned}
& ({\bf I}-\frac{1}{M}{\bf 1}{\bf 1}^\top) \bar {\bf X} \gamma \\
= &
\begin{bmatrix}
1-\frac{1}{M} & \cdots & -\frac{1}{M}\\
\vdots & \ddots & \vdots\\
-\frac{1}{M} & \cdots & 1-\frac{1}{M}
\end{bmatrix}
\begin{bmatrix}
{\bf X}_1^\top \gamma\\
{\bf X}_2^\top \gamma\\
\cdots\\
{\bf X}_M^\top \gamma
\end{bmatrix}
=
\begin{bmatrix}
\vdots\\
\big( {\bf X}_i - \frac{1}{M}\sum_{j=1}^M {\bf X}_j \big)^\top \gamma \\
\vdots
\end{bmatrix}\end{aligned}$$ Let $\mu=\frac{1}{M}\sum_{j=1}^M \mathbf{X}_j$ be the sample mean of the 3D coordinates and the horizontal plane loss $L_{HP}$ reads $$L_{HP}(\Omega_{HP})= \frac{1}{M}\sum_{i=1}^M \big( ({\bf X}_i - \mu)^\top \gamma \big)^2$$ which is the sample variance of the 3D coordinates projected to the direction of gravity $\gamma$ (coinciding with the surface normal for horizontal planes). To minimize $L_{HP}$ is to minimize the variance of the 3D coordinates along the surface normal.
Similarly, to minimize $L_{VP}$ Eq. is to minimize the variance of the 3D coordinates along some direction perpendicular to gravity. However, if the direction is unknown, one needs to jointly solve the direction while minimizing $L_{VP}$, which explains the constrained quadratic problem in $L_{VP}$. Though this can be solved via eigendecomposition, the gradients of the solver – needed in backpropagation – are non-trivial to compute. In fact, representing an optimization procedure as a layer of a neural network is an open research problem [@amos2017optnet]. To alleviate both numerical and implementation difficulties, we uniformly sample unit vectors from the null space of gravity and compute the minimum of the objective over the samples as an approximation to the loss. Empirically, we found using eight directions sampled every 45 degrees from 0 to 360 generally performs well.
View synthesis as supervision and baselines {#sect-baseline}
-------------------------------------------
To showcase the ability to improve upon existing self-supervised monocular depth prediction networks, we add our losses to two publicly available models – Godard [@godard2017unsupervised] ([`LR-Consistency`]{}) and Yin [@yin2018geonet] ([`GeoNet`]{}) – as baselines and perform both quantitative and qualitative comparisons. We additionally apply our losses to Zhan [@zhan2018unsupervised] ([`Stereo-Temporal`]{}) and Wang [@wang2017learning] ([`DDVO`]{}), the state-of-the-art methods in their respective training setting, stereo pairs/videos, and monocular videos. [`LR-Consistency`]{} is trained with rectified stereo image pairs, [`GeoNet`]{} and [`DDVO`]{} use monocular videos while [`Stereo-Temporal`]{} uses stereo videos. At test time, all training settings result in a system that takes a single image as input and predicts an inverse depth map as output. We show that by applying our losses to the baselines [`LR-Consistency`]{} and [`GeoNet`]{}, we achieve better performance than the state-of-the-art methods [`Stereo-Temporal`]{} and [`DDVO`]{}. Furthermore, we produce new state-of-the-art results by applying our losses to [`Stereo-Temporal`]{} and [`DDVO`]{}.
### Training with stereo pairs {#sect-method-stereo}
At training time, our first baseline model ([`LR-Consistency`]{}) takes a single left image as its input and predicts two disparity maps $D^L, D^R: \mathbb{R}^2 \supset \Omega \rightarrow \mathbb{R}_+$ for both left and right cameras. The network follows the fully convolutional encoder-decoder structure with skip connections. The total loss consists of three terms: Appearance loss, smoothness of disparity and left-right consistency, each of which is evaluated on both the left and the right streams across multiple scale levels. Here we address the view synthesis loss, which serves as the data term and is part of the appearance loss: $$\begin{aligned}
L_\text{vs}^L= \frac{1}{|\Omega|} \sum_{(x,y) \in \Omega} \| I^L(x,y) - I^R(x+D^L(x, y), y) \|_1.
\label{eq-stereo}\end{aligned}$$ The view synthesis loss is essentially the photometric difference of the left image $I^L(x, y)$ and the right image warped to the left view $I^R(x+D^L(x, y), y)$ according to the left disparity prediction $D^L(x, y)$. The right view synthesis loss is constructed in the same way. Though only one disparity map is needed at inference time, it has been shown that predicting both left and right disparity maps and including the left-right consistency loss Eq. are in general beneficial [@godard2017unsupervised].
$$L_\text{lr}^L = \frac{1}{|\Omega|} \sum_{(x,y) \in \Omega} \| D^L(x,y) - D^R(x+D^L(x, y), y)\|_1
\label{eq-consistency}$$
### Training with stereo videos {#sect-stereo-temporal}
In our second baseline [`Stereo-Temporal`]{}, stereo videos are used to train a monocular depth predictor, where two frames of a stereo pair and another frame one time step ahead are involved in constructing a stereo-temporal version of the photometric loss: For the stereo pair, Eq. is applied while for the temporal pair, Eq. (detailed below) is applied.
### Training with monocular videos {#sect-method-mono}
To train our third and fourth baseline models ([`GeoNet`]{} and [`DDVO`]{}), a single reference frame $I_t$ is fed into the depth network and frames $I_{t'}, t' \in W_t$ in a temporal window centered at $t$ are used to construct the view synthesis loss, also known as reprojection error: $$L_\text{vs}=\frac{1}{|W_t||\Omega|}
\sum_{t' \in W_t} \sum_{(x,y) \in \Omega}
\| I_{t}(x, y)-I_{t'}\big(\pi(\hat g_{t't} {\bf X})\big)\|_1
\label{eq-mono}$$ which is the difference between the reference frame $I_t$ and neighboring frames $I_{t'}$ warped to it. ${\bf X}$ is the back-projected point defined in Eq. , $\pi$ is a central (perspective) projection, and $\hat g_{t't}$ is the relative camera pose up to an unknown scale predicted by an auxiliary pose network which takes both $I_t$ and $I_{t'}$ as its input. Note that the pose and depth networks are coupled via the view synthesis loss at training time; at test time, the depth network alone is needed to perform depth prediction with a single image as its input. Interestingly, in Sect. \[sect-visma\] we found that replacing the pose network with pose estimation from VIO produces better results compared to the multi-task learning diagram where pose and depth networks are trained simultaneously, which sheds light on the use of classic SLAM/Odometry systems in developing better learning algorithms.
A detailed discussion about other losses serving as regularization terms is beyond the scope of this paper and can be found in [@godard2017unsupervised; @zhou2017unsupervised; @yin2018geonet; @wang2017learning].
<span style="font-variant:small-caps;">Implementation Details</span> {#sect-impl}
====================================================================
Semantic segmentation {#sect-pspnet}
---------------------
At training time, we use PSPNet [@zhao2017pyramid] pre-trained on the CityScapes dataset [@cordts2016cityscapes] provided by the authors to obtain per-pixel labeling. For every pixel $(x, y)\in \mathbb R^2$, a probability distribution over $19$ classes is predicted by PSPNet, of which the most likely class $c(x,y) \in C$ determines the orientation of the surface where the back-projected point ${\bf X}$ sits. We group the 19 classes into 7 categories[^4] according to the CityScapes benchmark and test our losses on all of them. Empirically, we found that it is most beneficial to apply our losses to the “flat”, “vehicle” and “construction” categories and therefore all the comparisons on KITTI against baseline methods are made with these categories regularized. The influence of other categories is studied in Sect. \[sect-ablation\].
Gravity {#sect-gravity}
-------
For imagery captured by a static platform equipped with an inertial measurement unit (IMU), one can use the gravity $\gamma_b \in \mathbb{R}^3$ measured in the body frame (coinciding with the IMU frame) and simply apply the body-to-camera rotation $R_{cb} \in \mathrm{SO}(3)$ to obtain the gravity in the camera frame $\gamma=R_{cb}\gamma_b$ which is then used in Eq. and . For moving platforms, one resorts to robust VIO, which is well studied [@mourikis2007multi; @tsotsos2015robust]. In Sect. \[sect-visma\], we demonstrated our approach on a visual-inertial odometry dataset, where both camera pose and gravity are estimated online by VIO.
For our experiments on the KITTI dataset, thanks to the GPS/IMU sensor package which provides linear acceleration of the sensor platform measured both in the body frame ($\alpha_b \in \mathbb{R}^3$) and the spatial frame ($\alpha_s \in \mathbb{R}^3$), we are able to compute the spatial-to-body rotation $R_{bs} \in \mathrm{SO}(3)$ and then bring the gravity $\gamma_s=[0,0,9.8]^\top$ from the spatial frame to the camera frame $\gamma=R_{cb}R_{bs}\gamma_s$. In all settings, $R_{cb}$ (rotational part of the body-to-camera transformation) is obtained via offline calibration procedures.
Training details
----------------
A GTX 1080 Ti GPU and Adam [@kingma2014adam] optimizer are used in our experiments. Depending on different model variants and input image sizes, training time varies from 8 hours to 16 hours. For [`LR-Consistency`]{} and [`GeoNet`]{} which were initially implemented in TensorFlow, we implemented our losses also in TensorFlow and applied them to the existing code bases. Code of [`Stereo-Temporal`]{} is available online, but in Caffe, thus we migrated their model to TensorFlow and applied our losses. We also implemented our losses in PyTorch, which were then applied to [`DDVO`]{} of which the PyTorch version was made available by the author. Our code is available at <https://github.com/feixh/GeoSup>.
<span style="font-variant:small-caps;">Experiments</span> {#sect-exp}
=========================================================
To enable quantitative evaluation, we exploit the KITTI benchmark, and test our approach against the state-of-the-art as described in detail below (Sect. \[sect-stereo\]&\[sect-mono\]). We also carried out ablation studies (Sect. \[sect-ablation\]) and tested the generalizability of our approach (Sect. \[sect-make3d\]). In addition to KITTI, which features planar motion in driving scenarios, we have conducted experiments on VISMA dataset [@fei2018visual] – an indoor visual-inertial odometry dataset captured under non-trivial ego-motion (Sect. \[sect-visma\]).
KITTI Eigen split
-----------------
We compare our approach with recent state-of-the-art methods on the monocular depth prediction task using the KITTI Eigen split [@eigen2014depth] in two training domains: stereo pairs/videos and monocular videos (Sect. \[sect-baseline\]). The Eigen split test set contains 697 test images selected from 29 of 61 scenes provided by the raw KITTI dataset. Of the remaining 32 scenes containing 23,488 stereo pairs, 22,600 pairs are used for training, and the rest is used for validation per the training split proposed by [@garg2016unsupervised]. To generate ground truth depth maps for validation and evaluation, we take the Velodyne data points associated with each image and project them from the Velodyne frame to the left RGB camera frame. Each resulting ground truth depth map covers approximately $5\%$ of the corresponding image and may be erroneous. To handle this, first, we use the cropping scheme proposed by [@garg2016unsupervised], which masks out the potentially erroneous extremities from the left, right and top areas of the ground truth depth map. Then we evaluate depth prediction only at pixels where ground truth depth is available. For visualization, we linearly interpolate each sparse depth map to cover the entire image (Fig. \[fig-kitti\]).
We additionally provide quantitative evaluations of variants of the models pre-trained on CityScapes and fine-tuned on KITTI. CityScapes dataset contains 22,973 training stereo pairs captured in various cities across Germany with a similar modality as KITTI. We cropped each input image to keep only the top 80% of the image, removing the reflective hood.
The error and accuracy metrics, which are initially proposed by [@eigen2014depth] and adopted by others, are used (Table \[tab-metric\]). Also as a convention in the literature, performances evaluated with depth prediction capped at 50 and 80 meters are reported as suggested by [@godard2017unsupervised]. The choice of 80 meters is two-fold: 1) maximum depth present in the KITTI dataset is on the order of 80 meters and 2) non-thresholded measures can be sensitive to the significant errors in depth caused by prediction errors at small disparity values. For the same reason, depth prediction is capped at 70 meters in the Make3D experiment. Prediction capped at 50 meters is also evaluated since depth at closer range is more applicable to real-world scenarios.
Metric Definition
------------- --------------------------------------------------------------------------------------------------------------------------------------
AbsRel $\frac{1}{|\Omega|}\sum_{(x,y) \in \Omega} \frac{|Z(x,y)-Z^\text{gt}(x,y)|}{Z^\text{gt}(x,y)}$
SqRel $\frac{1}{|\Omega|}\sum_{(x,y)\in\Omega}\frac{|Z(x,y)- Z^\text{gt}(x,y)|^2}{Z^\text{gt}(x,y)}$
RMSE $\sqrt{\frac{1}{|\Omega|} \sum_{(x,y) \in \Omega}|Z(x,y) - Z^\text{gt}(x,y) |^2}$
RMSE log $\sqrt{\frac{1}{|\Omega|} \sum_{(x,y) \in \Omega}|\log Z(x,y) - \log Z^\text{gt}(x,y)|^2}$
$\log_{10}$ $\frac{1}{|\Omega|}\sum_{(x,y)\in\Omega}|\log Z(x,y)-\log Z^\text{gt}(x,y)|$
Accuracy % of $Z(x,y)$ s.t. $\delta \doteq \max\big(\frac{Z(x,y)}{Z^\text{gt}(x,y)}, \frac{Z^\text{gt}(x,y)}{Z(x,y)}\big) < \text{threshold}$
: Error and Accuracy Metrics[]{data-label="tab-metric"}
$Z(x,y)$ is the predicted depth at $(x,y) \in \Omega$ and $Z^\text{gt}(z,y)$ is the corresponding ground truth. Three different thresholds ($1.25, 1.25^2$ and $1.25^3$) are used in the accuracy metric as a convention in the literature.
Training with stereo pairs {#sect-stereo}
--------------------------
The first baseline we adopt is Godard [@godard2017unsupervised] (with [`VGG`]{} [@simonyan2014very] as feature extractor), to which SIGL is imposed at training time along with the view synthesis loss Eq. and other generic regularizers used in [@godard2017unsupervised]. The model is trained from scratch with stereo pairs following the Eigen split and compared to both supervised [@eigen2014depth; @liu2016learning] and self-supervised methods [@godard2017unsupervised; @zhan2018unsupervised]. In addition, we apply our losses to variants of the baseline (with [`ResNet`]{} [@he2016deep] as feature extractor; w/ & w/o post-processing) and evaluate different training schemes (w/ & w/o pre-training on CityScapes). Quantitative comparisons can be found in Table \[tab-stereo\], where the results with SIGL added as an additional regularizer follow the results of the baseline models and variants. In the column marked “Data”, `K` refers to Eigen split benchmark on the KITTI dataset, and `CS` refers to the CityScapes dataset. Methods marked with `CS+K` are pre-trained on CityScapes and then fine-tuned on KITTI Eigen split. `pp` denotes post-processing. Cap $X$m means depth predictions are capped at $X$ meters. Results of Zhan [@zhan2018unsupervised] [`Stereo-Temporal`]{} are taken from their paper. The rest of the results are taken from [@godard2017unsupervised] unless otherwise stated.
We want to remind the reader that the first baseline model atop which we built ours is Godard [@godard2017unsupervised] [`VGG`]{} which initially performed worse than the [`Stereo-Temporal`]{} model of Zhan [@zhan2018unsupervised] by a large margin, but by applying our losses to the baseline at training time we managed to boost its performance and make it perform even better than the [`Stereo-Temporal`]{} model at test time. Note that the [`Stereo-Temporal`]{} model also exploits temporal information in addition to stereo pairs for training while our first baseline built atop Godard does not.
As a second baseline, we apply our losses additionally to the [`Stereo-Temporal`]{} model of Zhan to further push the state-of-the-art. Table \[tab-stereo\] shows that our losses improve the [`Stereo-Temporal`]{} model across all error metrics with the accuracy metrics $\delta < 1.25^2$ and $\delta < 1.25^3$ being comparable. Another variant of Zhan’s model pre-trains on NYU-V2 [@silberman2012nyuv2] in a fully supervised fashion and is therefore not pertinent to this comparison. Fig. \[fig-kitti\] shows a head-to-head qualitative comparison of ours and the baseline models.
\[tab-stereo\]
[l|c|cccc|ccc]{} Method & Data & &\
& & [AbsRel]{} & [SqRel]{} & [RMSE]{} & [RMSElog]{} & $1.25$ & $1.25^2$ & $1.25^3$\
\
TrainSetMean & `K` & 0.361 & 4.826 & 8.102 & 0.377 & 0.638 & 0.804 & 0.894\
Eigen [@eigen2014depth] [`Coarse`]{} & `K` & 0.214 & 1.605 & 6.563 & 0.292 & 0.673 & 0.884 & 0.957\
Eigen [@eigen2014depth] [`Fine`]{} & `K` & 0.203 & 1.548 & 6.307 & 0.282 & 0.702 & 0.890 & 0.958\
Liu [@liu2016learning] & `K` & 0.201 & 1.584 & 6.471 & 0.273 & 0.680 & 0.898 & 0.967\
Godard [@godard2017unsupervised] [`VGG`]{} & `K` & 0.148 & 1.344 & 5.927 & 0.247 & 0.803 & 0.922 & 0.964\
+SIGL & `K` & [**0.139**]{} & [**1.211**]{} & [**5.702**]{} & [**0.239**]{} & [**0.816**]{} & [**0.928**]{} & [**0.966**]{}\
Zhan [@zhan2018unsupervised] [`Stereo-Temporal`]{} & `K` & 0.144 & 1.391 & 5.869 & 0.241 & 0.803 & 0.928 & 0.969\
+SIGL & `K` & [**0.137**]{} & [**1.061**]{} & [**5.692**]{} & [**0.239**]{} & [**0.805**]{} & 0.928 & 0.969\
Godard [@godard2017unsupervised] [`VGG pp`]{} & `CS+K` & 0.124 & 1.076 & 5.311 & 0.219 & 0.847 & 0.942 & 0.973\
+SIGL & `CS+K` & [**0.114**]{} & [**0.885**]{} & [**4.877**]{} & [**0.203**]{} & [**0.858**]{} & [**0.950**]{} & [**0.978**]{}\
Godard [@godard2017unsupervised] [`ResNet pp`]{} & `CS+K` & 0.114 & 0.898 & 4.935 & 0.206 & 0.861 & 0.949 & 0.976\
+SIGL & `CS+K` & [**0.112**]{} & [**0.836**]{} & [**4.892**]{} & [**0.204**]{} & [**0.862**]{} & [**0.950**]{} & [**0.977**]{}\
\
Garg [@garg2016unsupervised] & `K` & 0.169 & 1.080 & 5.104 & 0.273 & 0.740 & 0.904 & 0.962\
Godard [@godard2017unsupervised] [`VGG`]{} & `K` & 0.140 & 0.976 & 4.471 & 0.232 & 0.818 & 0.931 & 0.969\
+SIGL & `K` & [**0.132**]{} & [**0.891**]{} & [**4.312**]{} & [**0.225**]{} & [**0.831**]{} & [**0.936**]{} & [**0.970**]{}\
Zhan [@zhan2018unsupervised] [`Stereo-Temporal`]{} & `K` & 0.135 & 0.905 & 4.366 & 0.225 & 0.818 & 0.937 & 0.973\
+SIGL & `K` & [**0.131**]{} & [**0.829**]{} & [**4.217**]{} & [**0.224**]{} & [**0.824**]{} & 0.937 & 0.973\
Godard [@godard2017unsupervised] [`VGG pp`]{} & `CS+K` & 0.112 & 0.680 & 3.810 & 0.198 & 0.866 & 0.953 & 0.979\
+SIGL & `CS+K` & [**0.108**]{} & [**0.658**]{} & [**3.728**]{} & [**0.192**]{} & [**0.870**]{} & [**0.955**]{} & [**0.981**]{}\
Godard [@godard2017unsupervised] [`ResNet pp`]{} & `CS+K` & 0.108 & 0.657 & 3.729 & 0.194 & 0.873 & 0.954 & 0.979\
+SIGL & `CS+K` & [**0.106**]{} & [**0.615**]{} & [**3.697**]{} & [**0.192**]{} & [**0.874**]{} & [**0.956**]{} & [**0.980**]{}\
With ground truth depth supervision.\
+SIGL: training with SIGL enabled
Training with monocular videos {#sect-mono}
------------------------------
To demonstrate the effectiveness of our loss in the second training setting (monocular videos), we impose SIGL to our third (Yin [@yin2018geonet]) and fourth (Wang [@wang2017learning]) baseline. Using the KITTI Eigen split, we follow the training and validation 3-frame sequence selection proposed by [@zhou2017unsupervised] where the first and third frames are treated as the source views and the central (second) frame is treated as the reference as in Eq. . Of the 44,540 total sequences, 40,109 are used for training and 4,431 for validation. We evaluate our system on the aforementioned 697 test images [@eigen2014depth]. The same training and evaluation scheme are also applied to other top-performing methods [@zhou2017unsupervised; @mahjourian2018unsupervised] in addition to the selected baselines.
Table \[tab-mono\] shows detailed comparisons against state-of-the-art self-supervised methods trained using monocular video sequences. We compare against best-performing model variants of Wang [@wang2017learning] ([`PoseCNN`]{} & [`PoseCNN+DDVO`]{}) and Yin [@yin2018geonet] ([`ResNet`]{}) with and without pre-training on CityScapes. By adding our losses to existing models, we observe systematic performance improvement across all metrics. Though initially performing worse than Wang [@wang2017learning] [`PoseCNN+DDVO`]{}, Yin [@yin2018geonet] [`ResNet`]{} with the proposed losses even outperforms the original [`PoseCNN+DDVO`]{}. Moreover, we achieve new state-of-the-art by adding our losses to [`PoseCNN+DDVO`]{} trained on both CityScapes and KITTI. Fig. \[fig-kitti\] illustrates representative image regions where we do better.
[l|c|cccc|ccc]{} Method & Data & &\
& & [AbsRel]{} & [SqRel]{} & [RMSE]{} & [RMSElog]{} & $1.25$ & $1.25^2$ & $1.25^3$\
\
Zhou [@zhou2017unsupervised] & `K` & 0.208 & 1.768 & 6.856 & 0.283 & 0.678 & 0.885 & 0.957\
Mahjourian [@mahjourian2018unsupervised] & `K` & 0.163 & 1.240 & 6.220 & 0.250 & 0.762 & 0.916 & 0.968\
Yin [@yin2018geonet] [`ResNet`]{} & `K` & 0.155 & 1.296 & 5.857 & 0.233 & 0.793 & 0.931 & 0.973\
+SIGL & `K` & [**0.142**]{} & [**1.124**]{} & [**5.611**]{} & [**0.223**]{} & [**0.813**]{} & [**0.938**]{} & [**0.975**]{}\
Wang [@wang2017learning] [`PoseCNN`]{} & `K` & 0.155 & 1.193 & [**5.613**]{} & 0.229 & 0.797 & 0.935 & 0.975\
+SIGL & `K` & [**0.147**]{} & [**1.076**]{} & 5.640 & [**0.227**]{} & [**0.801**]{} & 0.935 & 0.975\
Wang [@wang2017learning] [`PoseCNN+DDVO`]{} & `K` & 0.151 & 1.257 & 5.583 & 0.228 & [**0.810**]{} & 0.936 & 0.974\
+SIGL & `K` & [**0.146**]{} & [**1.068**]{} & [**5.538**]{} & [**0.224**]{} & 0.809 & [**0.938**]{} & [**0.975**]{}\
Zhou [@zhou2017unsupervised] & `CS+K` & 0.198 & 1.836 & 6.565 & 0.275 & 0.718 & 0.901 & 0.960\
Mahjourian [@mahjourian2018unsupervised] & `CS+K` & 0.159 & 1.231 & 5.912 & 0.243 & 0.784 & 0.923 & 0.970\
Yin [@yin2018geonet] [`ResNet`]{} & `CS+K` & 0.153 & 1.328 & 5.737 & 0.232 & 0.802 & 0.934 & 0.972\
+SIGL & `CS+K` & [**0.147**]{} & [**1.076**]{} & [**5.468**]{} & [**0.222**]{} & [**0.806**]{} & [**0.938**]{} & [**0.976**]{}\
Wang [@wang2017learning] [`PoseCNN+DDVO`]{} & `CS+K` & 0.148 & 1.187 & 5.496 & 0.226 & 0.812 & 0.938 & 0.975\
+SIGL & `CS+K` & [**0.142**]{} & [**1.094**]{} & [**5.409**]{} & [**0.219**]{} & [**0.821**]{} & [**0.941**]{} & [**0.976**]{}\
\
Zhou [@zhou2017unsupervised] & `K` & 0.201 & 1.391 & 5.181 & 0.264 & 0.696 & 0.900 & 0.966\
Mahjourian [@mahjourian2018unsupervised] & `K` & 0.155 & 0.927 & 4.549 & 0.231 & 0.781 & 0.931 & 0.975\
Yin [@yin2018geonet] [`ResNet`]{} & `K` & 0.147 & 0.936 & 4.348 & 0.218 & 0.810 & 0.941 & 0.977\
+SIGL & `K` & [**0.135**]{} & [**0.834**]{} & [**4.193**]{} & [**0.208**]{} & [**0.831**]{} & [**0.948**]{} & [**0.979**]{}\
Wang [@wang2017learning] [`PoseCNN`]{} & `K` & 0.149 & 0.920 & 4.303 & 0.216 & 0.813 & 0.943 & 0.979\
+SIGL & `K` & [**0.140**]{} & [**0.816**]{} & [**4.234**]{} & [**0.212**]{} & [**0.818**]{} & [**0.945**]{} & [**0.980**]{}\
Wang [@wang2017learning] [`PoseCNN+DDVO`]{} & `K` & 0.144 & 0.935 & 4.234 & 0.214 & [**0.827**]{} & 0.945 & 0.977\
+SIGL & `K` & [**0.139**]{} & [**0.808**]{} & [**4.180**]{} & [**0.209**]{} & 0.826 & [**0.948**]{} & [**0.980**]{}\
Zhou [@zhou2017unsupervised] & `CS+K` & 0.190 & 1.436 & 4.975 & 0.258 & 0.735 & 0.915 & 0.968\
Mahjourian [@mahjourian2018unsupervised] & `CS+K` & 0.151 & 0.949 & 4.383 & 0.227 & 0.802 & 0.935 & 0.974\
Yin [@yin2018geonet] [`ResNet`]{} & `CS+K` & / & / & / & / & / & / & /\
+SIGL & `CS+K` & [**0.141**]{} & [**0.837**]{} & [**4.160**]{} & [**0.209**]{} & [**0.823**]{} & [**0.947**]{} & [**0.980**]{}\
Wang [@wang2017learning] [`PoseCNN+DDVO`]{} & `CS+K` & 0.142 & 0.901 & 4.202 & 0.213 & 0.827 & 0.946 & 0.978\
+SIGL & `CS+K` & [**0.135**]{} & [**0.832**]{} & [**4.119**]{} & [**0.206**]{} & [**0.836**]{} & [**0.949**]{} & [**0.980**]{}\
Not available.
Evaluated with prediction released by the author.
+SIGL: training with SIGL enabled
-------------------------------------------------------------------- ------------------------------------------------------------------ ----------------------------------------------------------------- ------------------------------------------------------------------ ---------------------------------------------------------------------
{width="0.23\linewidth"} {width="0.23\linewidth"} {width="0.23\linewidth"} {width="0.23\linewidth"} {width="0.014\linewidth"}
-------------------------------------------------------------------- ------------------------------------------------------------------ ----------------------------------------------------------------- ------------------------------------------------------------------ ---------------------------------------------------------------------
Ablation study {#sect-ablation}
--------------
To study the contribution of each semantic category to the performance improvement, we performed an ablation study: We apply our losses to different semantic categories, one at a time, train the network until convergence, and show how the quality of depth prediction varies (Table \[tab-ablation\]). In Table \[tab-ablation\], Godard *et al*. [@godard2017unsupervised] is the baseline model where only the most generic regularizers, [*e.g.,*]{} smoothness and consistency, are used. The second column indicates the semantic category of which the depth prediction is regularized using our losses in addition to the generic regularizers. For the meaning of the semantic categories, see Sect. \[sect-pspnet\].
It turns out that the “flat” category contributes most to the performance gain over the baseline model, which is expected because most of the KITTI images contain a large portion of roads and sidewalks. We also observed that regularization of the “construction” and “vehicle” category provides reasonable improvement while the “nature” category (trees and hedges) helps a little. Applying our priors to the “human”, “sky” and “object” categories does not consistently improve over the baseline, for the following reasons: “sky” does not have well-defined surface normals; “human” has deformable surfaces of which normals can point arbitrarily; “object” category consists of thin structures which project to few pixels rendering it hard to apply segmentation and our losses. The best is achieved when we apply our losses to “vehicle”, “construction” and “flat” categories, denoted by `V+C+F` in Table \[tab-ablation\].
---------------------------------- -------------------- ------------ ----------- ---------- ------------- -------- ---------- ----------
Method Category
[AbsRel]{} [SqRel]{} [RMSE]{} [RMSElog]{} $1.25$ $1.25^2$ $1.25^3$
Godard [@godard2017unsupervised] `/` 0.148 1.344 5.927 0.247 0.803 0.922 0.964
Ours [`Human`]{} 0.152 1.394 5.945 0.251 0.801 0.921 0.963
Ours [`Sky`]{} 0.148 1.368 5.864 0.245 0.807 0.923 0.964
Ours [`Object`]{} 0.146 1.335 5.986 0.249 0.800 0.920 0.963
Ours [`Nature`]{} 0.146 1.292 5.826 0.247 0.804 0.923 0.964
Ours [`Vehicle`]{} 0.143 1.304 5.797 0.241 0.814 0.927 0.966
Ours [`Construction`]{} 0.142 1.252 5.729 0.240 0.810 0.928 0.967
Ours [`Flat`]{} 0.141 1.270 5.779 0.239 0.814 0.927 0.966
Ours `V+C+F` 0.139 1.211 5.702 0.239 0.816 0.928 0.966
---------------------------------- -------------------- ------------ ----------- ---------- ------------- -------- ---------- ----------
: Ablation study on KITTI.[]{data-label="tab-ablation"}
Generalize to other datasets: Make3D {#sect-make3d}
------------------------------------
To showcase the generalizability of our approach, we follow the convention of [@godard2017unsupervised; @zhou2017unsupervised; @yin2018geonet; @wang2017learning]: Our model trained *only* on KITTI Eigen split is directly tested on Make3D [@saxena2009make3d]. Make3D contains 534 images with $2272 \times 1707$ resolution, of which 134 are used for testing.[^5] Low resolution ground truth depths are given as $305 \times 55$ range maps and must be resized and interpolated for evaluation. We follow [@godard2017unsupervised] and [@zhou2017unsupervised] in applying a central cropping to generate a $852 \times 1707$ crop centered on the image. We use the standard $C1$ evaluation metrics for Make3D and measure our performance on depths less than 70 meters. Table \[tab-make3d\] shows a quantitative comparison to the competitors, both supervised and self-supervised, with two different training settings. Note that the results of [@karsch2012depth; @liu2016learning; @laina2016deeper] are directly taken from [@godard2017unsupervised]. Since the exact cropping scheme used in [@godard2017unsupervised] is not available, we re-implemented it closely following the description in [@godard2017unsupervised]. We trained our model on KITTI Eigen split and compared against models of [@godard2017unsupervised; @zhou2017unsupervised; @yin2018geonet; @wang2017learning] also trained on Eigen split (as provided by the authors) for a fair comparison.
A careful inspection of the baseline models (Godard [@godard2017unsupervised] in stereo and Yin [@yin2018geonet] in monocular supervision) versus ours reveals that the application of our losses does not hurt the generalizability of the baselines. Fig. \[fig-make3d\] shows some qualitative results on Make3D. Though our model registers some failure cases in texture-less regions, a rough scene layout is present in the prediction. Regarding that the model is only trained on KITTI, of which the data modality is very different from that of Make3D, the prediction is sensible. But after all, a single image only affords to hypothesize depth, so we expect that any method using such predictions would have mechanisms to handle model deficiencies.
Method Supervision AbsRel SqRel RMSE $\log_{10}$
-------------------------------------------- ------------- --------------- --------------- ---------------- ---------------
TrainSetMean Depth 0.893 15.517 11.542 0.223
Karsch [@karsch2012depth] Depth 0.417 4.894 8.172 0.144
Liu [@liu2016learning] Depth 0.462 6.625 9.972 0.161
Laina [@laina2016deeper] Depth [**0.198**]{} [**1.665**]{} [**5.461**]{} [**0.082**]{}
Godard [@godard2017unsupervised] [`VGG`]{} Stereo 0.468 9.236 12.525 0.165
**Ours** Stereo [**0.458**]{} [**8.681**]{} [**12.335**]{} [**0.164**]{}
Zhou [@zhou2017unsupervised] Mono 0.407 5.367 11.011 0.167
Yin [@yin2018geonet][`ResNet`]{} Mono 0.376 4.645 10.350 0.152
Wang [@wang2017learning][`PoseCNN+DDVO`]{} Mono 0.387 4.720 [**8.09**]{} 0.204
[**Ours**]{} Mono [**0.356**]{} [**4.517**]{} 10.047 [**0.144**]{}
: Generalizability test on Make3D.[]{data-label="tab-make3d"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0003_make3d_im.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0003_make3d_gt.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0003_make3d_d.jpg "fig:"){width="0.30\linewidth"}
![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0058_make3d_im.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0058_make3d_gt.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0058_make3d_d.jpg "fig:"){width="0.30\linewidth"}
![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0093_make3d_im.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0093_make3d_gt.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative results on Make3D.* Left to right, each row shows an input RGB image, the corresponding ground truth disparity map and our prediction. Our model is [*only*]{} trained on KITTI and directly applied to Make3D.[]{data-label="fig-make3d"}](figures/make3d/0093_make3d_d.jpg "fig:"){width="0.30\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Evaluation on indoor datasets {#sect-visma}
-----------------------------
To the best of our knowledge, none of the top-performing methods in self-supervised depth prediction have shown experimental results beyond planar motion, [*i.e.,*]{} driving scenarios such as KITTI and CityScapes, probably due to two reasons: Lack of rectified stereo pairs for training ([@godard2017unsupervised; @zhan2018unsupervised]) and difficulty to learn complex ego-motion along with depth prediction from video sequences ([@zhou2017unsupervised; @yin2018geonet; @wang2017learning]).
However, with two modifications to the [`GeoNet`]{} model of Yin [@yin2018geonet] – a multi-task learning approach where ego-motion and depth prediction are jointly learned, we managed to train our model and outperform [`GeoNet`]{} on publicly available VISMA [@fei2018visual] dataset which features monocular videos of indoor scenes captured by a hand-held visual-inertial sensor platform under challenging motion. As a first modification, we replace the pose network in [`GeoNet`]{} with pose estimation from a VIO system [@tsotsos2015robust], which makes the network easier to train (we call this model [`OursVIO`]{}). Second, to further improve the quality of predicted depth maps, we impose our gravity-induced regularization terms to [`OursVIO`]{}, where gravity is also estimated online by VIO. Our second model is named [`OursVIO++`]{}.
VISMA dataset contains time-stamped monocular videos ($30$ Hz) from a PointGrey camera and inertial measurements ($100$ Hz) from an Xsens unit, which are used in both VIO and network training. RGB-D reconstructions (dense point clouds) of the same scenes from a Kinect are also available, along with the spatial alignment $g_{\text{VIO}\leftarrow\text{RGBD}} \in {\mathrm{SE(3)}}$ from RGB-D to VIO provided by the author. To get ground truth depth for cross-modality validation, we apply $g_{\text{VIO}\leftarrow\text{RGBD}}$ to the dense point clouds which are then projected to the PointGrey video frames. PSPNet trained on ADE20K [@zhou2017scene] produces segmentation masks for training.[^6] Of the $10K$ frames in VISMA, we remove static ones and construct 3-frame sequences (triplet) which are five frames apart in the original video to ensure sufficient parallax, resulting $8,511$ triplets in total. We randomly sample $100$ triplets for validation and use the rest for training. Fig. \[fig-visma\] and Table \[tab-visma\] show comparisons of [`GeoNet`]{}, [`OursVIO`]{} and [`OursVIO++`]{}, all trained from scratch on VISMA until validation error stops decreasing. Both [`OursVIO`]{} and [`OursVIO++`]{} improve over the baseline model by a large margin. Moreover, [`OursVIO++`]{} trained with our gravity-induced losses has the capability to further refine results of [`OursVIO`]{} trained without our losses.
------------- --------------- --------------- --------------- --------------- --------------- --------------- ----------
Method
[AbsRel]{} [SqRel]{} [RMSE]{} [RMSElog]{} $1.25$ $1.25^2$ $1.25^3$
`GeoNet` 0.204 0.157 0.518 0.250 0.702 0.914 0.975
`OursVIO` 0.154 0.111 0.446 0.211 0.796 0.940 0.983
`OursVIO++` [**0.149**]{} [**0.105**]{} [**0.421**]{} [**0.202**]{} [**0.820**]{} [**0.947**]{} 0.983
------------- --------------- --------------- --------------- --------------- --------------- --------------- ----------
: Quantitative results on VISMA validation.[]{data-label="tab-visma"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![*Qualitative comparison on VISMA validation.* Top to bottom, each column shows an input RGB image, the corresponding ground truth inverse depth map, results of [`GeoNet`]{} (baseline), [`OursVIO`]{}, and [`OursVIO++`]{}. Both [`OursVIO`]{} and [`OursVIO++`]{} show largely improved results over the baseline, especially for images captured at extreme viewpoint (large in-plane rotation and top-down view). [`OursVIO++`]{} (with gravity-induced priors) further improves over [`OursVIO`]{} (without priors) at planar regions, [*e.g.,*]{} the chair backs, where holes have been filled.[]{data-label="fig-visma"}](figures/visma/visma_01_text.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative comparison on VISMA validation.* Top to bottom, each column shows an input RGB image, the corresponding ground truth inverse depth map, results of [`GeoNet`]{} (baseline), [`OursVIO`]{}, and [`OursVIO++`]{}. Both [`OursVIO`]{} and [`OursVIO++`]{} show largely improved results over the baseline, especially for images captured at extreme viewpoint (large in-plane rotation and top-down view). [`OursVIO++`]{} (with gravity-induced priors) further improves over [`OursVIO`]{} (without priors) at planar regions, [*e.g.,*]{} the chair backs, where holes have been filled.[]{data-label="fig-visma"}](figures/visma/visma_06_text.jpg "fig:"){width="0.30\linewidth"} ![*Qualitative comparison on VISMA validation.* Top to bottom, each column shows an input RGB image, the corresponding ground truth inverse depth map, results of [`GeoNet`]{} (baseline), [`OursVIO`]{}, and [`OursVIO++`]{}. Both [`OursVIO`]{} and [`OursVIO++`]{} show largely improved results over the baseline, especially for images captured at extreme viewpoint (large in-plane rotation and top-down view). [`OursVIO++`]{} (with gravity-induced priors) further improves over [`OursVIO`]{} (without priors) at planar regions, [*e.g.,*]{} the chair backs, where holes have been filled.[]{data-label="fig-visma"}](figures/visma/visma_20_text.jpg "fig:"){width="0.30\linewidth"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
<span style="font-variant:small-caps;">Discussion</span>
========================================================
Gravity informs the shape of objects populating the scene, which is a powerful prior to visual scene analysis. We have presented a simple illustration of this power by adding a prior to standard monocular depth prediction methods that biases the normals of surfaces of known classes to align to gravity or its complement. Far more can be done: While in this work we use known biases in the shape of certain object classes, such as the fact that roads tend to be perpendicular to gravity, in the future we could learn such biases directly.
[6]{}
B. Amos and J. Z. Kolter, “OptNet: Differentiable optimization as a layer in neural networks,” in *[ICML]{}*, 2017.
M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in *[CVPR]{}*, 2016.
D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,” in *[NIPS]{}*, 2014.
X. Fei and S. Soatto, “Visual-Inertial Object Detection and Mapping,” in *[ECCV]{}*, 2018.
R. Garg, V. K. BG, G. Carneiro, and I. Reid, “Unsupervised CNN for single view depth estimation: Geometry to the rescue,” in *[ECCV]{}*. Springer, 2016.
A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,” *[IJRR]{}*, 2013.
C. Godard, O. Mac Aodha, and G. J. Brostow, “Unsupervised monocular depth estimation with left-right consistency,” in *[CVPR]{}*, 2017.
C. Hane, C. Zach, A. Cohen, R. Angst, and M. Pollefeys, “Joint 3d scene reconstruction and class segmentation,” in *[CVPR]{}*, 2013.
K. He, X. Zhang, S. Ren, and S. Jian, “Deep residual learning for image recognition,” in *[CVPR]{}*, 2016.
O. H. Jafari, O. Groth, A. Kirillov, M. Y. Yang, and C. Rother, “Analyzing modular CNN architectures for joint depth prediction and semantic segmentation,” in *[ICRA]{}*, 2017.
K. Karsch, C. Liu, and S. B. Kang, “Depth extraction from video using non-parametric sampling,” in *[ECCV]{}*. Springer, 2012.
D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in *[ICLR]{}*, 2015.
J. Konrad, M. Wang, P. Ishwar, C. Wu, and D. Mukherjee, “Learning-based, automatic 2d-to-3d image and video conversion,” *[IEEE Trans. on Image Proc.]{}*, vol. 22, no. 9, pp. 3485–3496, 2013.
P. Kr[ä]{}henb[ü]{}hl and V. Koltun, “Efficient inference in fully connected crfs with gaussian edge potentials,” in *[NIPS]{}*, 2011.
A. Kundu, Y. Li, F. Dellaert, F. Li, and J. M. Rehg, “Joint semantic segmentation and 3d reconstruction from monocular video,” in *[ECCV]{}*. Springer, 2014.
I. Laina, C. Rupprecht, V. Belagiannis, F. Tombari, and N. Navab, “Deeper depth prediction with fully convolutional residual networks,” in *[3DV]{}*, 2016.
C. Liu, J. Yang, D. Ceylan, E. Yumer, and Y. Furukawa, “Planenet: Piece-wise planar reconstruction from a single rgb image,” in *[CVPR]{}*, 2018.
F. Liu, C. Shen, G. Lin, and I. Reid, “Learning depth from single monocular images using deep convolutional neural fields,” *[PAMI]{}*, vol. 38, no. 10, pp. 2024–2039, 2016.
J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in *[CVPR]{}*, 2015.
R. Mahjourian, M. Wicke, and A. Angelova, “Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints,” in *[CVPR]{}*, 2018.
N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox, “A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,” in *[CVPR]{}*, 2016.
A. I. Mourikis and S. I. Roumeliotis, “A multi-state constraint Kalman filter for vision-aided inertial navigation,” in *[ICRA]{}*, 2007.
X. Qi, R. Liao, Z. Liu, R. Urtasun, and J. Jia, “GeoNet: Geometric neural network for joint depth and surface normal estimation,” in *[CVPR]{}*, 2018.
C. Russell, P. Kohli, P. H. Torr, *et al.*, “Associative hierarchical crfs for object class image segmentation,” in *[ICCV]{}*, 2009.
A. Saxena, S. H. Chung, and A. Y. Ng, “Learning depth from single monocular images,” in *[NIPS]{}*, 2006.
A. Saxena, M. Sun, and A. Y. Ng, “Make3d: Learning 3d scene structure from a single still image,” *[PAMI]{}*, vol. 31, no. 5, pp. 824–840, 2009.
J. Shotton, M. Johnson, and R. Cipolla, “Semantic texton forests for image categorization and segmentation,” in *[CVPR]{}*, 2008.
N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmentation and support inference from RGBD images,” in *[ECCV]{}*, 2012.
K. Simonyan and A. Zisserman. “Very deep convolutional networks for large-scale image,” in *[ICLR]{}*, 2015.
K. Tsotsos, A. Chiuso, and S. Soatto, “Robust inference for visual-inertial sensor fusion,” in *[ICRA]{}*, 2015.
C. Wang, J. M. Buenaposada, R. Zhu, and S. Lucey, “Learning depth from monocular videos using direct methods,” in *[CVPR]{}*, 2018.
P. Wang, X. Shen, B. Russell, S. Cohen, B. Price, and A. L. Yuille, “Surge: Surface regularized geometry estimation from a single image,” in *[NIPS]{}*, 2016.
J. Xie, R. Girshick, and A. Farhadi, “Deep3d: Fully automatic 2d-to-3d video conversion with deep convolutional neural networks,” in *[ECCV]{}*. Springer, 2016.
Z. Yin and J. Shi, “Geonet: Unsupervised learning of deep depth, optical flow and camera pose,” in *[CVPR]{}*, 2018.
H. Zhan, R. Garg, C. S. Weerasekera, K. Li, H. Agarwal, and I. Reid, “Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction,” in *[CVPR]{}*, 2018.
H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” in *[CVPR]{}*, 2017.
T. Zhou, M. Brown, N. Snavely, and D. G. Lowe, “Unsupervised learning of depth and ego-motion from video,” in *[CVPR]{}*, 2017.
B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso and A. Torralba, “Scene parsing through ade20k dataset”,” in *[CVPR]{}*, 2017.
[^1]: This work was supported by ONR N00014-17-1-2072 and ARO W911NF-17-1-0304.
[^2]: The authors are with the Computer Science Department, University of California, Los Angeles, USA. Email: `{feixh, alexw, soatto}@cs.ucla.edu`
[^3]: For instance, one can compute the point-wise surface normal as the cross product of two vectors tangent to the surface, where the tangent vectors are approximated by connecting the underlying point to its nearest neighbors on the surface.
[^4]: “flat”: road, sidewalk; “human”: rider, person; “vehicle”: car, truck, bus, train, motorcycle, bicycle; “construction”: building, wall, fences; “object”: pole, traffic light, traffic sign; “nature”: vegetation, terrain; “sky”: sky.
[^5]: Ideally we want to test on the whole Make3D dataset since we do not train on Make3D, but other methods to which we compare train on it. For a fair comparison, we only use the 134 images for testing.
[^6]: Among the 91 categories in ADE20K which PSPNet is trained on, we select “floor”, “ceiling”, “wall”, “window”, “door”, “building”, “chair”, “cabinet”, “desk”, “table” to apply our losses.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Center for Radiophysics and Space Research, Cornell University,\
Ithaca, NY 14853
author:
- 'A. SEDRAKIAN & J. CORDES'
title: 'GENERATION OF PULSAR GLITCHES: A SUPERFLUID CORE MODEL [^1]'
---
Models of generation of pulsar glitches are supposed to explain the following observational facts: (i) the short spin-up time scales, which are less than 120 s in the Vela pulsar 0833–45 and less than an hour in the Crab pulsar 0531+21; (ii) the magnitudes of the jumps in the rotation and spin down rates, $\Delta\nu/\nu \sim 10^{-8}-10^{-6}$ and $\Delta\dot\nu/\dot\nu \sim 10^{-3}-10^{-2}$, respectively; and (iii) the origin of the instability driving a glitch along with characteristic intervals between glitches (typically of the order of several months to years). A number of existing generic models invoke as trigger crust- and core-quakes [@RUDERMAN] (discontinuous adjustments of the solid crust to the gradually changing oblateness of the star as it spins down); spontaneous quantum transitions of rotating superfluid between quasistationary eigenstates corresponding to different eigenvalues of total angular momentum [@PACKARD]; collective unpinning of a large number ($\sim 10^{13}$) of vortices in the neutron star crusts [@ANDERSON; @ALPARETAL] and thermal instabilities [@LINK]. The increasing bulk of observational evidence provides good chances of discrimination between the theoretical models in the future.
Here we shall give a brief account of a new trigger mechanism for generation of pulsar glitches in neutron star’s superfluid core. The complete discussion will be given in ref. 6. In the interjump epoch a neutron star is decelerating; consequently the vortex lattice in the superfluid core is expanding and the peripheral vortices attempt to cross the crust-core boundary. The crust-superfluid core interface acts as a potential barrier on the vortices in the superfluid core that approach this boundary. The barrier arises due to the magnetic interaction between the crustal magnetic field $H_0$ (which penetrates the superconducting core exponentially within a scale $\delta_p$ - the penetration depth) and quantum vortices whose magnetic field is governed by the generalized London equation \[LONDON\] \_p\^[-2]{}(B\_v)+B\_v = \_p \_0\^[(2)]{}(r-r\_p) +\_n\_1\^[(2)]{}(r-r\_n). Here ${{\vec\nu }}$’s are circulation unit vectors; subscripts $p$ and $n$ refer to protons and neutrons; $\Phi_0$ is the flux quantum carried by proton vortices; and $\Phi_1$ is the non-quantized flux of neutron vortices due to the entrainment effect (i.e. the effect of superconducting proton mass transport by the neutron superfluid circulation, see ref. 7). The total field $\vec B$ is the superposition of $\vec B_v$ of eq. (\[LONDON\]) and crustal filed $\vec B_{cr} = \vec H_0 {\rm exp} ( -\vec r\cdot\vec n/\delta_p)$, ($\vec n$ being the normal of the interface).
Suppose that the interface is the $(yz)$-plane of a Cartesian system of coordinates and the vectors of vortex circulation are in the positive $z$-direction. The half-plane $x<0$ corresponds to the crust while $x>0$ corresponds to the superfluid core. The knowledge of magnetic field distribution with the boundary condition at the interface $B_z=H_0$, allows one to calculate the relevant part of the Gibbs free energy of the system $G = F - (4\pi)^{-1}\int \vec B\cdot \vec H_0 dV$, where the free-energy is \[GIBBS\] F= dS + B dV . The force per single vortex of effective flux $\Phi_*(\equiv \Phi_0;
\, \Phi_1)$ derived from $G$ is \[FORCE\] f(x ) = . Here $K_1$ is the modified Bessel function. The first term in equation (\[FORCE\]) is the repulsive force acting between the vortex magnetic flux and the crustal magnetic field. It can also be interpreted as a Lorentz force resulting from superposition of velocity fields of the vortex and the surface Meissner currents. The second term is the attractive force acting between the vortex and the interface. For large distances the repulsive term dominates; (the attractive one goes to zero faster since $K_1(2x)\propto \sqrt{\pi/4x} ~e^{-2x}$ for $x\to \infty$). For small distances the second (attractive) term in eq. (\[FORCE\]) dominates. Thus the repulsive part of the vortex - crust-core interface interaction, which dominates at large distances, acts as a potential barrier on a vortex approaching the boundary. The Magnus force is then balanced by the vortex-interface force. When the disbalance between this forces reaches the value at which the vortices annihilate at the interface, the resultant rapid transfer of angular momentum from the superfluid spins-up the crust.
Let us next estimate the interaction. For relevant densities, $\rho\simeq 2\times 10^{14}$ g cm$^{-3}$, we have $\delta_p \simeq 100$ fm (e. g. ref. 7) and assuming a conventional value for the crustal magnetic field $H_0 = 10^{12}$ G, we find a maximal repulsive force $f^{\rm max} =3.13\times 10^{12}$ dyn cm$^{-1}$. In general, the magnitude of the crustal magnetic field at the crust-core interface for different objects can vary in a reasonable range $10^9< H_0<10^{13}$ G. We find that, the maximal force increases (decreases) by two orders of magnitude when the crustal magnetic field is increased (decreased) by an order of magnitude.
Further progress needs to specify the ground state structure of vortices in the superfluid core. For the vortex cluster model of ref. 7, 8 (no residual field when the superconducting state sets on or a complete Meissner expulsion of residual field) the number of proton vortices per neutron vortex interacting with the interface is $N\le 283$ and the maximal force is $f^{\rm max}_C = N~ f^{\rm max}
= 8.9\times 10^{14}$ dyn cm$^{-1}$. It should sustain the Magnus force excess (i.e. the part which is non-balanced along the normal $\vec n$) acting on a neutron vortex $\delta f^M =3.23\times 10^{17}~(\delta\omega_s/{\rm s}^{-1})
~ {\rm dyn}~{\rm cm}^{-1}, $ where $\delta\omega_s$ is the angular velocity difference between the superfluid and the normal components.[^2] From the balance condition $f^{\rm max}_C = \delta f^M$, the maximal departure that can be sustained by the boundary force on the cluster is $\delta\omega_s^{\rm max} \simeq 0.003$ s$^{-1}$. Then the angular momentum conservation in a Vela-type glitch implies that the ratio of the moment of inertia of the superfluid region to the normal component should be $I_s/I_n = 0.023$. This value is close to a previous estimate of the ratio $I_s/I_n = 0.02$ for a superfluid shell at the crust-core boundary with short dynamical coupling times $(\le 120$ s) [@PAPER2].
Acknowledgments {#acknowledgments .unnumbered}
===============
One of us (A.S.) gratefully acknowledges a research grant from the Max-Kade-Foundation, NY.
References {#references .unnumbered}
==========
[99]{} M. Ruderman, [*Nature*]{} [**223**]{}, 597 (1969); [*ApJ*]{} [**382**]{}, 576 (1991). R. E. Packard, [*Phys. Rev. Lett.*]{} [**28**]{}, 1080 (1972). P. W. Anderson and N. Itoh, [*Nature*]{} [**256**]{}, 25 (1975). A. Alpar, P. Anderson, D. Pines, and J. Shaham, [*ApJ*]{} [**249**]{}, L33 (1981). B. Link and R. I. Epstein, [*ApJ*]{} [**457**]{}, 844 (1996). A. D. Sedrakian and J. M. Cordes, in preparation. A. D. Sedrakian and D. M. Sedrakian, [*ApJ*]{} [**447**]{}, 305 (1995). A. D. Sedrakian, D. M. Sedrakian, J. M. Cordes, and Y. Terzian, [*ApJ*]{} [**447**]{}, 324 (1995).
[^1]: The Proceedings of the 18th Texas Symposium on Relativistic Astrophysics; eds. A. Olinto, J. Frieman, and D. Schramm; World Scientific Press.
[^2]: The neutron star model used in our estimates is discussed in detail in ref. 8.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We examine the effect of inhomogeneous broadening on the collective response of a planar metamaterial consisting of asymmetric split ring resonators. We show that such a response leads to a transmission resonance that can persist when the broadening of individual meta-atom resonance frequencies is roughly one half the frequency characterizing the split ring asymmetry. We also find that larger degrees of inhomogeneous broadening can drastically alter the cooperative response, destroying this resonance. The reduced effect of cooperative response due to inhomogeneous broadening may find applications in producing metamaterial samples that more closely mimic homogeneous magneto-dielectric medium with well-defined susceptibility and permittivity.'
author:
- 'S. D. Jenkins'
- 'J. Ruostekoski'
title: Resonance linewidth and inhomogeneous broadening in a metamaterial array
---
Introduction {#sec:introduction}
============
There exists increasing experimental evidence that metamaterial samples, consisting of arrays of sub-wavelength scale nano-structured circuit elements, can be prepared in the limit where collective interactions between the resonators play a vital role in determining their electromagnetic (EM) responses. For example, experiments on two-dimensional (2D) arrays of closely spaced asymmetric split ring (ASR) meta-molecules have indicated the presence of a high-quality transmission resonance, with a dramatic sensitivity of the resonance linewidth to the number of ASRs in the sample. [@FedotovEtAlPRL2010] The transmission resonance was also observed to depend strongly on the spatial distribution of the mutually interacting resonators. Where a sufficiently large, regular array of ASRs exhibits a high quality transmission resonance,[@FedotovEtAlPRL2010] introducing disorder in the elements’ positions destroys the observed spectral transmission window,[@papasimakis2009; @SavoEtAlPRB2012] further indicating the role collective EM interactions can play in metamaterial dynamics. We recently showed [@JenkinsLongPRB; @JenkinsLineWidthArxiv] that closely spaced ASR metamolecules interacting via a resonant EM field exhibit collective eigenmodes with strongly suppressed resonance linewidths. The cooperative response yields the characteristic feature in the experimentally observed enhanced quality factor of the transmission resonance in Ref. . Numerically analyzing the properties of collective modes with narrow radiative resonance linewidths provided a physical explanation of this phenomenon with an excellent agreement between the simulations and the measurements.
In our previous study[@JenkinsLineWidthArxiv] showing how the transmission resonance observed by Fedotov *et al*[@FedotovEtAlPRL2010] depends on the linewidth of a particular collective mode, all unit-cell resonators were assumed to respond to EM fields identically. In the preparation of metamaterial samples, fabrication defects, however, may in general lead to variation in the geometry of individual resonators. The current oscillations supported by the unit-cell resonators would therefore possess slightly different resonance frequencies, resulting in inhomogeneously broadened metamaterial arrays. Inhomogeneous broadening changes the conditions of the resonant interaction processes. This may impair the collective, coherent phenomena which are potentially important in several applications and physical effects of metamaterials such as lasing [@ZheludevEtAlNatPhot2008] and providing precise control and manipulation of EM fields on a sub-wavelength scale, as theoretically proposed in Ref. and experimentally observed in Ref. (for other related studies, see for example Ref. ).
In this work, we study how uncontrolled inhomogeneous broadening of plasmonic resonators limits the observation of collective phenomena in metamaterials. We will examine under what conditions these collective effects can still persist and potential implications of inhomogeneous broadening on metamaterial applications. In particular, we consider a rectangular 2D array of ASR meta-molecules whose resonance frequencies are stochastically varied. We numerically evaluate ensemble averages of the EM response of the metamaterial over the stochastic distributions of the resonator properties. These show how increased inhomogeneous broadening inhibits the response of the coherent collective modes responsible for the transmission resonance observed in Ref. . Moreover, we find the effects of interactions between different discrete resonator elements, which result in the narrowing of the radiative resonance linewidth, are diminished as a function of an increasing inhomogeneous broadening. Our results therefore illustrate how maintaining uniformity in the fabrication process is essential in designing new metamaterial based devices whose applications rely on strong interactions between the resonator elements and on a cooperative response.
The multiple scale spatial structure associated with nano-fabricated resonators in metamaterial arrays, along with the wave nature of scattered EM fields, poses a theoretical challenge for studies of the response of these systems to resonant EM fields. Interactions resulting from recurrent scattering events, in which a field is scattered more than once by the same resonator, frequently play a crucial role in the cooperative system responses. [@Ishimaru1978; @vantiggelen90; @MoriceEtAlPRA1995; @RuostekoskiJavanainenPRA1997L; @RuostekoskiJavanainenPRA1997; @JavanainenEtAlPRA1999; @devries98; @fermiline; @muller01; @pinheiro04; @JenkinsLongPRB; @JenkinsLineWidthArxiv; @optlattice; @dalibardexp] While in an infinite, regular lattice, the translational symmetry can be exploited to calculate approximate local field corrections in a medium of discrete scatterers [@Kastel07], recurrent scattering processes are generally more difficult to model in finite-sized systems with complex geometries.
However, since inhomogeneous broadening reduces the cooperative effects arising from recurrent scattering, engineering a controlled amount of inhomogeneous broadening into the metamaterial, e.g. via geometrical variation of the resonators, may provide a practical means to produce samples that are easier to design and theoretically analyze. In particular, we find that with an increasing inhomogeneous broadening the response of the system approaches that of standard continuous medium electrodynamics. Reducing cooperative effects is potentially important because several metamaterial applications, such as diffraction-free lenses formed from a medium with a negative refractive index,[@SmithEtAlPRL2000; @ShelbySci2001; @SmithEtAlSCI2004] are simplest to realize with a well-defined electric susceptibility and magnetic permeability that, in many systems with complex geometries, are only approximately achieved.[@szabo]
The effects of inhomogeneous broadening have previously been examined, e.g., by Gorkunov *et al* on the bulk properties of left-handed materials in periodic infinite lattices,[@PhysRevE.73.056605] and have been experimentally observed by Gollub *et al*.[@Gollub] In our study of ASR resonators we evaluate collective modes of a finite lattice in order to investigate the effects of the inhomogeneous broadening on the experimentally observed sample size dependent transmission resonant linewidth narrowing. [@FedotovEtAlPRL2010]
In our analysis we employ a general theoretical formalism of collective interactions between a discrete set of plasmonic resonators, or meta-atoms, mediated by the EM field that we developed in Ref. . In the model, we assume each meta-atom exhibits a single mode of current oscillation that possesses appropriate electric and magnetic dipole moments. Each meta-atom responds to EM fields exhibiting a specific resonance frequency and coupling strength that are determined by its characteristic design. Starting from the Lagrangian and Hamiltonian formalism describing the interaction of the EM field with polarization and magnetization densities created by a charge distribution, we then derived the coupled dynamics of the EM fields and the meta-atom dynamic variables.[@JenkinsLongPRB] In a collection of meta-atoms, interactions with the EM field mediate a dynamic coupling between the meta-atoms and determine the collective dynamics within the ensemble, resulting in distinct collective modes with corresponding resonance frequencies and linewidths. The analysis of collective response in terms of discrete resonators also points to the direction of an interesting analogy between resonators and a system of a cloud of atoms. In atoms the electron transitions driven by an EM field create an electric dipole moment, while in the case of circuit elements the oscillating current generates both the electric and magnetic dipole moments. The model of Ref. has previously been successful in providing an excellent agreement between the theory and experimental observations of cooperative transmission resonance linewidth narrowing of ASR metamolecules.[@JenkinsLineWidthArxiv]
The remainder of this article is organized as follows. We summarize the theoretical formalism we employ to describe collective interactions within the metamaterial [@JenkinsLongPRB] in Sec. \[sec:model-coll-resp\]. In Sec. \[sec:asymm-split-rings\], we describe the fundamental building block of our metamaterical, the ASR, in the context of this formalism. The main results of the article, describing the effects of inhomogeneous broadening on the collective response of the metamaterial are presented in Sec. \[sec:inhom-broad-\], and conclusions follow in Sec. \[sec:conclusion\].
A Model for the Collective Response in Metamaterials {#sec:model-coll-resp}
====================================================
In order to incorporate the effects of strongly heterogeneous metamaterial we describe the sub-wavelength structures of the medium as discrete scatterers.[@JenkinsLongPRB] Each unit-cell element, a meta-molecule, may also consist of sub-elements, which we call meta-atoms. While the general formalism of Ref. allows for multipole-field radiation of the resonator unit elements, as a first approximation here we consider each sub-wavelength-sized meta-atom simply as a radiating dipole and ignore its multipole-field contribution. Following our treatment in Ref. , we assume that each meta-atom $j$, with its position vector defined by ${\ensuremath{\mathbf{r}}}_j$, supports a single eigenmode of current oscillation. The dynamics of this current oscillation are determined by the dynamic variable $Q_j(t)$ with units of charge. Each meta-atom exhibits an electric and magnetic dipole moment. These may be expressed as
\[eq:eDipDef\] $$\begin{aligned}
{\ensuremath{\mathbf{d}}}_j &=& Q_j h_j {\ensuremath{\mathbf{\hat{d}}}}_j \,\textrm{,} \\
{\ensuremath{\mathbf{m}}}_j &=& I_j A_j{\ensuremath{\mathbf{\hat{m}}}}_j \,\textrm{,}
\end{aligned}$$
respectively. Here $I_j(t) = {\mathrm{d}}Q_j/ {\mathrm{d}}t$ denotes the current, and the directions of the dipole moments are specified by the unit vectors ${\ensuremath{\mathbf{\hat{d}}}}_j$ and ${\ensuremath{\mathbf{\hat{m}}}}_j$ with proportionality coefficients $h_j$ and $A_j$ (with units of length and area, respectively) that depend on the specific geometry of the resonators. We assume the meta-atoms are designed such that the electric quadrupole and higher order multipole contributions to the meta-atom dynamics can be ignored. Although each meta-atom possesses only electric and magnetic dipoles, a meta-molecule of two or more meta-atoms in our model would exhibit a non-vanishing quadrupole field. While, in general, this quadrupole contribution is inaccurately represented in the dipole approximation, in the case of the ASR meta-molecules considered in Refs. and in the present study, the generated quadrupole field is notably suppressed when compared to the corresponding dipolar field. This has been indicated by finite element simulations of Maxwell’s equations within a single meta-molecule.[@PapasimakisComm] Additionally, the fact that the size of the meta-atoms is often comparable to the spacing between them could result in a correction to the coupling strength between neighboring elements obtained in the point dipole approximation. Nonetheless, this model in the dipole approximation was employed in Ref. to characterize the cooperative linewidth narrowing responsible for the enhancement of quality factor with system size observed in Ref. , yielding excellent agreement with experimental results. In this section, we describe the key features of our theoretical formalism that are required to describe the collective response of an inhomogeneously broadened sample of ASRs to the EM field. Details of the derivation are presented in Ref. .
We write the polarization and magnetization densities as a sum of their contributions from the individual meta-atoms
\[eq:totalPolMagDens\] $$\begin{aligned}
{{\ensuremath{\mathbf{P}}}}({{\ensuremath{\mathbf{r}}}}) &=&\sum_j {{\ensuremath{\mathbf{P}}}}_j({{\ensuremath{\mathbf{r}}}}) \,\textrm{,} \\
{{\ensuremath{\mathbf{M}}}}({{\ensuremath{\mathbf{r}}}}) &=&\sum_j {{\ensuremath{\mathbf{M}}}}_j({{\ensuremath{\mathbf{r}}}}) \, \textrm{,}
\end{aligned}$$
where the polarization and the magnetization of the resonator $j$ in the dipole approximation read
$$\begin{aligned}
{\ensuremath{\mathbf{P}}}_j({{\ensuremath{\mathbf{r}}}},t) &\approx& {\ensuremath{\mathbf{d}}}_j \delta({{\ensuremath{\mathbf{r}}}}-{{\ensuremath{\mathbf{r}}}}_j) \,\textrm{,} \\
{\ensuremath{\mathbf{M}}}_j({{\ensuremath{\mathbf{r}}}},t) &\approx& {\ensuremath{\mathbf{m}}}_j \delta({{\ensuremath{\mathbf{r}}}}-{{\ensuremath{\mathbf{r}}}}_j) \, \textrm{,}
\end{aligned}$$
respectively.
An external beam with electric field ${\ensuremath{\mathbf{E}}}_{\mathrm{in}}({\ensuremath{\mathbf{r}}},t)$ and magnetic field ${\ensuremath{\mathbf{H}}}_{\mathrm{in}}({\ensuremath{\mathbf{r}}},t)$ with frequency $\Omega_0$ impinges on the ensemble of meta-atoms. The incident EM field drives the meta-atoms, generating dipole radiation from the oscillating electric and magnetic dipoles. The total radiation from the metamaterial array is the sum of the scattered electric and magnetic fields from all the meta-atoms
\[eq:TotalScatteredFields\] $$\begin{aligned}
{{\ensuremath{\mathbf{E}}}}_{\mathrm{S}}({{\ensuremath{\mathbf{r}}}},t)=\sum_j {{\ensuremath{\mathbf{E}}}}_{\mathrm{S},j}({{\ensuremath{\mathbf{r}}}},t) \,\textrm{,} \\
{{\ensuremath{\mathbf{H}}}}_{\mathrm{S}}({{\ensuremath{\mathbf{r}}}},t)=\sum_j {{\ensuremath{\mathbf{H}}}}_{\mathrm{S},j}({{\ensuremath{\mathbf{r}}}},t) \, \textrm{,}
\end{aligned}$$
where ${{\ensuremath{\mathbf{E}}}}_{\mathrm{S},j}({{\ensuremath{\mathbf{r}}}},t)$ and ${{\ensuremath{\mathbf{H}}}}_{\mathrm{S},j}({{\ensuremath{\mathbf{r}}}},t)$ denote the electric and magnetic field emitted by the meta-atom $j$. The Fourier components of the scattered fields have the familiar expressions of electric and magnetic dipole radiation,[@Jackson] $$\begin{aligned}
{\ensuremath{\mathbf{E}}}^+_{\mathrm{S},j}({\ensuremath{\mathbf{r}}},\Omega) &=&
\frac{k^3}{4\pi\epsilon_0} \int {\mathrm{d}}^3 r' \,
\Bigg[{\ensuremath{\boldsymbol{\mathbf{G}}}}({\ensuremath{\mathbf{r}}} - {\ensuremath{\mathbf{r}}}',\Omega) \cdot
{\ensuremath{\mathbf{P}}}^+_{j}({\ensuremath{\mathbf{r}}}',\Omega) \nonumber\\
&&\qquad+ \frac{1}{c}
{\ensuremath{\boldsymbol{\mathbf{G}}}}_\times({\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}',\Omega) \cdot
{\ensuremath{\mathbf{M}}}^+_j({\ensuremath{\mathbf{r}}}',\Omega) \Bigg] , \label{efield}\\
{\ensuremath{\mathbf{H}}}^+_{\mathrm{S},j}({\ensuremath{\mathbf{r}}},\Omega) &=&
\frac{k^3}{4\pi} \int {\mathrm{d}}^3 r' \,
\Big[{\ensuremath{\boldsymbol{\mathbf{G}}}}({\ensuremath{\mathbf{r}}} - {\ensuremath{\mathbf{r}}}',\Omega) \cdot
{\ensuremath{\mathbf{M}}}^+_j({\ensuremath{\mathbf{r}}}',\Omega) \nonumber\\
&&\qquad- c {\ensuremath{\boldsymbol{\mathbf{G}}}}_\times({\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}',\Omega) \cdot
{\ensuremath{\mathbf{P}}}^+_j({\ensuremath{\mathbf{r}}}',\Omega) \Big]\,,\label{mfield}\end{aligned}$$ where we have defined the positive and negative frequency components of a real time varying quantity $V(t)$ such that for a Fourier component of frequency $\Omega$, $V^{\pm}(\Omega) \equiv
\Theta(\pm\Omega)
V(\Omega)$, and hence $V(t) = V^+(t) +V^{-}(t)$ with $V^{-}(t) =
[V^{+}(t)]^*$. Here ${\ensuremath{\boldsymbol{\mathbf{G}}}}$ denotes the radiation kernel representing the electric (magnetic) field emitted from an electric (magnetic) dipole. [@Jackson] The explicit expression for the corresponding radiated field from a dipole ${\ensuremath{\mathbf{v}}}$ reads $$\begin{aligned}
{\ensuremath{\boldsymbol{\mathbf{G}}}}({\ensuremath{\mathbf{r}}},\Omega) &\cdot{\ensuremath{\mathbf{v}}} =
({\ensuremath{\mathbf{\hat{r}}}}\!\times\!{\ensuremath{\mathbf{v}}}
)\!\times\!{\ensuremath{\mathbf{\hat{r}}}}
\frac{{e}^{{i}kr}}{kr}+[3{\ensuremath{\mathbf{\hat{r}}}}({\ensuremath{\mathbf{\hat{r}}}}\cdot{\ensuremath{\mathbf{v}}})
-{\ensuremath{\mathbf{v}}}]\nonumber\\
&\times \left[ \frac{1}{(kr)^3} - \frac{{i}}{(kr)^2}\right]{e}^{{i}kr}
-{4\pi\over3}\delta(k{\ensuremath{\mathbf{r}}}) {\ensuremath{\mathbf{v}}}\,,
\label{eq:Green'sfunc}\end{aligned}$$ where ${\ensuremath{\mathbf{\hat{r}}}} \equiv {{\ensuremath{\mathbf{r}}}}/r$ and $k \equiv \Omega/c$. Similarly, ${\ensuremath{\boldsymbol{\mathbf{G}}}}_{\times}({\ensuremath{\mathbf{r}}},\Omega)$ represents the radiation kernel for the magnetic (electric) field of an electric (magnetic) dipole source.[@Jackson] Specifically, the corresponding radiated field from a dipole ${\ensuremath{\mathbf{v}}}$ yields $$\label{eq:CrossGreen}
{\ensuremath{\boldsymbol{\mathbf{G}}}}_{\times} ({\ensuremath{\mathbf{r}}},\Omega)\cdot {\ensuremath{\mathbf{v}}} =
\frac{{e}^{{i}kr}}{kr} \left(1-\frac{1} {{i}kr}\right) \,{\ensuremath{\mathbf{\hat{r}}}}
\times {\ensuremath{\mathbf{v}}}\,\textrm{.}$$
The polarization and magnetization densities appearing in Eqs. and are themselves driven by the incident and scattered EM fields. This driving, combined with the scattered EM fields, yield a coupled set of equations for the resonators and EM fields that we derived in Ref. . Current excitations in each meta-atom $j$ produces a field that interacts with the current and charge oscillations that generated it. Due to these self-generated fields, a meta-atom $j$ exhibits behavior similar to that of an LC circuit with resonance frequency[@JenkinsLongPRB] $$\label{eq:resFreqDef}
\omega_j \equiv \frac{1}{\sqrt{L_j C_j}} \,\textrm{,}$$ where $C_j$ is an effective self-capacitance and the effective self-inductance is $L_j$. In this work, we consider an inhomogeneously broadened sample of $N$ ASRs. Each ASR, $l$, ($l=1\ldots N$) consists of two meta-atoms whose resonance frequencies $\omega_j$ ($j=2l-1,2l$) are centered around the random frequency $\omega_0 + X_l$, where $X_l$ are independent identically distributed random variables. We assume the meta-atom resonance frequencies occupy a narrow band about $\omega_0$, i.e. $|\omega_j-\omega_0|,X_l \ll \omega_0$. The oscillating electric and magnetic dipoles of an isolated meta-atom radiate energy at respective rates $\Gamma_{{\mathrm{E}}}$ and $\Gamma_{{\mathrm{M}}}$,[@JenkinsLongPRB] $$\begin{aligned}
\Gamma_{\mathrm{E},j} &\equiv&
\frac{h_j^{2}C_j\omega_j^{4}}{6\pi\epsilon_{0}c^3} \,\textrm{,}
\label{eq:Gamma_EDef} \\
\Gamma_{\mathrm{M},j} &\equiv& \frac{\mu_{0}A_j^{2}\omega_j^{4}}{6\pi c^3
L_j} \label{eq:Gamma_MDef} \,\textrm{,}\end{aligned}$$ resulting in the scattered fields ${\ensuremath{\mathbf{E}}}_{\mathrm{S},j}$ and ${\ensuremath{\mathbf{H}}}_{\mathrm{S},j}$ \[see Eqs. and \]. For simplicity, we assume that these radiative emission rates $\Gamma_{{\mathrm{E}}}$ and $\Gamma_{{\mathrm{M}}}$ are independent of the resonator $j$ and that they are dominated by the meta-atom resonance frequencies, i.e. $\Gamma_{{\mathrm{E}},j},\Gamma_{{\mathrm{M}},j} \ll \Omega_0$. We further assume the resonance frequencies occupy a narrow bandwidth around the central frequency of the incident field.
The dynamics of current excitations in the meta-atom $j$ may then be described by $Q_j(t)$ \[introduced in Eq. \] and its conjugate momentum $\phi_j(t)$ (with units of magnetic flux [@JenkinsLongPRB]). In terms of the positive frequency components the equations of motion read [@JenkinsLongPRB] $$\begin{aligned}
\dot{Q}_j^{+} &=& \left( 1 - {i}\frac{\Gamma_{{\mathrm{M}}}}
{\omega_j} \right) \frac{\phi^{+}_j}{L_j}
- \frac{\mu_0A_j}{L_j} {\ensuremath{\mathbf{\hat{m}}}}_j \cdot {\ensuremath{\mathbf{H}}}^+_{j,\mathrm{ext}}({{\ensuremath{\mathbf{r}}}}_j,t)
\label{eq:saEqmQ} \\
\dot{\phi}_j^{+} &=& -\left(1 - {i}\frac{\Gamma_{{\mathrm{E}}}}{\omega_j}\right) \frac{Q^{+}_j}{C_j} +
h_j {\ensuremath{\mathbf{\hat{d}}}}_j \cdot {\ensuremath{\mathbf{E}}}^+_{j,\mathrm{ext}}({{\ensuremath{\mathbf{r}}}}_j,t) \, \textrm{,}
\label{eq:saEqmphi}\end{aligned}$$ where the fields generated externally to meta-atom $j$ that drive its dynamics, ${\ensuremath{\mathbf{E}}}^+_{j,\mathrm{ext}}({{\ensuremath{\mathbf{r}}}},t)$ and ${\ensuremath{\mathbf{H}}}^+_{j,\mathrm{ext}}({{\ensuremath{\mathbf{r}}}},t)$ are produced by the sums of the corresponding incident fields and the fields scattered by all other meta-atoms in the metamaterial sample, $\sum_{j'\ne j}
{\ensuremath{\mathbf{E}}}^+_{{\mathrm{S}},j'}({{\ensuremath{\mathbf{r}}}},t)$ and $\sum_{j'\ne j}
{\ensuremath{\mathbf{H}}}^+_{{\mathrm{S}},j'}({{\ensuremath{\mathbf{r}}}},t)$, respectively. The component of the external electric field ${\ensuremath{\mathbf{E}}}_{j,\mathrm{ext}}$ oriented along the dipole direction ${\ensuremath{\mathbf{\hat{d}}}}_j$ provides a net external electromotive force (EMF) $h_j
{\ensuremath{\mathbf{\hat{d}}}}_j \cdot {\ensuremath{\mathbf{E}}}_{j,\mathrm{ext}}^+({{\ensuremath{\mathbf{r}}}}_j,t)$ which drives $\phi_j(t)$. Similarly, the component of the external magnetic field ${\ensuremath{\mathbf{H}}}_{j,\mathrm{ext}}$ along the magnetic dipole direction ${\ensuremath{\mathbf{\hat{m}}}}_j$ provides a net applied magnetic flux $A_j\mu_0 {\ensuremath{\mathbf{\hat{m}}}}_j
\cdot {\ensuremath{\mathbf{H}}}^+_{j,\mathrm{ext}}({{\ensuremath{\mathbf{r}}}}_j,t)$ that drives $Q_j(t)$. In the absence of radiative emission and interactions with external fields, current and charge oscillate within the meta-atom at the resonance frequency $\omega_j$. The meta-atom dynamics are therefore naturally described by the slowly varying normal variables $$\label{eq:normalVariables}
b_j(t) \equiv \frac{{e}^{{i}\Omega_0 t}}{\sqrt{2}}
\left(\frac{Q_j(t)}{\sqrt{\omega_jC_j}} +{i}\frac{\phi_j(t)}{\sqrt{\omega_jL_j}}\right) \, \textrm{.}$$ In the absence of external field interactions and damping, $b_j$ oscillates with frequency $(\omega_j - \Omega_0)$, i.e. $b_j(t) = b_j(0)
\exp\left[-i \left(\omega_j - \Omega_0\right) t\right]$. For nonzero $\Gamma_{\mathrm{E}},\Gamma_{\mathrm{M}} \ll \Omega_0$, losses and driving from the external field act to perturb this oscillation.
The current oscillation dynamics in the meta-atom $j$, described by ${Q}_j(t)$ and $\phi_j(t)$ in Eqs. and , is driven by the incident field and the fields scattered from all the other meta-atoms and acts as a source of radiation that, in turn, drives the other meta-atoms. The expressions for the scattered fields by polarization and magnetization densities \[Eqs. and \] (generated by excitations in meta-atoms) and the expressions for the oscillating charge dynamics \[Eqs. and \] form a coupled set of equations, describing EM field mediated interactions between the resonators. In terms of the normal variables $b_j$, these interactions may be represented by the set of equations,[@JenkinsLongPRB] $$\dot{{\ensuremath{\mathrm{b}}}} = \mathcal{C} {\ensuremath{\mathrm{b}}} + {\ensuremath{\mathrm{f}}}_{\mathrm{in}} \textrm{ ,}
\label{eq:rwa_b_eqm}$$ where we have defined $$\label{eq:bColvec}
{\ensuremath{\mathrm{b}}}(t) \equiv \left(
\begin{array}{c}
b_1(t)\\
b_2(t)\\
\vdots\\
b_{nN}(t)
\end{array}
\right) \, \textrm{,} \qquad
{\ensuremath{\mathrm{f}}}_{\mathrm{in}}(t) \equiv \left(
\begin{array}{c}
f_{1,\mathrm{in}} (t) \\
f_{2,\mathrm{in}}(t) \\
\vdots \\
f_{nN,\mathrm{in}}(t)
\end{array}
\right) \,\textrm{.}$$ The driving $f_{j,\mathrm{in}}$ of each meta-atom $j$ results from the EMF and magnetic flux induced by the incident fields.[@JenkinsLongPRB] The component of the incident electric field ${\ensuremath{\mathbf{E}}}_{\mathrm{in}}$ parallel to the electric dipole orientation ${\ensuremath{\mathbf{\hat{d}}}}_j$ induces the EMF, while the component of ${\ensuremath{\mathbf{H}}}_{\mathrm{in}}$ along the magnetic dipole orientation ${\ensuremath{\mathbf{\hat{m}}}}_j$ provides an incident magnetic flux. Here we assume that the meta-atom magnetic dipoles are aligned perpendicular to the incident magnetic field, and thus only the EMF contributes to the driving of each meta-atom, which is given by $$\label{eq:f_driving}
{e}^{-{i}\Omega_0 t} f_{j,\mathrm{in}}(t) = {i}\frac{h_j}{\sqrt{2\omega_j L_j}} {\ensuremath{\mathbf{\hat{d}}}}_j \cdot
{\ensuremath{\mathbf{E}}}_{\mathrm{in}}^+({{\ensuremath{\mathbf{r}}}}_j,t)
\, \textrm{.}$$ The coupling matrix between the meta-atoms in Eq. reads $$\mathcal{C} = -{i}\mathrm{\Delta} - \frac{\Gamma}{2} \mathrm{I}
+ \frac{1}{2} \left( {i}\mathcal{C}_{\mathrm{E}} +
{i}\mathcal{C}_{\mathrm{M}}
+ \mathcal{C}_\times
+ \mathcal{C}_\times^{T} \right) \textrm{,}
\label{eq:C_rwa}$$ where $\mathrm{I}$ represents the identity matrix. Here the detunings of the incident field from the meta-atom resonances are contained in the diagonal matrix $\mathrm{\Delta}$ with elements $$\Delta_{j,j'} \equiv \delta_{j,j'} \left(\omega_j-\Omega_0\right)\,\textrm{,}
\label{eq:DeltaMatDef}$$ and the energy carried away from individual meta-atoms by the scattered fields manifests itself in the decay rate $$\label{eq:GammaMatDef}
\Gamma \equiv \Gamma_{{\mathrm{E}}} +
\Gamma_{{\mathrm{M}}} + \Gamma_{\mathrm{O}}$$ appearing in the diagonal elements of $\mathcal{C}$. We account for non-radiative, e.g. ohmic losses, by introducing phenomenological decay rate $\Gamma_{\mathrm{O}}$. The multiple scattering processes are included in the terms $\mathcal{C}_{{\mathrm{E}}}$, $\mathcal{C}_{{\mathrm{M}}}$, and $\mathcal{C}_{\times}$, which generate interaction between the meta-atom dynamic variables The matrices $\mathcal{C}_{{\mathrm{E}}}$ and $\mathcal{C}_{{\mathrm{M}}}$ characterize the electric dipole-dipole and magnetic dipole-dipole interactions, respectively. Additionally, the interaction embodied by $\mathcal{C}_\times$ arises from the *electric* field emitted by the *magnetic* dipole of one atom driving the electric dipoles of the others. Similarly, $\mathcal{C}_{\times}^T$ results from the *magnetic* field produced by the meta-atoms’ *electric* dipoles impinging on the magnetic dipoles of all the other meta-atoms. Because the interaction matrices $\mathcal{C}_{{\mathrm{E}}|{\mathrm{M}}|\times}$ govern interactions between distinct meta-atoms, their diagonal elements are zero. Their off diagonal elements are given by [@JenkinsLongPRB] $$\begin{aligned}
\label{eq:GreensElecMat}
\left[\mathcal{C}_{{\mathrm{E}}}\right]_{j,j'} &=& \frac{3}{2}\Gamma_{{\mathrm{E}}}\,{\ensuremath{\mathbf{\hat{d}}}}_j \cdot
{\ensuremath{\boldsymbol{\mathbf{G}}}}({\ensuremath{\mathbf{r}}}_j - {\ensuremath{\mathbf{r}}}_{j'},\Omega_0 ) \cdot
{\ensuremath{\mathbf{\hat{d}}}}_{j'},\\
\label{eq:GreenMagMat}
\left[\mathcal{C}_{{\mathrm{M}}}\right]_{j,j'}& =& \frac{3}{2} \Gamma_{{\mathrm{M}}} \, {\ensuremath{\mathbf{\hat{m}}}}_j \cdot
{\ensuremath{\boldsymbol{\mathbf{G}}}}({\ensuremath{\mathbf{r}}}_j - {\ensuremath{\mathbf{r}}}_{j'},\Omega_0 ) \cdot {\ensuremath{\mathbf{\hat{m}}}}_{j'},\\
\label{eq:emCrossMat}
\left[\mathcal{C}_\times \right]_{j,j'} &=& \frac{3}{2}\bar{\Gamma}\, {\ensuremath{\mathbf{\hat{d}}}}_j\cdot
{\ensuremath{\boldsymbol{\mathbf{G}}}}_\times({\ensuremath{\mathbf{r}}}_j -
{\ensuremath{\mathbf{r}}}_{j'},\Omega_0) \cdot
{\ensuremath{\mathbf{\hat{m}}}}_{j'}\,,\end{aligned}$$ where $\bar{\Gamma} \equiv \sqrt{\Gamma_{{\mathrm{E}}} \Gamma_{{\mathrm{M}}}}$ is the geometric mean of the electric and magnetic dipole emission rates.
Asymmetric Split Ring Resonators {#sec:asymm-split-rings}
================================
In order to investigate the effects of inhomogeneous broadening of meta-atom resonance frequencies on a metamaterial’s collective EM response, we consider an ensemble of asymmetric meta-molecules arranged in a regular lattice. To facilitate our description of this EM response, in this section we summarize the behaviour of a single ASR in the context of the model presented in Sec. \[sec:model-coll-resp\].
An ASR is a variation on the split ring resonator used to produce bulk metamaterials with negative indices of refraction.[@SmithEtAlPRL2000; @ShelbySci2001; @SmithEtAlSCI2004] The meta-atoms of an ASR consist of two separate concentric circular arcs labeled by $j \in\{{\mathrm{l}},{\mathrm{r}}\}$ and separated by ${\ensuremath{\mathbf{u}}} \equiv {\ensuremath{\mathbf{r}}}_{{\mathrm{r}}} - {\ensuremath{\mathbf{r}}}_{{\mathrm{l}}}$. The current oscillations in meta-atoms produce electric dipoles with orientation ${\ensuremath{\mathbf{\hat{d}}}}_{{\mathrm{r}}} = {\ensuremath{\mathbf{\hat{d}}}}_{{\mathrm{l}}} = {\ensuremath{\mathbf{\hat{d}}}}$ associated with charge oscillating between the ends of the arcs. Owing to the curvature of the meta-atoms, these currents also produce magnetic dipoles with opposite orientations ${\ensuremath{\mathbf{\hat{m}}}}_{{\mathrm{r}}} =
-{\ensuremath{\mathbf{\hat{m}}}}_{{\mathrm{l}}} = {\ensuremath{\mathbf{\hat{m}}}}$ where ${\ensuremath{\mathbf{\hat{d}}}}\perp{\ensuremath{\mathbf{u}}}$ and ${\ensuremath{\mathbf{\hat{m}}}} \perp{\ensuremath{\mathbf{u}}},{\ensuremath{\mathbf{\hat{d}}}}$. An asymmetry between the rings, e.g., resulting from a difference in arc length, manifests itself as a difference in resonance frequencies with $$\begin{aligned}
\label{eq:4}
\omega_{{\mathrm{r}}} = \omega_0 + \delta\omega \\
\omega_{{\mathrm{l}}} = \omega_0 - \delta\omega\end{aligned}$$
To analyze the dynamics of a single ASR unit-cell resonator consisting of two meta-atoms, we apply the formalism presented in Sec. \[sec:model-coll-resp\]. According to (\[eq:rwa\_b\_eqm\]), the normal variables $b_{{\mathrm{r}}}$ and $b_{{\mathrm{l}}}$ that describe the current oscillations in the right and left meta-atoms, respectively, are coupled by the EM fields so that they evolve according to $$\label{eq:SSRDynEq}
\left(
\begin{array}{c}
\dot{b}_r \\
\dot{b}_l
\end{array}
\right) = {\mathcal{C}^{(\mathrm{ASR})}}\left(
\begin{array}{c}
b_r \\
b_l
\end{array}
\right)
+
\left(
\begin{array}{c}
f_{r,\mathrm{in}} \\
f_{l,\mathrm{in}}
\end{array}
\right)\,.$$ Here ${\mathcal{C}^{(\mathrm{ASR})}}$ denotes the specific coupling matrix between the two meta-atoms that depends of the radiative electric dipole – electric dipole, magnetic dipole – magnetic dipole, and electric dipole – magnetic dipole interaction processes between the two meta-atoms \[See Eqs. -.\]. On the other hand, the incident field produces the driving terms $f_{j,\mathrm{in}}$ for each meta-atom $j={\mathrm{l}},{\mathrm{r}}$ \[See Eq. .\].
To analyze the modes of the ASR, we consider the dynamics of symmetric $c_+$ and antisymmetric $c_-$ modes of oscillation defined by $$\label{eq:c_pm_def}
c_\pm \equiv \frac{1}{\sqrt{2}} \left(b_{{\mathrm{r}}} \pm b_{{\mathrm{l}}} \right) \textrm{.}$$ The oscillations $c_{\pm}$ represent the eigenmodes of the ASR in the absence of asymmetry $\delta\omega=0$. By diagonalizing ${\mathcal{C}^{(\mathrm{ASR})}}$ with $\delta\omega=0$, one finds the eigenvalues of the modes $c_\pm$, $$\label{eq:CASR_eigenvalues}
\lambda_\pm = -i\left(\omega_0-\Omega_0 \pm \Delta\right) -
\frac{\gamma_\pm}{2} \,\textrm{.}$$ The interaction between the elements shifts the two collective resonance frequencies by equal and opposite amounts $\Delta$ and results in the decay rates $\gamma_{\pm}$, where the coefficients $\Delta$ and $\gamma_\pm$ depend on the radiative interactions between the two meta-atoms.[@JenkinsLongPRB; @JenkinsLineWidthArxiv] When the spacing between the arcs $u \ll \lambda$ ( $\lambda = 2\pi
c/\Omega_0$ ), the decay rates simplify to
$$\begin{aligned}
\label{eq:gamma_p}
\gamma_+ &=& 2\Gamma_{\mathrm{E}}+\Gamma_{\mathrm{O}} \textrm{ ,}
\\
\label{eq:gamma_m}
\gamma_- &=& 2\Gamma_{\mathrm{M}} + \Gamma_ {\mathrm{O}} \textrm{ .}
\end{aligned}$$
In this limit, the symmetric mode, possessing a net electric dipole, emits electric dipole radiation, and the antisymmetric mode, possessing a net magnetic dipole, emits magnetic dipole radiation. We therefore may refer to symmetric and antisymmetric oscillations as electric and magnetic dipole excitations, respectively.
A nonzero asymmetry, $\delta\omega \ne 0$, tends to couple the symmetric and anti-symmetric oscillations. One finds that, when driven by an external field, these oscillations in a single ASR evolve as[@JenkinsLongPRB; @JenkinsLineWidthArxiv] $$\label{eq:cpmEqs}
\dot{c}_{\pm} = \left[-{i}\left(\omega_0\pm\Delta - \Omega_0\right)
- \frac{\gamma_\pm}{2}
\right]c_\pm - {i}\delta\omega c_{\mp} + F_\pm \,\textrm{,}$$ where the driving terms $F_\pm = (f_{{\mathrm{r}},\mathrm{in}} \pm f_{{\mathrm{l}},\mathrm{in}})/\sqrt{2}$. The symmetric and antisymmetric oscillations are driven purely by the electric and magnetic fields, respectively, and when the meta-atom separation $u\ll\lambda$, $F_+
\propto {\ensuremath{\mathbf{\hat{d}}}} \cdot {{\ensuremath{\mathbf{E}}}}_{\mathrm{in}}^+({\ensuremath{\mathbf{R}}},t)$ and $F_-
\propto {\ensuremath{\mathbf{\hat{m}}}} \cdot {{\ensuremath{\mathbf{B}}}}_{\mathrm{in}}^+({\ensuremath{\mathbf{R}}},t)$, where ${\ensuremath{\mathbf{R}}}$ is the center of mass of the ASR. Therefore, an incident field with ${\ensuremath{\mathbf{E}}}_{\rm in} \parallel {\ensuremath{\mathbf{\hat{d}}}}$ and ${\ensuremath{\mathbf{B}}}_{\rm in} \perp
{\ensuremath{\mathbf{\hat{m}}}}$ only excites the symmetric mode when $\delta\omega =0$. However for $\delta\omega \ne 0$, the asymmetry couples the symmetric and antisymmetric modes, and this incident field can resonantly pump the anti-symmetric magnetic mode via an effective two-photon transition.[@JenkinsLineWidthArxiv]
![A schematic illustration of an array of ASR metamolecules excited in the uniform phase magnetic mode. The meta-atom currents in each ASR oscillate with opposite phases, producing magnetic dipoles represented by the red arrows. []{data-label="fig:mag_mode_schematic"}](fig1_d_v2.eps){width="0.95\columnwidth"}
In this article, we consider a 2D metamaterial comprised of ASRs arranged in a regular array. The fields scattered from each ASR then mediate interactions between them, resulting in collective modes of oscillation distributed over the array, each with its own resonance frequency and decay rate. Figure \[fig:mag\_mode\_schematic\] provides a schematic illustration of a mode consisting primarily of magnetic dipoles oscillating in phase throughout the metamaterial.[@JenkinsLongPRB; @JenkinsLineWidthArxiv] We showed in Ref. that, for sufficiently large array, such a mode radiates more slowly than the magnetic excitation of a single ASR and is responsible for the transmission resonance observed by Fedotov *et al*.[@FedotovEtAlPRL2010]. When an incident EM field whose magnetic field is perpendicular to the ASR magnetic dipoles impinges on the array, this mode cannot be excited directly. But, the presence of an asymmetry provides a coupling between electric and magnetic dipoles allowing it to be driven. Driving of the metamaterial’s collective modes are responsible for the cooperative response that yields phenomena such as the observed transmission resonance.[@FedotovEtAlPRL2010; @JenkinsLineWidthArxiv]
Inhomogeneous Broadening {#sec:inhom-broad-}
=========================
In this section, we study the effects of inhomogeneous broadening on the cooperative EM response of an array of ASRs. Here inhomogeneous broadening refers to a statistical uncertainty in the resonance frequencies in individual ASR meta-molecules. Such an uncertainty may result, for example, from imperfections in the manufacturing processes which yield meta-atoms whose shape varies slightly from the design specifications. Collective modes that are phase matched with an incident EM field have been shown to be responsible for transmission resonances [@JenkinsLineWidthArxiv] that have been observed experimentally.[@FedotovEtAlPRL2007; @FedotovEtAlPRL2010] We will illustrate the response of a regular array of ASRs to an incident plane wave, and show that broadening adversely affects the characteristics of the response responsible for observed resonances.
We consider an ensemble of ASR meta-molecules whose constituent meta-atoms are separated by ${\ensuremath{\mathbf{u}}}=u{\ensuremath{\mathbf{\hat{e}}}}_x$ with electric dipoles oriented along ${\ensuremath{\mathbf{\hat{d}}}} = {\ensuremath{\mathbf{\hat{e}}}}_y$ such that a symmetric oscillation in a single ASR produces an electric dipole along ${\ensuremath{\mathbf{\hat{d}}}}$, and an antisymmetric oscillation produces a magnetic dipole along ${\ensuremath{\mathbf{\hat{m}}}}={\ensuremath{\mathbf{\hat{e}}}}_z$. The ASRs are arranged in an $N_x \times N_y$ 2D square lattice with lattice spacing $a$ and lattice vectors ${\ensuremath{\mathbf{a}}}_1=a{\ensuremath{\mathbf{\hat{e}}}}_x$ and ${\ensuremath{\mathbf{a}}}_2 =a {\ensuremath{\mathbf{\hat{e}}}}_y$. The sample is illuminated by a cw plane wave ${{\ensuremath{\mathbf{E}}}}_{\rm in}^+({{\ensuremath{\mathbf{r}}}})=
{\hbox{$1\over2$}}{\cal E} {\ensuremath{\mathbf{\hat{e}}}}_y e^{i{{\ensuremath{\mathbf{k}}}}\cdot {{\ensuremath{\mathbf{r}}}}}$ with ${\ensuremath{\mathbf{k}}} =k
{\ensuremath{\mathbf{\hat{e}}}}_z$, coupling exclusively to the electric dipole moments of the ASRs.
The vector of $2N_xN_y$ normal variables describing the state of current oscillations in each meta-atom obeys the coupled equations of motion in Eq. , where the matrix $\mathcal{C}$ \[See Eq. .\] arises from the meta-atom interactions mediated by the EM field. The metamaterial therefore exhibits $2N_x N_y$ collective modes of oscillation corresponding to the eigenvectors ${\ensuremath{\mathrm{v}}}_i$ ($i=1\ldots 2N_xN_y)$ of interaction matrix $\mathcal{C}$.[@JenkinsLongPRB] Each eigenmode $i$ possesses a particular resonance frequency $\Omega_i$ and decay rate $\gamma_i$ given in terms of the eigenvalue $\lambda_i$ by
\[eq:eigenvals\] $$\begin{aligned}
\Omega_i &=& -\operatorname{Im}(\lambda_i) + \Omega_0\,\textrm{,} \\
\gamma_i &=& -2\operatorname{Re}(\lambda_i) \,\textrm{,}
\end{aligned}$$
respectively.
Since the incident field drives all ASRs uniformly, it couples most effectively to the collective modes in which all of the metamolecules oscillate in phase. The two modes of particular interest are the uniform electric and uniform magnetic modes. In the absence of an asymmetry ($\delta\omega=0$), the incident field drives the uniform electric mode which, owing to the electric dipole orientations, emits strongly into the $\pm {\ensuremath{\mathbf{\hat{e}}}}_z$ directions. The electric dipoles oscillating in phase are responsible for reflection from the metamaterial. By contrast, the magnetic dipoles in the uniform magnetic mode, illustrated in Fig. \[fig:mag\_mode\_schematic\], emit into the plane of the meta-material array, and for sufficiently large lattices of closely spaced ASRs, the magnetic dipole radiation becomes trapped. This results in a suppressed radiative decay rate of the uniform magnetic mode.[@JenkinsLineWidthArxiv] Introduction of an asymmetry ($\delta\omega \ne 0$) provides an effective coupling between these two collective modes similar to the coupling between the symmetric and antisymmetric modes in a single ASR. The collective magnetic mode can thus be resonantly excited at the expense of the electric dipoles resulting in a transmission resonance [@JenkinsLineWidthArxiv; @FedotovEtAlPRL2007] whose quality factor increases with the size of the array, as observed by Fedotov *et al*. [@FedotovEtAlPRL2010]
![The uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$ of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The excitations $|c_{\pm,l}|^2$ are normalized to the peak ASR excitation $E_0 = \max_l (|c_{+,l}|^2+|c_{-,l}|^2)$. The vertical scale of panel (b) was amplified by a factor of ten to render the magnetic dipole excitations $|c_{-,l}|^2$ visible. The vertices on the plots correspond to the ASR positions. The lattice spacing $a=0.28\lambda$, meta-atom separation within an ASR $u=0.12\lambda$, $\Gamma_{\mathrm{E}}=\Gamma_{\mathrm{M}}$, and asymmetry parameter $\delta\omega = 0.3\Gamma$.[]{data-label="fig:uniformElectricMode"}](ElecModeElec.eps "fig:"){width="0.9\columnwidth"} ![The uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$ of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The excitations $|c_{\pm,l}|^2$ are normalized to the peak ASR excitation $E_0 = \max_l (|c_{+,l}|^2+|c_{-,l}|^2)$. The vertical scale of panel (b) was amplified by a factor of ten to render the magnetic dipole excitations $|c_{-,l}|^2$ visible. The vertices on the plots correspond to the ASR positions. The lattice spacing $a=0.28\lambda$, meta-atom separation within an ASR $u=0.12\lambda$, $\Gamma_{\mathrm{E}}=\Gamma_{\mathrm{M}}$, and asymmetry parameter $\delta\omega = 0.3\Gamma$.[]{data-label="fig:uniformElectricMode"}](ElecModeMag.eps "fig:"){width="0.9\columnwidth"}
![The uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$ of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The vertical scale of (a) was amplified by a factor of ten to render the electric dipole excitations $|c_{+,l}|^2$ visible. All parameters are as in Fig. \[fig:uniformElectricMode\]. []{data-label="fig:uniformMagneticMode"}](magModeElec.eps "fig:"){width="0.9\columnwidth"} ![The uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$ of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the uniform electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The vertical scale of (a) was amplified by a factor of ten to render the electric dipole excitations $|c_{+,l}|^2$ visible. All parameters are as in Fig. \[fig:uniformElectricMode\]. []{data-label="fig:uniformMagneticMode"}](magModeMag.eps "fig:"){width="0.9\columnwidth"}
Specifically, the uniform magnetic mode is the eigenmode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$ of $\mathcal{C}$ \[See Eq. .\] which maximizes the overlap $O_{{\textrm{m}}}({\ensuremath{\mathrm{b}}}_A)$ with the pure magnetic excitation ${\ensuremath{\mathrm{b}}}_{\mathrm{A}}$ in which all meta-atoms are excited with equal amplitude and the two meta-atoms in each split ring oscillate out of phase. Similarly, the uniform electric mode is the eigenvector ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$ of $\mathcal{C}$ that maximizes the overlap $O_{{\mathrm{e}}}({\ensuremath{\mathrm{b}}}_{{\mathrm{S}}})$ with the pure electric excitation ${\ensuremath{\mathrm{b}}}_{{\mathrm{S}}}$ for which all current oscillations oscillate in phase with equal amplitude. Explicitly, these column vectors of $2N_xN_y$ elements are given by $$\label{eq:b_A_and_b_S_def}
{\ensuremath{\mathrm{b}}}_{\mathrm{A}} \equiv
\left(
\begin{array}{c}
+1 \\
-1 \\
\vdots \\
+1 \\
-1
\end{array}
\right)\textrm{,}
\qquad
{\ensuremath{\mathrm{b}}}_{\mathrm{S}} \equiv
\left(
\begin{array}{c}
1 \\
1 \\
\vdots \\
1 \\
1
\end{array}
\right) \textrm{ .}$$ The alternating signs of the elements of $b_{\mathrm{A}}$ indicate the relative phase of the oscillations in each meat-atom of an ASR. We define the overlap of mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}/{\mathrm{e}}}$ with an arbitrary excitation ${\ensuremath{\mathrm{b}}}$ as $$O_{{\textrm{m}}/{\mathrm{e}}}({\ensuremath{\mathrm{b}}}) \equiv \frac{ | {\ensuremath{\mathrm{v}}}_{{\textrm{m}}/{\mathrm{e}}}^T {\ensuremath{\mathrm{b}}} |^2}
{\sum_i | {\ensuremath{\mathrm{v}}}_i^T {\ensuremath{\mathrm{b}}} |^2} \label{eq:overlapDef} \,\textrm{,}$$ where the index $i$ is summed over all the eigenmodes of the interaction matrix $\mathcal{C}$. The uniform electric and uniform magnetic modes for a $21 \times 21$ array of ASRs with a nonzero asymmetry parameter $\delta\omega=0.3\Gamma$ are shown in Figs. \[fig:uniformElectricMode\] and \[fig:uniformMagneticMode\], respectively. We used the experimental value for the lattice spacing from Ref. and the estimate the asymmetry parameter $\delta\omega \simeq 0.3 \Gamma$ from the relative arc lengths of the ASR meta-atoms studied by Fedotov *et al*.[@FedotovEtAlPRL2007; @FedotovEtAlPRL2010] The ohmic loss rate $\Gamma_{\mathrm{O}}$ was fitted so that the quality factor of the uniform magnetic mode as a function of system size matched the experimental observations. [@JenkinsLineWidthArxiv] Where the state of the ensemble is characterized by the vector of meta-atom normal variables ${\ensuremath{\mathrm{b}}}$ \[See Eq. .\], the symmetric (electric) and antisymmetric (magnetic) oscillations of an ASR $l$ ( $l=1,\ldots
N_xN_y$ ) are represented by $c_{+,l}$ and $c_{-,l}$, respectively, where $$\label{eq:2}
c_{\pm,l} \equiv \frac{1}{\sqrt{2}}(b_{2l-1} \pm b_{2l}) \textrm{ .}$$ The respective symmetric (electric dipole) and antisymmetric (magnetic dipole) excitation energies in ASR $l$ are proportional to $|c_{+,l}|^2$ and $|c_{-,l}|^2$. The asymmetry in the ASRs causes a mixing of the electric and magnetic dipoles, producing a slight electric dipole excitation in the uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$ and a slight magnetic excitation of the electric mode ${\ensuremath{\mathrm{v}}}_{{\mathrm{e}}}$. In the example illustrated in Figs. \[fig:uniformElectricMode\] and \[fig:uniformMagneticMode\], when the ohmic loss rate in each meta-atom is $\Gamma_{\mathrm{O}} \simeq 0.14\Gamma$, the electric mode has an enhanced decay rate $\gamma_{\mathrm{e}} = 2.7 \Gamma$ and the magnetic mode has a suppressed decay rate $\gamma_{{\textrm{m}}} = 0.31\Gamma$ with respect to the total isolated single meta-atom decay rate $\Gamma$. Due to the suppressed decay rate of ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$ and its small electric dipole component, this mode can be resonantly excited by the incident field.
![The response of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice to an incident plane wave electric field resonant on the uniform collective magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$, showing the excitation of a uniform magnetic dipole excitation at the expense of electric dipoles. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the response. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The ohmic loss rate is $\Gamma_{\mathrm{O}} = 0.14\Gamma$. All other parameters are as in Fig. \[fig:uniformElectricMode\]. []{data-label="fig:response-nonBroadened"}](response_hom_elec.eps "fig:"){width="0.9\columnwidth"} ![The response of a homogeneously broadened ( $\sigma=0$ ) $21 \times 21$ ASR square lattice to an incident plane wave electric field resonant on the uniform collective magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$, showing the excitation of a uniform magnetic dipole excitation at the expense of electric dipoles. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the response. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The ohmic loss rate is $\Gamma_{\mathrm{O}} = 0.14\Gamma$. All other parameters are as in Fig. \[fig:uniformElectricMode\]. []{data-label="fig:response-nonBroadened"}](response_hom_mag.eps "fig:"){width="0.9\columnwidth"}
We illustrate this phenomenon in Fig. \[fig:response-nonBroadened\], where we show the steady state response \[See Eq. .\], $$\label{eq:b_r_def}
{\ensuremath{\mathrm{b}}}_{\mathrm{r}} \equiv -
\mathcal{C}^{-1}{\ensuremath{\mathrm{f}}}_{\mathrm{in}} \,\textrm{,}$$ of an array whose resonance frequencies are not inhomogeneously broadened. The right and left meta-atoms of ASR $l$ in such an array have respective resonance frequencies $\omega_{2l-1} = \omega_{0} + \delta\omega$ and $\omega_{2l} =
\omega_0 - \delta\omega$ centered around $\omega_0$. The driving field is resonant on the uniform magnetic mode, and the asymmetry in the split rings $\delta\omega = 0.3 \Gamma$ facilitates the phase-coherent excitation of the magnetic dipoles at the expense of the electric dipoles. Panels (a) and (c) of Fig. \[fig:response-nonBroadened\] illustrate that the magnetic dipoles are much more strongly excited than the electric dipoles in the bulk of the array, and that these magnetic dipoles oscillate in phase. The more excited of the weak electric dipoles also oscillate in phase, thus facilitating the driving of this excitation by the uniform incident field. Although other collective modes of the system are excited, more than $60\%$ of the excitation energy resides in the uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$. In the absence of ohmic losses, one can optimize the asymmetry parameter $\delta\omega$ in large lattices so that over $98\%$ of the excitation energy resides in the uniform magnetic mode. [@JenkinsLineWidthArxiv]
![Response of inhomogeneously broadened square lattice to an incident plane wave resonant on the homogeneously broadened uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the response. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The individual ASR resonance frequencies are shifted by a independent identically distributed Gaussian random variables with standard deviation $\sigma = 0.8\delta\omega$. All other parameters are as in Fig. \[fig:response-nonBroadened\]. []{data-label="fig:response-Broadened"}](response_inhom_elec.eps "fig:"){width="0.9\columnwidth"} ![Response of inhomogeneously broadened square lattice to an incident plane wave resonant on the homogeneously broadened uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}$. (a) The electric dipole excitations $|c_{+,l}|^2$ and (b) the magnetic dipole excitations $|c_{-,l}|^2$ of the response. The phases of the electric ($c_{+,l}$) and magnetic ($c_{-,l}$) dipole excitations are indicated by the color of the surfaces in (a) and (b), respectively. The black dots indicate the positions of the ASRs in the array. The individual ASR resonance frequencies are shifted by a independent identically distributed Gaussian random variables with standard deviation $\sigma = 0.8\delta\omega$. All other parameters are as in Fig. \[fig:response-nonBroadened\]. []{data-label="fig:response-Broadened"}](response_inhom_mag.eps "fig:"){width="0.9\columnwidth"}
The introduction of inhomogeneous broadening alters the collective interactions and can destroy the characteristics of the metamaterial response that produces the transmission resonance. We model the inhomogeneous broadening by shifting the central resonance frequency of each ASR $l$ by independent identically distributed Gaussian random variables $X_l$ with zero mean and standard deviation $\sigma$. With the asymmetry characterized by $\delta\omega$, the right and left circular arcs in ASR $l$ possess resonance frequencies $\omega_{2l-1}
= \omega_0 + X_l +\delta\omega$ and $\omega_{2l}=\omega_0 + X_l -
\delta\omega$, respectively. The deleterious effects of this broadening on the response are illustrated in Fig. \[fig:response-Broadened\], which shows a much less uniform magnetic response in addition to localized electric dipole excitations. This non-uniformity inhibits the coherent reflection and transmission through the meta-material array.
![(a) The overlap of the uniform magnetic mode with the excitation driven by an incident EM plane wave and (b) the decay rate $\gamma_{{\textrm{m}}}(\sigma)$ of the uniform magnetic mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}(\sigma)$ as a function of inhomogeneous broadening $\sigma$. These quantities are plotted for $\delta\omega = 0.3\Gamma$ and ohmic losses $\Gamma_{\mathrm{O}} = 0.14\Gamma$ (solid black line), $\delta\omega = 0.3 \Gamma$ and $\Gamma_{\mathrm{O}} = 0$ (dashed blue line), and $\delta\omega = 0.1 \Gamma$ and $\Gamma_{\mathrm{O} } = 0$ (dot dashed-red line). The lines indicate the average over $240$ sample realizations, and the error bars indicate the standard deviations. The incident wave has electric field polarization ${\ensuremath{\mathbf{\hat{e}}}}_y$ aligned with the ASR electric dipoles and is resonant on the mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}(\sigma=0)$ of the homogeneously broadened array. The decay rates of the uniform magnetic modes for a homogeneously broadened array $\gamma_{{\textrm{m}}}(\sigma=0)$ are $0.308\Gamma$ (black line), $0.145\Gamma$ (blue line) and $0.0371\Gamma$ (red line). This shows that excitation of the magnetic mode, and hence the transmission resonance, vanishes as the inhomogeneous broadening becomes comparable to $\delta\omega$. []{data-label="fig:modeOverlap"}](mModeOverlapAndGamma.eps)
We quantify the effects of inhomogeneous broadening by examining the overlap of the metamaterial steady state response ${\ensuremath{\mathrm{b}}}_{\mathrm{r}}$ with the magnetic mode ${\ensuremath{\mathrm{v}}}_{\textrm{m}}(\sigma=0)$ of a homogeneously broadened array (corresponding to the case in which all ASR meta-molecules are identical). If the excitation is purely in the mode ${\ensuremath{\mathrm{v}}}_{\textrm{m}}(0)$, then the overlap $O_{\textrm{m}}({\ensuremath{\mathrm{b}}}_{\mathrm{r}})$ \[Eq. \] is unity. Figure \[fig:modeOverlap\](a) shows the overlap $O_{\textrm{m}}({\ensuremath{\mathrm{b}}}_{\mathrm{r}})$ of the response to an incident field resonant on the homogeneously broadened mode ${\ensuremath{\mathrm{v}}}_{{\textrm{m}}}(\sigma=0)$ averaged over $240$ realizations. The solid black line was calculated for the same parameters as in Figs. \[fig:uniformElectricMode\] through \[fig:response-Broadened\] with varying degrees of broadening. The blue dashed curve shows the corresponding overlap in the absence of ohmic losses, while for the red dashed-dot curve, the asymmetry parameter was reduced to $\delta\omega=0.1$ and $\Gamma_O=0$. In all cases, as the broadening standard deviation $\sigma$ becomes comparable to $\delta\omega$, the ability to excite the uniform magnetic mode drastically decreases. The other modes that are excited either contain electric dipole components or are not phase matched. This either results in scattering of the field or in absorption of the field due to ohmic losses. As a result, the coherent collective response responsible for the transmission resonance observed in by Fedotov *et al* [@FedotovEtAlPRL2010] becomes unobservable when the inhomogeneous broadening is larger than $\delta\omega$. These collective effects do persist, however, for $\sigma$ roughly half $\delta\omega$. We show the effect of inhomogeneous broadening in the decay rate $\gamma_{{\textrm{m}}}(\sigma)$ of the magnetic mode itself in Fig. \[fig:modeOverlap\](b) as a function of broadening. Randomization of the ASR resonance frequencies apparently has little effect on the collective linewidth of the magnetic mode for $\sigma <
\delta\omega$. However, for larger degrees of broadening, the decay rate of this mode can be increased several times over and the resonance linewidth narrowing, that results from the cooperative response of the metamaterial array, disappears. Furthermore, the large standard deviations of $\gamma_{\textrm{m}}(\sigma)$ indicate that the width of the uniform magnetic mode is highly sensitive to the particular realization of ASR resonance frequencies. A larger decay rate renders the magnetic mode more difficult to excite since any excitation of this mode is more quickly radiated away. In order to for an array of ASRs to exhibit a transmission resonance, a large fraction of the excitation created by the driving field must be in the uniform magnetic mode distributed over the array. In low loss metamaterials, this can be achieved in conjunction with a higher quality for that resonance for larger arrays and smaller values of $\delta\omega$. [@JenkinsLineWidthArxiv] However, as Fig. \[fig:modeOverlap\] indicates, reduction in the asymmetry to achieve this quality factor enhancement correspondingly reduces the tolerance for inhomogeneous broadening in the resonance frequency.
The observation that the cooperative metamaterial response to EM fields can be suppressed in the presence of sufficiently strong inhomogeneous broadening could potentially also be exploited in design of metamaterial samples that would benefit from well-defined homogeneous properties for electric susceptibility and magnetic permeability, such as diffraction-free lenses due to negative refractive index.[@SmithEtAlPRL2000; @ShelbySci2001; @SmithEtAlSCI2004] One could prepare a controlled amount of inhomogeneous broadening for the metamaterial sample in order to generate an EM response that more closely mimics standard continuous medium electrodynamics with suppressed contribution from recurrent scattering events.
Conclusion {#sec:conclusion}
==========
In conclusion, we have analyzed the collective modes of a finite-sized 2D metamaterial array of ASR resonators and how they are influenced by an inhomogeneous broadening of the resonance frequencies of the individual resonators. The study was motivated by recent experimental observations of transmission resonance linewidth narrowing as a function of the size of the system.[@FedotovEtAlPRL2010] This effect can be understood by analyzing the resonance linewidths of collective modes of the system that undergo dramatic narrowing due to strong EM field mediated interactions between the resonators. As demonstrated by previous comparisons between the numerical simulation results and the experimental observations,[@JenkinsLineWidthArxiv] the response can be analyzed by a simplified model in which each meta-atom is treated as a discrete element supporting a single mode of current oscillation possessing electric and magnetic dipole moments. Collective interactions between the meta-atoms are mediated by the scattered EM fields. The excellent agreement between the theory and the experiment can be understood by a relatively weak higher-order multiple radiation of individual ASR metamolecules.[@PapasimakisComm]
We examined in detail how inhomogeneous broadening of resonator resonance frequencies impairs the coherent collective phenomena that are expected to find important applications in metamaterial systems.[@ZheludevEtAlNatPhot2008; @KAO10; @SentenacPRL2008; @FedotovEtAlPRL2007] While the transmission resonance experimentally observed in Ref. persists for inhomogeneous broadening that is a fraction of the ASR asymmetry parameter $\delta\omega$, the cooperative response vanishes when the broadening begins to exceed that parameter. Production of high quality resonances with low loss materials requires the reduction of $\delta\omega$.[@JenkinsLineWidthArxiv] Figure \[fig:modeOverlap\] indicates that, in order to produce such high quality resonances, the uniformity in the production of meta-molecules will need to become correspondingly small.
We acknowledge discussions with N. Papasimakis, V. Fedotov, and N. Zheludev and financial support from the EPSRC and the Leverhulme Trust.
[30]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [**** ()]{} @noop [****, ()]{} , , , , ****, (). [****, ()](\doibase 10.1103/PhysRevLett.101.013901) @noop [**]{}, Vol. (, , ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/RevModPhys.70.447) [****, ()](\doibase 10.1103/PhysRevLett.82.4741) [****, ()](\doibase 10.1103/PhysRevA.64.053804) [****, ()](\doibase 10.1103/PhysRevE.69.026605) @noop [ ****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.76.062509) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop (), @noop [**]{} (, , )
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In recent years Planck-scale modifications of the dispersion relation have been attracting increasing interest also from the viewpoint of possible applications in astrophysics and cosmology, where spacetime curvature cannot be neglected. Nonetheless the interplay between Planck-scale effects and spacetime curvature is still poorly understood, particularly in cases where curvature is not constant. These challenges have been so far postponed by relying on an [*ansatz*]{}, first introduced by Jacob and Piran. We here propose a general strategy of analysis of the effects of modifications of dispersion relation in FRW spacetimes, applicable both to cases where the relativistic equivalence of frames is spoiled (“preferred-frame scenarios") and to the alternative possibility of “DSR-relativistic theories", theories that are fully relativistic but with relativistic laws deformed so that the modified dispersion relation is observer independent. We show that the Jacob-Piran [*ansatz*]{} implicitly assumes that spacetime translations are not affected by the Planck-scale, while under rather general conditions the same Planck-scale quantum-spacetime structures producing modifications of the dispersion relation also affect translations. Through the explicit analysis of one of the effects produced by modifications of the dispersion relation, an effect amounting to Planck-scale corrections to travel times, we show that our concerns are not merely conceptual but rather can have significant quantitative implications.'
author:
- Giacomo ROSATI
- 'Giovanni AMELINO-CAMELIA'
- Antonino MARCIANÒ
- Marco MATASSA
title: 'Planck-scale-modified dispersion relations in FRW spacetime'
---
Introduction
============
The possibility that Planck-scale effects might modify the “dispersion relation", the on-shell requirement linking energy and momentum of a particle, has attracted quite some interest in the recent quantum-gravity literature (see, [*e.g.*]{}, Refs. [@gacLRR; @mattinglyLRR] and references therein). Motivation for the study of this possibility also comes from the fact that it provides a rare case of conjectured Planck-scale feature that could produce effects observable already with presently-available experimental technologies. These opportunities for testing are not for controlled Earth-bound laboratory setups, where the effects would still be too small, but rather they arise in some contexts of astrophysics and cosmology where the ultralarge propagation distances act as amplifier of the minute Planck-scale effects [@gacLRR; @mattinglyLRR].
These opportunities from astrophysics and cosmology however also bring about a challenge for theories, since in the relevant quantum-spacetime pictures very little has been so far understood about the interplay between curvature and Planck-scale effects. This is particularly true for cases where the analysis does not allow schematization in terms of a constant spacetime curvature, as it is indeed the case when tests are performed exploiting cosmological distances.
These challenges have been so far mostly postponed, assuming the applicability of an [*ansatz*]{}, first formulated by Jacob and Piran [@jacobpiran] (also see Refs.[@Ellis:2005wr; @PiranMartinez]) for this interplay between Planck-scale modifications of the dispersion relation and spacetime curvature.
We here propose a general strategy of analysis of the effects of modifications of dispersion relation applicable when a non-constant curvature of spacetime is to be taken into account. We adopt a phenomenological approach: rather than attempting to establish a specific form of interplay between Planck-scale effects and spacetime curvature within one or another quantum-spacetime picture, we use what is presently known about the various possibilities for formalizing a quantum spacetime as guidance for modeling in a very general way the possible forms of this interplay.
Of particular interest is the fact that we allow for Planck-scale features to be present not only in the dispersion relation but also in the description of translation transformations. This is important since in some of the most studied quantum spacetimes, such as the “$\kappa$-Minkowski noncommutative spacetime [@majidruegg; @lukieKAPPAmink; @bobKAPPAplb], dedicated analyses have shown that the Planck scale does have this double role, affecting both the dispersion relation and translations.
In this respect we uncover a particularly significant characterization of the Jacob-Piran [*ansatz*]{}: we show that this [*ansatz*]{} implicitly assumes that translations are unaffected by Planck-scale structures. By this we mean that in particular specializing the Jacob-Piran [*ansatz*]{} to the case of constant rate of spacetime expansion (de Sitter spacetime) one gets a picture that is invariant under ordinarily classical space and time translations. Analyses depicted as full explorations of the implications of modified dispersion relations, which instead rely exclusively on the Jacob-Piran [*ansatz*]{}, should be more carefully described as testing the dispersion relation under the restrictive requirement that translations should be unaffected by Planck-scale features.
Another qualifying aspect of our analysis is that it considers both the case of modifications of the dispersion relation that signal “LIV" (Lorentz Invariance Violation, [*i.e.*]{} preferred-frame scenarios) and the very different case where the modified dispersion relation is implemented within a fully relativistic picture. This latter possibility of course requires deforming the relativistic laws of transformation among observers in a way suitable for enforcing the modified dispersion relation as an observer-independent law, a possibility that can be formalized in terms of the relativistic theories of type “DSR" (doubly-special, or, for some authors, deformed-special, relativity), first introduced in Ref. [@gacdsr1IJMPD] (also see Refs. [@jurekDSRfirst; @joaoleePRDdsr]).
As well established in the recent literature, the DSR case imposes taking into account of the novel relativistic effect of “relative locality" [@principle] (also see Refs.[@bob; @bobKAPPAplb]). Within the scopes of our analysis this is tedious but straightforward. However, it should be noted that we here report the most advanced phenomenological analysis to date of the interplay between relative locality and (non-constant) spacetime curvature.
Most of the issues we are concerned with are already present in the constant-curvature case, as we shall show explicitly. But extending the analysis to cases with non-constant curvature leads to encounter additional challenges. The constant-curvature case allows us to analyze most aspects of the problem at hand in terms of pure symmetry considerations. This is not directly available in cases with non-constant curvature; however, we recover a role for symmetries also in the case of non-constant curvature by using the fact that locally the symmetries of the constant-curvature case reemerge. One can in particular describe a finite path within a FRW-type spacetime by gluing infinitesimal paths within a suitable series of deSitter-type spacetimes, a strategy of analysis we refer to as “thick slicing".
While the conceptual significance of our concerns should become clear to our readers very early in the analysis here reported, it is important for us to show that our concerns can also have significant quantitative implications. It is for this reason that we devote much effort to the illustrative application provided by the description of the path of a particle in a “quantum-FRW spacetime" and the evaluation of travel times from a given source to a given detector.
In light of the conceptual and quantitative complexity of the issues we are dealing with we opt for focusing on the case of a 1+1-dimensional spacetime and obtaining results only at leading order in the ultrasmall Planck length.
Preliminaries on classical de Sitter spacetime and translations in a quantum spacetime
======================================================================================
Before starting with the main part of our analysis we find convenient to collect in this section some known facts that we shall then use. In the first part of this section we shall review some known facts about classical de Sitter spacetime, the ones most relevant for our later discussion of propagation of particles in a quantum spacetime with curvature. In the second part we remind our readers about the interconnection found in some much-studied quantum spacetimes between Planck-scale modifications of the dispersion relation and Planck-scale modifications of translation transformations.
Covariant mechanics in de Sitter spacetime {#sec:deSitter}
------------------------------------------
In this section we present a covariant Hamiltonian formulation for the motion of a classical point particle in de Sitter spacetime. Taking $E,p,N$ to be respectively the generators of time translations, space translations, and boosts, the algebra of spacetime symmetries for 2D de Sitter spacetime can be described in terms of Poisson brackets as $$\left\{ E,p\right\} =Hp, ~~
\left\{ N , E \right\} =p+HN, ~~
\left\{ N , p\right\} =E.
\label{algebraDS}$$ The spacetime-symmetry generators leave invariant the Casimir $${\cal C} = E^2 - p^2 -2 H N p.
\label{casimirDS}$$
Let us then consider the “conformal time coordinatization” with the spatial coordinate $x$ the conformal-time coordinate $\eta$, related to the comoving time $t$ by $$\eta = H^{-1}\left( 1 - e^{-Ht}\right) \, ,$$ and their canonically conjugate variables $\Omega,\Pi$: $$\begin{gathered}
\left\lbrace \Omega, \eta \right\rbrace = 1, ~~~ \left\lbrace \Omega, x \right\rbrace = 0, ~~~ \left\lbrace \Omega, \Pi \right\rbrace = 0, \nonumber \\
\left\lbrace \Pi, \eta \right\rbrace = 0, ~~~ \left\lbrace \Pi, x \right\rbrace = -1, ~~~ \left\lbrace \eta, x \right\rbrace = 0.
\label{CanPhSpConf}\end{gathered}$$ We can represent the symmetry generators (\[algebraDS\]) in terms of conformal-time coordinates as $$\begin{gathered}
E= \Omega (1- H \eta) + H x \Pi, ~~ p=\Pi, \nonumber \\
N = x \Omega (1-H\eta) -\Pi \left(\eta -\frac{ H}{2} \eta^2 - \frac{H}{2} x^2\right).
\label{representDS}\end{gathered}$$
The invariant ${\cal H} = {\cal C} - m^2$, with $m$ the mass of the particle, can be taken to describe the particles on-shell relation (${\cal H}=0$), and then can be used as Hamiltonian constraint, in the spirit of a “covariant formulation" of classical mechanics, generating equations of motion in terms of an auxiliary affine parameter, which we denote by $\tau$. In terms of conformal-time coordinates (\[representDS\]) the mass-shell relation ${\cal H}=0$ assumes the “conformal” aspect $${\cal H} = (1-H\eta)^2 \left( \Omega^2 -\Pi^2\right) - m^2 = 0.
\label{on-shell-dS}$$ The fact that the Poisson brackets between ${\cal H}$ and $E,p,N$ vanish, implies that $E,p,N$ are conserved charges, i.e. they are constant along the evolution generated by the Hamiltonian $\dot{E}=\dot{p}=\dot{N}=0$, where $\dot{f} = d f / d \tau = \{{\cal H}, f\}$. Then $E,p,N$ generate symmetry transformations (time and space translations and boosts) on the system by Poisson brackets.
One can derive the velocity of particles from the Hamiltonian constraint and the Poisson brackets (\[CanPhSpConf\]). Indeed by the chain rule[^1] $$v(\eta) \! = \frac{dx(\eta)}{d\eta} \!= \frac{dx/d\tau}{d\eta/d\tau}\Big|_{{\cal H}=0} \!= \frac{\left\lbrace {\cal H},x\right\rbrace}{\left\lbrace {\cal H}, \eta\right\rbrace}\Big|_{{\cal H}=0} .$$ By using Eqs. (\[on-shell-dS\]) and (\[CanPhSpConf\]) one thus finds the velocity in conformal-time coordinates (assuming $\Pi >0$) $$v(\eta) = \frac{\Pi}{\sqrt{\Pi^2 - \frac{m^2}{\left( 1 - H\eta\right)^2}}} ~~~ \stackrel{m\rightarrow 0}{\longrightarrow} ~~~ 1.
\label{velocityConf}$$ Notice that in the equivalent description given in terms of the comoving time the velocity is $$v(t) \! = \! \frac{dx}{dt} \!=\! e^{\!-Ht} v(\eta(t)) \!=\! \frac{ e^{-Ht} \Pi}{\sqrt{\Pi^2 \!-\! e^{2Ht} m^2 }} \stackrel{m\rightarrow 0}{\longrightarrow} e^{-Ht} .
\label{velocityCom}$$
Preliminaries on Planck-scale deformations of translations
----------------------------------------------------------
As announced, we shall find that there are important issues at stake particularly in the understanding of how Planck-scale effects may affect translation transformations. It is valuable for us to provide early on in this manuscript some intuition to our readers concerning the reason why these aspects of translation transformations become so significant when spacetime curvature is taken into account, while they are instead largely irrelevant in the Minkowski limit. We shall be satisfied in this subsection with offering only a few remarks and exclusively for the DSR-relativistic case. In the following sections the relevant issues will of course be analyzed in detail, both for the DSR-relativistic case and the LIV case.
In the Minkowski limit the realm of possibilities we could consider is rather tightly constrained by the fact that the characteristic scale of quantum-gravity effects with the dimensions of an inverse-momentum (which we denote by $\ell$ for DSR), is the only scale available for deforming the mass-shell relation. This in particular implies that at leading order in this deformation scale one can only have two possible new terms, one proportional to the cubic power of energy $\ell E^3$ and one linear in energy and quadratic in momentum $\ell E p^2 $, so that[^2] $$m^2=E^2-p^2+\alpha\ell E^3+\beta\ell Ep^2$$
It is then easy to see why such modifications of the mass-shell relation have no impact on translation transformations in the Minkowski limit: both the correction term $\ell E^3$ and the correction term $\ell E p^2 $ have vanishing Poisson brackets with the generators of translations, which are $E$ and $p$ themselves, and have vanishing Poisson brackets among them in the Minkowski limit: $$\{E,p\}=0$$ so that the modified mass-shell is still invariant under the standard action of space and time translation generators.
As we underlined in the previous subsection, the situation is very different when spacetime is curved: the translation generators no longer have vanishing Poisson brackets, as shown by Eq. (\[algebraDS\]) $$\{E,p\}=Hp \, .$$
It turns out that in the DSR case and in presence of spacetime curvature, the assumption of undeformed translation transformations is only compatible with the term $\ell E p^2$ whereas allowing for the term $\ell E^3$ imposes a rather severe modification of translation transformations.
Since the quantum-gravity literature provides no argument favouring the term $\ell E p^2$ over the term $\ell E^3$ evidently any systematic study of quantum-gravity modifications of the mass-shell relation should take into account what we here find on the implications of the translation sector.
LIV with constant curvature {#sec:LIVdeSitter}
===========================
Evidently the general idea of a LIV modification of de Sitter kinematics could be realized in an infinity of ways: if one allows breaking relativistic symmetries, as is the premise of the LIV scenario, the framework is totally unconstrained. It should be understood that a fully general analysis of the LIV scenario is accordingly impossible. It is nonetheless useful for our purposes to consider at least a few possibilities, also as a way to show that the differences between alternative LIV scenarios are not “merely academic” but rather have tangible (potentially observable) consequences. This objective is pursued efficiently by taking as starting point for the LIV modification the de Sitter property coded in (\[on-shell-dS\]) and adding a few parametrized terms of LIV type: $$\begin{split}
{\cal H} \!= & -m^2 + (1 \!-\! H\eta)^2 \!\left( \Omega^2 \!-\! \Pi^2\right) \!+\! \left( \tilde{\alpha} \Omega^3 \!\!+\! \tilde{\beta} \Omega \Pi^2 \right) \! \times\\ &
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \times \!\! \left[ \! \lambda' \!\! \left( 1 \!-\! H\eta \right) \!+\! \lambda'' \!\! \left( 1 \!-\! H\eta \right)^2 \!\!+\! \lambda \! \left( 1 \!-\! H\eta \right)^3 \!\!+\!\! \lambda''' \!\! \left( 1 \!-\! H\eta \right)^4 \! \right]\!\!.
\end{split}
\label{dispLIVrep}$$ Here $\lambda'$, $\lambda''$, $\lambda$ and $\lambda'''$ are different choices of the inverse-momentum scale characterizing the contribution of different LIV terms, which differ essentially for their time dependence, while $\tilde{\alpha},\tilde{\beta}$ are parameters governing the dependence of these terms on the particle’s energy and momentum. Of course, one could add not only other forms of time dependence of the LIV terms, but also LIV terms of completely different form (for example involving spatial dependence), but these few terms we introduced will suffice for exposing the strength of our concerns.
Important for the main objective of this study is the fact that the LIV modification coded in (\[dispLIVrep\]) is not in general translationally invariant. One can check by using (\[algebraDS\]) and (\[CanPhSpConf\]) that (\[dispLIVrep\]) is invariant under the action of spatial translations $p$, but is not in general invariant under the action of time translations generated by $E$. The Poisson bracket between $E$ and ${\cal H}$ is $$\begin{split}
& \left\{{\cal H}, E\right\} = H \left(\tilde{\alpha}\Omega^{3}
+\tilde{\beta}\Omega\Pi^{2}\right) \times \\ &
\times \! \left(\! -2\lambda'\left(1 \!-\! H\eta\right)
-\lambda''\left(1 \!-\! H\eta\right)^{2}
+\lambda'''\left(1 \!-\! H\eta\right)^{4}\right)
\end{split}
\label{joclambda}$$ Thus, in general, not only (evidently) boost symmetry but also (more implicitly) time-translational symmetries are broken by a LIV scenario. Among the LIV parameters we introduced only $\lambda$ is not present in (\[joclambda\]), meaning that $\lambda$ is compatible with translational invariance, while all other parameters ($\lambda'$,$\lambda''$,$\lambda'''$) are incompatible with translational invariance. These differences are particularly striking since from the “LIV perspective" (the perspective of Lorentz Invariance Violation) the four parameters $\lambda$,$\lambda'$,$\lambda''$,$\lambda'''$ should be viewed at exactly the same level, since their Lorentz-transformation properties are completely analogous.
As a way to quantify the differences produced by different LIV parameters we perform a travel time analysis. We start by noticing that from the Hamiltonian constraint (\[dispLIVrep\]), for massless particles, one gets the velocity (for $\Pi>0$) $$v(\eta) \! =\! 1 \!-\! \Pi \! \left(\!\! \frac{\lambda'}{\left(1 \!-\! H\eta\right)} \!+\!\! \lambda'' \!\!\!+\! \lambda \! \left(1 \!-\! H\eta\right) \!+\!\! \lambda''' \!\!\left(1 \!-\! H\eta\right)^{2} \!\! \right)
\label{velocityLIVdeSitter}$$ where we have set the parameters $\tilde{\alpha} \!+\! \tilde{\beta} \!=\! 1$, noticing that for massless particles they can be both reabsorbed in the definition of the parameters $\lambda$,$\lambda'$,$\lambda''$,$\lambda'''$.
In looking for the travel time for a particle to go from one observer to another it is important to notice that the breakdown of (time) translational invariance implies that the description of a particle’s Hamiltonian given in (\[dispLIVrep\]), and the velocity law (\[velocityLIVdeSitter\]), can only hold for one of the two distant observers: if (\[dispLIVrep\]) holds for Bob then Alice’s description is given by the non-trivially translated version of (\[dispLIVrep\]). Suppose that (\[dispLIVrep\]) and (\[velocityLIVdeSitter\]) hold in the frame of Bob who is local to a detector, and suppose that this detector reveals two photons, one “soft” (whose energy is small enough that the LIV effects are negligible) and one “hard”, emitted simultaneously at a distant source, local to an observer, say Alice, at rest relatively to Bob ([*i.e.*]{} Alice and Bob are related by a pure translation). The relation between Bob and Alice’s coordinates, connected by a finite deSitter-translation, can be derived by exponentiating the action by Poisson brackets of the translation generators $E,p$, Eqs. (\[representDS\]) and (\[CanPhSpConf\]). Taking Bob to be connected to Alice by a finite spatial translation followed by a finite time translation, we write symbolically $$\left( \eta,x \right)^B= e^{-\xi p}\triangleright e^{-{\zeta}E}\triangleright \left( \eta,x \right)^A,
\label{TTR}$$ where $\zeta$ and $\xi$ are the finite translation parameters, and $\triangleright$ stands for the action by nested Poisson brackets[^3]. One finds $$\begin{gathered}
\eta^{B} \!=\! e^{H\zeta} \eta^A - \frac{e^{H{\zeta}} \!-\! 1}{H}, \qquad
x^{B} = e^{H{\zeta}} (x^{A} - \xi ), \\
\Omega^{B} = e^{-H{\zeta}} \Omega^{A}, \qquad
\Pi^{B} = e^{-H{\zeta}} \Pi^{A},
\end{gathered}
\label{TransCoord-dS}$$
On the basis of (\[velocityLIVdeSitter\]) one deduces that Bob describes the photons trajectories to be $$x^B(\eta^B) = x^B_{O_A} + \int_{\eta^B_{O_A}}^{\eta^B} v^B(\eta) d\eta,
\label{trajectoryLIVdeSitter}$$ where $v^B$ is given by Eq. (\[velocityLIVdeSitter\]) written in Bob coordinates, and $x^B_{O_A},\eta^B_{O_A}$ are the coordinates that Bob assigns to the event of emission, which coincides with Alice’s origin, i.e. $\eta^B_{O_A} = \eta^B(x^A=0,\eta^A=0)$, $x^B_{O_A} = x^B(x^A=0,\eta^A=0)$. Suppose that the event of detection of the soft photon coincides with Bob frame’s origin ($x^B = \eta^B = 0$). Since for the soft photon $v(\eta)\simeq 1$, (\[trajectoryLIVdeSitter\]) implies that $x^B_{O_A}=\eta^B_{O_A}$. One then obtains the time Bob assigns to the arrival of the hard photon by setting to zero Eq. (\[trajectoryLIVdeSitter\]), and solving for $\eta^B$. From (\[trajectoryLIVdeSitter\]) and (\[velocityLIVdeSitter\]), and using that $x^B_{O_A}=\eta^B_{O_A}$, one finds that the hard photon is detected with a delay $$\Delta\eta \!=\! p^B_h \!\! \left[\lambda'T \!+\! \lambda''\frac{e^{HT} \!\!\!-\! 1}{H} \!+\!\lambda\frac{e^{2HT} \!\!\!-\! 1}{2H} \!+\! \lambda'''\frac{e^{3HT} \!\!\!-\! 1}{3H}\right]
\label{delayLIVdeSitter}$$ where we called $p^B_h$ the hard particle’s momentum measured by Bob, we used Eq. (\[TransCoord-dS\]) to derive $\eta^B_{O_A}$, and we called $T=\zeta$ the (comoving) time distance between Alice and Bob. We can express the delay in terms of the redshift of the source (relative to Bob) $z \!=\! -\! H \eta^B_{O_A} \!=\! e^{HT} \!\!-\! 1$: $$\!\!\!\!\!\!\!\!\!\!\!
\Delta\eta \!=\! \frac{p^B_h}{H} \!\! \left[\lambda' \! \ln\left(1 \!+\! z\right) \!+\!\! \lambda'' \! z \!+\!\! \lambda \! \left( \! z \!+\! \frac{z^{2}}{2} \!\right) \!\!+\!\! \lambda''' \!\! \left( \! z \!+\! z^{2} \!+\! \frac{z^{3}}{3}\right) \!\right]
\label{dsLIVmain}$$ where we also expressed the delay in terms of comoving time. Notice that, since $\Delta \eta = O(\ell \Pi)$, it follows that $$\Delta t = \int_0^{\Delta \eta} d\eta a(\eta) = \Delta \eta
+ O\left(\ell ^2\right),
\label{DtDeta}$$ so that the expression of the delay in conformal and comoving time coincide at the level accurately described by the approximations we are relying on.
The term in the delay (\[dsLIVmain\]) proportional to $\lambda$ coincides with the one advocated by Jacob and Piran in [@jacobpiran]. Jacob and Piran provided as motivation an intuitive argument in favour of switching on only the $\lambda$ parameter in LIV phenomenology. Here we exposed the fact that the Jacob-Piran intuition unknowingly reflected a preference for assuming that the quantum-gravity effects leave translations unaffected[^4]. However, as stressed above, the quantum-gravity literature providing motivation for LIV research does not justify the assumption that quantum-gravity effects should leave translations unaffected. This makes us particularly concerned for the fact that limits claiming general applicability to LIV scenarios have been presented in the literature [@Bolmont:2010np; @Ellis:2011ek; @HESS:2011aa; @Aharonian:2008kz; @Albert:2007qk; @Abdo:2009zza; @Ackermann:2009aa], even though they were based exclusively on the Jacob-Piran [*ansatz*]{}. The observations reported in this section, strengthened by what we shall find in the following sections, show instead that those limits only apply to a particular case of LIV. They should be viewed as “conditional limits", the condition being indeed that translations are unaffected by the quantum-gravity effects producing the breakdown of Lorentz invariance.
In closing this section we offer one more remark that can provide some intuition for the significance of our concerns, a remark which applies to the description of the equations of motion. For this we consider a third observer Bob$'$, at rest relatively to Alice and Bob, whose origin is along the soft photon worldline connecting Alice and Bob, at some point between Alice and Bob, a point which is at (comoving) time $T'$ from the origin of the reference frame of Bob. The coordinates of Bob$'$ will be related to the ones of Bob by expression (\[TransCoord-dS\]) replacing the coordinates of Alice with the ones of Bob$'$ and replacing $\zeta$ with $T'$. Then, from (\[TransCoord-dS\]) and (\[velocityLIVdeSitter\]), it follows that Bob$'$ describes the photons traveling with velocity $$\begin{split}
& v^{B'}\left(\eta^{B'}\right)=1-\Pi^{B'} \bigg(\frac{\lambda'e^{-2HT'}}{\left(1-H\eta^{B'}\right)}+\lambda''e^{-HT'} \\ &
~~~~~~~~~ + \lambda\left(1-H\eta^{B'}\right)+\lambda'''e^{HT'}\left(1-H\eta^{B'}\right)^{2}\bigg)
\end{split}
\label{velocityLIVdeSitter-2}$$ Here it is particularly important to notice that the only term which maintains the same form of (\[velocityLIVdeSitter\]) is the one proportional to $\lambda$, which indeed corresponds to the translational invariant term in the Hamiltonian (\[dispLIVrep\]).
DSR-relativistic picture and relative locality with constant curvature {#sec.DSR-dS}
======================================================================
In the DSR approach, the (inverse-momentum) scale $\ell$ at which the dispersion relation is modified is a relativistic invariant (it plays a role completely analogous to the role of $c$, the velocity of light, in ordinary special relativity). This requirement enforces the ($\ell$-deformed) dispersion relation to be expressible as a combination of only the (charges) generators of the relativistic symmetries; in fact, it must be a Casimir of an algebra of charges/generators. As first observed in Ref. [@gacdsr1IJMPD] (also see Refs. [@jurekDSRfirst; @joaoleePRDdsr]), the relativistic invariance of a modified dispersion relation requires that the algebra of relativistic-symmetry generators must also be deformed by the scale $\ell$.
In light of these considerations, it is preferable, for the purposes of the DSR analysis, to start from a modification, rather then of the dispersion relation (\[on-shell-dS\]), of the Casimir equation (\[casimirDS\]). The motion of particles is then generated by the deformed Hamiltonian constraint ${\cal H} = {\cal C} - m^2$. We take the following deformation[^5] $${\cal C}=E^{2}-p^{2}-2H N p+\ell\left(\alpha E^{3}+\beta Ep^{2}\right) ~,
\label{casimirDSR-DS}$$ The following $\ell$-deformed (2D) de Sitter algebra of charges is compatible with the invariance of ${\cal C}$: $$\begin{gathered}
\left\{ E,p\right\} =Hp-\ell (\alpha -\gamma) HEp \ ,\nonumber \\
\left\{ N , E \right\} =p+HN-\ell (\alpha -\gamma) E(p+HN)-\ell\beta Ep\ ,\nonumber \\
\left\{ N , p\right\} =E+\frac{1}{2}\ell (\alpha + 2 \gamma) E^{2}+\frac{1}{2}\ell\beta p^{2}\ .
\label{algebraDSR-DS}\end{gathered}$$ This generalizes the results of Ref. [@DSR-DS], and reproduces the results of Ref. [@DSR-DS] for $\gamma = 0$.
Notice that, as stressed already earlier in this manuscript, the deformation term going like $E p^2$ does not require a modification of translation transformations while for the deformation term going like $E^3$ a modification of translation transformations is required. This is particularly clearly by looking at Eqs.(\[casimirDSR-DS\])-(\[algebraDSR-DS\]): the parameter $\alpha$ of the term going like $E^3$ has implications not only for transformations involving boosts ($\alpha$ affects $\left\{ N , E \right\}$ and $\left\{ N , p\right\}$) but also for transformations involving exclusively translations ($\alpha$ also affects $\left\{ E,p\right\}$), whereas the parameter $\beta$ of the term going like $E p^2$ does not affect $\left\{ E,p\right\}$.
To study the motion of particles, we want to express the deformed Hamiltonian ${\cal H} = {\cal C} - m^2$ in terms of a set of spacetime coordinates and conjugate momenta, and derive the particles velocity in terms of these coordinates, as done in Subsec. \[sec:deSitter\] and Sec.\[sec:LIVdeSitter\]. At this stage one must expect a crucial difference with respect to the LIV case, a feature known in the literature as “relative locality" [@principle] (also see Refs.[@bob; @bobKAPPAplb]. This is another consequence of the fact that, in the DSR case, the algebra of the generators of spacetime relativistic-symmetry transformations is deformed by the inverse-momentum scale $\ell$: the result is that the action of the whole set of the symmetry generators on spacetime coordinates must be necessarily momentum-dependent, leading to the presence of momentum-dependent misleading inferences for the description by a given observer of events occurring at a from that observer (large distance from the origin of the reference frame of that observer). This core feature of “relative locality" [@principle] is usually rather surprising for newcomers, but actually it can be easily understood via an analogy with the one case of deformation of relativistic symmetries we all know, the one of the deformation from Galileian to special-relativistic symmetries: according to the mindset of a Galileian-era physicist the special-relativistic transformations introduce misleading inferences concerning the simultaneity of events, misleading inferences of increasing severity when the compared notions of simultaneity concern observers connected by a .
We shall deal with these challenges drawing from previous results on relative locality, but we shall at the same time provide sufficient details for a self-contained description of the relative-locality effects.
We start by seeking a representation[^6] of the charges (\[algebraDSR-DS\]) suitable for a comparison with the LIV scenario of Sec. \[sec:LIVdeSitter\]. We choose the following representation in terms of the canonically conjugate “conformal-time coordinates" $\Omega,\Pi,\eta,x$ defined in (\[CanPhSpConf\]): $$\begin{split}
& \!\! E \! = \!(1 \!-\! H \eta) \Omega \!+\! H x \Pi \!-\!\! \frac{\ell}{2} \!\left( \alpha \!-\! \gamma \right) \! \left( (1 \!-\! H \eta) \Omega \!+\! H x \Pi \right) \!,
\\ & p=\Pi, \\ &
N = x \Omega (1-H\eta) -\Pi \left(\eta -\frac{ H}{2} \eta^2 - \frac{H}{2} x^2\right) \\ &
+\frac{\ell}{2} \beta \left(\eta \left(2 - 3 H\eta + H^2\eta^2 \right) \Omega \Pi + x \Pi ^2\right) \\ &
\!+\! \frac{\ell}{2} \gamma x \left(H^2 x^2 \Pi^2 \!\!+\! 3 \Omega (1-H\eta) \left(\Omega (1 \!-\! H\eta) \!+\! H x \Pi \right)\right)
\end{split}
\label{repDSR-DS}$$ With this representation, the Casimir (\[casimirDSR-DS\]) takes the form $${\cal C} \!=\! (1-H\eta)^2 \!\left( \Omega^2 \!-\! \Pi^2\right) \!+ \ell \left( 1 \!-\! H\eta \right)^3 \! \left( \gamma \Omega^3 \!\!+\! \beta \Omega \Pi^2 \right)\!,$$ and it is interesting to notice that the corresponding form of the Hamiltonian constraint ${\cal H} = {\cal C} - m^2$ reproduces the one in the LIV scenario (\[dispLIVrep\]) for the special choice $\lambda \neq 0$, $\lambda' = \lambda'' = \lambda'''=0 $, which is the particular LIV case in which translational symmetry is not broken. Much insight on the differences between the LIV and the DSR scenarios will be here provided by contrasting these two cases with the same Hamiltonian constraint. We shall see that\
$\star$ even when they rely on the same Hamiltonian constraint a LIV and a DSR scenario do not in general lead to the same predictions,\
$\star$ and in particular even a case where the Hamiltonian constrain requires nothing new of translation transformations in the LIV scenario can require in the DSR scenario a deformation of translation-symmetry transformations.
This last point is already evident in (\[repDSR-DS\]) which shows how the representation of (time-)translation generators depends in the DSR case on the parameters $\alpha$ and $\gamma$, through the combination $\alpha - \gamma$. So translations are unaffected by the parameter $\beta$ but do depend on $\gamma$, the other parameter present in the Hamiltonian constraint. They also depend on $\alpha$ which is a parameter coding properties specifically of boost-symmetry transformations.
For massless particles, setting to zero the Hamiltonian constraint (\[casimirDSR-DS\]), one finds the velocity (for $\Pi>0$) $$v(\eta) = 1 - \ell \left( \gamma + \beta \right)\left( 1- H\eta \right) \Pi
\label{velocityDSRdS}$$ Differently from the LIV scenario of the previous section, in the DSR scenario the (deformed) relativistic symmetries are preserved, and all the relativistic observers describe (massless) particles moving, in their coordinates, with the same expression for the velocity, given in Eq. (\[velocityDSRdS\]). Indeed the generators of relativistic-symmetry transformations, $E,p$ and $N$, are conserved charges, and therefore their Poisson bracket with the Casimir/Hamiltonian is null, as one can easily verify.
We are interested, as in Sec. \[sec:LIVdeSitter\], in the time of arrival at a distant detector of two photons, one “hard" ($x_h,\eta_h,p_h$) and one “soft" ($x_s,\eta_s,p_s$), emitted simultaneously at a distant source. Again, take Alice and Bob to be the observers local respectively to the source and to the detector. As in (\[trajectoryLIVdeSitter\]), Bob describes photons to move along the trajectories $$\begin{split}
x^B(\eta^B) & = x^B_{O_A} \!+\!\! \int_{\eta^B_{O_A}}^{\eta^B} \!\!\!\!\! d\eta\ v^B(\eta) = x^B_{O_A} \!+\! (\eta^B \!\!-\! \eta^B_{O_A}) \\ &
~ -\! \ell (\eta^B \!\!-\! \eta^B_{O_A}) (\gamma \!+\! \beta)\Pi^B \!\! \left( 1 \!-\! H(\eta^B \!\!+\! \eta^B_{O_A}) \right)
\end{split}
\label{worDSRdS}$$ where $v^B$ now is given by Eq. (\[velocityDSRdS\]). For simplicity we focus on the case such that the soft photon reaches the detector in the origin of Bob’s frame, i.e. its worldline crosses $\eta^B_s = x^B_s =0$. In order to derive the measured time of arrival at the detector of the hard photon we need to enforce the event of emission to be local for Alice, who is at the source. This is most clearly investigated by focusing on the case such that both particles are emitted, according to Alice, at $\eta^A = x^A = 0$. As mentioned above, differently from the LIV case, since the translations (\[PhSpDSRdS\]) depend on momenta, they are affected by relative locality. Then the coordinates $x^B_{O_A} \!\!\!=\! x^B \!(x^A\!\!\!= \!\!0,\eta^A \!\!\!= \!\!0)$, $\eta^B_{O_A} \!\!\!=\! \eta^B \!(x^A\!\!\!= \!\!0,\eta^A \!\!\!= \!\!0)$ that Bob assigns to the (distant) events of emission do not coincide, as a result of the fact that the particles have different momenta (and, as stressed and shown above, translation transformations act in momentum-dependent manner). In order to evaluate $\eta^B_{O_A}, x^B_{O_A}$ we need to calculate the finite translations connecting Alice and Bob. As shown in [@DSR-DS], the relation between Bob and Alice’s coordinates can be derived by exponentiating the action by Poisson brackets of the translation generators $E,p$, which is $$\begin{gathered}
\left\lbrace E, \eta \right\rbrace = (1-H\eta) \left( 1 - \ell (\alpha -\gamma) E \right), ~~~ \left\lbrace p, \eta \right\rbrace = 0, \\
\left\lbrace E, x \right\rbrace = - H x \left( 1 - \ell (\alpha -\gamma) E \right), ~~~ \left\lbrace p, x \right\rbrace = -1.
\end{gathered}
\label{PhSpDSRdS}$$
As in Sec. \[sec:LIVdeSitter\], taking Bob to be connected to Alice by a finite spatial translation followed by a finite time translation, Eq. (\[TTR\]), one finds $$\begin{gathered}
\eta^{B} \!=\! \frac{1 \!-\! e^{H{\zeta}}}{H} \!+ e^{H{\zeta}}\eta^{A}
\!+ \ell\,(\alpha \!-\! \gamma) \, {\zeta} e^{H{\zeta}} (1 \!-\! H\eta^{A}) E^{A}_{H,\xi}, \nonumber\\
x^{B} = e^{H{\zeta}} (x^{A} - \xi) - \ell (\alpha \!-\! \gamma) {\zeta} e^{H{\zeta}} H(x^{A}-\xi) E^{A}_{H,\xi},
\nonumber\\
\Omega^{B} = e^{-H{\zeta}} \left(\Omega^{A} + \ell (\alpha \!-\! \gamma) H {\zeta} \Omega^{A} E^{A}_{H,\xi}\right), \nonumber\\
\Pi^{B} = e^{-H{\zeta}}\left(\Pi^{A} + \ell(\alpha \!-\! \gamma) H{\zeta} \Pi^{A}E^{A}_{H,\xi} \right),
\label{TransCoord-DSR}\end{gathered}$$ where we found convenient to introduce $$E^{A}_{H,\xi} \equiv \Omega^A\left( 1-H\eta^A \right)+H\Pi^A\left( x^A -\xi\right)$$
Notice that the condition for Bob’s origin to coincide with the event of detection of the soft photon is enforced by setting $\zeta = T$ and $\xi \!=\! H^{-1}(1 \!-\! e^{-HT})$, where $T$ is the (comoving) time distance between Alice and Bob. From these relations one gets (also using the fact that, for a massless particle on shell, one has $\Omega = \Pi + O(\ell \Pi)$) $$\begin{gathered}
\eta^{B}_{O_A} \!=\! - \frac{e^{H{T}} \!-\! 1}{H} \!+ \ell\,(\alpha \!-\! \gamma) \, {T} e^{H{T}} \Pi^B \\
x^{B}_{O_A} = - \frac{e^{HT} \!-\! 1}{H} + \ell (\alpha \!-\! \gamma) {T} \left( e^{HT} \!-\! 1\right) \Pi^B.
\end{gathered}$$ Substituting $\eta^B_{O_A}, x^B_{O_A}$ in (\[worDSRdS\]) and solving for $\eta_h^B(x^B_h \!\!=\! 0)$ we get the time of arrival at the detector of the hard photon: $$\!\!\!\Delta \eta^B \!\!=\! \eta_h^B(x^B_h \!\!=\! 0) \!=\! \ell p_h^{B} \!\! \left( \!\!(\alpha \!-\! \gamma) T \!\!+\! (\beta\!+\!\gamma)\frac{e^{2H T}\!\!\!-\!1}{2H} \!\right).
\label{delayBOBold}$$ At this stage it is convenient to remove the suffixes indicating the observer and express the detected delay in terms of the redshift $z$ characterizing the source, giving us the result in the form $$\Delta t=\ell p_h \left(\left( \alpha - \gamma \right)\frac{\ln\left(1+z\right)}{H}+(\beta+\gamma)\frac{z+\frac{z^{2}}{2}}{H}\right),
\label{delayBOBz}$$ where we denoted by $p_h$ the momentum of the hard particle observed at the detector, and we expressed the delay in terms of comoving time (keeping again in mind the observation reported in (\[DtDeta\])).
FRW spacetime combining slices of de Sitter spacetime {#sec.Slicing}
=====================================================
So far we focused on spacetimes with constant curvature (de Sitter), providing for the LIV case a generalization of the results of Ref. [@jacobpiran] and providing for the DSR case a genealization of the results of Ref. [@DSR-DS]. For our purposes it is crucial to further generalize these result to the case of non-constant curvature, specifically the case of FRW-type expansion, since our analysis is aimed at applications of modified dispersion relations in astrophysics and cosmology. For the constant-curvature case our strategy of analysis proved to be very powerful thanks to its reliance on techniques that exploit the high amount of symmetries present in the de Sitter case. This is of course particularly evident for DSR-relativistic scenarios, but also plays a role for the LIV scenarios (where the relativistic symmetries are broken, but at zero-th order of analysis one has them all). The fact that for FRW expansion one looses quite a bit of those symmetries poses a challenge from our perspective. We shall argue however that the connection between de Sitter spacetimes and FRW spacetimes is nonetheless strong enough to provide a clear path for the generalization we are seeking. Specifically we shall use the fact that FRW spacetimes can be described in terms of a suitable sequence of “thick slices" of de Sitter spacetimes.
In the following sections we shall use this notion of “de-Sitter slicing of FRW spacetimes" for the purpose of deriving predictions for the implications of modified dispersion relations in presence of FRW-type spacetime expasion at non-constant rate. In preparation for that we first show in this section how the “thick slicing" can be used in the standard case of the propagation of a classical particle in a classical FRW spacetime (no modifications of the dispersion relation). In doing so we shall build on the strength of related results on “de Sitter slicing" already reported in Ref. [@interplay]. We refine the proposal put forward in [@interplay] by strengthening the role of symmetry generators in characterizing each de-Sitter slice and by making reference to observers associated to each slice. This refinements prove valuable in order to deal with the effects of relative locality, which require physical observations to be described by observers local to the spacetime events [@principle].
We find that it is convenient to perform the analysis first in terms of the comoving time $t$. The FRW spacetime is described by the metric $$ds^2 = dt^2 - a^2(t) dx^2,
\label{metricFRW}$$ where $a(t)$ is the scale factor of the universe, defined by the relation $H(t) \!=\! \dot{a}(t)/a(t)$, $H(t)$ being the (time-dependent) expansion rate, from which we also get $$\frac{a\left(t_{f}\right)}{a\left(t_{i}\right)}=\exp\left( \int_{t_{i}}^{t_{f}}dt\ H\left(t\right) \right)
\label{scaleH}$$ We reconstruct the trajectory of a FRW massless particle from its source to a distant detector considering a first observer, Alice, local to the emission, and a final observer, Bob, local to the detection. If the particle moves in a FRW spacetime, it is straightforward from (\[metricFRW\]) to derive Bob’s description of its trajectory as $$x^B\left( t^B \right) = x^B_{O_A} + \int_{t^B_{O_A}}^{t^B} \frac{dt}{a\left( t \right)}
\label{FRWworld}$$ where we are assuming $a(t^B=0) = 1$.
Since in this section we just consider a classical particle in a classical FRW spacetime, the “thick slices" here of interest will have the geometry of a classical de Sitter spacetime, which in particular affords us also the luxury of analyzing each slice using classical translational invariance (this will change in the later section for the cases where the deformation affects translational invariance).
Our “thick slices" of FRW spacetime are introduced by dividing the time interval between the event of emission and the event of detection in $N$ time intervals of temporal size[^7] $\Delta t_n$, $n \!=\! 1,..,N$. In each slice spacetime is described,to good approximation, by a constant expansion rate $H_n \!=\! H\left( t_n \right)$, where $t_n$ is the initial time of the $n$-th slice. For simplicity we divide the time interval in slices of equal size $\Delta t_n \!=\! T/N$, where $T$ is the time of flight of the photon from the source to the detector. Then of course we have that $t_n \!=\! t_i + n T/N $. Starting from the observer Alice, who is at the source, we contemplate a set of intermediate observers Bob$_n$, $n=1,..,N$, such that each $n$-th observer crosses the photon’s trajectory at the time $t_n$. Each observer Bob$_n$, in the corresponding $n$-th slice, which goes from $t_{n-1}$ to $t_{n}$, will describe the motion of particles in terms of a constant expansion rate $H_n$.
We are of course ultimately interested in taking the limit $N \rightarrow \infty$, so that we have an infinity of infinitesimally small slices and the assumption of constant expansion rate in each slice becomes fully justified.
We can also describe each Bob$_n$ as the observer connected to Alice by a set of $n$ spatial translations followed by a set of $n$ time translations, with each $k$-th translation characterized by the relative constant expansion rate $H_k$ and finite translation parameters $\zeta_k,\xi_k$, i.e. $$\left(t,x\right)^{B_{n}}\!\! = e^{-\sum_{k=1}^n \xi_k p}
\triangleright e^{-\sum_{k=1}^n \zeta_kE_{H_k}}\triangleright\left(t,x\right)^A \!\!.
\label{FiniteTransSlicing}$$ These give the relation between Bob$_n$ and Alice’s coordinates $$\begin{gathered}
t^{B_{n}}\left(t^{A},x^{A}\right)=t^{A}-\sum_{k=1}^n\zeta_k,\\
x^{B_{n}}\left(t^{A},x^{A}\right)=e^{\sum_{k=1}^{n}H_{k}\zeta_k}\left(x^{A}-\sum_{k=1}^{n}\xi_k\right).
\label{BobnAl}
\end{gathered}$$ The requirement for each observer Bob$_n$ to be along the photons trajectories at the time $t_n$, is then ensured by imposing that the translation parameters satisfy the conditions $$\zeta_n = \zeta = T/N , \qquad
\xi_n = e^{-\sum_{k=1}^{n}H_{k} \zeta_n} \frac{e^{H_{n} \zeta_n} - 1}{H_{n}},
\label{param}$$ and Alice describes the slices to be of sizes $\Delta t_n^A \!=\! \zeta$, $\Delta x_n^A \!=\! \xi_n$.
We are interested in the worldline described by the observer Bob$_N$, which is our Bob, local to the detection. From the relations (\[BobnAl\]) we can derive the relation between the coordinates of Bob$_n$ and the coordinates of Bob$_N$ $$\begin{gathered}
t^{B_{N}} = t^{B_{n}}-\left(N-n\right)\zeta \\
x^{B_{N}} \!\! =\! e^{\sum_{k=n+1}^{N} \! H_{k}\zeta}x^{B_{n}} \!-\! \!\!\! \sum_{k=n+1}^{N}e^{\sum_{s=k}^{N} \! H_{s}\zeta}\frac{1 \!-\! e^{-H_{k}\zeta}}{H_{k}}.
\end{gathered}
\label{BobnBobN}$$ Since, as seen using these relations, $$v^{B_{N}} \!=\! \frac{dx^{B_{N}}}{dt^{B_{N}}} \!=\! e^{\sum_{k=n+1}^{N} \!\! H_{k}\zeta}\frac{dx^{B_{n}}}{dt^{B_{n}}} \!=\! e^{\sum_{k=n+1}^{N} \!\! H_{k}\zeta}v^{B_{n}} \!,$$ Bob$_N$, in each $n$-th slice, describes photons to move with velocity $$\begin{split}
v_{n}^{B_{N}} \!\! \left(t^{B_{N}} \! \right) & \!\!=\! e^{\sum_{k=n \!+\! 1}^{N} \!\! H_{k}\zeta} v^{B_{n}}_n \!\! \left(t^{B_{n}} \!\! \left(t^{B_{N}} \!\right) \! \right) \\
& \!\!=\! e^{\sum_{k=n \!+\! 1}^{N} \!\! H_{k}\zeta} e^{ -H_{n} \! \left( \! t^{B_{\!N}} \! \!-t_{O_n}^{B_{\!N}} \!\right)},
\end{split}$$ where, here and in the following, we denote by $t^{B_N}_{O_n}$ and $x^{B_N}_{O_n}$, the value of the coordinates that the observer Bob$_N$ attributes to the position of the particle at crossing of the spatial origin of observer Bob$_n$’s frame.
Then, the photon trajectories in the $n$-th slice, are described by Bob$_N$ to be $$x^{B_{N}} \! \left(t^{B_{N}} \!\right)_{n} \!\!=\! x_{O_A}^{B_{N}} \!+\! \sum_{k=1}^{n-1} \int_{t_{O_{k-1}}^{B_{N}}}^{t_{O_{k}}^{B_{N}}} \!\!\!\! dt\ \!v_{k}^{B_{N}} \!+\!\! \int_{t_{O_{n-1}}^{B_{N}}}^{t^{B_{N}}} \!\!\!\!\!\!\! dt\ \! v_{n}^{B_{N}} .
\label{trajectoryFRW-N}$$ Since the observers Bob$_n$ are defined so that the photons cross the origin of their reference frame, the second term on the right-hand side can be obtained using relations (\[BobnBobN\]) with $t^{B_N}_{O_n} \!\!\!=\! t^{B_N}(t^{B_n} \!\!=\! 0) $, $x^{B_N}_{O_n} \!\!\!=\! x^{B_N}(x^{B_n} \!\!=\! 0) $, is $$\sum_{k=1}^{n-1} \! e^{\sum_{s =k + 1}^{N} \!\! H_{s}\zeta} \! \frac{e^{H_{k}\zeta} \!\!-\! 1}{H_{k}} .
\label{term-1}$$ Considering that $\zeta \!=\! T/N \!=\! \Delta T_n$, in the limit $N \! \rightarrow\infty$, $$\sum_{k=n_i+1}^{n_f} \zeta \rightarrow \int_{t_{n_i}}^{t_{n_f}} dt.
\label{Riemann}$$ It follows from (\[scaleH\]) that $$e^{\sum_{s=k+1}^{n}H_{s}} \rightarrow \frac{a\left( t_n \right)}{a\left( t_k \right)},
\label{Riemann-2}$$ Then the term (\[term-1\]) tends to $$\sum_{k=1}^{n-1}\frac{\zeta}{a\left(t_{k}\right)} \rightarrow \int_{t_{O_A}^{B_{N}}}^{t_{O_{n-1}}^{B_{N}}} \!\!\!\! \frac{dt}{a\left(t\right)}
\label{term-1-2}$$ where we considered that for $N \!\rightarrow \! \infty $ one has that $\frac{e^{H_{k}\zeta} - 1}{H_{k}} \!\rightarrow \zeta$ and that $a\left( t_N \right) \!=\! 1$ since $t_N \!=\! t_{O_N}$. For what concerns the last term in (\[trajectoryFRW-N\]), consider that in the slice $t_{O_{n-1}}^{B_{n}} \!\! \leq t \leq t_{O_n}^{B_{n}}$, where $H\left(t\right)=H_{n}=\text{const.}$, $$e^{-H_{n}\left(t-t_{O_n}^{B_{N}}\right)}\simeq\frac{a\left(t_{n}\right)}{a\left(t\right)},$$ so that, using also (\[Riemann-2\]), the term becomes $$e^{\sum_{s=n + 1}^{N} \!\! H_{s}\zeta} \!\! \int_{t_{O_{n-1}}^{B_{N}}}^{t^{B_{N}}} \!\!\!\!\!\!\!\!\! dt\ \!e^{- \! H_{n} \!\left(\! t - t_{O_n}^{B_{N}} \!\right)} \rightarrow \int_{t_{O_{n-1}}^{B_{N}}}^{t^{B_{N}}} \!\!\! \frac{dt}{a\left(t\right)}
\label{term-2}$$ and, substituting (\[term-1-2\]) and (\[term-2\]) in (\[trajectoryFRW-N\]), we obtain that as $N \! \rightarrow \infty$ the trajectory of Bob$_N$ tends to the trajectory of Bob, given in (\[FRWworld\]).
The fact that in the $N \! \rightarrow \infty$ limit our thick-slicing procedure would match exactly the results obtained in a FRW spacetime was assured by construction. It was useful for what follows to see, as done above, the technical details of how this convergence takes shape. On the quantitative side it is rather impressive how quickly this convergence takes place: as shown in Fig.1 even for relatively small values of N our slicing procedure gives results that are already in pretty good agreement with the results of the analysis done in the corresponding FRW spacetime.
![Here we show the worldline of a massless particle emitted at a distant source at the time $-T$, from the point of view of an observer at the detector. We choose for our pictorial example the scale factor to obey the power law behavior $a(t^B) = ((t^B+t_0)/t_0)^{2/3}$, were $t_0$ is a constant indicating the “Big-Bang” time. We show respectively the trajectory (black curve) described by a FRW observer, and the trajectories described by observers (Bob$_N$) obtained through our thick-slicing procedure, Eq. (\[trajectoryFRW-N\]), with different levels of approximations (curves magenta, purple and blue). The dots indicate the spacetime position of each Bob$_n$ observer, as described by Bob$_N$. We see that as the slicing approximation gets finer (as $N$ increases) the particle’s worldline, as well as the source position, converge to the ones of FRW spacetime.](SlicingFRW.pdf)
Further insight can be gained by working with conformal-time coordinates. Conformal time is defined by the relation $dt = a(\eta)d\eta$, where $a(\eta) = a(t(\eta))$. In conformal-time coordinates, the FRW spacetime metric becomes $$ds^2 = a^2(\eta)\left(d\eta^2 - dx^2 \right).$$ Notice also that the expansion rate can then be expressed as follows $$H(t) = \frac{1}{a(t)}\frac{da(t)}{dt} = \frac{1}{a^2(\eta)}\frac{da(\eta)}{d\eta}
\label{expRate}$$ Bob’s description of the worldline (Eq. (\[FRWworld\])) is $$x^B(\eta^B) = x^B_{O_A} - \eta^B_{O_A} + \eta^B = \eta^B,
\label{FRWworld-CON}$$ where the requirement for Bob’s origin to coincide with the detection of the photon enforces $$\eta^B_{O_A} = x^B_{O_A} = - \int_{-T}^{0} \frac{dt}{a(t)}.
\label{emissFRW}$$
The slices are of size $\Delta\eta_n \!=\! H_n^{-1} \!\left( 1 \!-\! e^{-H_n T/N}\right)$and the relations between Alice and Bob$_n$ coordinates, generated by the action (\[FiniteTransSlicing\]), are $$\begin{gathered}
\eta^{B_{n}} \! =\! e^{\sum_{k=1}^n \!\! H_k \zeta_k} \eta^A \!\!-\! \sum_{k=1}^n e^{\sum_{s=k+1}^n \!\! H_s \zeta_s} \frac{ e^{H_k \zeta_k} \!-\! 1}{H_k}, \\
x^{B_{n}} \!=e^{\sum_{k=1}^n H_k \zeta_k} \left( x^A - \sum_{k=1}^N \xi_k \right).
\label{BobnAl-CON}
\end{gathered}$$ Imposing the conditions (\[param\]), one finds from (\[BobnAl-CON\]) the relations between Bob$_n$ and Bob$_N$ coordinates:$$\begin{gathered}
\eta^{B_{N}} \!\!=\! e^{\sum_{k=n+1}^{N} \!\! H_{k} \zeta} \eta^{B_{n}} \!\!-\!\!\!\! \sum_{k=n+1}^{N} \!\!\! e^{\sum_{s=k+1}^{N} \!\! H_{s}a_{t}} \frac{e^{H_{k}\zeta} \!\!-\! 1}{H_{k}},\\
x^{B_{N}} \!\!=\! e^{\sum_{k=n+1}^{N} \!\! H_{k} \zeta} x^{B_{n}} \!\!-\!\!\!\! \sum_{k=n+1}^{N} \!\!\! e^{\sum_{s=k+1}^{N} \!\! H_{s}a_{t}} \frac{e^{H_{k}\zeta} \!\!-\! 1}{H_{k}} \\
\Pi_{B_{\!N}} \!\!\!=\! e^{-\!\sum_{k=n+1}^{N} \!\! H_{k}\zeta} \Pi_{\!B_{n}}, ~~~ \Omega_{B_{\!N}} \!\!\!=\! e^{-\!\sum_{k=n+1}^{N} \!\! H_{k}\zeta} \Omega_{B_{n}}
\label{BobnBobN-CON}
\end{gathered}$$ Each Bob$_n$ describes, in each $n$-th slice, the photons to move with velocity $v_{n}^{B_{n}} \! \left(\eta^{B_{n}}\!\right) \!=\! 1$, since, using Eq. (\[BobnBobN-CON\]), $$v^{B_{N}}=\frac{dx^{B_{N}}}{d\eta^{B_{N}}}=\frac{dx^{B_{n}}}{d\eta^{B_{n}}}=v^{B_{n}}=1,
\label{velBNBn}$$ Then, it is straightforward to see that Bob$_N$’s description of the photons trajectories in the $n$-th slice coincides with Eq. (\[FRWworld-CON\]): $$\begin{split}
& x^{B_{N}}\!\! \left(\eta^{B_{N}} \!\right)_{\!N} \!\!=\! x_{O_A}^{B_{N}} \!+\!\! \sum_{k=1}^{n-1} \! \int_{\eta_{O_{k-1}}^{B_{N}}}^{\eta_{O_{k}}^{B_{N}}} \!\!\!\!\! d\eta\ \! v_{k}^{B_{N}} \!\! +\!\! \int_{\eta_{O_{n-1}}^{B_{N}}}^{\eta^{B_{N}}} \!\!\!\!\!\!\!\!\!\! d\eta\ \! v_{n}^{B_{N}}\\
& = x_{O_A}^{B_{N}} \!\!+\! \sum_{k=1}^{n-1}\left(\eta_{O_{k}}^{B_{N}} \!\!-\! \eta_{O_{k-1}}^{B_{N}}\right) \!+\!\eta^{B_{N}} \!\!-\! \eta_{O_{n-1}}^{B_{N}}\\
& ~~ = x_{O_A}^{B_{N}}-\eta_{O_A}^{B_{N}}+\eta^{B_{N}}=\eta^{B_{N}}.
\end{split}$$ For a standard (undeformed) FRW picture all the non-triviality of our thick slicing is in Bob$_N$’s description of the point of emission $\left( \eta^{B_N}_{O_A},x^{B_N}_{O_A} \right)$, which in a FRW spacetime is Eq. (\[emissFRW\]). From (\[BobnAl-CON\]), considering also (\[param\]), one finds that$$\begin{split}
& \eta^{B_N}_{O_A} = x^{B_N}_{O_A} = \!-\! \sum_{k=1}^N e^{\sum_{s=k+1}^N \!\! H_s \zeta_s} \frac{ e^{H_k \zeta_k} \!-\! 1}{H_k} \\ &
\rightarrow \!-\! \sum_{k=1}^N \frac{\zeta}{a(t_k)} \rightarrow \!-\! \int_{-T}^{0} \frac{dt}{{a(t)}}.
\end{split}$$
As a last remark we notice that, as we discussed at the beginning of this section, the translational invariance of de Sitter spacetime allowed us to define our thick slices in such a way that each observer Bob$_n$ describes the particle’s motion, in the corresponding $n$-th slice (defined in the interval $\eta_{O_{n-1}} \!\! \leq \eta \leq \! \eta_{O_n} $), through the same de Sitter Hamiltonian, the only difference being the value of the expansion rate $H_n$. Explicitly, in the $n$-th slice, $${\cal H}_{n}^{B_{n}}=\left(1-H_{n}\eta^{B_{n}}\right)^{2}\left(\Omega_{B_{n}}^{2}-\Pi_{B_{n}}^{2}\right)-m^{2}.$$ Using Eqs. (\[BobnBobN-CON\]) we get the Hamiltonian that Bob$_N$ attributes to each $n$-th slice: $$\begin{split}
& {\cal H}_{n}^{B_{N}} \!\! + m^{2} \!=\! \left(\Omega_{B_{N}}^{2}-\Pi_{B_{N}}^{2}\right) \times \\ &
\times \! \left(\!\! e^{\sum_{k=n+1}^{N} \!\! H_{k}a_{t}} \!\!-\! H_{n} \!\!\!\!\! \sum_{k=n+1}^{N} \!\!\!\! e^{\sum_{s=k+1}^{N} \!\! H_{s}a_{t}}\frac{e^{H_{k}\zeta} \!\!-\! 1}{H_{k}} \!-\! H_{n}\eta^{B_{N}} \!\!\!\right)^{\!\!\!2}
\end{split}$$ One can see that the slices are patched together in such a way that the Hamiltonian is continuous in the point of juncture of two consecutive slices. Indeed $${\cal H}_{n}^{B_{N}}(\eta^{B_N}_{O_n}) = {\cal H}_{n+1}^{B_{N}}(\eta^{B_N}_{O_n}).$$
LIV WITH FRW EXPANSION {#LIVFRW}
======================
The thick-slicing setup introduced in the previous section is ideally suited for our objective of studying LIV and DSR-relativistic effects in spacetimes with FRW expansion. We start in this section with the LIV case. Through our slicing we will be able to rely on the results for LIV in de Sitter spacetime reported in sec. \[sec:LIVdeSitter\], thereby obtaining a LIV-FRW scenario.
In the previous Sec. \[sec.Slicing\] we described FRW expansion in terms of thick slices of de Sitter expansion by also relying on the the translational invariance available in each de Sitter slice. For the LIV-FRW case we shall of course involve LIV-de Sitter slices, for which, as shown at the end of sec. \[sec:LIVdeSitter\], the breakdown of translational invariance is such that distant observers, at relative rest, describe the motion of particles through different laws, i.e. different functional expressions for the Hamiltonian and velocities (see Eq. (\[velocityLIVdeSitter-2\])).
Just as in Sec. \[sec:LIVdeSitter\], we want to find the time of arrival of a soft and a hard photon emitted simultaneously at a distant source. As in the previous section, we take Alice to be local to the event of emission, and Bob$_N$ to be local to the detector. We assume that the soft photon has been emitted at the (comoving) Bob$_N$ time $-T$, and that Bob$_N$’s origin coincides with the event of detection of the soft photon. Moreover we take the intermediate observers Bob$_n$ such that the soft photon crosses the origin of their reference frame. Since in the LIV scenario the laws of transformation among observers are not deformed, the relation between Bob$_n$ and Bob$_N$ coordinates is still given by Eqs. (\[BobnBobN-CON\]). We start by assuming that Bob$_N$ describes, in the $N$-th slice, the velocity to be the one given by the Hamiltonian (\[dispLIVrep\]), Eq. (\[velocityLIVdeSitter\]). In order to reconstruct the velocity in the other slices, we can focus on the scale factor in conformal-time coordinates $$a_{n}\left(\eta\right)=\frac{1}{1-H_{n}\eta}
\label{an}$$ so that (as in Sec. \[sec:LIVdeSitter\] we assume $\tilde{\alpha}\!+\!\tilde{\beta}\!=\!1$) $$v_{N}^{B_{\!N}} \!\! \left(\eta^{B_{\!N}} \!\right) \!=\! 1 - \Pi^{B_{\!N}} \!\! \left( \!\!\lambda'a_{\!N} \!\left(\eta^{B_{\!N}} \!\right) \!+\! \lambda'' \!\!\!+\! \frac{\lambda}{a_{\!N} \! \left(\eta^{B_{\!N}} \!\right)}+\!\frac{\lambda'''}{a_{\!N}^{2} \!\left(\eta^{B_{\!N}} \!\right)} \! \right)$$ Let us consider first the observer Bob$_{N-1}$. Using relations (\[BobnBobN-CON\]) to get the relations between Bob$_{N}$ and Bob$_{N-1}$ coordinates, $$\begin{split}
& \eta^{B_{N}}\left(\eta^{B_{N-1}}\right)
=e^{H_{N}\zeta}\left(\eta^{B_{N\!-\!1}}+\eta_{O_{N-1}}^{B_{N}}\right) \\ &
~~~~ = e^{H_{N}\zeta}\left(\eta^{B_{N\!-\!1}}-\frac{1-e^{-H_{N}\zeta}}{H_{N}}\right),
\end{split}$$ we derive Bob$_{N-1}$’s description of the scale factor $a_{N}$ in the $N$-th slice as $$a_{N}\left(\eta^{B_{N}}\left(\eta^{B_{N\!-\!1}}\right)\right)
= \frac{e^{-H_{N}\zeta}}{1-H_{N}\eta^{B_{N\!-\!1}}}=e^{-H_{N}\zeta}a_{N}\left(\eta^{B_{N\!-\!1}}\right)$$ In its origin, since $a_{N}\left(0\right)=1$, Bob$_{N-1}$ describes the scale factor to be $e^{-H_{N}\zeta}$. We define the $N\!-\!1$-th slice to be such that the scale factor remains constant at the value it has in Bob$_{N-1}$’s origin, the point of contact between the two slices. This condition can be imposed assuming that in the $N-1$-th slice, governed by the constant expansion rate $H_{N-1}$, Bob$_{N-1}$ describes the scale factor to be $$e^{-H_{N}\zeta}a_{N-1}\left(\eta^{B_{N-1}}\right).$$ With this condition, Bob$_N$ describes the scale factor to be, in the $N \!-\!1$-th slice, $$e^{-H_{\!N}\zeta}a_{\!N\!-\!1} \!\left(\eta^{B_{\!N\!-\!1}} \!\!\left(\eta^{B_{\!N}}\!\right) \! \right)
\!=\! e^{-H_{\!N}\zeta}a_{\!N\!-\!1} \!\left(\!e^{\!-H_{\!N}\zeta} \!\!\left( \!\eta^{B_{\!N}} \!\!-\! \eta_{O_{\!N\!-\!1}}^{B_{\!N}} \!\right) \!\!\right)$$ Iterating this construction one finds that Bob$_N$ describes the scale factor in the $n$-th slice to be $$\begin{split}
& a_{n}^{B_{\!N}} \!\!\left(\eta^{B_{\!N}} \!\right) \!\!=\! e^{-\!\!\sum_{k=n+1}^{N} \!\!H_{k}\zeta} a_{n} \! \left(\! e^{-\!\!\sum_{k=n+1}^{N} \!\! H_{k}\zeta} \!\left( \! \eta^{B_{\!N}} \!\!\! -\! \eta_{O_n}^{B_{\!N}} \!\right) \!\!\right)\\ &
~~~~~~~ = \frac{e^{-\sum_{k=n+1}^{N}H_{k}\zeta}}{1-H_{n}e^{-\sum_{k=n+1}^{N}H_{k}\zeta}\left(\eta^{B_{N}}-\eta_{O_n}^{B_{N}}\right)}.
\end{split}
\label{eq:anBN}$$
Consistently with this setup we have that the velocity assigned by Bob$_N$ to the photons in each $n$-th slice is $$v_{n}^{\!B_{\!N}} \!\!\!=\!\! 1 \!-\! \Pi^{B_{N}} \!\!\!\left(\!\! \lambda' \!a_{n}^{\!B_{\!N}} \!\!\left( \!\eta^{B_{\!N}} \!\right) \!+\!\! \lambda'' \!\!\!+\! \frac{\lambda}{\!a_{n}^{\!B_{\!N}} \!\!\left(\eta^{B_{\!N}}\! \right)\!} \!+\! \frac{\lambda'''}{\!\!\left( \! a_{n}^{\!B_{\!N}} \!\! \left(\eta^{B_{\!N}} \!\right)\!\! \right)^{\!\!2} \!} \!\!\right)
\label{velBN-LIV-N}$$ Notice that in this way the velocity (and then the Hamiltonian) is continuous in the point of junction of two contiguous slices. The fact that, as discussed in Sec. [\[sec:LIVdeSitter\]]{}, the term proportional to $\lambda$ in (\[dispLIVrep\]) plays a special role in relation to translational invariance in the de Sitter case, still has a role in the FRW case, even though time-translations are not a symmetry of the FRW case. We shall see this by first observing that, as a result of (\[velBNBn\]),(\[BobnBobN-CON\]) and (\[an\]), each observer Bob$_n$ describes the velocity, in the $n$-th slice, to be $$\begin{split}
& v_{n}^{B_{n}}\left(\eta^{B_{n}},\Pi^{B_{n}}\right)= v_{n}^{B_{N}}\left(\eta^{B_{N}}\left(\eta^{B_{n}}\right),\Pi^{B_{N}}\left(\Pi^{B_{n}}\right)\right)\\
& = 1-\Pi^{B_{n}}\bigg(\lambda'e^{-2\sum_{k=n+1}^{N}H_{k}\zeta}a_{n}\left(\eta^{B_{n}}\right)+\lambda''e^{-\sum_{k=n+1}^{N}H_{k}\zeta} \\&
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +\frac{\lambda}{a_{n}\left(\eta^{B_{n}}\right)}+\lambda'''\frac{e^{\sum_{k=n+1}^{N}H_{k}\zeta}}{a_{n}^{2}\left(\eta^{B_{n}}\right)}\bigg)
\end{split}$$ One can see that for the term proportional to $\lambda$ the velocity takes the same form in all slices (though of course with a different expansion rate $H_n$).
In order to evaluate the arrival time at the detector, we consider that the trajectory that Bob$_N$ assigns to photons in the $N$-th slice is $$x^{B_{N}} \! \left(\eta^{B_{N}} \!\right)_{N} \!\! =\! x_{O_A}^{B_{N}} \!+\!\!\! \sum_{n=1}^{N-1} \!\! \int_{\eta_{O_{n-1}}^{B_{N}}}^{\eta_{O_n}^{B_{N}}} \!\!\!\!\!\! v_{n}^{B_{N}}d\eta^{B_{N}} \!+\!\! \int_{\eta_{O_{N-1}}^{B_{N}}}^{\eta^{B_{N}}} \!\!\!\!\!\!\!\!\!\!\!\! v_{N}^{B_{N}}d\eta^{B_{N}}
\label{trajectoryBobN-LIV}$$ The expression (\[velBN-LIV-N\]) can be easily integrated in each slice.
The intermediate observers Bob$_n$ are along the worldline of the soft photon, but the hard photon crosses the spatial origin of these intermediate observers at $\eta^{B_n} \neq 0$, $$\begin{gathered}
\eta^{B_N}_{O_n}=\eta^{B_N}(\eta^{B_n}=0,x^{B_n}=0)+O(\lambda),\\
x^{B_N}_{O_n}=x^{B_N}(\eta^{B_n}=0,x^{B_n}=0)+O(\lambda).
\end{gathered}
\label{XBN-orBn}$$ However in Eq. (\[trajectoryBobN-LIV\]) the coordinates $\eta^{B_N}_{O_n},x^{B_N}_{O_n}$ appear only multiplied by factors of $\lambda$ and therefore we can neglect higher orders and obtain from (\[BobnBobN-CON\]) the relations $$\begin{gathered}
\eta_{O_n}^{B_{N}} = -\sum_{k=n+1}^{N}e^{\sum_{s=k}^{N}H_{s}\zeta}\frac{1-e^{-H_{k}\zeta}}{H_{k}}, \label{O-1}\\
\eta_{O_n}^{B_{N}}-\eta_{O_{n-1}}^{B_{N}} = e^{\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-H_{n}\zeta}}{H_{n}}, \label{O-2}\\
\eta_{O_{n-1}}^{B_{N}}+\eta_{O_n}^{B_{N}}=\eta_{O_{n-1}}^{B_{N}}-\eta_{O_n}^{B_{N}}+2\eta_{O_n}^{B_{N}}, \label{O-3}\end{gathered}$$ Using also that $x_{O_A}^{B_{N}}=\eta_{O_A}^{B_{N}}$ one gets, after some algebra, the trajectory $$\begin{split}
& x^{B_{N}}\left(\eta^{B_{N}}\right)_{N} = \eta^{B_{N}}-\Pi^{B_{N}}\sum_{n=1}^{N}\Bigg[\lambda'\zeta+\lambda''e^{\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-H_{n}\zeta}}{H_{n}}\\
& +\lambda e^{2\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-2H_{n}\zeta}}{2H_{n}}+\lambda'''e^{3\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-3H_{n}\zeta}}{3H_{n}}\Bigg]\\
& -\Pi^{B_{N}}\Bigg[\lambda'\frac{1}{H_{N}}\ln\left[1+H_{n}e^{-\sum_{k=n+1}^{N}H_{k}\zeta}\left(\eta_{O_n}^{B_{N}}-\eta^{B_{N}}\right)\right] \\ & +\left(\lambda''+\lambda+\lambda'''\right)\eta^{B_{N}}\\
& -\left(\frac{\lambda}{2}+\lambda'''\right)H_{N}\left(\eta^{B_{N}}\right)^{2}+\frac{\lambda'''}{3}H_{N}^{2}\left(\eta^{B_{N}}\right)^{3}\Bigg]
\end{split}
\label{trajectoryBobN-2}$$ The time of arrival of the hard photon is obtained by solving $\eta^{B_{N}}\left(x^{B_{N}}=0\right)$. Since the delay is $O\left(\lambda\right)$, we can disregard all the terms $O\left(\left(\eta^{B_{N}}\right)^{2}\right)$ in the right-hand side of (\[trajectoryBobN-2\]), and we get the expression for the delay $$\begin{split}
& \Delta\eta^{B_{N}} = \Pi^{B_{N}}\sum_{n=1}^{N}\Bigg[\lambda'\zeta+\lambda''e^{\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-H_{n}\zeta}}{H_{n}}\\
& +\lambda e^{2\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-2H_{n}\zeta}}{2H_{n}}+\lambda'''e^{3\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-3H_{n}\zeta}}{3H_{n}}\Bigg]
\end{split}$$ In the limit in which $N\rightarrow \infty$, using Eqs. (\[Riemann\]) and (\[Riemann-2\]), the delay takes the form (omitting the suffix $B_{N}$) $$\Delta t = \Delta \eta = p_h \int_{-T}^{0}dt\left[\lambda'+\frac{\lambda''}{a\left(t\right)}+\frac{\lambda}{a^{2}\left(t\right)}+\frac{\lambda'''}{a^{3}\left(t\right)}\right],$$ where again we expressed the delay in terms of comoving time and we denoted by $p_h$ the momentum of the hard particle observed at the detector. We can finally reexpress the delay in terms of the redshift of the source $z\equiv z(-T)$ noting that for $\bar{z} \equiv z(t) $ $$a\left(t\right)=\frac{1}{1+\bar{z}}, \qquad
dt=-\frac{d\bar{z}}{H\left(\bar{z}\right)\left(1+\bar{z}\right)},
\label{redshift}$$ so that the delay becomes $$\Delta t=p_h\int_{0}^{z}\frac{d\bar{z}}{H\left(\bar{z}\right)}\left[\frac{\lambda'}{\left(1+\bar{z}\right)}+\lambda''+\lambda\left(1+\bar{z}\right)+\lambda'''\left(1+\bar{z}\right)^{2}\right]
\label{gacextratoo}$$ For the choice of parameters $\lambda'=\lambda''=\lambda'''=0$, $\lambda\neq 0$, the delay coincides with the one reported in [@jacobpiran], as it should have been expected on the basis of the observations we offered above. Indeed one can show that the trajectory (\[trajectoryBobN-2\]), in the limit $N\rightarrow\infty$, is consistent with the Hamiltonian $${\cal H} \!+\! m^2 \!\!= \!\! \frac{\Omega^2 \!\!-\! \Pi^2 \!\!}{a^2(\eta)} + \frac{\tilde{\alpha} \Omega^3 \!\!+\! \tilde{\beta} \Omega \Pi^2 \!}{a^2(\eta)} \!\! \left( \!\! \lambda' \! a(\eta) \!+\! \lambda'' \!\!\!+\! \frac{\lambda}{\! a(\eta)} \!+\! \frac{\lambda'''\!}{ \!a^2(\eta)} \!\! \right)
\label{dispLIV-FRW}$$ which coincides, for the same choice of parameters $\lambda'=\lambda''=\lambda'''=0$, $\lambda\neq 0$, with the one presented in [@jacobpiran] and [@PiranMartinez].
DSR and Relative locality WITH FRW EXPANSION
============================================
Our final task is to perform a DSR-relativistic analysis with FRW expansion. We use again as illustrative application an analysis of the times of arrival of a soft and a hard photon emitted simultaneously at a distant source, with Alice local to the event of emission, and Bob$_N$ local to the event of detection of the soft photon, emitted at the (comoving) Bob$_N$ time $-T$. Evidently for a DSR-relativistic analysis our “thick slicing" of FRW must involve slices described by the DSR-deformed de Sitter scenario of sec. \[sec.DSR-dS\]. Analogously to sec. \[sec.Slicing\], the translational invariance (under the $\ell$-deformed translations (\[TransCoord-DSR\])) of the DSR-de Sitter setup of sec. \[sec.DSR-dS\], allows us to construct our slices choosing the Hamiltonian in the $n$-th slice, for each observer Bob$_n$, to have the same functional expression as the one of sec. \[sec.DSR-dS\], but with the corresponding value of the expansion rate $H_n$. It follows that in the $n$-th slice, Bob$_n$ describes the photons moving with velocity (for $\Pi>0$) $$v_{n}^{B_{n}}=1-\ell\left(\gamma+\beta\right)\left(1-H_{n}\eta^{B_{n}}\right)\Pi_{B_{n}}.$$ Moreover, each observer Bob$_n$ is connected to Alice, who is at the source, by a transformation (\[FiniteTransSlicing\]), but with the $\ell$-deformed translation generators given in Eqs. (\[repDSR-DS\]),(\[PhSpDSRdS\]) (with the $H_n$ appropriate for each $n$-th slice). That is Bob$_n$ is connected to Alice by a series of finite $\ell$-deformed space translations (each with the corresponding parameter $\xi_k$) followed by a series of finite $\ell$-deformed time translations (each with the corresponding parameter $\zeta_k,H_k$). Then, the relation between Bob$_N$’s and Alice’s coordinates is obtained computing the transformation (\[FiniteTransSlicing\]), for $n=N$, where each intermediate step is described by Eq. (\[TransCoord-DSR\]), with the relative translation parameter $\zeta_k$ or $\xi_k$ and expansion rate $H_k$. This leads to $$\begin{gathered}
\eta^{B_N} \!\!= \eta_N^{(0)} \!+\! \ell (\alpha \!-\! \gamma) \sum_{n=1}^N \zeta_n e^{\sum_{k=n}^N H_k \zeta_k} \!\left(\! 1 - H_n \eta_{n-1}^{(0)} \right) \! E_{H_n,n-1}^{(0)} , \nonumber \\
\begin{split}
x^{B_N} \!\!= x_N^{(0)} - & \ell (\alpha \!-\! \gamma) \sum_{n=1}^N e^{\sum_{k=n}^N H_k \zeta_k} H_n x_{n-1}^{(0)} E_{H_n,n-1}^{(0)},
\end{split}
\nonumber \\
\begin{split}
\Omega^{B_N} \!\!=\! \Omega_N^{(0)} + \, & \ell (\alpha \!-\! \gamma) \sum_{n=1}^N e^{-\sum_{k=n}^N H_k \zeta_k} H_n \zeta_n \Omega_n^{(0)} E_{H_n,n-1}^{(0)},
\end{split}
\nonumber\\
\Pi^{B_N} \!\!\!=\! \Pi_N^{(0)} \!\!+\! \ell (\alpha \!-\! \gamma) \!\sum_{n=1}^N \!e^{-\!\sum_{k=n}^N \! H_k \zeta_k} H_n \zeta_n \Pi^{(0)}_{n-1} E_{H_n,n-1}^{(0)},
\label{TransCoord-N}\end{gathered}$$ where $$E_{H_n,k}^{(0)}=\Omega_k^{(0)}\left( 1-H_n\eta_k^{(0)} \right)+H_n\Pi_k^{(0)}x_k^{(0)},
\label{TransCoord-N-2}$$ and $$\begin{gathered}
\eta_k^{(0)} = e^{\sum_{s=1}^k H_s \zeta_s} \eta^A + \sum_{s=1}^k e^{\sum_{r=s+1}^k H_r \zeta_r} \frac{1-e^{H_s \zeta_s}}{H_s} , \\
x_k^{(0)} = e^{\sum_{s=1}^k H_s \zeta_s} \left( x^A - \sum_{s=1}^N \xi_s \right) , \\
\Omega_k^{(0)} = e^{-\sum_{s=1}^k H_s \zeta_s}\Omega^A , ~~~
\Pi_k^{(0)} = e^{-\sum_{s=1}^k H_s \zeta_s}\Pi^A .
\end{gathered}
\label{TransCoord-N-3}$$ From these relations we see that, as in Sec. \[sec.DSR-dS\], relative locality affects the coordinates that Bob$_N$ assigns to the point of emission $\eta^{B_N}_{O_A},x^{B_N}_{O_A}$, which now depend on the particles momenta.
We impose again the conditions (\[param\]): $$\zeta_n = \zeta = T/N , \qquad
\xi_n = e^{-\sum_{k=1}^{n}H_{k} \zeta_n} \frac{e^{H_{n} \zeta_n} - 1}{H_{n}},
\label{param-2}$$ This amounts to enforcing that every observer Bob$_n$ is local, at the time $\eta(t_n)$, to the trajectory of a soft photon, for which the effects of the deformation can be neglected, emitted at Alice.
To evaluate the time of arrival of the hard photon, we consider the trajectory that Bob$_N$ assigns to photons in the $N$-th slice, which is given by $$x^{B_{N}} \! \left(\eta_{B_{N}} \!\right)_{N} \!\! =\! x_{O_A}^{B_{N}} \!+\!\!\! \sum_{n=1}^{N-1} \!\! \int_{\eta_{O_{n-1}}^{B_{N}}}^{\eta_{O_n}^{B_{N}}} \!\!\!\!\!\! v_{n}^{B_{N}}d\eta \!+\!\! \int_{\eta_{O_{N-1}}^{B_{N}}}^{\eta^{B_{N}}} \!\!\!\!\!\!\!\!\! v_{N}^{B_{N}}d\eta .
\label{worldlineBobN}$$ It is straightforward to verify, using Eqs. (\[TransCoord-N\]) to establish the relation between the coordinates of Bob$_N$ and the coordinates of Bob$_n$, that the velocity of the photons in the $n$-th slice is described by Bob$_N$ by simply re-expressing $v^{B_n}_n$ in terms of the coordinates of Bob$_N$: $$v^{B_N}_n(X^{B_N}) = v^{B_n}_n(X^{B_n}(X^{B_N})),$$ so that $$\begin{split}
& v_{n}^{B_{N}}(\eta^{B_N}) = 1-\ell\left(\gamma+\beta\right)\Bigg(e^{\sum_{s=n+1}^{N}H_{S}\zeta} \\ &
- H_{n}\sum_{k=n+1}^{N}e^{\sum_{s=k}^{N}H_{S}\zeta}\frac{1-e^{-H_{k}\zeta}}{H_{k}}-H_{n}\eta^{B_{N}}\Bigg)\Pi_{B_{N}}
\end{split}
\label{velBNBn-DSR}$$
Eq. (\[velBNBn-DSR\]) can be easily integrated in each $n$-th slice, and, after some algebra, the trajectory (\[worldlineBobN\]) can be rearranged as follows $$\begin{split}
& x^{B_N} \!\!=\! x_{O_A}^{B_{N}} \!\!-\! \eta_{O_A}^{B_{N}} \!\!+\! \eta^{B_{N}} \!\!-\! \ell\left(\!\gamma \!+\! \beta\right)\Pi_{B_{N}} \!\left( \!\eta^{B_{N}} \!\!-\! \frac{H_{n}}{2} \!\left(\eta^{B_{N}} \!\right)^{2}\right)\\
& +\ell\left(\gamma+\beta\right)\Pi_{B_{N}}\sum_{n=1}^{N}\left(\eta_{O_{n-1}}^{B_{N}}-\eta_{O_n}^{B_{N}}\right) \bigg[e^{\sum_{k=n+1}^{N}H_{k}a_{t}}-\\&
H_{n}\sum_{k=n+1}^{N}e^{\sum_{s=k}^{N}H_{s}a_{t}}\frac{1-e^{-H_{k}a_{t}}}{H_{k}}-\frac{H_{n}}{2}\left(\eta_{O_{n-1}}^{B_{N}}+\eta_{O_n}^{B_{N}}\right)\bigg]
\end{split}
\label{trajectoryBobN-DSR}$$ Again, the time of arrival of the hard photon is obtained by solving $\eta^{B_N}(x^B_N=0)$. We can identify two contributions to the delay: one comes from the terms proportional to $\ell(\gamma +\beta)$ (second and third raw of Eq. (\[trajectoryBobN-DSR\])); the other comes from the term $\eta_{O_A}^{B_{N}}-x_{O_A}^{B_{N}}$, which, as we pointed out above, contains a non-vanishing contribution proportional to $\ell (\alpha - \gamma)$, as one can see by using Eqs. (\[TransCoord-N\]) (recall that $\eta^{B_N}_{O_A} \!\!\!=\! \eta^{B_N}(\eta^{A} \!\!=\! 0, x^{A} \!\!=\! 0) $ and $x^{B_N}_{O_A} \!\!\!=\! x^{B_N}(\eta^{A} \!\!=\! 0, x^{A} \!\!=\! 0) $). Thus $$\Delta \eta = \ell (\gamma +\beta) \Delta \eta_{\gamma+\beta} + \ell (\alpha - \gamma) \Delta \eta_{\alpha - \gamma}$$ Considering that at 0-th order in $\ell$ the relations (\[O-1\]), (\[O-2\]), (\[O-3\]) hold, we get $$\Delta \eta_{\gamma+\beta} = p_h \sum_{n=1}^{N} e^{2\sum_{s=n}^{N}H_{s}\zeta}\frac{1-e^{-2H_{n}\zeta}}{2H_{n}},$$ where we denoted again with $p_h$ the momentum of the hard particle observed at the detector. From relations (\[TransCoord-N\]), (\[TransCoord-N-2\]), (\[TransCoord-N-3\]), (\[param-2\]), one gets (after a tedious but straightforward derivation) $$\begin{split}
& \Delta \eta_{\alpha - \gamma} = p_h \sum_{n=1}^{N}\zeta_{n}e^{2\sum_{k=n}^{N}H_{k}\zeta} \\&
\times \left(1-H_{n}\sum_{k=n}^{N}e^{-\sum_{r=n}^{k}H_{r}\zeta}\frac{e^{H_{k}\zeta}-1}{H_{k}}\right)^{2}.
\end{split}$$ Using again the relations (\[Riemann\]), (\[Riemann-2\]) and (\[DtDeta\]) we get that in the limit $N\rightarrow \infty$, the delay is $$\begin{split}
\Delta t = ~& \ell p_h \Bigg( \ell (\beta+\gamma) \int_{-T}^{0}\frac{dt}{a^2(t)} \\&
+ (\alpha-\gamma) \int_{-T}^{0}dt\left(\frac{1}{a\left(t\right)}+H\left(t\right)\int_{0}^{t}\frac{dt'}{a\left(t'\right)}\right)^{2} \Bigg).
\end{split}$$ Using relations (\[redshift\]), the delay can be written in terms of the redshift of the source as $$\begin{split}
& \Delta t \!=\! \ell p_h \Bigg( \! (\beta+\gamma) \int_{0}^{z}\frac{d\bar{z}\left(1+\bar{z}\right)}{H\left(\bar{z}\right)} \\ &
~~~~ + \! (\alpha-\gamma) \int_{0}^{z} \frac{d\bar{z}}{
\left( 1 \!+\! \bar{z}\right) \! H \! \left(\bar{z}\right)} \! \left( \! 1 \!+\! \bar{z} \!-\! H \! \left(\bar{z}\right) \!\! \int_{0}^{
\bar{z}\left(t\right)} \!\!\!\!\! \frac{d\bar{z}'}{H\left(\bar{z}'\right)}\right)^{\!\!2} \Bigg)
\end{split}
\label{gsacextra}$$
One of the interesting applications of this DSR-relativistic result (\[gsacextra\]) is to compare it to the corresponding result (\[gacextratoo\]) obtained in the LIV case. This comparison shows that, as expected, in general DSR scenarios and LIV scenarios produce completely different results, even when applied to the same class of modified dispersion relations. It is also noteworthy however that there is special case where the two pictures give the same result: by fixing $\alpha = \gamma$ in the DSR case of (\[gsacextra\]) one gets the same formula obtained by fixing $\lambda' = \lambda''= \lambda'''=0$ in the LIV case of (\[gacextratoo\]). This also provides a conceptual perspective on our results: on the DSR side the choice $\alpha = \gamma$ is such that translation transformations are unaffected (see Eq. (\[algebraDSR-DS\])) by the quantum-gravity scenario just like on the LIV side the choice $\lambda' = \lambda''= \lambda'''=0$ is such that translational invariance is unaffected (see Eq. (\[joclambda\])) by the quantum-gravity scenario. So both in the LIV case and in the DSR case the formula advocated by Jacob and Piran [@jacobpiran] $$\Delta t \propto \int_{0}^{z}\frac{d\bar{z}\left(1+\bar{z}\right)}{H\left(\bar{z}\right)}$$ is applicable only if the quantum-gravity scenario has no implications for translational invariance and translation transformations. When the quantum-gravity scenario does have implications for translational invariance and/or translation transformations the Jacob-Piran [*ansatz*]{} does not apply and the DSR scenario gives results different from the LIV scenario.
Conclusions
===========
Our research work was motivated by the fact that there are at this point literally hundreds of publications on applications of Planck-scale-modified dispersion relations in contexts involving FRW expansion, but all these studies have an exclusively heuristic basis. Over the last decade significant progress was made on the understanding of Planck-scale-modified dispersion relations in the flat-spacetime limit and we now do have a partial but satisfactory understanding of the generalization to the case of expansion at constant rate, but for FRW expansion before our investigations we only had heuristic analyses and very limited understanding of the conceptual issues at stake.
We feel we here provided a significant step forward toward raising the standards of this phenomenology. Of particular significance is the understanding that the much-used Jacob-Piran [*ansatz*]{} implicitly assumes that spacetime translations are unaffected by the quantum-gravity scenario, and therefore that [*ansatz*]{} is applicable exclusively to a corresponding subset of possible quantum-gravity scenarios. We obtained predictions for both the LIV case and the DSR case applicable when instead quantum gravity affects spacetime translations, thereby finally providing a target for those interested in testing Planck-scale modifications of the dispersion relation in full generality.
For some of our results an important role was played by our setup describing FRW expansion as a series of stages of de-Sitter expansion, with appropriate conditions for gluing the different “thick slices". This provided a safe path for generalizing to the case of FRW expansion the results recently obtained for de-Sitter expansion. While quantum-gravity research is certainly full of surprises, we cannot imagine any quantum-gravity picture that would prevent one from describing FRW expansion as a series of de-Sitter expansions, and indeed to our knowledge there is no quantum-gravity result in the literature that would suggest this could be prevented. Since we have here shown that this “thick-slicing setup" can play a pivotal role in phenomenology, of course this issue should attract even more interest in the future.
Acknowledgements {#acknowledgements .unnumbered}
================
GAC is supported by a grant from the John Templeton Foundation. GR is supported by funds provided by the National Science Center under the agreement DEC-2011/02/A/ST2/00294. AM wish to acknowledge support by the Shanghai Municipality, through the grant No. KBH1512299, and by Fudan University, through the grant No. JJH1512105.
[50]{}
G. Amelino-Camelia, Living Rev. Rel. [**16**]{} (2013) 5 \[arXiv:0806.0339 \[gr-qc\]\].
D. Mattingly, Living Rev. Rel. [**8**]{} (2005) 5 \[gr-qc/0502097\].
U. Jacob and T. Piran, JCAP [**0801**]{} (2008) 031 \[arXiv:0712.2170 \[astro-ph\]\].
M. Rodriguez Martinez and T. Piran, JCAP [**0604**]{} (2006) 006 \[astro-ph/0601219\].
J. R. Ellis, N. E. Mavromatos, D. V. Nanopoulos, A. S. Sakharov and E. K. G. Sarkisyan, Astropart. Phys. [**25**]{} (2006) 402 \[Astropart. Phys. [**29**]{} (2008) 158\] \[arXiv:0712.2781 \[astro-ph\], astro-ph/0510172\].
S. Majid and H. Ruegg, Phys. Lett. B [**334**]{} (1994) 348 \[hep-th/9405107\].
J. Lukierski, A. Nowicki and H. Ruegg, Phys. Lett. B [**293**]{} (1992) 344; J. Lukierski, H. Ruegg, A. Nowicki and V. N. Tolstoi, Phys. Lett. B [**264**]{} (1991) 331.
G. Amelino-Camelia, N. Loret and G. Rosati, Phys. Lett. B [**700**]{} (2011) 150 \[arXiv:1102.4637 \[hep-th\]\].
G. Amelino-Camelia, Int. J. Mod. Phys. D [**11**]{} (2002) 35 \[gr-qc/0012051\].
Phys. Rev. D [**67**]{} (2003) 044017 \[gr-qc/0207085\].
J. Kowalski-Glikman, Phys. Lett. A [**286**]{} (2001) 391 \[hep-th/0102098\].
G. Amelino-Camelia, L. Freidel, J. Kowalski-Glikman and L. Smolin, Phys. Rev. D [**84**]{} (2011) 084010 \[arXiv:1101.0931 \[hep-th\]\].
G. Amelino-Camelia, M. Matassa, F. Mercati and G. Rosati, Phys. Rev. Lett. [**106**]{} (2011) 071301 \[arXiv:1006.2126 \[gr-qc\]\].
G. Amelino-Camelia, A. Marcianò, M. Matassa and G. Rosati, [*Phys. Rev. D*]{} [**86**]{} (2012) 124035, arXiv:1206.5315 \[hep-th\].
J. Ellis and N. E. Mavromatos, Astropart. Phys. [**43**]{} (2013) 50 \[arXiv:1111.1178 \[astro-ph.HE\]\].
A. A. Abdo [*et al.*]{} \[Fermi-LAT and Fermi GBM Collaborations\], Science [**323**]{} (2009) 1688.
M. Ackermann [*et al.*]{} \[Fermi GBM/LAT Collaboration\], Nature [**462**]{} (2009) 331 \[arXiv:0908.1832 \[astro-ph.HE\]\].
J. Bolmont and A. Jacholkowska, Adv. Space Res. [**47**]{} (2011) 380 \[arXiv:1007.4954 \[astro-ph.HE\]\].
J. Albert [*et al.*]{} \[MAGIC and Other Contributors Collaborations\], Phys. Lett. B [**668**]{} (2008) 253 \[arXiv:0708.2889 \[astro-ph\]\].
A. Abramowski [*et al.*]{} \[HESS Collaboration\], Astropart. Phys. [**34**]{} (2011) 738 \[arXiv:1101.3650 \[astro-ph.HE\]\].
F. Aharonian [*et al.*]{} \[HESS Collaboration\], Phys. Rev. Lett. [**101**]{} (2008) 170402 \[arXiv:0810.3475 \[astro-ph\]\].
G. Amelino-Camelia, M. Arzano, J. Kowalski-Glikman, G. Rosati and G. Trevisan, Class. Quant. Grav. [**29**]{} (2012) 075007 \[arXiv:1107.1724 \[hep-th\]\].
A. Marcianò, G. Amelino-Camelia, N. R. Bruno, G. Gubitosi, G. Mandanici and A. Melchiorri, JCAP [**1006**]{} (2010) 030 \[arXiv:1004.1110 \[gr-qc\]\].
[^1]: One can show, using the Poisson brackets (\[CanPhSpConf\]) and the chain rule that the velocity can be also expressed, after enforcing ${\cal H}=0$, as $v(\eta) \equiv \partial\Omega\left( \Pi \right)/\partial \Pi$.
[^2]: As usually done in this research area, we exclude terms going with odd powers of momentum, since (in their application to spacetimes with two or more spatial dimensions) they would bring in implications also for spatial rotations, implications which are not usually expected in the quantum-gravity literature.
[^3]: For a generator $G$ with parameter $a$, the finite action on a coordinate $x$ is $e^{{a}G} \triangleright x \equiv \sum_{n=0}^{\infty} \frac{a^n}{n!}\left\lbrace G,x\right\rbrace_n $, where $ \lbrace G,x \rbrace_n = \lbrace G, \lbrace G,x \rbrace_{n-1}\rbrace$, $\lbrace G, x \rbrace_0 = x $. In this formalism, the composed action of a spatial translation followed by a time translation is given by $e^{-\xi p}\triangleright e^{-{\zeta}E}\triangleright x$ (cfr. also [@DSR-DS]).
[^4]: The Jacob-Piran proposal [@jacobpiran] applies to the general case of an arbitrary expansion rate, but it is for us particularly insightful to analyze it in the special case of constant expansion rate. With constant expansion rate the original theory is fully translationally invariant, whereas for non-constant expansion rate time translations are not a symmetry. It is noteworthy that the Jacob-Piran proposal is such that the LIV effects are fully translationally invariant (including invariance under time translations) when the expansion rate is constant, whereas other possible LIV terms would not have this property. This is a key aspect of what we label as a case where translations are unaffected by the quantum-gravity effects.
[^5]: Consistently with our objectives, we enforce the requirement that in the “flat limit”, $H\rightarrow 0$, one should get the velocity $v \simeq 1 - \ell E$. We also do not make room for deformation terms proportional to $N$, since our flat-limit requirements would only allow such terms to be of the type $\ell H N E p$ or $\ell H N E^2$, producing negligibly small effects. In general we should proceed with full awareness of the smallness of both $\ell$ and $H$, so among possible terms going with $\ell H$ we shall only keep track of those involving $\ell H x $ or $\ell H t$, [*i.e.*]{} cases where the cosmological travel times and travel distances here of interest can compensate for the smallness of $\ell H$.
[^6]: Previous related studies of relative locality have shown [@bobKAPPAplb; @anatomy; @DSR-DS] that the physical results do not depend on the choice of the representation.
[^7]: Above we used the symbol $\Delta t$ to indicate the delay in time of arrival of a hard photon with respect to a soft photon. Here with the symbol $\Delta t$ we indicate the thickness, in time, of each slice. Which notion of $\Delta t$ we refer to at a given point in the manuscript should always be clear from the context.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate Hörmander spectral multiplier theorems as they hold on $X = L^p(\Omega),\: 1 < p < \infty,$ for many self-adjoint elliptic differential operators $A$ including the standard Laplacian on ${\mathbb{R}}^d.$ A strengthened matricial extension is considered, which coincides with a completely bounded map between operator spaces in the case that $X$ is a Hilbert space. We show that the validity of the matricial Hörmander theorem can be characterized in terms of square function estimates for imaginary powers $A^{it}$, for resolvents $R(\lambda,A),$ and for the analytic semigroup $\exp(-zA).$ We deduce Hörmander spectral multiplier theorems for semigroups satisfying generalized Gaussian estimates.'
address: |
Ch. Kriegler\
Laboratoire de Mathématiques (CNRS UMR 6620)\
Université Blaise-Pascal (Clermont-Ferrand 2)\
Campus des Cézeaux\
63177 Aubière Cedex\
France
author:
- 'Ch. Kriegler'
---
Introduction {#Sec 1 Intro}
============
Let $f$ be a bounded function on $(0,\infty)$ and $u(f)$ the operator on $L^p({\mathbb{R}}^d)$ defined by $[u(f)g]\hat{\phantom{i}}(\xi) = f(|\xi|^2) \hat{g}(\xi).$ Hörmander’s theorem on Fourier multipliers [@Hor Theorem 2.5] asserts that $u(f) : L^p({\mathbb{R}}^d) \to L^p({\mathbb{R}}^d)$ is bounded for any $p \in (1,\infty)$ provided that for some integer $N$ strictly larger than $\frac{d}{2}$ $$\label{Equ Hormander original}
\sup_{R > 0} \int_{R/2}^{2R} \left|t^k f^{(k)}(t)\right|^2 \frac{dt}{t} < \infty \quad \left(k = 0,1,\ldots,N\right).$$
This theorem has many refinements and generalisations to various similar contexts. For $\alpha > \frac12,$ let ${W^\alpha}_2({\mathbb{R}}) = \{ f \in L^2({\mathbb{R}}):\: \|f\|_{{W^\alpha}_2({\mathbb{R}})} = \| (1 + \xi^2)^{\alpha/2} \hat{f}(\xi)\|_{L^2({\mathbb{R}})} < \infty \}$ denote the usual Sobolev space, and ${W^\alpha}= \{ f : (0,\infty) \to {\mathbb{C}}:\: f \circ \exp \in {W^\alpha}_2({\mathbb{R}})\},$ which is a Banach algebra with respect to $\|f\|_{{W^\alpha}} = \|f \circ \exp\|_{{W^\alpha}_2({\mathbb{R}})}.$ Let $\phi_0 \in C^\infty_c(\frac12,2).$ For $n \in {\mathbb{Z}},$ let $\phi_n = \phi_0(2^{-n} \cdot)$ and assume that $\sum_{n \in {\mathbb{Z}}} \phi_n(t) = 1$ for any $t > 0.$ Such a function exists [@BeL Lemma 6.1.7] and we call $(\phi_n)_{n \in {\mathbb{Z}}}$ a dyadic partition of unity. We define the Banach algebra $${{\mathcal{H}}^\alpha}= \left\{ f : (0,\infty) \to {\mathbb{C}}:\: \| f\|_{{{\mathcal{H}}^\alpha}} = \sup_{n \in {\mathbb{Z}}} \|\phi_n f\|_{{W^\alpha}} < \infty \right\}.$$ The definition of ${{\mathcal{H}}^\alpha}$ is independent of the dyadic partition of unity, different choices resulting in equivalent norms [@Kr Section 4.2]. The space ${{\mathcal{H}}^\alpha}$ refines , more precisely, $f \in {{\mathcal{H}}^\alpha}$ implies that $f$ satisfies for $N \leq \alpha,$ and the converse holds for $N \geq \alpha$ [@Kr Proposition 4.11].
Now if $A$ is a self-adjoint positive operator on some $L^2(\Omega,\mu),$ then its functional calculus assigns to any bounded measurable function $f$ on $(0,\infty)$ an operator $f(A)$ on $L^2(\Omega,\mu).$ In particular, if $A = - \Delta$ and $(\Omega,\mu) = ({\mathbb{R}}^d,dx),$ then $f(A)$ equals the above $u(f).$ A theorem of Hörmander type holds true for many elliptic differential operators $A,$ including sublaplacians on Lie groups of polynomial growth, Schrödinger operators and elliptic operators on Riemannian manifolds [@DuOS; @Alex; @Bluna; @Chri; @Duon]. By this, we mean that $$\label{Equ Intro Hormander calculus}
u : {{\mathcal{H}}^\alpha}\to B(X),\, f \mapsto f(A)\text{ is a bounded homomorphism,}$$ where $X = L^p(\Omega),\,p \in (1,\infty),$ $\alpha$ is the differentiation parameter typically larger than $\frac{d}{2},$ where $d$ is the dimension of $\Omega,$ and $f(A)$ is (the unique bounded $L^p$-extension of) the self-adjoint functional calculus.\
The aim of this article is to characterize the validity of the Hörmander multiplier theorem for $A$ in terms of square function estimates.
The latter have been introduced in Stein’s classical book [@Ste1] and have since then been used widely with applications to functional calculi and multiplier theorems. Note that $\|(\cdot)^{it}\|_{{{\mathcal{H}}^\alpha}} \cong ( 1 + |t|^2)^{\alpha/2}$ [@Kr Proposition 4.12 (4)], so that for this particular function, implies $\|A^{it}\| \leq C ( 1 + |t|^2 )^{\alpha/2}.$ Then a natural square function estimate for our situation is $$\label{Equ Intro Square Function}
\| (1 + t^2)^{-\alpha/2} A^{it} x \|_{\gamma({\mathbb{R}},X)} \leq C \| x\|_X ,$$ where $\gamma({\mathbb{R}},X)$ is given by $$\| x(t) \|_{\gamma({\mathbb{R}},X)} \cong \left\| \left( \int_{\mathbb{R}}\left|x(t)\right|^2 dt \right)^{\frac12} \right\|_X$$ for $X = L^p(\Omega,\mu)$ and $p \in [1,\infty),$ which explains the name square function. The general definition of the space $\gamma({\mathbb{R}},X)$ involves Gaussian random sums in the Banach space $X,$ see Section \[Sec 2 Prelims\].
Our setting, developed in Section \[Sec 2 Prelims\], is as follows: We let $X$ be a Banach space having Pisier’s property $(\alpha),$ which a geometric property playing an important role for the theory of spectral multipliers. It is natural to assume the operator $A$ to be $0$-sectorial i.e. a negative generator of an analytic semigroup $(\exp(-zA))_{\operatorname{Re}z > 0}$ which is uniformly bounded on the sector $\Sigma_\omega = \{ z \in {\mathbb{C}}\backslash \{ 0 \} :\: | \arg z | < \omega \}$ for each $\omega < \frac{\pi}{2}.$ Indeed, $\exp(-z \cdot)$ belongs to ${{\mathcal{H}}^\alpha}$ with uniform norm bound on such sectors. Further, for simplicity we assume throughout that $A$ has dense range.
We shall base the definition of $u$ in on the well-known ${H^\infty}$ functional calculus [@CDMY; @KuWe]. This means that for $f$ belonging to ${H^\infty}_0(\Sigma_\omega) = \{ f \in {H^\infty}(\Sigma_\omega) :\: \exists\: \epsilon,\, C > 0 \text{ s.th. }|f(z)| \leq C \min ( |z|^{\epsilon}, |z|^{-\epsilon} ) \}$ which is a subclass of ${H^\infty}(\Sigma_\omega) = \{ f : \Sigma_\omega \to {\mathbb{C}}:\: f\text{ is analytic, }\|f\|_{\infty,\omega} = \sup_{z \in \Sigma_\omega} | f(z) | < \infty \},$ $f(A) \in B(X)$ is defined by a certain Cauchy integral formula, see . Secondly, under certain conditions, $A$ has a bounded ${H^\infty}$ calculus, which means that there is an extension to a bounded homomorphism ${H^\infty}(\Sigma_\omega) \to B(X),\,f \mapsto f(A).$ Note that ${H^\infty}(\Sigma_\omega)$ is a subclass of ${{\mathcal{H}}^\alpha}.$ In Lemma \[Lem A u\] it will be shown in particular that an extension of the ${H^\infty}$ calculus to a bounded homomorphism $u : {{\mathcal{H}}^\alpha}\to B(X)$ is unique.
For any such mapping $u$ and $n\in {\mathbb{N}},$ we now consider the linear tensor extension $$u_n : \begin{cases} M_n \otimes {{\mathcal{H}}^\alpha}& \to M_n \otimes B(X) \\ a \otimes f & \mapsto a \otimes u(f) \end{cases},$$ where $M_n$ is the space of $n \times n$ scalar matrices. We will equip both $M_n \otimes {{\mathcal{H}}^\alpha}$ and $M_n \otimes B(X)$ with suitable norms. In fact, ${{\mathcal{H}}^\alpha}$ will become an operator space (see Section 4), $M_n \otimes B(X) \cong B(\ell^2_n \otimes_2 X)$ if $X$ is a Hilbert space, and if $X$ is a Banach space, $M_n \otimes B(X) \cong B(\operatorname{Gauss}_n(X))$ carries the norm induced by an action on $X$-valued Gaussian random sums. We call $u$ matricially $\gamma$-bounded in this article if $$\label{Equ Intro mat-gamma}
\|u\|_{{\text{mat-}\gamma}}= \sup_{n \in {\mathbb{N}}} \| u_n \| < \infty.$$ This is in general strictly stronger than $\|u\| < \infty$ (see Proposition \[Prop bounded vs mat-gamma bounded\]), and is related to the following two well-known boundedness notions, explained in Section \[Sec 2 Prelims\]. First, if $X$ is a Hilbert space, then is equivalent to the complete boundedness of $u,$ and second, if $X$ is a Banach space, then entails that the set of spectral multipliers $\{ u(f) : \: \|f\|_{{{\mathcal{H}}^\alpha}} \leq 1\}$ is $\gamma$-bounded.
The main result reads as follows.
\[Thm Intro Main\] Let $X$ be a space with property $(\alpha).$ Let $A$ be a $0$-sectorial operator on $X$ with bounded ${H^\infty}$ calculus. Let $\alpha > \frac12.$ Then the following are equivalent.
1. The square function estimate holds.
2. The ${H^\infty}$ calculus mapping $f\mapsto f(A)$ extends to a homomorphism $u : {{\mathcal{H}}^\alpha}\to B(X)$ which is matricially $\gamma$-bounded.
Theorem \[Thm Intro Main\] entails a spectral multiplier theorem in the following situations: The space $X = L^p(\Omega)$ for $p \in (1,\infty)$ has property $(\alpha).$ If $(\Omega,\mu)$ is a $d$-dimensional space of homogeneous type, e.g. a sufficiently regular open subset of ${\mathbb{R}}^d$ with Lebesgue measure $\mu,$ and $A$ is self-adjoint positive on $L^2(\Omega)$ such that the corresponding semigroup $\exp(-tA)$ has an integral kernel $k_t(x,y)$ that satisfies the Gaussian estimate for some $m \in {\mathbb{N}}$ $$\label{Equ Intro GE}
|k_t(x,y)| \leq C \mu(B(x,t^{\frac1m}))^{-1} \exp\left( -c (\text{dist}(x,y)/t^{\frac1m})^{\frac{m}{m-1}} \right) \quad (x,y \in \Omega,\: t > 0),$$ then $A$ has a bounded ${H^\infty}$ calculus on $X$ [@DR Theorem 3.4], [@Blun Corollary 2.3]. This is indeed the case for many operators listed before [@Bluna Section 2]. Moreover, the mappings $u$ from and Theorem \[Thm Intro Main\] (2) are the same, so that we obtain as a corollary
\[Cor Intro\] Assume that $A$ is a self-adjoint positive operator on $L^2(\Omega)$ satisfying . Let $\alpha > \frac12$ and $p \in (1,\infty).$ If $A$ satisfies the square function estimate $$\label{Equ Intro Cor}
\left\| \left( \int_{\mathbb{R}}\left|(1 + t^2 )^{-\alpha/2} A^{it} x \right|^2 dt \right)^{\frac12} \right\|_p \leq C \|x\|_p,$$ then for any $f \in {{\mathcal{H}}^\alpha},$ the spectral multiplier $f(A)$ is bounded $ L^p(\Omega) \to L^p(\Omega).$
In Proposition \[Prop bounded vs mat-gamma bounded\], we will show a partial converse of Corollary \[Cor Intro\]. More precisely, implies that a restriction to a smaller Hörmander space ${\mathcal{H}}^\beta$ is matricially $\gamma$-bounded.
Let us close the introduction with an overview of the rest of the article. In Section \[Sec 2 Prelims\], we give the necessary background of the above mentioned notions of matricial norms, square functions, Gaussian random sums and functional calculus. Matricially $\gamma$-bounded mappings and the connection to square functions are explained in Section \[Sec 3 Main Thm\]. Section \[Sec 4 Hormander operator space\] is devoted to homomorphisms $u : {{\mathcal{H}}^\alpha}\to B(X)$ and the connection to ${H^\infty}$ functional calculus. Moreover Theorem \[Thm Intro Main\] is proved. A main ingredient is to deduce a spectral decomposition of Paley-Littlewood type, see , under the hypotheses of Theorem \[Thm Intro Main\]. In Section \[Sec 5 Examples\], we discuss some extensions and applications. Firstly, the square function estimate in terms of imaginary powers $A^{it}$ in Theorem \[Thm Intro Main\] has several equivalent and almost equivalent rewritings in terms of other typical square functions, involving the analytic semigroup $$\begin{aligned}
\left\| A^{\frac12} \exp(-te^{i\theta} A)x \right\|_{\gamma({\mathbb{R}}_+,X)} & \leq C (\frac{\pi}{2} - |\theta| )^{-\beta} \|x\| & \quad \left(\theta \in (-{\pi}/{2},{\pi}/{2}) \right),
\label{Equ Intro Sgr Square Function} \\
\intertext{or resolvents}
\|A^{\frac12} R(e^{i\theta}t,A) x \|_{\gamma({\mathbb{R}}_+,X)} & \leq C |\theta|^{-\beta} \|x\| & \quad \left(\theta \in (-\pi,\pi) \backslash \{ 0 \} \right). \label{Equ Intro Res Square Function}\end{aligned}$$ We have $\Rightarrow$ and for $\alpha \leq \beta,$ and conversely, or $\Rightarrow$ for $\alpha > \beta.$
Secondly, we discuss Theorem \[Thm Intro Main\] in the presence of generalized Gaussian estimates (see Assumption \[Ass Examples\]), which in particular covers semigroups satisfying . This is a well-studied property in connection with (Hörmander) functional calculus, see e.g. [@Duon; @DuOS]. In particular, we show the square function assumption of Corollary \[Cor Intro\] in the form of and improve the derivation order of the Hörmander theorem from $\alpha > \frac{d}{2} + \frac12$ as proved in [@Bluna] to $\alpha > d \left| \frac{1}{p_0} - \frac12 \right| + \frac12.$ We finally discuss the connections and differences between matricially $\gamma$-bounded Hörmander calculus and bounded Hörmander calculus. The last Section \[Sec 6 Proofs Lemmas\] contains some technical proofs of Section \[Sec 4 Hormander operator space\].
Preliminaries on Operator spaces, Gaussian sums, Square functions and Functional calculus {#Sec 2 Prelims}
=========================================================================================
We will need in different contexts cross norms on a tensor product of two Banach spaces.
Operator spaces {#operator-spaces .unnumbered}
---------------
A Banach space $E$ is called operator space if it is isometrically embedded into $B(H),$ where $H$ is a Hilbert space. Let $M_n$ denote the space of scalar $n \times n$ matrices. What makes operator spaces different from mere Banach spaces is that there is a specific collection of norms on $M_n \otimes E$, the operator space structure of $E.$ Namely for all $n \in {\mathbb{N}},$ it is equipped with the norm arising from the embedding $M_n \otimes E \hookrightarrow B(\ell^2_n(H)),\, [a_{ij}] \otimes x \mapsto \left( (h_i)_{i=1}^n \mapsto (\sum_{j=1}^n a_{ij}x(h_j))_{i=1}^n \right).$
Let $E$ and $F$ be operator spaces and $u : E \to F$ a linear mapping. For any $n \in {\mathbb{N}},$ let $u_n$ be the linear mapping $M_n \otimes E \to M_n \otimes F,\, a \otimes x \mapsto a \otimes u(x).$ Then $u$ is called completely bounded (completely isometric) if $\|u\|_{cb} = \sup_{n \in {\mathbb{N}}} \|u_n\| < \infty$ (for any $n \in {\mathbb{N}},\:u_n$ is isometric).
Clearly, any space $B(H)$ itself is an operator space, so in particular $M_m = B(\ell^2_m)$ is. Further we will consider the Hilbert row space $\ell^2_r = \{h \mapsto {\langle h, x \rangle}e :\:x \in \ell^2\} \subset B(\ell^2)$ where $e \in \ell^2$ is a fixed element of norm $1$ and ${\langle h, x \rangle}$ is the scalar product. Different choices of $e$ give isometric norms of $M_n \otimes \ell^2_r$ and $\ell^2_r$ is isometric to $\ell^2$ as a Banach space. We shall also consider the $m$-dimensional subspaces $\ell^2_{m,r} \subset \ell^2_r.$ These are completely isometrically determined by the following embedding, which also explains the name of row space:
$$\label{Equ i_m}
i_m : \ell^2_m \hookrightarrow M_m,\,(a_1,\ldots,a_m) \mapsto
\left(
\begin{array}{ccc}
a_1 & \ldots & a_m \\
0 & \ldots & 0 \\
\vdots & \ddots & \vdots \\
0 & \ldots & 0 \\
\end{array}
\right).$$
We refer to the books [@ER; @Pis2] for further information on operator spaces.
$\gamma$-bounded sets, property $(\alpha)$ and square functions {#gamma-bounded-sets-property-alpha-and-square-functions .unnumbered}
---------------------------------------------------------------
We let $\Omega$ be a probability space and $(\gamma_k)_{k \in {\mathbb{Z}}}$ a sequence of independent standard Gaussian random variables on $\Omega.$ For a Banach space $X,$ we let $\operatorname{Gauss}(X)$ be the closure of $\operatorname{span}\{ \gamma_k \otimes x_k :\: k \in {\mathbb{Z}}\}$ in $L^2(\Omega;X)$ with respect to the norm $$\label{Equ Def Gauss}
{\bigl\Vert \sum_k \gamma_k \otimes x_k \bigr\Vert}_{\operatorname{Gauss}(X)} = \left( \int_\Omega \left\| \sum_{k=1}^n \gamma_k(\omega) x_k \right\|^2 d\omega \right)^{\frac12}.$$ It will be convenient to denote $\operatorname{Gauss}_n(X)$ the subspace of $\operatorname{Gauss}(X)$ of elements of the form $\sum_{k=1}^n \gamma_k \otimes x_k.$
Note that if $X$ is a Hilbert space, then $$\label{Equ Hilbert Gaussian}
{\bigl\Vert \sum_k \gamma_k \otimes x_k \bigr\Vert}_{\operatorname{Gauss}_n(X)}^2 = \sum_{k=1}^n \|x_k\|^2.$$
A collection $\tau \subset B(X)$ is called $\gamma$-bounded if there exists $C > 0$ such that $$\left\| \sum_k \gamma_k \otimes T_k x_k \right\|_{\operatorname{Gauss}(X)} \leq C \left\| \sum_k \gamma_k \otimes x_k \right\|_{\operatorname{Gauss}(X)}$$
for any finite families $T_1,\ldots,T_n \in \tau$ and $x_1,\ldots, x_n \in X.$ The least admissible constant is denoted by $\gamma(\tau)$ (and $\gamma(\tau) := \infty$ if such a $C$ does not exist). Note that a $\gamma$-bounded set is automatically uniformly norm bounded, since one has $\gamma(\tau) \geq \sup_{T \in \tau} \|T\|.$ For $\sigma,\tau \subset B(X)$ and $\sigma \circ \tau = \{S \circ T :\: S \in \sigma,\,T \in \tau\},$ one has $\gamma(\sigma\circ \tau) \leq \gamma(\sigma)\gamma(\tau).$ The set $\tau = \{ a \operatorname{id}_X :\: a \in {\mathbb{C}},\, |a| \leq 1\}$ is $\gamma$-bounded with constant 1.
We say that $X$ has property $(\alpha)$ if there is a constant $C \geq 1$ such that for any finite family $(x_{ij})$ in $X$, we have $$\label{Equ Property Gauss alpha}
\frac1C {\bigl\Vert \sum_{i,j} \gamma_{ij} \otimes x_{ij} \bigr\Vert}_{\operatorname{Gauss}(X)} \leq {\bigl\Vert\sum_{i,j} \gamma_i \otimes \gamma_j \otimes x_{ij} \bigr\Vert}_{\operatorname{Gauss}(\operatorname{Gauss}(X))} \leq C {\bigl\Vert \sum_{i,j} \gamma_{ij} \otimes x_{ij} \bigr\Vert}_{\operatorname{Gauss}(X)},$$ where $\gamma_{ij}$ is a doubly indexed family of independent standard Gaussian variables. Property $(\alpha)$ is inherited by closed subspaces and isomorphic spaces. The $L^p$ spaces have property $(\alpha)$ for $1 \leq p < \infty$ and moreover, if $X$ has property $(\alpha),$ then also $L^p(\Omega;X)$ has. Property $(\alpha)$ is usually defined in terms of independent Rademacher variables $\epsilon_i,$ i.e. $\operatorname{Prob}(\epsilon_i = \pm 1) = \frac12$ instead of Gaussian variables [@Pis]. In analogy with , we define $\operatorname{Rad}(X) \subset L^2(\Omega;X)$ by $${\bigl\Vert \sum_k \epsilon_k \otimes x_k \bigr\Vert}_{\operatorname{Rad}(X)} = \left( \int_\Omega \left\| \sum_{k=1}^n \epsilon_k(\omega) x_k \right\|^2 d\omega \right)^{\frac12}.$$ It turns out that the two definitions are the same:
\[Rem Property Gauss alpha\] The property is equivalent to the following equivalence uniform in finite families $(x_{ij})$ in $X.$ $$\label{Equ Property Rad alpha}
{\bigl\Vert \sum_{i,j} \epsilon_i \otimes \epsilon_j \otimes x_{ij} \bigr\Vert}_{\operatorname{Rad}(\operatorname{Rad}(X))} \cong {\bigl\Vert \sum_{i,j} \epsilon_{ij} x_{ij} \bigr\Vert}_{\operatorname{Rad}(X)}.$$
First observe that the Schatten classes $S^p$ for $p \in (1,\infty) \backslash \{ 2 \}$ are spaces which do not satisfy nor . This is shown in [@Pis] for the Rademachers. On the other hand, $S^p$ has finite cotype, which implies that on this space, Rademacher sums and Gaussian sums are equivalent [@DiJT Theorem 12.27], i.e. $$\label{Equ Gauss Rad}
\|\sum_{k \in F} \gamma_k \otimes x_k\|_{\operatorname{Gauss}(X)} \cong \|\sum_{k \in F} \epsilon_k \otimes x_k\|_{\operatorname{Rad}(X)},$$ uniformly in $F \subset {\mathbb{Z}}.$ From this one easily deduces that does not hold.
Next observe that by the Banach-Mazur theorem and [@DiJT Theorem 3.2], $S^p$ (in fact any Banach space) has the property that all finite dimensional subspaces are isomorphic to a subspace of some $\ell^\infty_n,$ with one fixed isomorphism constant. This implies that $\ell^\infty$ does not satisfy nor .
Therefore, by the characterization of finite cotype in [@DiJT Theorem 14.1], a space $X$ satisfying or has finite cotype. As cited above, Rademacher and Gaussian sums are then equivalent, so the corresponding expressions in and are, which shows the lemma.
We recall the construction of Gaussian function spaces from [@KaW2], see also [@KNVW Section 1.3].
Let $H$ be a separable Hilbert space. We consider the tensor product $H \otimes X$ as a subspace of $B(H,X)$ in the usual way, i.e. by identifying $\sum_{k=1}^n h_k \otimes x_k \in H \otimes X$ with the mapping $u : h \mapsto \sum_{k=1}^n {\langle h, h_k \rangle} x_k$ for any finite families $h_1,\ldots,h_n \in H$ and $x_1,\ldots,x_n \in X.$ Choose such families with corresponding $u$, where the $h_k$ shall be orthonormal. Let $\gamma_1,\ldots,\gamma_n$ be independent standard Gaussian random variables over some probability space. We equip $H \otimes X$ with the norm $${\bigl\Vert u \bigr\Vert}_{\gamma(H,X)} = {\bigl\Vert \sum_k \gamma_k \otimes x_k \bigr\Vert}_{\operatorname{Gauss}(X)}.$$ By [@DiJT Corollary 12.17], this expression is independent of the choice of the $h_k$ representing $u.$ We let $\gamma(H,X)$ be the completion of $H \otimes X$ in $B(H,X)$ with respect to that norm. Then for $u \in \gamma(H,X),$ $\|u\|_{\gamma(H,X)} = {\bigl\Vert\sum_k \gamma_k \otimes u(e_k)\bigr\Vert}_{\operatorname{Gauss}(X)},$ where the $e_k$ form an orthonormal basis of $H$ [@KaW2 Remark 4.2].\
A particular subclass of $\gamma(H,X)$ will be important, which is obtained by the following procedure. Assume that $(\Omega,\mu)$ is a $\sigma$-finite measure space and $H = L^2(\Omega).$ Denote $P_2(\Omega,X)$ the space of Bochner-measurable functions $f: \Omega \to X$ such that $x' \circ f \in L^2(\Omega)$ for all $x' \in X'.$ We identify $P_2(\Omega,X)$ with a subspace of $B(L^2(\Omega),X'')$ by assigning to $f$ the operator $u_f$ defined by $$\label{Equ u_f}
{\langle u_f h, x' \rangle} = \int_\Omega {\langle f(t), x' \rangle} h(t) d\mu(t).$$ An application of the uniform boundedness principle shows that, in fact, $u_f$ belongs to $B(L^2(\Omega),X)$ [@KaW2 Section 4], [@Frohl Section 5.5]. Then we let $$\gamma(\Omega,X) = \left\{ f \in P_2(\Omega,X):\: u_f \in \gamma(L^2(\Omega),X)\right\}$$ and set $$\|f\|_{\gamma(\Omega,X)} = \|u_f\|_{\gamma(L^2(\Omega),X)}.$$ The space $\{u_f:\:f \in \gamma(\Omega,X)\}$ is a dense and in general proper subspace of $\gamma(L^2(\Omega),X).$ Resuming the above, we have the following embeddings of spaces, cf. also [@LM10 Section 3]. $$L^2(\Omega)\otimes X \to \gamma(\Omega,X) \to \gamma(L^2(\Omega),X) \to B(L^2(\Omega),X).$$
In some cases, $\gamma(L^2(\Omega),X)$ and $\gamma(\Omega,X)$ can be identified with more classical spaces. If $X$ is a Banach function space with finite cotype, e.g. an $L^p$ space for some $p \in [1,\infty),$ then for any step function $f = \sum_{k=1}^n x_k \chi_{A_k}: \Omega \to X,$ where $x_k \in X$ and the $A_k$ are measurable and disjoint with $\mu(A_k) \in (0,\infty),$ we have (cf. [@KaW2 Remark 3.6, Example 4.6])
$$\|f\|_{\gamma(\Omega,X)} = {\Bigl\Vert \sum_k \gamma_k \otimes \mu(A_k)^{\frac12} x_k \Bigr\Vert}_{\operatorname{Gauss}(X)} \!\!\!\!\!\!\!\!
\cong {\Bigl\Vert \left(\sum_k \mu(A_k) |x_k|^2 \right)^{\frac12}\Bigr\Vert}_X \label{Equ classical square function}
= {\Bigl\Vert \left( \int_\Omega |f(t)(\cdot)|^2 d\mu(t) \right)^{\frac12} \Bigr\Vert}_X.$$
The second equivalence follows from [@DiJT Theorem 16.18]. The last expression above is a classical square function (see e.g. [@CDMY Section 6]), whence for an arbitrary space $X,\,\|u\|_{\gamma(H,X)}$ is called (generalized) square function [@KaW2 Section 4]. In particular, if $X$ is a Hilbert space, then $\gamma(\Omega,X) = L^2(\Omega,X)$ with equal norms.
We have the following well-known properties of square functions.
\[Lem Folklore square functions\] Let $(\Omega,\mu)$ be a $\sigma$-finite measure space and $X$ a Banach space with property $(\alpha).$
1. Suppose that $f_n,f \in P_2(\Omega,X)$ and $f_n(t) \to f(t)$ for almost all $t \in \Omega.$ Then $\|f\|_{\gamma(\Omega,X)} \leq \liminf_n \|f_n\|_{\gamma(\Omega,X)}.$
2. Let $K \in B(H_2,H_1),$ where $H_1,H_2$ are separable Hilbert spaces. Then for $u \in \gamma(H_1,X)$ we have $u \circ K \in \gamma(H_2,X)$ and $\|u \circ K\|_{\gamma(H_2,X)} \leq \|u\|_{\gamma(H_1,X)} \|K\|.$
3. If $\Omega \to B(X),\: t \mapsto N(t)$ is a strongly continuous map such that $\tau = \{ N(t) :\: t \in \Omega \}$ is $\gamma$-bounded, and $f \in \gamma(\Omega,X),$ then $\|N \cdot f \|_{\gamma(\Omega,X)} \leq \gamma(\tau) \|f\|_{\gamma(\Omega,X)}.$
As $X$ has property $(\alpha),$ it does not contain $c_0$ isomorphically. Using this fact, a proof of (1) can be found in [@KaW2 Lemma 4.10], or in [@vN Proposition 3.18]. For (2), we refer to [@KaW2 Proposition 4.3] or [@vN Corollary 6.3]. Finally, (3) is proved in [@KaW2 Proposition 4.11], see also [@vN Theorem 5.2].
Sectorial operators and ${H^\infty}$ functional calculus {#sectorial-operators-and-hinfty-functional-calculus .unnumbered}
--------------------------------------------------------
Let $\theta \in (0,\pi)$ and $A :\: D(A) \subset X \to X$ a densely defined linear mapping on some Banach space $X.$ $A$ is called $\theta$-sectorial, if
1. The spectrum $\sigma(A)$ is contained in $\overline{\Sigma_\theta}.$
2. For all $\omega > \theta$ there is a $C_\omega > 0$ such that $\| \lambda (\lambda - A)^{-1} \| \leq C_\omega$ for all $\lambda \in \overline{\Sigma_\omega}^c.$
3. $R(A)$ is dense in $X.$
We call $A$ $0$-sectorial if it is $\theta$-sectorial for all $\theta > 0.$ In the literature, property (3) is sometimes omitted. It entails that $A$ is injective [@KuWe Proposition 15.2]. For such an operator $A$ and $f \in {H^\infty}_0(\Sigma_\omega),\: \omega \in (\theta , \pi),$ one defines the operator
$$\label{Equ Cauchy integral formula}
f(A) = \frac{1}{2\pi i} \int_{\partial \Sigma_{(\theta + \omega)/2}} f(\lambda) (\lambda - A)^{-1} d\lambda,$$
where $\partial\Sigma_{(\theta + \omega)/2}$ is the sector boundary which is parametrized as usual counterclockwise. It is easy to check that $f(A)$ is bounded and that $u : {H^\infty}_0(\Sigma_\omega) \to B(X)$ is a linear and multiplicative mapping. Suppose that there exists $C > 0$ such that $$\label{Equ Bounded HI calculus}
\| f(A) \| \leq C \|f\|_{\infty,\omega} \quad (f \in {H^\infty}_0(\Sigma_\omega)).$$ Then there exists an extension of $u$ to a bounded mapping ${H^\infty}(\Sigma_\omega) \to B(X),\,f \mapsto f(A),$ satisfying the so-called Convergence Lemma [@CDMY Lemma 2.1].
\[Lem Convergence Lemma\] Let $(f_n)_{n \in {\mathbb{N}}}$ be a sequence in ${H^\infty}(\Sigma_\omega)$ such that $\sup_{n \in {\mathbb{N}}} \| f_n \|_{\infty,\omega} < \infty$ and $f_n(\lambda) \to f(\lambda)$ for all $\lambda \in \Sigma_\omega$ and some $f$ (which then necessarily belongs to ${H^\infty}(\Sigma_\omega)$). Then $f(A)x = \lim_{n \to \infty} f_n(A) x$ for any $x \in X.$
Note that the extension is uniquely determined by Lemma \[Lem Convergence Lemma\] since for any $f \in {H^\infty}(\Sigma_\omega),$ $$\label{Equ Convergence lemma}
f_n(\lambda) = f(\lambda) \left( \lambda / (1 + \lambda)^2 \right)^{\frac1n}$$ is a sequence in ${H^\infty}_0(\Sigma_\omega)$ approximating $f$ in the sense of that lemma. As a consequence, if is satisfied, then it also holds for any $f \in {H^\infty}(\Sigma_\omega).$ In this case, we say that $A$ has a bounded ${H^\infty}(\Sigma_\omega)$ calculus, or without precising the angle $\omega \in (\theta,\pi),$ a bounded ${H^\infty}$ calculus.
Square function estimate and matricially bounded homomorphism {#Sec 3 Main Thm}
=============================================================
Throughout the section, we let $X$ be a Banach space. For any $n \in {\mathbb{N}},$ we identify $M_n \otimes B(X)$ with $B(\operatorname{Gauss}_n(X))$ by associating $[a_{ij}] \otimes T \in M_n \otimes B(X)$ with the operator $$\label{Equ M_n otimes B(X)}
\sum_{k=1}^n \gamma_k \otimes x_k \mapsto \sum_{k,j=1}^n \gamma_k \otimes a_{kj} T(x_j).$$ Via this identification, we get a norm on the tensor product space, which we note by $M_n \otimes_\gamma B(X).$
\[Def matricially gamma bounded\] Let $E$ be an operator space. Let further $u : E \to B(X)$ be a linear mapping. We call $u$ matricially $\gamma$-bounded, if $\operatorname{id}_{M_n} \otimes u : M_n\otimes E \to M_n \otimes_\gamma B(X)$ is bounded uniformly in $n \in {\mathbb{N}},$ i.e. if there is a constant $C > 0$ such that for any $n \in {\mathbb{N}},$ $${\bigl\Vert\sum_{i,j = 1}^n \gamma_i \otimes u(f_{ij}) x_j \bigr\Vert}_{\operatorname{Gauss}_n(X)} \leq C {\bigl\Vert \left[f_{i,j}\right]\bigr\Vert}_{M_n \otimes E} {\bigl\Vert\sum_{i=1}^n \gamma_i \otimes x_i\bigr\Vert}_{\operatorname{Gauss}_n(X)} .$$ We denote the least admissible constant $C$ by $\|u\|_{{{\text{mat-}\gamma}}}.$
\[Rem mat\]
1. If $u : E \to B(X)$ is matricially $\gamma$-bounded then $$\label{Equ u gamma bounded}
\{ u(x) : \: \|x\|_E \leq 1\}$$ is $\gamma$-bounded. Indeed, from the definition of $\gamma$-boundedness in Section \[Sec 2 Prelims\], we immediately deduce that is satisfied if and only if $\operatorname{id}_{M_n} \otimes u|_{D_n \otimes E}$ is bounded uniformly in $n \in {\mathbb{N}},$ where $D_n \subset M_n$ denotes the subspace of diagonal matrices. We call a linear mapping $u$ $\gamma$-bounded if holds.
2. Assume that $X$ is a Hilbert space. Then by , $u$ is matricially $\gamma$-bounded if and only if $u$ is completely bounded, and in this case, $\|u\|_{cb} = \|u\|_{{\text{mat-}\gamma}}.$
A first example for Definition \[Def matricially gamma bounded\] is given by
\[Prop sigma(m,X)\] For a given space $X$ and $m \in {\mathbb{N}},$ consider $$\sigma_{m,X} : M_m \to M_m \otimes_\gamma B(X),\,[a_{ij}] \mapsto [a_{ij}\operatorname{id}_X].$$ Then $\sigma_{m,X}$ is matricially $\gamma$-bounded with $\sup_{m \in {\mathbb{N}}} \|\sigma_{m,X}\|_{{\text{mat-}\gamma}}< \infty$ if and only if $X$ has property $(\alpha).$
It is shown in [@KrLM Lemma 4.3] that the $\sigma_{m,X}$ are $\gamma$-bounded uniformly in $m \in {\mathbb{N}}$ if and only if $X$ has property $(\alpha)$ (with Rademachers in place of Gaussians). Actually the same proof applies to Proposition \[Prop sigma(m,X)\].
Mappings which are $\gamma$-bounded or matricially $\gamma$-bounded have been studied so far in connection with functional calculi and unconditional decompositions [@KrLM; @dPR] where $E$ is a $C(K)$-space and representations of amenable groups [@LM10], where $E$ is a nuclear $C^*$-algebra. We shall focus in this section on the row Hilbert space $E = \ell^2_r.$
\[Thm Main 1\] Let $u : \ell^2 \to B(X)$ be a bounded linear mapping. Assume that $X$ has Pisier’s property $(\alpha).$ For $n \in {\mathbb{N}},$ denote by $C_n \subset M_n$ the subspace of matrices vanishing outside the first column. Then the following conditions are equivalent:
1. ${\bigl\Vertu(\cdot)x\bigr\Vert}_{\gamma(\ell^2,X)} \leq C \| x \|.$
2. $u : \ell^2_r \to B(X)$ is matricially $\gamma$-bounded.
3. The restriction $\operatorname{id}\otimes u: C_n \otimes \ell^2_r \to M_n \otimes_\gamma B(X)$ is bounded uniformly in $n \in {\mathbb{N}}.$
We fix an orthonormal basis $(e_k)_k$ of $\ell^2.$ Write $T_k = u(e_k).$ Then condition (1) of the statement rewrites $$\label{Equ 1 Thm}
{\bigl\Vert u(\cdot)x \bigr\Vert}_{\gamma} = \sup_{n \in {\mathbb{N}}} {\bigl\Vert \sum_{k=1}^n \gamma_k \otimes T_k x \bigr\Vert}_{\operatorname{Gauss}_n(X)} \leq C \| x \|.$$ On the other hand, for $[f_{ij}] \in M_n \otimes \ell^2,$ we have with $f^{(k)}_{ij} = {\langle f_{ij}, e_k \rangle},$ $${\bigl\Vert(\operatorname{id}_{M_n}\otimes u)[f_{ij}]\bigr\Vert}_{M_n \otimes_\gamma B(X)} = \lim_m {\bigl\Vert \left[ \sum_{k=1}^m f^{(k)}_{ij} T_k \right] \bigr\Vert}_{M_n \otimes_\gamma B(X)}.$$ Thus, condition (2) is equivalent to $$\label{Equ 2 Thm}
{\bigl\Vert \sum_{i=1}^n \sum_{k=1}^m \gamma_i \otimes f_{ij}^{(k)} T_k x_j \bigr\Vert}_{\operatorname{Gauss}_n(X)} \leq C \|[f_{ij}]\|_{M_n \otimes \ell^2_{m,r}} {\bigl\Vert \sum_{i=1}^n \gamma_i \otimes x_i \bigr\Vert}_{\operatorname{Gauss}_n(X)} ,$$ where $f_{ij} = (f_{ij}^{(k)})_k \in \ell^2_m,$ and $C$ is independent of $n$ and $m.$ Denote the linear bounded mapping $\operatorname{Gauss}_n(X) \to \operatorname{Gauss}_n(X)$ arising from by $u_{n,m}([f_{ij}]).$ Finally, condition (3) is equivalent to with $f_{ij} = 0$ for $j \geq 2.$\
(1) $\Longrightarrow$ (2)
For $m \in {\mathbb{N}}$ fixed, let $Y = \operatorname{Gauss}_m(X)$ and define the operators $$V : X \to Y, \, x\mapsto \sum_{k=1}^m \gamma_k \otimes T_k x,\quad W : Y \to X,\, \sum_{k=1}^m \gamma_k \otimes x_k \mapsto x_1.$$ By assumption , $V$ is bounded with constant $C$ independent of $m.$ Further, $W$ is bounded (see e.g. [@Kr (2.13)] for a simple proof). For $n \in {\mathbb{N}},$ denote $$V_n = \operatorname{id}_{\ell^2_n} \otimes V : \operatorname{Gauss}_n(X) \to \operatorname{Gauss}_n(Y),\, \sum_{k=1}^n \gamma_k \otimes x_k \mapsto \sum_{k=1}^n \gamma_k \otimes V(x_k).$$ It is easy to check that $\|V_n\| = \|V\|.$ Similarly, defining $W_n = \operatorname{id}_{\ell^2_n} \otimes W : \operatorname{Gauss}_n(Y) \to \operatorname{Gauss}_n(X),$ one has $\|W_n\| = \|W\|.$ Let $i_m : \ell^2_{m,r} \to M_m$ be the first row identification as in which is completely bounded of $cb$-norm 1. Then by Remark \[Rem mat\] and Proposition \[Prop sigma(m,X)\], along with property $(\alpha),$ $\pi_m = \sigma_{m,X} \circ i_m : \ell^2_{m,r} \to B(M_m \otimes X)$ is a matricially $\gamma$-bounded mapping and $\sup_{m \in {\mathbb{N}}} \|\pi_m\|_{{\text{mat-}\gamma}}< \infty.$ For $f = [f_{ij}]_{ij} \in M_n \otimes \ell^2_{m,r},$ we have the identity $$u_{n,m}(f) = \left[\sum_{k=1}^m f_{ij}^{(k)} T_k \right] = [W \pi_m(f_{ij}) V] = W_n [\pi_m(f_{ij})] V_n.$$ Therefore, $$\|u_{n,m}(f)\| \leq \|W_n\| \, \|V_n\| \, \| [\pi_m(f_{ij})] \| \leq \|W\| \, \|V\| \, \|\pi_m\|_{{\text{mat-}\gamma}}\|[f_{ij}]\|_{M_n \otimes \ell^2_{m,r}},$$ so follows.\
(2) $\Longrightarrow$ (3)
This is clear, since (3) is an obvious restriction of (2).\
(3) $\Longrightarrow$ (1)
Choose $n = m \in {\mathbb{N}}$ and $f = [f_{ij}] \in M_n \otimes \ell^2_{n,r}$ with $f_{ij} = \delta_{j1} e_i,$ where $(e_i)$ is the standard basis of $\ell^2_n.$ By definition of the row norm, we have $$\begin{aligned}
\|f\|_{M_n \otimes \ell^2_{n,r}} & = {\bigl\Vert \left[ \sum_k {\langle f_{ik}, f_{jk} \rangle} \right] \bigr\Vert}^{\frac12}_{M_n} \\
& = {\bigl\Vert \left[ \sum_k \delta_{k1} {\langle e_i, e_j \rangle} \right] \bigr\Vert}_{M_n}^{\frac12} \\
& = {\bigl\Vert [\delta_{ij}] \bigr\Vert}_{M_n}^{\frac12} \\
& = 1.\end{aligned}$$ As $f$ is supported by the first column, by assumption there is some $C < \infty$ such that $$\begin{aligned}
C & \geq \|u_{n,n}(f)\|_{M_n \otimes_\gamma B(X)} \\
& = {\bigl\Vert \left(
\begin{array}{cccc}
T_1 & 0 & \ldots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
T_n & 0 & \ldots & 0 \\
\end{array}
\right)\bigr\Vert}_{M_n \otimes_\gamma B(X)}
\\
& = \sup \left\{ {\bigl\Vert \sum_k \gamma_k \otimes T_k x_1 \bigr\Vert}_{\operatorname{Gauss}_n(X)} :\: {\bigl\Vert \sum_k \gamma_k \otimes x_k \bigr\Vert}_{\operatorname{Gauss}_n(X)} \leq 1 \right\} \\
& = \sup \left\{ {\bigl\Vert \sum_k \gamma_k \otimes T_k x \bigr\Vert}_{\operatorname{Gauss}_n(X)} :\: \|x\| \leq 1 \right\}.\end{aligned}$$ Letting $n \to \infty$ shows that holds.
Theorem \[Thm Main 1\] is a generalization of [@LM04 Proposition 3.3], where $X$ is an $L^p$-space, and [@HaKu Corollary 3.19], where more generally $X$ has property $(\alpha).$ There it is shown that condition (1) of the theorem implies that $u : \ell^2 \to B(X)$ is $\gamma$-bounded. (In these two references, $u$ maps to $B(Y,X)$ instead of $B(X).$ A corresponding version of Theorem \[Thm Main 1\] with $B(Y,X)$ in place of $B(X)$ holds with the same proof).
The Hörmander functional calculus {#Sec 4 Hormander operator space}
=================================
Recall the spaces ${W^\alpha}$ and ${{\mathcal{H}}^\alpha}$ and the dyadic partition of unity $(\phi_n)_{n \in {\mathbb{Z}}}$ from the introduction. Clearly the space ${W^\alpha}$ is a Hilbert space. We equip ${{\mathcal{H}}^\alpha}$ with an operator space structure by putting $$\label{Equ Ha OSS}
\|[f_{ij}]\|_{M_n \otimes {{\mathcal{H}}^\alpha}} = \sup_{k \in {\mathbb{Z}}} \| [\phi_k f_{ij}] \|_{M_n \otimes {W^\alpha}_r},$$ where the index $r$ refers to the row space structure. It is easy to check that indeed defines an operator space, arising from the embedding $${{\mathcal{H}}^\alpha}\hookrightarrow B(\bigoplus^2_{k \in {\mathbb{Z}}} {W^\alpha}),\, f \mapsto \left((g_k)_{k \in {\mathbb{Z}}} \mapsto ({\langle g_k, \phi_k f \rangle} e)_{k \in {\mathbb{Z}}} \right),$$ where $\displaystyle \bigoplus^2_{k \in {\mathbb{Z}}} {W^\alpha}$ is the Hilbert sum, and $e$ is some fixed element in ${W^\alpha}$ of norm 1.
In this section we focus on (unital) homomorphisms $$\label{Equ u} u : {{\mathcal{H}}^\alpha}\to B(X) .$$ We give a characterization when such mappings are matricially $\gamma$-bounded. An example of a bounded homomorphism of this type is given by Hörmander’s classical theorem mentioned in the introduction, which states that for $\alpha > \frac{d}{2},\: X = L^p({\mathbb{R}}^d)$ and $p \in (1,\infty),$ the radial Fourier multiplier representation $u_{-\Delta} : {{\mathcal{H}}^\alpha}\to B(X) $ given by $$\label{Equ u Delta}
u_{-\Delta}(f)g = \left[ f(|\cdot|^2) \hat{g} \right]\check{\phantom{i}} = f(-\Delta) g$$ is bounded. In fact, by means of our characterization, we will show in Section \[Sec 5 Examples\] that $u_{-\Delta}$ is even matricially $\gamma$-bounded provided that $\alpha > \frac{d+1}{2}.$
For $n \in {\mathbb{N}},$ let $\operatorname{M}^n$ be the space consisting of $n$-times continuously differentiable functions $f$ defined on $(0,\infty)$ such that $\|f\|_{\operatorname{M}^n} = \sum_{k=0}^n \sup_{t > 0} |t^k f^{(k)}(t)|$ is finite. Let us record how ${W^\alpha},\,{{\mathcal{H}}^\alpha},\,{H^\infty}(\Sigma_\omega)$ and the auxiliary space $\operatorname{M}^n$ compare. The proof is an easy verification, see also [@Kr Lemma 4.15, Proof of Proposition 4.22, Proposition 4.9].
\[Lem Function spaces\] Let $\omega \in (0,\pi)$ and $\alpha > \frac12.$
1. ${H^\infty}(\Sigma_\omega) \hookrightarrow {{\mathcal{H}}^\alpha}.$
2. ${W^\alpha}\hookrightarrow {{\mathcal{H}}^\alpha},$ where the embedding is completely bounded.
3. For any $\omega \in (0,\pi),$ ${H^\infty}(\Sigma_\omega) \cap {W^\alpha}$ is a dense subset of ${W^\alpha}.$
4. $\operatorname{M}^n \hookrightarrow {{\mathcal{H}}^\alpha}$ for $n > \alpha.$
5. ${H^\infty}(\Sigma_\omega)$ is a dense subset of $\operatorname{M}^n.$ Moreover, any $f \in {W^\alpha}\cap \operatorname{M}^n$ can be simultaneously approximated by a sequence $(f_k)_{k \in {\mathbb{N}}} \subset {H^\infty}(\Sigma_\omega) \cap {W^\alpha}\cap \operatorname{M}^n.$
The main interest of $\operatorname{M}^n$ is the following convergence lemma, which is proved in [@Kr Section 4.2.4].
\[Lem Convergence Mn\] Let $u : \operatorname{M}^n \to B(X)$ be bounded such that $u(f) = f(A)$ for some $0$-sectorial operator $A$ and any $f \in \bigcup_{\theta \in (0,\pi)} {H^\infty}(\Sigma_\theta).$ Let $(\phi_n)_{n \in {\mathbb{Z}}}$ be a dyadic partition of unity and $(a_n)_{n \in {\mathbb{Z}}}$ a bounded sequence. Then $\sum_{n \in {\mathbb{Z}}} a_n \phi_n$ belongs to $\operatorname{M}^n$ and $$\label{Equ Lem A u}\sum_{n \in {\mathbb{Z}}} a_n u(\phi_n) x = u \left( \sum_{n \in {\mathbb{Z}}} a_n \phi_n \right)x \quad (x \in X).$$
Many spectral multiplier theorems for Laplace type operators $A$ consist in the boundedness of $u$ in , which in turn is the functional calculus $u_A$ of some $0$-sectorial operator. For example, in the case of one has $A = - \Delta.$ In the sequel we will only consider homomorphisms of the form $u = u_A.$ The next lemma gives a criterion when this is the case.
\[Lem A u\] Let $\omega \in (0,\pi).$
1. Let $u$ be a bounded homomorphism $u : {{\mathcal{H}}^\alpha}\to B(X).$ There exists a $0$-sectorial operator $A$ such that $$\label{Equ Lem A u 2}
u(f) = f(A) \quad (f \in {H^\infty}(\Sigma_\omega)),$$ if and only if the restriction of $u$ to ${H^\infty}(\Sigma_\omega)$ satisfies the Convergence Lemma \[Lem Convergence Lemma\], i.e. for any $f_n,\: f \in {H^\infty}(\Sigma_\omega)$ with $\sup_n \|f_n\|_{\infty,\omega} < \infty$ and $f_n(\lambda) \to f(\lambda)$ pointwise, we have $u(f)x = \lim_n u(f_n)x$ for any $x \in X.$ In this case, we write $f(A)$ in place of $u(f)$ for any $f \in {{\mathcal{H}}^\alpha}.$
2. Let $A$ be a $0$-sectorial operator. Then there exists a bounded homomorphism $u : {{\mathcal{H}}^\alpha}\to B(X)$ satisfying if and only if $\|f(A)\| \leq C \|f\|_{{{\mathcal{H}}^\alpha}}\quad (f \in {H^\infty}(\Sigma_\omega)).$ Moreover, $u$ is uniquely determined by .
\(1) By Lemma \[Lem Function spaces\] (1), only the “if” part has to be shown. Suppose that $u : {{\mathcal{H}}^\alpha}\to B(X)$ satisfies the Convergence Lemma \[Lem Convergence Lemma\]. Since $u$ is a homomorphism, one has $u((\lambda - \cdot)^{-1}) - u((\mu - \cdot)^{-1}) = (\mu - \lambda) u((\lambda - \cdot)^{-1}) u((\mu - \cdot)^{-1})$ for any $\lambda,\mu \in {\mathbb{C}}\backslash [0,\infty),$ i.e. $u((\lambda - \cdot)^{-1})$ is a pseudo resolvent [@Paz Definition 9.1]. By Lemma \[Lem Convergence Lemma\], $u(\lambda (\lambda - \cdot)^{-1} )x \to x$ for any $x \in X$ and $|\lambda|\to \infty.$ Thus, by [@Paz Corollary 9.5], there exists a densely defined operator $A$ such that $u((\lambda - \cdot)^{-1}) = (\lambda-A)^{-1}$ for $\lambda \in {\mathbb{C}}\backslash [0,\infty).$ Again by Lemma \[Lem Convergence Lemma\], $-\frac1n (\frac1n + A)^{-1}x \to 0$ for any $x \in X$ and $n \to \infty.$ Thus, $A$ has dense range [@KuWe Proposition 15.2]. As $u$ is a bounded homomorphism, it now follows that $A$ is $0$-sectorial and for any rational function $r \in {{\mathcal{H}}^\alpha},$ $\|r(A)\| \lesssim \|r\|_{{{\mathcal{H}}^\alpha}} \lesssim \|r\|_{\infty,\omega}.$
We claim that $A$ has an ${H^\infty}$ calculus coinciding with $u.$ Indeed, a given $f \in {H^\infty}_0(\Sigma_\omega),$ we write $$f = \frac{1}{2\pi i} \int_{\partial\Sigma_{\omega/2}} f(\lambda) (\lambda - \cdot)^{-1} d\lambda.$$ As $f(\lambda) (\lambda - \cdot)^{-1} : \partial\Sigma_{\omega/2} \to {H^\infty}(\Sigma_{\omega/4})$ is continuous, we find a sequence $r_n = \sum_{k=1}^K c_k f(\lambda_k) (\lambda_k-\cdot)^{-1}$ such that $r_n \to f$ in ${H^\infty}(\Sigma_{\omega/4}),$ so in particular in ${{\mathcal{H}}^\alpha}.$ Clearly, $r_n$ are rational functions. Inserting formally $(\cdot) = A$ in the Cauchy integral, the same arguments apply, and $r_n(A) \to f(A).$ We conclude $u(f) = \lim_n u(r_n) = \lim_n r_n(A) = f(A).$
We have shown that $u(f) = f(A)$ for any $f \in {H^\infty}_0(\Sigma_\omega).$ For a general $f \in {H^\infty}(\Sigma_\omega)$ we use the approximation .\
(2) The “only if” part is clear and the “if” part is shown in [@Kr Remark 4.27]. Using density and Lemma \[Lem Function spaces\], $u$ is uniquely determined on ${W^\alpha}$ and $\operatorname{M}^n$ for any $n > \alpha.$ Thus $u$ satisfies the decomposition . Then for $f \in {{\mathcal{H}}^\alpha},$ we have $u(f)x = u(f) \sum_{k \in {\mathbb{Z}}} u(\phi_k) x = \sum_{k \in {\mathbb{Z}}} u(f \phi_k) x.$ As $f \phi_k \in {W^\alpha},$ we conclude the uniqueness of $u.$
The strategy to prove matricial $\gamma$-boundedness of a mapping from ${{\mathcal{H}}^\alpha}$ to $B(X)$ will be to show the matricial $\gamma$-boundedness from ${W^\alpha}$ to $B(X),$ and then to pass to ${{\mathcal{H}}^\alpha}$ by means of a spectral decomposition, given by in the following theorem. The restriction of the ${H^\infty}$ calculus angle $\omega$ to $(0,\pi/4)$ is only for technical reasons.
\[Thm Paley Littlewood\] Let $X$ be a Banach space with property $(\alpha).$ Let $\alpha > \frac12,\,\omega \in (0,\pi/4)$ and $A$ be a $0$-sectorial operator on $X$ having a bounded ${H^\infty}(\Sigma_\omega)$ calculus. Assume that $$\|f(A)\| \leq C \|f\|_{{W^\alpha}} \quad (f \in {H^\infty}(\Sigma_\omega) \cap {W^\alpha})$$ and that the extension resulting from density $u : {W^\alpha}\to B(X),\, f \mapsto f(A)$ is matricially $\gamma$-bounded. Let $(\phi_n)_{n \in {\mathbb{Z}}}$ be a dyadic partition of unity. Then $$\label{Equ Paley Littlewood}
\|x\| \cong \| \sum_{n \in {\mathbb{Z}}} \gamma_n \otimes \phi_n(A)x \|_{\operatorname{Gauss}(X)},$$ the sum on the right hand side being convergent in $\operatorname{Gauss}(X).$
As the proof is rather long we separate four preliminary lemmas, whose proofs are annexed in Section \[Sec 6 Proofs Lemmas\].
\[Lem 1\] Let $X$ have property $(\alpha)$, let $\alpha > \frac12.$ Let $A$ be as in Theorem \[Thm Paley Littlewood\]. Then for $\beta \in {\mathbb{N}},\,\beta > \alpha,$ $$\label{Equ NT}
\left\{\exp(-2^k z A) :\: k \in {\mathbb{Z}}\right\} \text{ is }\gamma\text{-bounded with constant }\lesssim \left| \frac{z}{\operatorname{Re}z}\right|^\beta \quad (\operatorname{Re}z > 0).$$
\[Lem 2\] Let $A$ be a $0$-sectorial operator on some Banach space $X$ such that for some $\beta > 0,$ holds. Then for $\gamma = \beta + 1,$ we have $$\label{Equ Res}
\left\{ \lambda^{\frac12} (2^k A)^{\frac12}(\lambda - 2^k A)^{-1} :\: k \in {\mathbb{Z}}\right\}\text{ is }\gamma\text{-bounded with constant }\lesssim
|\arg \lambda|^{-\gamma} \quad (\operatorname{Re}\lambda > 0).$$
\[Lem 3\] Let $A$ be a $0$-sectorial operator on some space $X$ with property $(\alpha)$ having a bounded ${H^\infty}(\Sigma_\omega)$ calculus. Assume that $A$ satisfies for some $\gamma > 0.$ Then for any $n > \gamma,$ $$\|f(A)\| \leq C \|f\|_{\operatorname{M}^n} \quad \left(f \in {H^\infty}(\Sigma_\omega)\right).$$
\[Lem 4\] Let $n \in {\mathbb{N}}.$ Let $(g_k)_{k \in {\mathbb{Z}}}$ satisfy $\sup_{k \in {\mathbb{Z}}} \|g_k\|_{\operatorname{M}^n} < \infty.$ Suppose that the supports of $g_k$ satisfy the following overlapping condition $$\sup_{x > 0}\#\{ k \in {\mathbb{Z}}:\: \operatorname{supp}g_k \cap [\frac12 x,2x] \not= \emptyset \} < \infty.$$ Then $\sum_{k \in {\mathbb{Z}}} g_k,$ which is consequently pointwise a finite sum belongs to $\operatorname{M}^n,$ and $$\|\sum_{k \in {\mathbb{Z}}} g_k\|_{\operatorname{M}^n} \lesssim \sup_{k\in{\mathbb{Z}}} \|g_k\|_{\operatorname{M}^n} < \infty.$$
Using Lemmas \[Lem 1\], \[Lem 2\] and \[Lem 3\] one after another shows that $\|f(A)\| \lesssim \|f\|_{\operatorname{M}^n}$ for $n$ sufficiently large $(n > \lfloor \alpha \rfloor + 2).$ For any $k \in {\mathbb{Z}},$ let $a_k \in \{ 1 , - 1\}.$ Apply Lemma \[Lem 4\] with $g_k = a_k \phi_k.$ It is easy to check that $\|g_k\|_{\operatorname{M}^n}$ is independent of $k \in {\mathbb{Z}}.$ Further, the overlapping condition is clearly satisfied with constant 2. Thus we have, for any finite $F \subset {\mathbb{Z}},$ $$\| \sum_{k \in F} a_k \phi_k(A) x\| \lesssim \|\sum_{k \in F} a_k \phi_k \|_{\operatorname{M}^n} \|x\| \lesssim \|x\|.$$ Replacing $a_k$ by independent Rademacher variables $\epsilon_k$ and taking expectation gives $$\|\sum_{k \in F} \epsilon_k \otimes \phi_k(A) x\|_{\operatorname{Rad}(X)} \lesssim \|x\|.$$ Since $X$ has property $(\alpha),$ the equivalence of Gaussian and Rademacher sums holds. By , $\sum_{k \in {\mathbb{Z}}} a_k \phi_k(A)x$ converges in $X.$ By dominated convergence (resp. ), convergence holds also in $\operatorname{Rad}(X)$ (resp. $\operatorname{Gauss}(X)$), when $a_k$ is replaced by $\epsilon_k$ (resp. $\gamma_k$).
We have shown $$\| \sum_{k \in {\mathbb{Z}}} \gamma_k \otimes \phi_k(A) x \|_{\operatorname{Gauss}(X)} \lesssim \|x\|.$$
For the reverse inequality, we argue by duality. Let $x' \in X',$ write ${\mathbb{E}}$ for expectation and $\widetilde\phi_l = \sum_{k=l-1}^{l+1}\phi_k.$ By the support condition on the $\phi_k,$ $\widetilde\phi_l \phi_l = \phi_l.$ Then using the independence of the $\gamma_k,$ we have $$\begin{aligned}
\left|{\langle x, x' \rangle}\right| & = \left|{\mathbb{E}}{\langle \sum_{k \in {\mathbb{Z}}} \gamma_k \phi_k(A)x, \sum_{l \in {\mathbb{Z}}} \gamma_l \widetilde\phi_l(A)'x' \rangle}\right| \\
& \leq \|\sum_{k \in {\mathbb{Z}}} \gamma_k \otimes \phi_k(A) x\|_{\operatorname{Gauss}(X)} \| \sum_{l \in {\mathbb{Z}}} \gamma_l \otimes \widetilde\phi_l(A)'x'\|_{\operatorname{Gauss}(X')}.\end{aligned}$$ We conclude the proof by the same argument as above which shows that $$\| \sum_{l \in {\mathbb{Z}}} \gamma_l \otimes \widetilde\phi_l(A)' x'\|_{\operatorname{Gauss}(X')} \lesssim \|x'\|.$$
The main result of this section reads as follows.
\[Thm Ha\] Let $A$ be a $0$-sectorial operator with bounded ${H^\infty}(\Sigma_\theta)$ calculus for some $\theta \in (0,\pi)$ on a space $X$ with property $(\alpha).$ Let $\alpha > \frac12.$ Then the following are equivalent.
1. For any $x \in X,$ $(1+t^2)^{-\alpha/2} A^{it}x$ belongs to $\gamma({\mathbb{R}},X)$ and $$\| (1 + t^2)^{-\alpha/2} A^{it} x\|_{\gamma({\mathbb{R}},X)} \leq C \|x\|.$$
2. The ${H^\infty}$ calculus of $A$ extends to a matricially $\gamma$-bounded mapping $u : {{\mathcal{H}}^\alpha}\to B(X).$
Assume first that the theorem is shown under the additional assumption that $\theta < \frac{\pi}{4}.$ For a general $\theta \in (0,\pi),$ we can reduce to this case by considering $B = A^{\frac14}.$ Namely, by [@Haa Theorem 2.4.2], $B$ has a bounded ${H^\infty}(\Sigma_\omega)$ calculus for some $\omega \leq \frac{\theta}{4} < \frac{\pi}{4}.$ Moreover, $$(1 + t^2)^{-\alpha/2} B^{it} x = \left( \frac{1 + (\frac{t}{4})^2}{1 + t^2}\right)^{\alpha/2} \cdot \left( 1 + \frac{t}{4} \right)^{-\frac{\alpha}{2}} A^{i\frac{t}{4}} x.$$ The first factor is bounded, so its multiplication with an $L^2({\mathbb{R}})$ function is a bounded operation on $L^2({\mathbb{R}}).$ The same holds for its inverse, and also for the change of variables $f \mapsto f(\frac{\cdot}{4}),$ and its inverse. Thus, by Lemma \[Lem Folklore square functions\], if $A$ satisfies (1) then so does $B,$ so $B$ satisfies (2). As ${H^\infty}(\Sigma_\omega) \hookrightarrow {{\mathcal{H}}^\alpha},$ $B$ has an ${H^\infty}(\Sigma_\omega)$ calculus actually for any $\omega >0,$ so by [@Haa Theorem 2.4.2], $A$ has an ${H^\infty}(\Sigma_\theta)$ calculus for some $\theta < \frac{\pi}{4}.$ The same is true, provided that $A$ satisfies (2). Thus (1) or (2) imply that the assumption of the theorem actually holds with $\theta < \frac{\pi}{4}.$ We suppose from now on that $\theta < \frac{\pi}{4}.$\
$(1) \Longrightarrow (2).$\
Fix an arbitrary orthonormal basis $(f_k)_k$ of $L^2({\mathbb{R}}).$ Let $T_k \in B(X)$ be defined by $${\langle T_k x, x' \rangle} = \frac{1}{2 \pi} \int_{\mathbb{R}}f_k(t) (1 + t^2)^{-\alpha/2} {\langle A^{it} x, x' \rangle} dt$$ and $v : \ell^2 \to B(X)$ the linear mapping given by $e_k \mapsto T_k.$ Then $(1)$ implies $$\|v(\cdot)x\|_{\gamma(\ell^2,X)} = \| \sum_k \gamma_k \otimes v(e_k)x \|_{\operatorname{Gauss}(X)} = \| \sum_k \gamma_k \otimes T_k x \|_{\operatorname{Gauss}(X)} = \|(1 + t^2)^{-\frac{\alpha}{2}} A^{it}x\|_\gamma \lesssim \|x\|,$$ so that by Theorem \[Thm Main 1\], $v : \ell^2_r \to B(X)$ is matricially $\gamma$-bounded. Consider the mapping $w : {W^\alpha}\to \ell^2,\, f \mapsto \left( \frac{1}{2 \pi} \int_{\mathbb{R}}(f\circ \exp)\hat{\phantom{i}}(t) (1 + t^2)^{\alpha/2} \overline{f_k}(t) dt \right)_k.$ It is easy to check that $w$ is unitary and consequently, $\tilde{u} = v \circ w : {W^\alpha}_r \to B(X)$ is also matricially $\gamma$-bounded. On the other hand, $\tilde{u}(f) = f(A)$ for any $f \in {H^\infty}(\Sigma_\theta) \cap {W^\alpha}.$ Indeed, by the representation formula in [@Kr Proposition 4.22], $$\begin{aligned}
2 \pi {\langle f(A)x, x' \rangle} & = \int_{\mathbb{R}}(f \circ \exp)\hat{\phantom{i}}(t) {\langle A^{it}x, x' \rangle} dt\\
& = \int_{\mathbb{R}}(f \circ \exp)\hat{\phantom{i}}(t) (1 + t^2)^{\alpha/2} (1 + t^2)^{-\alpha/2} {\langle A^{it}x, x' \rangle} dt\\
& = \sum_{k} \int_{\mathbb{R}}(f\circ\exp)\hat{\phantom{i}}(t) (1 + t^2)^{\alpha/2} \overline{f_k}(t) dt {\langle T_k x, x' \rangle}\\
& = 2 \pi {\langle \tilde{u}(f)x, x' \rangle}.\end{aligned}$$
Let $n \in {\mathbb{N}}$ and $F = [f_{ij}] \in M_n \otimes {H^\infty}(\Sigma_\theta).$ We show that $\| [ f_{ij}(A) ] \|_{M_n \otimes_\gamma B(X)} \lesssim \|[f_{ij}]\|_{M_n \otimes {{\mathcal{H}}^\alpha}}.$ For $N \in {\mathbb{N}}$ consider $$F_N = \left[
\begin{array}{cccc}
[\phi_{-N} f_{ij}] & 0 & \ldots & 0 \\
0 & [\phi_{-N + 1} f_{ij}] & & \vdots \\
\vdots & & \ddots & 0 \\
0 & \ldots & 0 & [\phi_N f_{ij}] \\
\end{array}
\right] \in M_{(2N + 1)n} \otimes {W^\alpha}.$$ By , we have $\sup_N \|F_N\|_{M_{(2N+1)n} \otimes {W^\alpha}} = \|F\|_{M_n \otimes {{\mathcal{H}}^\alpha}}.$ Observe first that for any scalars $g_1,\ldots,g_n,$ by Theorem \[Thm Paley Littlewood\], with $\tilde\phi_k = \sum_{l = k - 1}^{k + 1} \phi_l,$ $$\| \sum_{i,j = 1}^n g_i f_{ij}(A)x_j \| \cong \| \sum_{i,j = 1}^n \sum_{k \in {\mathbb{Z}}} \gamma_k \otimes g_i \tilde{u}(f_{ij} \phi_k) x_j \|
\cong \| \sum_{i,j = 1}^n \sum_{k \in {\mathbb{Z}}} \gamma_k \otimes g_i \tilde{u}(f_{ij} \phi_k) \tilde{u}(\widetilde\phi_k) x_j \|.$$ Replacing $g_i$ by Gaussian variables and taking expectations shows that $$\label{Equ Proof Main Thm 1}
\|\sum_{i,j=1}^n \sum_{k \in {\mathbb{Z}}} \gamma_i \otimes \tilde{u}(f_{ij}) x_j\|_{\operatorname{Gauss}_n(X)} \cong \| \sum_{i,j=1}^n \sum_{k \in {\mathbb{Z}}} \gamma_i \otimes \gamma_k \otimes
\tilde{u}(f_{ij}\phi_k) \tilde{u}( \widetilde\phi_k) x_j\|_{\operatorname{Gauss}_n(\operatorname{Gauss}(X))}.$$ Further we have $$\begin{aligned}
& \| \sum_{i,j=1}^n \sum_{k=-N}^N \gamma_i \otimes \gamma_k \otimes \tilde{u}(f_{ij} \phi_k) \tilde{u}(\widetilde\phi_k) x_j \|_{\operatorname{Gauss}_n(\operatorname{Gauss}(X))} \nonumber \\
& \cong \| \sum_{i,j,k} \gamma_{ik} \otimes \tilde{u}(f_{ij} \phi_k) \tilde{u}(\widetilde\phi_k) x_j\|_{\operatorname{Gauss}(X)} \nonumber \\
& \lesssim \| F_N \|_{M_{(2N+1)n} \otimes {W^\alpha}} \| \sum_{i,k} \gamma_{i,k} \otimes \tilde{u}(\widetilde\phi_k) x_i \|_{\operatorname{Gauss}(X)} \nonumber \\
& \lesssim \|F\|_{M_n \otimes {{\mathcal{H}}^\alpha}} \|\sum_i \gamma_i \otimes x_i \|_{\operatorname{Gauss}_n(X)}. \label{Equ Proof Main Thm 2}\end{aligned}$$ Finally taking the supremum over $N \in {\mathbb{N}},$ and give $$\| [f_{ij}(A)] \|_{M_n \otimes_\gamma B(X)} \lesssim \| [f_{ij}] \|_{M_n \otimes {{\mathcal{H}}^\alpha}} .$$ In particular, $\|f(A)\| \lesssim \|f\|_{{{\mathcal{H}}^\alpha}},$ so that by Lemma \[Lem A u\], there exists a bounded mapping $u : {{\mathcal{H}}^\alpha}\to B(X)$ extending the ${H^\infty}$ calculus in the sense of . Now repeat the above argument with an arbitrary $F = [f_{ij}] \in M_n \otimes {{\mathcal{H}}^\alpha}$ and $u(f_{ij})$ in place of $f_{ij}(A),$ to deduce that $u$ is matricially $\gamma$-bounded.\
$(2) \Longrightarrow (1).$ Denote $\tilde{u}$ the restriction of $u$ to ${W^\alpha}$ which by Lemma \[Lem Function spaces\] is again matricially $\gamma$-bounded. Thus also the mapping $v= \tilde{u} \circ w^{-1}$ from the first part of the proof is matricially $\gamma$-bounded and by Theorem \[Thm Main 1\], $\| (1 + t^2)^{-\alpha/2} A^{it} x \|_{\gamma({\mathbb{R}},X)} \leq C \|x\|.$
Extensions and Applications {#Sec 5 Examples}
===========================
We have characterized in Theorem \[Thm Ha\] the matricially $\gamma$-bounded Hörmander calculus in terms of square functions of $A.$ In fact, the imaginary powers $A^{is}$ appearing in these square functions can be replaced by several other typical operator families associated with $A,$ such as resolvents $R(\lambda,A)$ for $\lambda \in {\mathbb{C}}\backslash [0,\infty)$ and the semigroup generated by $-A,$ $T(z) = \exp(-zA)$ for $\operatorname{Re}z > 0.$ This gives (almost) equivalent conditions, see Proposition \[Prop BIP Sgr Res\] below. Subsequently, we use the semigroup condition of this proposition to apply Theorem \[Thm Ha\] to some examples. The starting point for us will be semigroups that satisfy (generalized) Gaussian estimates (see ).
The following lemma serves as a preparation for Proposition \[Prop BIP Sgr Res\].
\[Lem square function integral transform\] For $i = 1,2,$ let $(\Omega_i,\mu_i)$ be $\sigma$-finite measure spaces and $K \in B(L^2(\Omega_1),L^2(\Omega_2)).$
1. Assume that $f \in \gamma(\Omega_1,X)$ and that there exists a Bochner-measurable $g : \Omega_2 \to X$ such that $${\langle g(\cdot), x' \rangle} = K({\langle f(\cdot), x' \rangle}) \quad (x' \in X').$$ Then $g \in \gamma(\Omega_2,X)$ and $$\|g\|_{\gamma(\Omega_2,X)} \leq \|K\|\,\|f\|_{\gamma(\Omega_1,X)}.$$
2. Let $\Omega_1 \to B(X),\,t \mapsto N(t)$ and $\Omega_2 \to B(X),\,t \mapsto M(t)$ be weakly measurable. Assume that $\|N(\cdot)x\|_\gamma \leq C \|x\|$ and that there is $K \in B(L^2(\Omega_1),L^2(\Omega_2))$ such that $K \left[ {\langle N(\cdot)x, x' \rangle} \right] = {\langle M(\cdot)x, x' \rangle}$ for $x \in D,$ where $D$ is some dense subset of $X.$ Then $M(\cdot)x \in \gamma(\Omega_2,X)$ for any $x \in X$ and $\|M(\cdot)x\|_\gamma \lesssim \|N(\cdot)x\|_\gamma.$
3. Let $(\Omega,\mu)$ be a measure space and $g :\Omega \to X$ measurable. For $n \in {\mathbb{N}},$ let $\varphi_n:\Omega \to [0,1]$ measurable with $\sum_{n = 1}^\infty \varphi_n(t) = 1$ for all $t \in \Omega.$ Then $$\|g\|_{\gamma(\Omega,X)} \leq \sum_{n = 1}^\infty \| \varphi_n g \|_{\gamma(\Omega,X)}.$$
\(1) By assumption, $g \in P_2(\Omega_2,X).$ Consider the associated operator $u_g : H \to X$ as in . We have $u_g = u_f \circ K'.$ Thus, by Lemma \[Lem Folklore square functions\], $\|u_g\|_{\gamma(L^2(\Omega_2),X)} \leq \|K'\|\,\|u_f\|_{\gamma(L^2(\Omega_1),X)},$ which proves (1).\
(2) We show first that $M(\cdot)x$ belongs to $P_2(\Omega_2,X)$ for any $x \in X.$ For $x \in D,$ this follows immediately from the assumption. For $x \in X,$ we let $x_n \in D$ such that $x_n \to x$ $(n \to \infty).$ Then ${\langle M(t)x, x' \rangle} = \lim_n {\langle M(t)x_n, x' \rangle}$ for any $t \in \Omega_2.$ On the other hand, ${\langle M(\cdot)x_n, x' \rangle}$ is convergent in $L^2(\Omega_2).$ Indeed, $$\begin{aligned}
\| {\langle M(\cdot)(x_n - x_m), x' \rangle} \|_{L^2(\Omega_2)} & = \|K\left[ {\langle N(\cdot)(x_n - x_m), x' \rangle} \right] \|_{L^2(\Omega_1)} \\
& \lesssim \| N(\cdot)(x_n - x_m)\|_\gamma \, \|x'\| \lesssim \|x_n - x_m\| \, \|x'\|,\end{aligned}$$ which converges to $0\: (n,m \to \infty).$ Thus, ${\langle M(\cdot)x, x' \rangle}$ the pointwise limit, so necessarily equal to the $L^2$ limit, belongs to $L^2(\Omega_2).$ Consequently, by (1) and Lemma \[Lem Folklore square functions\], $\|M(\cdot)x\|_\gamma = \lim_n \|M(\cdot)x_n\|_\gamma \leq \|K\| \lim_n \|N(\cdot)x_n\|_\gamma = \|K\| \,\|N(\cdot)x\|_\gamma,$ which shows (2).\
(3) For $n \in {\mathbb{N}},$ put $\phi_n = \sum_{k=1}^n \varphi_k.$ Then $\phi_n: \Omega \to [0,1]$ and $\phi_n(t) \to 1$ monotonically for all $t \in \Omega.$ Then $\sup_n \|\phi_n g\|_\gamma \leq \sup_n \sum_{k=1}^n \| \varphi_k g\|_\gamma = \sum_{k=1}^\infty \|\varphi_k g \|_\gamma.$ It remains to show $\|g\|_\gamma \leq \sup_n \| \phi_n g\|_\gamma.$ Let us show first that $g \in P_2(\Omega,X),$ i.e. for any $x' \in X',\, {\langle g(\cdot), x' \rangle} \in L^2(\Omega).$ By assumption, we have $|{\langle g(t), x' \rangle}| = \lim_n \phi_n(t) |{\langle g(t), x' \rangle}|$ for any $t \in \Omega,$ and this convergence is monotone. Then by Beppo Levi’s theorem, $$\|{\langle g(\cdot), x' \rangle}\|_{L^2(\Omega)} = \lim_n \|{\langle \phi_n(\cdot) g(\cdot), x' \rangle}\|_{L^2(\Omega)}
\leq \limsup_n \|\phi_n \cdot g\|_{\gamma(\Omega,X)} \|x'\|.$$ where we have used that $\|{\langle f(\cdot), x' \rangle}\|_{L^2(\Omega)} \leq \|f\|_{\gamma(\Omega,X)} \, \|x'\|$ for any $f \in \gamma(\Omega,X).$ Thus we have shown that $g \in P_2(\Omega,X).$ Then by Lemma \[Lem Folklore square functions\], $$\|g\|_{\gamma(\Omega,X)} \leq \liminf_n \|\phi_n \cdot g\|_{\gamma(\Omega,X)} \leq \sup_n \|\phi_n \cdot g\|_{\gamma(\Omega,X)}.$$
\[Prop BIP Sgr Res\] Let $A$ be a $0$-sectorial operator having a bounded ${H^\infty}$ calculus on some space $X$ with property $(\alpha).$ Let $\alpha > \frac12.$ Consider the following conditions. $$\begin{array}{ll}
& \textit{H\"ormander functional calculus}\\
(1) & \text{The }{H^\infty}\text{ calculus of }A\text{ extends to a matricially }\gamma\text{-bounded mapping }{{\mathcal{H}}^\alpha}\to B(X).\vspace{0.2cm}\\
& \textit{Imaginary powers}\\
(2) &\| (1 + t^2)^{-\alpha/2} A^{it} x \|_{\gamma({\mathbb{R}},X)} \leq C \|x\|.\vspace{0.2cm}\\
& \textit{Resolvents}\\
(3)&\text{For some }\beta \in (0,1)\text{ and }\theta \in (-\pi,\pi) \backslash \{ 0 \} :\: \| t^\beta A^{1 - \beta} R(e^{i\theta} t,A) x \|_{\gamma({\mathbb{R}}_+,dt/t,X)} \lesssim |\theta|^{-\alpha} \|x\|.\\
(4)&\text{For some }\beta \in (0,1),\:\theta_0 \in (0,\pi] :\: \| \, \|\theta|^{\alpha - \frac12} t^\beta A^{1 - \beta}R(e^{i \theta} t,A)x \|_{\gamma({\mathbb{R}}_+ \times [-\theta_0,\theta_0], dt/t d\theta,X)} \lesssim \|x\|.\vspace{0.2cm}\\
& \textit{Analytic semigroup}\\
(5)&\text{For }\theta \in (-\frac{\pi}{2},\frac{\pi}{2}),\:\|A^{1/2}T(e^{i\theta}t)x\|_{\gamma({\mathbb{R}}_+,dt,X)} \lesssim (\frac{\pi}{2}-|\theta|)^{-\alpha} \|x\|.\\
(6)& \|( 1+ \left| \frac{a}{b} \right|^2)^{\alpha/2} |a|^{-\frac12} A^{1/2} T(a+ib)x\|_{\gamma({\mathbb{R}}_+\times {\mathbb{R}},da db,X)} \lesssim \|x\|.
\end{array}$$ Then the conditions (1), (2), (4), (6) are equivalent. Further these conditions imply the remaining ones (3), (5), which conversely imply that the ${H^\infty}$ calculus of $A$ extends to a matricially $\gamma$-bounded homomorphism ${{\mathcal{H}}^{\alpha + \epsilon}}\to B(X)$ for any $\epsilon > 0.$
$(1) \Longleftrightarrow (2).$\
This is Theorem \[Thm Ha\].\
$(2) \Longleftrightarrow (4).$\
Consider $$\label{Equ Proof Charact K}
K : L^2({\mathbb{R}},ds) \to L^2({\mathbb{R}}\times (-\pi,\pi),ds d\theta),\,
f(s) \mapsto (\pi-|\theta|)^{\alpha-\frac12}\frac{1}{\sin \pi(\beta + is)} e^{\theta s} \langle s \rangle^{\alpha} f(s),$$ where we write in short $$\langle s \rangle = (1 + |s|^2)^{\frac12}.$$ Note that $|\sin \pi(\beta + is)| \cong \cosh(\pi s)$ for $\beta \in (0,1)$ fixed. $K$ is an isomorphic embedding. Indeed, $$\|Kf\|_2^2 =
\int_{\mathbb{R}}\int_{-\pi}^\pi \left((\pi-|\theta|)^{\alpha-\frac12} e^{\theta s}\right)^2 d\theta \frac{1}{|\sin^2(\pi(\beta + i s))|} \langle s \rangle^{2\alpha} |f(s)|^2 ds$$ and $$\begin{aligned}
\int_{-\pi}^\pi (\pi-|\theta|)^{2\alpha -1} e^{2\theta s} d\theta
& \cong \int_0^\pi \theta^{2\alpha -1} e^{2(\pi - \theta) |s|} d\theta
\nonumber \\
& \cong \cosh^2(\pi s) \int_0^\pi \theta^{2\alpha-1} e^{2\theta |s|} d\theta.
\nonumber\end{aligned}$$ The last integral is bounded from below uniformly in $s \in {\mathbb{R}},$ and for $|s|\geq 1,$ $$\int_0^\pi \theta^{2\alpha -1} e^{2\theta |s|} d\theta = (2|s|)^{-2\alpha} \int_0^{2|s|\pi} \theta^{2\alpha -1} e^{\theta}d\theta \cong |s|^{-2\alpha}.$$ This clearly implies that $\|Kf\|_2 \cong \|f\|_2.$ Applying Lemma \[Lem square function integral transform\], we get $$\| \langle s \rangle^{-\alpha} A^{is}x \|_{\gamma({\mathbb{R}},ds,X)}
\cong \left\|(\pi-|\theta|)^{\alpha-\frac12} \frac{1}{\cosh (\pi s)} e^{\theta s}A^{is}x \right\|_{\gamma({\mathbb{R}}\times (-\pi,\pi),ds d\theta,X)}.$$ In [@KuWe p. 228 and Theorem 15.18], the following formula is derived for $x \in A(D(A^2))$ and $|\theta| < \pi:$ $$\label{Equ Res A Mellin BIP}
\frac{\pi}{\sin \pi (\beta + i s)} e^{\theta s} A^{is} x =
\int_0^\infty t^{is}\left[t^{\beta} e^{i\theta \beta} A^{1 - \beta}(e^{i\theta}t+A)^{-1}x\right] \frac{dt}{t}.$$ Note that $A(D(A^2))$ is a dense subset of $X.$ As the Mellin transform $f(s) \mapsto \int_0^\infty t^{is} f(s) \frac{ds}{s}$ is an isometry $L^2({\mathbb{R}}_+,\frac{ds}s)\to L^2({\mathbb{R}},dt),$ we get by Lemma \[Lem square function integral transform\] (2) $$\begin{aligned}
\| \langle s \rangle^{-\alpha} A^{is} x\|_{\gamma({\mathbb{R}},X)}
& \cong \|(\pi -|\theta|)^{\alpha-\frac12} t^{\beta}A^{1-\beta}(e^{i\theta}t+A)^{-1} x\|_{\gamma({\mathbb{R}}_+\times(-\pi,\pi),\frac{dt}{t}d\theta,X)}
\nonumber \\
& \cong \| |\theta|^{\alpha-\frac12} t^\beta A^{1 - \beta} R(e^{i\theta}t,A) x\|_{\gamma({\mathbb{R}}_+ \times (0,2\pi),dt/t d\theta,X)}.
\nonumber\end{aligned}$$
so that (2) $\Leftrightarrow$ (4) for $\theta_0 = \pi.$
For a general $\theta_0 \in (0,\pi],$ consider $K$ from with restricted image, i.e. $$K : L^2({\mathbb{R}},ds) \to L^2({\mathbb{R}}\times (-\pi,-(\pi-\theta_0)] \cup [\pi-\theta_0,\pi),ds d\theta).$$ Then argue as in the case $\theta_0 = \pi.$\
(4) $\Longleftrightarrow$ (6).\
The proof of (2) $\Leftrightarrow$ (4) above shows that condition (4) is independent of $\theta_0 \in (0,\pi]$ and $\beta \in (0,1).$ Put $\theta_0 = \pi$ and $\beta = \frac12.$ The equivalence follows again from Lemma \[Lem square function integral transform\], using the fact that for $\theta \in (-\frac{\pi}{2},\frac{\pi}{2})$ and $\mu > 0,$ $$\label{Equ Proof BIP Sgr Res}
(e^{i\theta} \mu + it)^{-1} = K[\exp(-(\cdot)e^{i\theta} \mu)\chi_{(0,\infty)}(\cdot)](t),$$ where $K : L^2({\mathbb{R}},ds) \to L^2({\mathbb{R}},dt)$ is the Fourier transform.\
(3) $\Longleftrightarrow$ (5) for $\beta = \frac12.$\
We use the same argument as right above.\
(2) $\Longrightarrow$ (3).\
We use a similar $K_\theta$ as in the proof of $(2) \Leftrightarrow (4),$ fixing $\theta \in (-\pi,\pi):$ $$K_\theta: L^2({\mathbb{R}},ds) \to L^2({\mathbb{R}},ds),\, f(s) \mapsto (\pi - |\theta|)^{\alpha} \frac{1}{\sin\pi (\beta + is)} e^{\theta s} \langle s \rangle^{\alpha} f(s).$$ We have $$\sup_{|\theta| < \pi} \|K_\theta\| = \sup_{|\theta| < \pi,\,s\in{\mathbb{R}}} \langle s \rangle^\alpha (\pi-|\theta|)^\alpha
\frac{e^{\theta s}}{|\sin \pi (\beta + is)|} \lesssim \sup_{\theta,s} \langle s(\pi-|\theta|) \rangle^\alpha e^{-|s|(\pi-|\theta|)} < \infty.$$ Thus, by , $$\begin{aligned}
\sup_{0 < |\theta| \leq \pi} |\theta|^{\alpha} \|t^\beta A^{1-\beta}R(te^{i\theta},A)x\|_{\gamma({\mathbb{R}}_+,dt/t,X)}
& = \sup_{|\theta| < \pi} (\pi-|\theta|)^{\alpha} \|t^{\beta} A^{1 - \beta} (e^{i\theta} t +A)^{-1}x\|_{\gamma({\mathbb{R}}_+,\frac{dt}{t},X)}
\nonumber \\
& = \sup_{|\theta| < \pi} (\pi-|\theta|)^{\alpha} \| \frac{\pi}{\sin\pi (\beta + is)}e^{\theta s}A^{is}x \|_{\gamma({\mathbb{R}},ds,X)}
\label{Equ Proof Thm Charact} \\
& \lesssim \|\langle s \rangle^{-\alpha} A^{is}x\|_{\gamma({\mathbb{R}},ds,X)}.
\nonumber\end{aligned}$$
(3), $\alpha$ $\Longrightarrow$ (2), $\alpha + \epsilon.$\
First we consider $\langle s \rangle^{-(\alpha+\epsilon)}A^{is}x$ for $s\geq 1.$ $$\begin{aligned}
\|\langle s \rangle^{-(\alpha+\epsilon)}A^{is}x\|_{\gamma([1,\infty),X)}
& \leq \sum_{n = 0}^\infty 2^{-n\epsilon} \| \langle s \rangle^{-\alpha} A^{is}x\|_{\gamma([2^n,2^{n+1}],X)}.
\label{Equ Proof Thm Charact 2}\end{aligned}$$ For $s \in [2^n,2^{n+1}],$ we have $$\langle s \rangle^{-\alpha} \lesssim 2^{-n\alpha} \lesssim 2^{-n\alpha} e^{-2^{-n}s} \lesssim (\pi-\theta_n)^\alpha \frac{e^{\theta_n s}}{\sin \pi(\beta + i s)} ,$$ where $\theta_n = \pi - 2^{-n}.$ Therefore $$\begin{aligned}
\| \langle s \rangle^{-\alpha} A^{is}x\|_{\gamma([2^n,2^{n+1}],X)}
& \lesssim (\pi-\theta_n)^{\alpha} \| \frac{\pi}{\sin\pi (\beta + is)}e^{\theta_n s}A^{is}x \|_{\gamma({\mathbb{R}},X)} \nonumber \\
& \overset{\eqref{Equ Proof Thm Charact}}{\lesssim} \sup_{0<|\theta| \leq \pi} |\theta|^{\alpha} \| t^\beta A^{1 - \beta}R(te^{i\theta},A) x\|_{\gamma({\mathbb{R}}_+,dt/t,X)} < \infty.
\nonumber\end{aligned}$$ Thus, the sum in is finite.
The part $\langle s \rangle^{-(\alpha + \epsilon)}A^{is}x$ for $s \leq -1$ is treated similarly, whereas $\|\langle s \rangle^{-\alpha}A^{is}x\|_{\gamma((-1,1),X)} \cong \|A^{is}x\|_{\gamma((-1,1),X)}.$ It remains to show that the last expression is finite. We have assumed that $X$ has property $(\alpha).$ Then the fact that $A$ has an ${H^\infty}$ calculus implies that $\{A^{is}:\: |s| < 1\}$ is $\gamma$-bounded [@KaW2 Corollary 6.6]. Then by Lemma \[Lem Folklore square functions\] (3), we have $\|A^{is}x\|_{\gamma((-1,1),X)} \leq \gamma\left(\left\{ A^{is} : |s| < 1 \right\}\right) \|1\|_{L^2(-1,1)} \|x\|.$
Condition (5) of the preceding proposition can be checked in the following way.
\[Lem Sgr\] Let $A$ be a $0$-sectorial operator on a space $X$ with property $(\alpha)$ having an ${H^\infty}$ calculus. If for some $\beta > 0$ $$\label{Equ Lem Sgr}
\left\{ T \left(te^{\pm i(\frac{\pi}{2} - \theta)} \right) :\: t > 0\right\}\text{ is }\gamma\text{-bounded with constant }\lesssim \theta^{-\beta},$$ then $\| A^{\frac12} T(e^{\pm i (\frac{\pi}{2} - \theta)}t) x \|_{\gamma({\mathbb{R}}_+,X)} \lesssim \theta^{-\alpha} \|x\|$ with $\alpha = \beta + \frac12.$
Decompose $$t \exp \left(\pm i (\frac{\pi}{2} - \theta)\right) = s(t,\theta) + r(t,\theta) \exp\left(\pm i (\frac{\pi}{2} - \frac{\theta}{2})\right)$$ where the reals $s$ and $r$ are uniquely determined by $t$ and $\theta.$ We have $s(t,\theta) = \kappa(\theta) t$ with $\kappa(\theta) \cong \theta.$ Then by Lemma \[Lem Folklore square functions\], $$\begin{aligned}
\| A^{\frac12} T(t e^{\pm i (\frac{\pi}{2} - \theta)}) x \|_{\gamma({\mathbb{R}}_+,X)} & = \| T(r e^{\pm i (\frac{\pi}{2} - \frac{\theta}{2})}) A^{\frac12} T(s) x\|_{\gamma({\mathbb{R}}_+,X)} \\
& \leq \gamma\left(\left\{ T(r e^{\pm i (\frac{\pi}{2} - \frac{\theta}{2})}) :\: r > 0 \right\} \right) \| A^{\frac12} T(s(t,\theta)) x \|_{\gamma({\mathbb{R}}_+,X)} \\
& \lesssim (\theta/2)^{-\beta} \theta^{-\frac12} \| A^{\frac12} T(t) x \|_{\gamma({\mathbb{R}}_+,X)}.\end{aligned}$$ By and [@KaW2 Theorem 7.2], $\| A^{\frac12} T(t) x\|_{\gamma({\mathbb{R}}_+,X)} = \| A^{\frac12} (it - A)^{-1} x\|_{\gamma({\mathbb{R}},X)} \leq C \|x\|,$ which finishes the proof.
Let us now turn to some examples.
Let $\Omega$ be a topological space which is equipped with a distance $\rho$ and a Borel measure $\mu.$ Let $d \geq 1$ be an integer. $\Omega$ is called a homogeneous space of dimension $d$ if there exists $C > 0$ such that for any $x \in \Omega,\,r> 0$ and $\lambda \geq 1:$ $$\mu(B(x,\lambda r)) \leq C \lambda^d \mu(B(x,r)).$$
Typical cases of homogeneous spaces are open subsets of ${\mathbb{R}}^d$ with Lipschitz boundary and Lie groups with polynomial volume growth, in particular stratified nilpotent Lie groups (see e.g. [@FoSt]).
We will consider operators satisfying the following assumption.
\[Ass Examples\] $A$ is a self-adjoint positive (injective) operator on $L^2(\Omega),$ where $\Omega$ is a homogeneous space of a certain dimension $d.$ Further, there exists some $p_0 \in [1,2)$ such that the semigroup generated by $-A$ satisfies the so-called generalized Gaussian estimate (see e.g. [@Bluna (GGE)]): $$\tag{GGE}\label{Equ generalized Gaussian estimate}
\| \chi_{B(x,r_t)} e^{-tA} \chi_{B(y,r_t)} \|_{p_0 \to p_0'} \leq C \mu(B(x,r_t))^{\frac{1}{p_0'}-\frac{1}{p_0}} \exp \left(-c \left(\rho(x,y)/r_t \right)^\frac{m}{m-1} \right) \quad (x,y \in \Omega,\,t>0).$$ Here, $p_0'$ is the conjugated exponent to $p_0,\,C,c > 0,\,m \geq 2$ and $r_t = t^{\frac1m},$ $\chi_B$ denotes the characteristic function of $B,$ $B(x,r)$ is the ball $\{y \in \Omega:\: \rho(y,x) < r\}$ and $\|\chi_{B_1} T \chi_{B_2}\|_{p_0 \to p_0'} = \sup_{\|f\|_{p_0} \leq 1} \|\chi_{B_1} \cdot T(\chi_{B_2} f)\|_{p_0'}.$
If $p_0 = 1,$ then it is proved in [@BlKub] that is equivalent to the usual Gaussian estimate, i.e. $e^{-tA}$ has an integral kernel $k_t(x,y)$ satisfying the pointwise estimate (cf. e.g. [@DuOS Assumption 2.2]) $$\tag{GE}\label{Equ Gaussian estimate}
|k_t(x,y)| \lesssim \mu(B(x,t^{\frac1m}))^{-1} \exp\left(-c \left(\rho(x,y)/t^{\frac1m}\right)^{\frac{m}{m-1}}\right)\quad (x,y \in \Omega,\,t>0).$$ This is satisfied in particular by sublaplacian operators on Lie groups of polynomial growth [@Varo] as considered e.g. in [@MaMe; @Chri; @Alex; @MSt; @Duon], or by more general elliptic and sub-elliptic operators [@Davia; @Ouha], and Schrödinger operators [@Ouhaa]. It is also satisfied by all the operators in [@DuOS Section 2].
Examples of operators satisfying a generalized Gaussian estimate for $p_0 > 1$ are higher order operators with bounded coefficients and Dirichlet boundary conditions on domains of ${\mathbb{R}}^d,$ Schrödinger operators with singular potentials on ${\mathbb{R}}^d$ and elliptic operators on Riemannian manifolds as listed in [@Bluna Section 2] and the references therein.
\[Thm R-bounded Blunck Hormander thm\] Let Assumption \[Ass Examples\] hold. Then for any $p \in (p_0,p_0'),$ the ${H^\infty}$ calculus of $A$ extends to a matricially $\gamma$-bounded homomorphism ${{\mathcal{H}}^\alpha}\to B(L^p(\Omega))$ with $$\alpha > d\left| \frac1{p_0} - \frac12\right|+\frac12.$$
We show that holds with $\beta = d (\frac{1}{p_0} - \frac12).$ By [@BlKua Proposition 2.1], the assumption implies that $$\| \chi_{B(x,r_t)} e^{-tA} \chi_{B(y,r_t)}\|_{p_0\to 2} \leq C_1 \mu(B(x,r_t))^{\frac{1}{2}-\frac{1}{p_0}} \exp(-c_1 (\rho(x,y)/r_t)^{\frac{m}{m-1}}) \quad (x,y \in \Omega,\,t>0)$$ for some $C_1,c_1>0.$ By [@Blun Theorem 2.1], this can be extended from real $t$ to complex $z = te^{i\theta}$ with $\theta \in (-\frac{\pi}{2},\frac{\pi}{2}):$ $$\|\chi_{B(x,r_z)} e^{-zA} \chi_{B(y,r_z)} \|_{p_0 \to 2} \leq C_2 \mu(B(x,r_z))^{\frac{1}{2}-\frac{1}{p_0}} (\cos \theta )^{-d(\frac{1}{p_0} -\frac{1}{2})} \exp(-c_2 (\rho(x,y)/r_z)^{\frac{m}{m-1}}),$$ for $r_z = (\cos \theta )^{-{\frac{m-1}{m}}} t^{\frac1m},$ and some $C_2,c_2 >0.$ By [@BlKua Proposition 2.1 (i) (1) $\Rightarrow$ (3) with $R = e^{-zA},\,\gamma =\alpha = \frac{1}{p_0} - \frac12,\,\beta = 0,\,r = r_z,\,u=p_0$ and $v = 2$], this gives for any $x \in \Omega,\,\operatorname{Re}z > 0$ and $k \in {\mathbb{N}}_0$ $$\|\chi_{B(x,r_z)} e^{-zA} \chi_{A(x,r_z,k)}\|_{p_0 \to 2} \leq C_3 \mu(B(x,r_z))^{\frac12 - \frac{1}{p_0}} (\cos \theta )^{-d(\frac{1}{p_0} -\frac{1}{2})} \exp(-c_3 k^{\frac{m}{m-1}}),$$ where $A(x,r_z,k)$ denotes the annular set $B(x,(k+1)r_z)\backslash B(x,k r_z).$ By [@Kuns Theorem 2.2 with $q_0 = p_0,\,q_1 = s = 2,\rho(z) = r_z$ and $S(z) = (\cos \theta)^{d(\frac{1}{p_0}-\frac{1}{2})} e^{-zA}$] and property $(\alpha),$ we deduce that $$\{ (\cos \theta)^{d(\frac{1}{p_0}-\frac{1}{2})} e^{-zA} :\: \operatorname{Re}z > 0\}$$ is $\gamma$-bounded. Now apply Lemma \[Lem Sgr\] and Proposition \[Prop BIP Sgr Res\], noting that $A$ has an ${H^\infty}$ calculus on $L^p(\Omega)$ [@Blun Corollary 2.3].
\[Rem R-bounded Blunck Hormander thm\]
1. Theorem \[Thm R-bounded Blunck Hormander thm\] improves on [@Bluna Theorem 1.1] in that it includes the matricial $\gamma$-boundedness of the Hörmander calculus. Note that [@Bluna] obtains also a weak-type result for $p = p_0.$ If $p_0$ is strictly larger than $1,$ then Theorem \[Thm R-bounded Blunck Hormander thm\] improves the order of derivation $\alpha$ of the calculus from $\displaystyle \frac{d}{2} + \frac{1}{2} +\epsilon$ in [@Bluna] to $\displaystyle d \left|\frac{1}{p_0} - \frac{1}{2} \right| + \frac12 + \epsilon.$ In [@Uhl Theorem 6.4 a)], under the assumptions of Theorem \[Thm R-bounded Blunck Hormander thm\], a ${\mathcal{H}}^\beta_r$ calculus with with $\beta > (d+1) | \frac{1}{p_0} - \frac{1}{2} |$ and $r > | \frac12 - \frac1p |^{-1}$ is derived. Here ${\mathcal{H}}^\beta_r$ is defined similarly to ${{\mathcal{H}}^\alpha}$ by $${\mathcal{H}}^\beta_{r} = \{ f : (0,\infty)\to {\mathbb{C}}:\: \sup_{k \in {\mathbb{Z}}} \|(f \circ \exp )\phi_k\|_{W^\beta_r} < \infty\}.$$ Note that ${\mathcal{H}}^\beta_r$ is larger than ${\mathcal{H}}^\alpha.$ In the classical case of Gaussian estimates, i.e. $p_0 = 1,$ [@DuOS] yields a ${\mathcal{H}}^{\alpha_2}_\infty$ calculus under Assumption \[Ass Examples\] and even a ${\mathcal{H}}^{\alpha_2}$ calculus for many examples, e.g. homogeneous operators, with the better derivation order $\alpha_2 > \frac{d}{2}.$
2. The theorem also holds for the weaker assumption that $\Omega$ is an open subset of a homogeneous space $\tilde{\Omega}.$ In that case, the ball $B(x,r_t)$ on the right hand side in is the one in $\tilde{\Omega}.$ This variant can be applied to elliptic operators on irregular domains $\Omega \subset {\mathbb{R}}^d$ as discussed in [@Bluna Section 2].
In Theorem \[Thm R-bounded Blunck Hormander thm\], the operator $A$ was assumed to be self-adjoint, and thus, admits a functional calculus $L^\infty \to B(L^2(\Omega)).$ The space $L^\infty = L^\infty((0,\infty);d\mu_A)$ is larger than ${{\mathcal{H}}^\alpha},$ and one can use this fact to ameliorate the functional calculus of $A$ on $L^q(\Omega)$ by complex interpolation.
Let $A$ satisfy Assumption \[Ass Examples\]. Then for $q \in (p_0,p_0'),\, \alpha > d \left| \frac{1}{p_0}- \frac12 \right| + \frac12$ and $\theta \in (0,1)$ with $\theta > \left| \frac1q -\frac{1}{p_0}\right| / \left|{\frac12 - \frac1{p_0}} \right|,$ the functional calculus of $A$ on $L^q$ $$\label{Equ Proof Rem Blunck}
u_{L^q} : (L^\infty,{{\mathcal{H}}^\alpha})_\theta \to B(L^q(\Omega)) \text{ is matricially }\gamma\text{-bounded.}$$ Here $(L^\infty,{{\mathcal{H}}^\alpha})_\theta$ is the complex interpolation space which is given an operator space structure [@Pis2 p. 56] by $$M_n \otimes (L^\infty,{{\mathcal{H}}^\alpha})_{\theta} = (M_n \otimes L^\infty,M_n \otimes {{\mathcal{H}}^\alpha})_\theta.$$
The self-adjoint calculus $u_{L^2} : L^\infty \to B(L^2(\Omega))$ is completely bounded since it is a $\ast$-representation [@Pis2 Proposition 1.5], so by Remark \[Rem mat\], $u_{L^2}$ is matricially $\gamma$-bounded. Moreover, we have $(\operatorname{Gauss}(L^p),\operatorname{Gauss}(L^2))_\theta = \operatorname{Gauss}((L^p,L^2)_\theta)$ [@KaKW Proposition 3.7]. Then by bilinear interpolation between $$M_n \otimes L^\infty \times \operatorname{Gauss}_n(L^2) \to \operatorname{Gauss}_n(L^2),\, ([a_{ij}] \otimes f ,\sum_k \gamma_k \otimes x_k) \mapsto \sum_{k,j} \gamma_k a_{kj} u_{L^2}(f)x_j$$ and, with the mapping $u_{L^p}$ resulting from Theorem \[Thm R-bounded Blunck Hormander thm\], $p$ given by $\theta = |\frac1q - \frac1p| / |\frac12 - \frac1p|,$ $$M_n \otimes {{\mathcal{H}}^\alpha}\times \operatorname{Gauss}_n(L^p) \to \operatorname{Gauss}_n(L^p),\, ([a_{ij}] \otimes f , \sum_k \gamma_k \otimes x_k) \mapsto \sum_{k,j} \gamma_k a_{kj} u_{L^p}(f)x_j,$$ one deduces .
Note that the space $(L^\infty,{{\mathcal{H}}^\alpha})_{\theta}$ contains ${\mathcal{H}}^\beta_{r,0},$ where $\frac1r > \frac{\theta}{2},$ $\beta > \alpha \theta + (\frac1r - \frac{\theta}{2}).$ Here ${\mathcal{H}}^\beta_{r,0} = \left\{ f \in {\mathcal{H}}^{\beta}_r :\: \|(f \circ \exp) \phi_k \|_{W^\beta_r} \to 0\text{ for }|k| \to \infty \right\}.$ Then implies that in particular, ${\mathcal{H}}^\beta_{r,0} \to B(L^q),\,f \mapsto f(A)$ is (norm) bounded and by [@Kr Section 4.6.1], this extends moreover boundedly to ${\mathcal{H}}^\beta_r \to B(L^q).$
In [@DuOS Section 7], for many examples of operators $A$ satisfying , it is shown that the functional calculus $$\label{Equ DuOS}
u: {{\mathcal{H}}^\alpha}\to B(L^p(\Omega)),\,f \mapsto f(A)\text{ is bounded for }1 < p < \infty\text{ and }\alpha > \frac{d}{2}.$$ Moreover, for the fundamental example $A = - \Delta$ on $L^p({\mathbb{R}}^d),$ the critical order $\frac{d}{2}$ in is optimal [@Ste IV.7.4],[@Kr Proposition 4.12 (2)]. Note that the derivation order for the matricially $\gamma$-bounded calculus obtained in Theorem \[Thm R-bounded Blunck Hormander thm\] under the assumption (i.e. $p_0 = 1$) is only $\frac{d+1}{2},$ and therefore gives a weaker result in the derivation order compared to .
Thus the question arises if an arbitrary $A$ that has a norm-bounded Hörmander calculus also has a matricially $\gamma$-bounded Hörmander calculus. In contrast to the self-adjoint $L^\infty$ calculus on Hilbert space, which is always matricially $\gamma$-bounded (see the proof above), we have the following result.
\[Prop bounded vs mat-gamma bounded\] Let $A$ be a $0$-sectorial operator on a space $X$ with property $(\alpha).$ Let $\alpha > \frac12$ and $\beta > \alpha + 1.$ Suppose that its functional calculus $f \mapsto f(A)$ is bounded $u_\alpha : {{\mathcal{H}}^\alpha}\to B(X),$ and denote $u_\beta$ the restriction of $u_\alpha$ to ${\mathcal{H}}^\beta.$ Then $u_\beta : {\mathcal{H}}^\beta \to B(X)$ is matricially $\gamma$-bounded.
On the other hand, for any $\alpha > 0,$ there exists some $A$ on a Hilbert space $X$ such that $u_\alpha : {{\mathcal{H}}^\alpha}\to B(X)$ is bounded (even $\gamma$-bounded because of ), but its restriction $u_{\alpha + \frac12} : {\mathcal{H}}^{\alpha + \frac12} \to B(X)$ is not matricially $\gamma$-bounded.
For $t \in {\mathbb{R}},$ let $f_t(\lambda) = \lambda^{it}.$ It is easy to check that $\|f_t\|_{{{\mathcal{H}}^\alpha}} \lesssim \langle t \rangle^{\alpha}$ [@Kr Lemma 4.12 (4)]. By Lemma \[Lem Function spaces\] (1), $A$ has an ${H^\infty}$ calculus. By [@Kr1 Corollary 6.3], the set $\left\{ T(te^{\pm i (\frac{\pi}{2} - \theta)}) :\: t > 0 \right\}$ is $\gamma$-bounded with constant $\leq C \theta^{-\alpha - \frac12}\quad (\theta \in (0,\frac{\pi}{2})).$ By Lemma \[Lem Sgr\], condition (5) of Proposition \[Prop BIP Sgr Res\] is satisfied with $\alpha + 1$ in place of $\alpha$ and therefore, $u_{\beta} : {\mathcal{H}}^{\beta} \to B(X)$ is matricially $\gamma$-bounded.
For the second statement, let $\alpha > \frac12.$ Consider $X = {W^\alpha}$ and the group $U(t) g = (\cdot)^{it} g$ on $X.$ Note that $$\| (\cdot)^{it} g \|_X = \| (g \circ \exp)\hat{\phantom{i}}(\cdot - t) \langle \cdot \rangle^\alpha \|_2 = \| (g \circ \exp)\hat{\phantom{i}} (\cdot) \langle (\cdot) + t \rangle^\alpha \|_2 \cong \langle t \rangle^\alpha \|g\|_X.$$ In particular, $\|U(t)\| \cong \langle t \rangle^\alpha.$ It is easy to check that $U(t) = A^{it}$ are the imaginary powers of a $0$-sectorial operator $A$ and that $f(A) g = f g$ for any $g \in X$ and $f \in \bigcup_{\omega > 0} {H^\infty}(\Sigma_\omega).$ By [@Str], one has $\|fg\|_{{W^\alpha}} \lesssim \|f\|_{{{\mathcal{H}}^\alpha}} \|g\|_{{W^\alpha}}.$ Thus, $A$ has a bounded ${{\mathcal{H}}^\alpha}$ calculus.
On the other hand, since $X$ is a Hilbert space, the square function condition of Theorem 4.10 reads $$\| \langle t \rangle^{-\beta} A^{it} x \|_{\gamma({\mathbb{R}},X)} = \| \langle t \rangle^{-\beta} A^{it} x\|_{L^2({\mathbb{R}},X)} \cong \left( \int_{\mathbb{R}}\langle t \rangle^{-2 \beta + 2 \alpha} dt \right)^{\frac12} \|x\|,$$ which is finite if and only if $\beta > \alpha + \frac12.$
Proofs of Lemmas \[Lem 1\] - \[Lem 4\] {#Sec 6 Proofs Lemmas}
======================================
Since $X$ has property $(\alpha),$ the fact that $A$ has a bounded ${H^\infty}(\Sigma_\omega)$ calculus implies [@KuWe Theorem 12.8] that for any $\theta > \omega,$ $$\left\{ g(A) :\: \|g\|_{\infty,\theta} \leq 1 \right\}\text{ is }\gamma\text{-bounded}.$$ We fix some $\theta \in (\omega, \frac{\pi}{4}).$ As the mapping $u : \: {W^\alpha}\to B(X)$ is matricially $\gamma$-bounded, by Remark \[Rem mat\], $$\left\{ h(A) :\: \|h\|_{{W^\alpha}} \leq 1 \right\}\text{ is }\gamma\text{-bounded}.$$ The lemma stated that $\gamma\left(\left\{ f_{2^k z}(A) :\: k \in {\mathbb{Z}}\right\}\right) \lesssim \left| z /\operatorname{Re}z\right|^\beta,$ where $f_{2^k z}(\lambda) = \exp(-2^k z \lambda).$ Thus it suffices to decompose $f_{2^k z} = g + h,$ where $\|g\|_{\infty,\theta},\,\|h\|_{{W^\alpha}} \lesssim \left| z / \operatorname{Re}z \right|^\beta.$
As $\Psi : f \mapsto f(r \cdot)$ is an isomorphism ${H^\infty}(\Sigma_\theta) \to {H^\infty}(\Sigma_\theta)$ and ${W^\alpha}\to {W^\alpha},$ with $\|\Psi \| \cdot \|\Psi^{-1}\| \leq C,\:C$ independent of $r > 0,$ it suffices to have the above decomposition for $|z| = 1$ and $k = 0.$ We choose $g(\lambda) = \exp(-(z+1) \lambda)$ and $h(\lambda) = \exp(-z\lambda) (1 - e^{-\lambda}).$ As $|\arg(z+1)| + \theta \leq \frac{\pi}{4} + \frac{\pi}{4} = \frac{\pi}{2},$ we actually have $\|g\|_{\infty,\theta} \leq 1 \lesssim \left |\operatorname{Re}z \right|^{-\beta}.$ Further it is a simple matter to check that $\|h\|_{{W^\alpha}} \lesssim \left |\operatorname{Re}z \right|^{-\beta}$ for any $\beta > \alpha.$
The assumption of the lemma was $$\label{Equ Proof Lem 2}
\gamma\left(\left\{ \exp(-2^k z A) :\: k \in {\mathbb{Z}}\right\}\right) \lesssim \left| \frac{z}{\operatorname{Re}z} \right|^\beta \quad (\operatorname{Re}z > 0).$$ We first show that $$\label{Equ 2 Proof Lem 2}
\gamma \left( \left\{ (2^k t A)^{\frac12} \exp(-2^k t e^{\pm i (\frac{\pi}{2} - \omega) } A ) :\: k \in {\mathbb{Z}}\right\} \right) \lesssim \omega^{-(\beta + \frac12)} \quad (\omega \in (0,\frac{\pi}{2})).$$ Decompose $$e^{\pm i (\frac{\pi}{2} - \omega)} t = s + e^{\pm i (\frac{\pi}{2} - \frac{\omega}{2})} r,$$ where $s,r > 0$ are uniquely determined by $t$ and $\omega.$ Then $$(2^k t A)^{\frac12} \exp(-e^{\pm i (\frac{\pi}{2} - \omega)} 2^k t A) = \left( \frac{t}{s} \right)^{\frac12} (2^k s A)^{\frac12} \exp(-2^k s A) \exp(- 2^k r e^{\pm i (\frac{\pi}{2} - \frac{\omega}{2})}A),$$ and consequently, $$\begin{aligned}
\gamma\left(\left\{(2^k t A)^{\frac12} \exp(-e^{\pm i (\frac{\pi}{2} - \omega)}2^k t A) :\: k \in {\mathbb{Z}}\right\} \right) & \leq \sup_t (t/s)^{\frac12} \times \gamma\left(\left\{ (2^k s A)^{\frac12} \exp(-2^k s A) :\: k \in {\mathbb{Z}}\right\} \right) \nonumber \\
& \times \gamma \left(\left\{ \exp(-2^k r e^{\pm i (\frac{\pi}{2} - \frac{\omega}{2})}A):\: k \in {\mathbb{Z}}\right\} \right).
\label{Equ 3 Proof Lem 2}\end{aligned}$$ We will show that the right hand side of can be estimated by $\lesssim \omega^{-\frac12} \times 1 \times \omega^{-\beta}.$ The estimate for the first factor follows from the law of sines $$t/s = \sin (\frac{\pi}{2} + \omega/2)/\sin(\omega/2) \cong \omega^{-1}.$$ For the second estimate, note that by [@KuWe Example 2.16], implies that $\{ \exp(-zA) :\: z \in \Sigma_\delta \}$ is $\gamma$-bounded for any $\delta < \frac{\pi}{2}$ and consequently, by [@KuWe Theorem 2.20, $(iii) \Longrightarrow (i)$], $\{ \lambda (\lambda - A)^{-1} :\: - \lambda \in \Sigma_\theta \}$ is $\gamma$-bounded for any $\theta \in (\frac{\pi}{2},\pi).$ Then with $f(\lambda) = \lambda^{\frac12} e^{-\lambda},$ the Cauchy integral formula gives $$\begin{aligned}
(2^k t A)^{\frac12} \exp(-2^k t A) & = \frac{1}{2\pi i} \int_{\partial \Sigma_{\pi - \theta}} f(\lambda) (\lambda - 2^ktA)^{-1} d\lambda \\
& = \frac{1}{2\pi i} \int_{\partial \Sigma_{\pi - \theta}} \frac{f(\lambda)}{\lambda} \times \frac{\lambda}{2^k t} (\frac{\lambda}{2^k t} - A)^{-1} d\lambda\end{aligned}$$ The first factor in the last integral belongs to $L^1(\partial \Sigma_{\pi - \theta},|d\lambda|)$ and the second factor is $\gamma$-bounded by the above for any $\theta < \pi.$ Thus by the well-known integral lemma for $\gamma$-bounds [@KuWe Corollary 2.14], the second factor in is finite.
The estimate for the third factor in follows from the assumption , so that we have shown .
Now we will write the expression in as an integral of the expression in . Let $\theta \in (0,\frac{\pi}{2}),\,\lambda = t e^{i \theta}$ and set $\phi = \frac{\pi}{2} - \frac{\theta}{2},$ so that $\operatorname{Re}( e^{i \phi} \lambda ) < 0.$ Then $$\begin{aligned}
\lambda^{\frac12} (2^k A)^{\frac12} ( \lambda - 2^k A)^{-1} & = \lambda^{\frac12} (2^k A)^{\frac12} e^{i \phi} (e^{i \phi} \lambda - e^{i \phi} 2^k A)^{-1} \\
& = \int_0^\infty -e^{i \phi} s^{-\frac12} \lambda^{\frac12} \exp( e^{i \phi} \lambda s) \times (2^k s A)^{\frac12} \exp(-2^k e^{i \phi} s A) ds.\end{aligned}$$ The second factor of the integral is $\gamma$-bounded by and the first factor is integrable, as the following lines show. $$\begin{aligned}
\int_0^\infty s^{-\frac12} |\lambda^{\frac12} \exp(e^{i\phi} \lambda s)| ds & = \int_0^\infty s^{-\frac12} | \exp(e^{i \phi} e^{i\theta} s)| ds \\
& = \int_0^\infty s^{-\frac12} \exp( \cos(\frac{\pi}{2} + \frac{\theta}{2}) s) ds \\
& = \int_0^\infty s^{-\frac12} \exp(-s) ds \,|\cos(\frac{\pi}{2} + \frac{\theta}{2})|^{-\frac12} \\
& \lesssim \theta^{-\frac12}\end{aligned}$$ Then $\tau = \left\{ \lambda^{\frac12}(2^k A)^{\frac12} (\lambda - 2^k A)^{-1} :\: k \in {\mathbb{Z}}\right\}$ is $\gamma$-bounded since by [@Kr Proposition 2.6 (5)], we have $$\begin{aligned}
\gamma(\tau) & \leq \int_0^\infty s^{-\frac12} | \lambda^{\frac12} \exp(e^{i\phi} \lambda s) | ds \times
\sup_{ t > 0} \gamma\left( \left\{ (2^k t A)^{\frac12} \exp(-2^k t e^{\pm i (\frac{\pi}{2} - \omega)} A) :\: k \in {\mathbb{Z}}\right\} \right) \\
& \lesssim |\arg \lambda |^{- \frac12} \times |\arg \lambda |^{-\beta - \frac12}\end{aligned}$$ The same reasoning applies for $\lambda = t e^{i\theta}$ and $\theta \in (-\frac{\pi}{2},0).$
By [@Kr Proposition 4.18] and [@CDMY p. 73], it suffices to show that for some $\delta \in (\gamma,n),$ $$\label{Equ 1 Lem Intermediate bad Mihlin calculus}
\|f(A)\| \lesssim \theta^{-\delta} \|f\|_{\infty,\theta}\text{ for any }f \in \bigcup_{\theta > 0} {H^\infty}_0(\Sigma_\theta).$$ To show , we use the Kalton-Weis characterization of the bounded ${H^\infty}(\Sigma_\theta)$ calculus in terms of $\gamma$-bounded operator families ([@KaW], see also [@KuWe Theorem 12.7]). More precisely, we follow that characterization in the form of the proof of [@KuWe Theorem 12.7] and keep track of the dependence of appearing constants on the angle $\theta$. It is shown there that for $f \in {H^\infty}_0(\Sigma_{2\theta}),\,x \in X$ and $x' \in X',$ $$\begin{aligned}
|{\langle f(A)x, x' \rangle}| & = |\frac{1}{2\pi i} \int_{\partial\Sigma_\theta} {\langle \lambda^{-\frac12}f(\lambda) A^{\frac12} (\lambda- A)^{-1}x, x' \rangle} d\lambda|
\nonumber \\
& \leq \frac{1}{2\pi} \sum_{j = \pm 1} \int_0^\infty |{\langle f(te^{ij\theta}) (tA)^{\frac12}(e^{ij\theta}- tA)^{-1} x, x' \rangle}|\frac{dt}{t}
\nonumber \\
& = (*).
\nonumber\end{aligned}$$ We put $$\phi_{j \theta}(\lambda) = \frac{\lambda^{\frac14}(1+\lambda)^{\frac12}}{e^{ij\theta}-\lambda}
\text{ and }\psi(\lambda)= \left(\frac{\lambda}{(1+\lambda)^2}\right)^{\frac18},$$ so that $(tA)^{\frac12} (e^{ij\theta}-tA)^{-1} = \phi_{j\theta}(tA)\psi(tA)\psi(tA).$ By [@KuWe Lemma 12.6], the integral $(*)$ can be controlled by $\operatorname{Gauss}$-norms. More precisely, we have $$\begin{aligned}
(*) & \lesssim \sup_{j=\pm 1} \sup_{t > 0} \sup_N \| \sum_{k=-N}^N \gamma_k \otimes f(2^k t e^{ij \theta}) \phi_{j \theta}(2^k t A) \psi(2^k t A) x\|_{\operatorname{Gauss}(X)}
\label{Equ 2 Lem Intermediate bad Mihlin calculus} \\
& \cdot \|\sum_{k=-N}^N \gamma_k \otimes \psi(2^k t A)'x'\|_{\operatorname{Gauss}(X')}
\nonumber \\
& \lesssim \|f\|_{\infty,\theta} \sup_{j,t} \gamma\left(\{ \phi_{j \theta}(2^k t A) :\: k \in {\mathbb{Z}}\}\right) \sup_{N,t} \| \sum_{k=-N}^N \gamma_k \otimes \psi(2^k t A)x\|_{\operatorname{Gauss}(X)}
\nonumber \\
& \cdot \sup_{N,t}\| \sum_{k = -N}^N \gamma_k \otimes \psi(2^k t A)'x'\|_{\operatorname{Gauss}(X')}.
\nonumber\end{aligned}$$ By [@KuWe Theorem 12.2], the fact that $A$ has a bounded ${H^\infty}$ calculus implies that $\sup_{N,t} \| \sum_{k=-N}^N \gamma_k \otimes \psi(2^k t A) x\|_{\operatorname{Gauss}(X)} \lesssim \|x\|$ and $\sup_{N,t} \| \sum_{k=-N}^N \gamma_k \otimes \psi(2^k t A)' x'\|_{\operatorname{Gauss}(X')} \lesssim \|x'\|.$ Note that there is no dependence on $\theta$ in these two inequalities. It remains to show that $$\label{Equ R-bound of phi_theta}
\sup_{j = \pm 1,t> 0}\gamma(\{ \phi_{j \theta}(2^k t A) :\: k \in {\mathbb{Z}}\}) \lesssim \theta^{-\delta}.$$ We have $$\begin{aligned}
\phi_{j \theta}(2^ktA) & = \frac{1}{2\pi i} \int_{\partial\Sigma_{\frac{\theta}{2}}} \phi_{j\theta}(\lambda) \lambda^{\frac12} (2^ktA)^{\frac12} (\lambda-2^ktA)^{-1} \frac{d\lambda}{\lambda}
\nonumber \\
& = \frac{1}{2\pi i} \int_{\partial\Sigma_{\frac{\theta}{2}}} \phi_{j\theta}(t\lambda) \lambda^{\frac12} (2^kA)^{\frac12} (\lambda-2^kA)^{-1} \frac{d\lambda}{\lambda}.
\nonumber\end{aligned}$$ By [@Kr Proposition 2.6 (5)], $$\begin{aligned}
\sup_{j= \pm 1,t>0}\gamma(\{\phi_{j\theta}(2^ktA):\:k \in {\mathbb{Z}}\}) & \lesssim \sup_{j = \pm 1,t > 0} \|\phi_{j \theta}(t\lambda)\|_{L^1(\partial\Sigma_{\frac{\theta}{2}},|\frac{d\lambda}{\lambda}|)} \\
& \times
\sup_{\lambda \in \partial\Sigma_{\theta/2} \backslash \{ 0 \} }\gamma\left(\{ \lambda^{\frac12} (2^kA)^{\frac12} (\lambda - 2^kA)^{-1} :\: k \in {\mathbb{Z}}\}\right).\end{aligned}$$ By assumption, it suffices to show that for any $\epsilon > 0$ $$\sup_{t > 0} \|\phi_{j \theta}(t\lambda)\|_{L^1(\partial\Sigma_{\frac{\theta}{2}},|\frac{d\lambda}{\lambda}|)} \leq C_\epsilon \theta^{-\epsilon}.$$ $$\begin{aligned}
\int_{\partial\Sigma_{\frac{\theta}{2}}} |\phi_{j \theta}(t\lambda)| \left|\frac{d\lambda}{\lambda}\right|
& = \int_{\partial\Sigma_{\frac{\theta}{2}}} | \phi_{j \theta}(\lambda)| \left| \frac{d\lambda}{\lambda}\right|
= \sum_{l = \pm 1}\int_0^\infty \left|\frac{s^\frac14 (1 + e^{i l \frac{\theta}{2}} s)^{\frac12}}{e^{i j\theta} - e^{il\frac{\theta}{2}} s}\right| \frac{ds}{s}.
\nonumber\end{aligned}$$ The denominator is estimated from below by $$\begin{aligned}
|e^{ij \theta} - e^{il \frac{\theta}{2}} s | & = |e^{i\theta (j - \frac{l}{2})} - s|
\gtrsim |\cos(\theta (j- \frac{l}{2})) - s| + |\sin(\theta (j-\frac{l}{2}))|
\nonumber \\
& \gtrsim |1-s| - |\cos(\theta (j-\frac{l}{2}))-1| + \theta
\nonumber \\
& \gtrsim |1-s| - \theta^2 + \theta
\gtrsim |1-s| + \theta
\nonumber\end{aligned}$$ for the crucial case of small $\theta.$ Thus $$\int_{\partial\Sigma_{\frac{\theta}{2}}} |\phi_{j \theta}(\lambda)| \left|\frac{d\lambda}{\lambda}\right|
\lesssim \int_0^\infty \frac{s^{\frac14} (1+s)^{\frac12}}{\theta + |1-s|} \, \frac{ds}{s}.$$ We split the integral into the parts $\int_0^\infty = \int_0^{\frac12} + \int_{\frac12}^{1-\theta} + \int_{1-\theta}^{1+\theta} + \int_{1+\theta}^2 + \int_2^\infty.$ $$\int_0^{\frac12} \frac{s^{\frac14}(1+s)^{\frac12}}{\theta + |1-s|} \, \frac{ds}{s} \leq \int_0^{\frac12} \frac{s^{\frac14}(1+s)^{\frac12}}{|1-s|} \, \frac{ds}{s} < \infty$$ is independent of $\theta.$ The same estimate applies to $\int_2^\infty.$ $$\int_{\frac12}^{1-\theta} \frac{s^{\frac14}(1+s)^{\frac12}}{\theta + |1 - s|} \, \frac{ds}{s} \lesssim \int_{\frac12}^{1-\theta} \frac{1}{\theta + |1 -s|} ds \leq \int_{\frac12}^{1-\theta} \frac{1}{1-s} ds \lesssim | \log \theta |.$$ Similarly, $$\int_{1 + \theta}^2 \frac{s^{\frac14}(1+s)^{\frac12}}{\theta + |1 - s|} \, \frac{ds}{s} \lesssim \int_{1 + \theta}^2 \frac{1}{s-1} ds \lesssim |\log \theta|.$$ Finally, $$\int_{1-\theta}^{1 + \theta} \frac{s^{\frac14}(1+s)^{\frac12}}{\theta + |1 - s|} \, \frac{ds}{s} \lesssim \int_{1-\theta}^{1 + \theta} \frac{1}{\theta} ds \lesssim 1.$$ Since $1 + |\log \theta| \leq C_\epsilon \theta^{-\epsilon},$ the lemma is shown.
Denote $N = \sup_{x > 0} \# \{ k \in {\mathbb{Z}}: \: \operatorname{supp}g_k \cap [\frac12 x, 2x] \neq \emptyset \} < \infty.$ Fix $x > 0$ and $j \in \{ 0, 1 ,\ldots, n\}.$ Then almost all $g_k$ vanish in a neighborhood of $x,$ and thus $$\left|x^j \frac{d^j}{dx^j} \left(\sum_{k \in {\mathbb{Z}}} g_k\right)(x)\right| = \left|\sum_{k \in {\mathbb{Z}}} x^j \frac{d^j}{dx^j}g_k(x)\right| \leq N \sup_{k \in {\mathbb{Z}}} |x^j \frac{d^j}{dx^j}g_k(x)| \leq N \sup_{k \in {\mathbb{Z}}} \|g_k\|_{\operatorname{M}^n} .$$ Taking the supremum over $x$ and $j$ gives $\|\sum_{k \in {\mathbb{Z}}} g_k\|_{\operatorname{M}^n} \leq N \sup_{k \in {\mathbb{Z}}} \|g_k\|_{\operatorname{M}^n}.$
Acknowledgment {#acknowledgment .unnumbered}
==============
I would like to thank Lutz Weis for several discussions on the subject of this article. Further, I acknowledge financial support from the Karlsruhe House of Young Scientists KHYS and the Franco-German University DFH-UFA.
[100]{}
G. Alexopoulos. Spectral multipliers on Lie groups of polynomial growth. 120(3):973–979, 1994.
J. Bergh and J. Löfström. Grundlehren der mathematischen Wissenschaften, 223. Berlin etc.: Springer, 1976.
S. Blunck. A Hörmander-type spectral multiplier theorem for operators without heat kernel. 2(3):449–459, 2003.
S. Blunck. Generalized Gaussian estimates and Riesz means of Schrödinger groups. 82(2):149–162, 2007.
S. Blunck and P. Kunstmann. Generalized Gaussian estimates and the Legendre transform. 53(2):351–365, 2005.
S. Blunck and P. C. Kunstmann. Calderón-Zygmund theory for non-integral operators and the $H^\infty$ functional calculus. 19(3):919–942, 2003.
M. Christ. $L\sp p$ bounds for spectral multipliers on nilpotent groups. 328(1):73–81, 1991.
M. Cowling, I. Doust, A. McIntosh and A. Yagi. Banach space operators with a bounded $H^\infty$ functional calculus. 60(1):51–89, 1996.
E. Davies. Cambridge Tracts in Mathematics, 92. Cambridge etc.: Cambridge University Press, 1989.
J. Diestel, H. Jarchow and A. Tonge. Cambridge Studies in Advanced Mathematics, 43. Cambridge: Cambridge Univ. Press, 1995.
B. de Pagter and W. Ricker. $C(K)$-representations and $R$-boundedness. (2) 76:498–512, 2007.
X. T. Duong. From the $L\sp 1$ norms of the complex heat kernels to a Hörmander multiplier theorem for sub-Laplacians on nilpotent Lie groups. 173(2):413–424, 1996.
X. Duong and D. Robinson. Semigroup kernels, Poisson bounds, and holomorphic functional calculus. 142 No. 1:89–128, 1996.
X. T. Duong, E. M. Ouhabaz and A. Sikora. Plancherel-type estimates and sharp spectral multipliers. 196(2):443–485, 2002.
E. Effros and Zh.-J. Ruan. London Mathematical Society Monographs, New Series, 23. The Clarendon Press, Oxford University Press, New York, 2000.
G. Folland and E. Stein. Mathematical Notes, 28. Princeton, NJ: Princeton University Press, University of Tokyo Press, 1982.
A. Fr[ö]{}hlich. . PhD thesis, Universität Karlsruhe, 2003.
B. Haak and P. Kunstmann. Admissibility of unbounded operators and wellposedness of linear systems in Banach spaces. 55 No. 4:497–533, 2006.
M. Haase. Operator Theory: Advances and Applications, 169. Basel: Birkhäuser, 2006.
L. Hörmander. Estimates for translation invariant operators in $L^p$ spaces. 104:93–140, 1960.
N. Kalton, P. Kunstmann and L. Weis. Perturbation and interpolation theorems for the $H^\infty$-calculus with applications to differential operators. 336, no. 4:747–801, 2006.
N. Kalton, J. van Neerven, M. Veraar and L. Weis. Embedding vector-valued Besov spaces into spaces of $\gamma$-radonifying operators. 281(2):238–252, 2008.
N. Kalton and L. Weis. The $H^\infty$-calculus and square function estimates, preprint.
N. Kalton and L. Weis. The $H^\infty$-calculus and sums of closed operators. 321(2):319–345, 2001.
C. Kriegler. Spectral multipliers, $R$-bounded homomorphisms, and analytic diffusion semigroups.
C. Kriegler. Functional calculus and dilation for $c_0$-groups of polynomial growth, submitted.
C. Kriegler and C. Le Merdy. Tensor extension properties of $C(K)$-representations and applications to unconditionality. 88:205–230, 2010.
P. C. Kunstmann. On maximal regularity of type $L^p-L^q$ under minimal assumptions for elliptic non-divergence operators. 255(10):2732–2759, 2008.
P. C. Kunstmann and L. Weis. Maximal $L_p$-regularity for parabolic equations, Fourier multiplier theorems and $H^\infty$-functional calculus. Berlin: Springer, Lect. Notes Math. 1855, 65–311, 2004.
C. Le Merdy. On square functions associated to sectorial operators. 132 (1):137–156, 2004.
C. Le Merdy. $\gamma$-bounded representations of amenable groups. 224 No. 4:1641–1671, 2010.
G. Mauceri and S. Meda. Vector-valued multipliers on stratified groups. 6(3-4):141–154, 1990.
D. Müller and E. Stein. On spectral multipliers for Heisenberg and related groups. 73(4):413–440, 1994.
E. M. Ouhabaz. London Mathematical Society Monographs, 31. Princeton, NJ: Princeton University Press, 2005.
E. M. Ouhabaz. Sharp Gaussian bounds and $L^p$-growth of semigroups associated with elliptic and Schrödinger operators. 134(12):3567–3575, 2006.
A. Pazy. Semigroups of linear operators and applications to partial differential equations. 44, New York: Springer Verlag, 1983.
G. Pisier. Some results on Banach spaces without local unconditional structure. 37:3–19, 1978.
G. Pisier. London Math. Soc. Lecture Notes Series, 294: Cambridge University Press, Cambridge, 2003.
E. Stein. Ann. Math. Stud. 63, Princeton Univ. Press, 1970.
E. Stein. Princeton Mathematical Series, No. 30 Princeton University Press, 1970.
R. Strichartz. Multipliers on fractional Sobolev spaces. 16, 1031–1060, 1967.
J. van Neerven. $\gamma$-radonifying operators: a survey. 44:1–61, 2010.
M. Uhl. Spectral multiplier theorems of Hörmander type via generalized Gaussian estimates.
N. Varopoulos. Analysis on Lie groups. 76(2):346–410, 1988.
| {
"pile_set_name": "ArXiv"
} |
[**Ising Models for Inferring Network Structure From Spike Data**]{}
Abstract {#abstract .unnumbered}
--------
Now that spike trains from many neurons can be recorded simultaneously, there is a need for methods to decode these data to learn about the networks that these neurons are part of. One approach to this problem is to adjust the parameters of a simple model network to make its spike trains resemble the data as much as possible. The connections in the model network can then give us an idea of how the real neurons that generated the data are connected and how they influence each other. In this chapter we describe how to do this for the simplest kind of model: an Ising network. We derive algorithms for finding the best model connection strengths for fitting a given data set, as well as faster approximate algorithms based on mean field theory. We test the performance of these algorithms on data from model networks and experiments.
[**1 Introduction**]{}
Now that we can record the spike trains of large numbers of neurons simultaneously, we have a chance, for the first time in the history of neuroscience, to start to understand how networks of neurons work. But how are we to proceed, once we have such data? In this chapter, we will review some ideas we have been developing. The reader will recognize that we are only describing the very first steps in a long journey. But we hope that they will help point the way toward real progress some time in the not-too-distant future.
What is it we want to learn from multi-spiketrain data? Here we will focus on finding the connectivity in the network under study. We hope that ultimately this can help us understand the dynamics of the network and how it implements computations. For example, in the hippocampus and entorhinal cortex, many interesting kinds of neurons (place cells, grid cells, etc.) have been identified, but we do not know how they interact. If we could find out how they are connected, we might be able to get a much better idea of how they compute the animal’s estimate of its location.
Of course, we will not be able to find the full connectivity unless we can record from [*all*]{} neurons, which is impossible, at least with any current techniques. Several orders of magnitude separate the number of electrodes in a recording array and the number of neuron in the area it covers, so any apparent effect of one neuron on another might actually be mediated by unobserved neurons. Thus, we (and everyone else in the field) are forced into using terms like “effective connections” or “functional connectivity.” In this chapter, for brevity, we will generally just write “connections" or “connectivity", but we will always mean that they are “effective” or “functional” connections.
Given this fundamental limitation on our project, it is nevertheless worth remarking that while all connections are functional one, some kinds are more functional than others. Modeling approaches that are closer to biophysical reality stand a better chance of recovering something like the real synaptic connectivity than those which are farther away. The approach we describe here is at least a first step, albeit a modest one, in the right direction.
One can take a rather different attitude, proceeding formally and making a purely statistical model: a mathematical object with some parameters that characterize the data. However, even if such a model fits the data well, one has in general no way to assign biological meaning to the parameters. This is not to say that we don’t care about fit quality – we do. It is only to say that fitting, for example, equal-time neuronal firing correlations in the data perfectly with some model is not very satisfying if the parameters of the model cannot be related fairly directly to the connectivity of the network.
To present the models and formal approaches that we study, we have to start with our notation for the data. We work with time-binned spike trains under the assumption that firing rates are low enough that there is (almost) never more than one spike per bin. Time will be measured in units of the bin size. We denote the spike train of neuron $i$ by $S_i(t)$, where $S_i(t)=+1$ if it fires in bin $t$ and $S_i(t)=-1$ if it does not. (One can equivalently use $S_i = 0$ for no spike, but we won’t do that here.) Thus we can visualize the data as a big array, the “spike matrix” ($N \times T$ if there are $N$ neurons and $T$ time bins), of $+1$s and $-1$s.
This representation of the data lends itself to formulating the problem in terms of a very simple, perhaps the simplest possible model: a McCulloch-Pitts network (McCulloch and Pitts, 1943), or what in physics is called an Ising model. In this chapter we will use several different kinds of Ising models to treat the problem of finding the connections in a network from its spiking history.
[**2 The Gibbs Equilibrium Model**]{}
A possible formal approach to the problem is this (Schneidman et al, 2006): One considers the data as the set of columns on the spike matrix, i.e., all the [*spike patterns*]{} observed, ignoring their temporal order, and asks what distribution they could have been sampled from. Since one is treating all the patterns as coming from the same distribution, one is implicitly assuming stationarity in the data. If one seeks the distribution with the largest entropy, consistent with the sample means $m_i = \langle S_i\rangle$ and correlations $\langle
(S_i-m_i)(S_j-m_j)\rangle$,[^1] one finds something with a familiar look (at least if one is a statistical physicist): P\[[**S**]{}\] = This is the Gibbs equilibrium distribution for the (pairwise) Ising model. Its parameters, $J_{ij}$ and $h_i$, are Lagrange multipliers that are used in the constrained maximization. The parameter $J_{ij}$ is the strength with which unit (neuron) $j$ influences unit $i$, and the bias “field” $h_i$ controls how likely unit $i$ is to be +1 (“fire”) in the absence of the other units. The expectation values of the $S_i$ under the distribution (\[Ising\]), denoted $m_i$ (“magnetizations” in the original context of the model), are related to the neuronal firing rates $r_i$ (in units of spikes/time bin) through $m_i = 2r_i - 1$.
One might like to think of $J_{ij}$ as something like a synaptic strength. However, the $\sf J$ matrix in (\[Ising\]) is necessarily symmetric, $J_{ij} = J_{ji}$. This is true for the couplings in magnets, but it is generally not true for synapses, so we have to be cautious about what $J_{ij}$ means.
A few remarks: In physics, the normalizing denominator $Z$ in (\[Ising\]) is called the partition function, and its log is the negative of the free energy. We are everywhere setting the temperature equal to 1, since changing the temperature just amounts to rescaling the $J_{ij}$s and $h_i$s by a constant factor.
Normally in statistical mechanics, one is given a model and its parameters, and the task is to find the moments $\langle S_i \rangle$, $\langle S_iS_j \rangle$, etc., which are the quantities that can be measured in experiments. But here we have an [*inverse problem*]{}: we are given the measured correlation functions and want to find the parameters of the model.
It has been known for some time how to solve this problem, at least in principle. One should adjust the $h_i$ and $J_{ij}$ to maximize the probability (\[Ising\]), evaluated on the data. Doing this by gradient ascent gives learning rules h\_i &=& ( S\_i \_[data]{} - S\_i \_[[J]{}, [**h**]{}]{}),\
J\_[ij]{} &=& ( S\_i S\_j \_[data]{} - S\_i S\_j\_[[J]{}, [**h**]{}]{}). The averages in the second terms are under the model with the current values of the $J_{ij}$ and $h_i$, and $\eta$ is a learning rate, to be chosen small enough to get smooth convergence. To evaluate them exactly, we have to know the probabilities (\[Ising\]), which requires evaluating the partition function $Z$. To do this exactly one has to sum $2^N$ terms, which is feasible for systems up to $N \approx 20$. For larger $N$ one has to estimate the averages by Monte Carlo simulations of the model. This fact makes the learning slow, especially when working with data from long recordings. One wants the estimates of the model averages to be as good as the direct data averages in the first terms, so the length of each Monte Carlo run has to be about equal to the number of time bins in the data set. It may take many iterations to reach stationary values of the $J_{ij}$ and $h_i$, so this can be a lot of computing. Just what limit this places on the size of systems that can usefully be studied this way depends on patience and the computing resources available, but in our work we found it impractical to try to work with $N > 100$.
These learning rules are a special case of what is called Boltzmann learning (Ackley et al, 1985), which also covers systems with “hidden”, i.e., unrecorded units. In this case, the first terms on the right-hand sides of (\[Boltzh\]) and (\[BoltzJ\]), when evaluated for hidden units, are averages over the model with the visible units clamped to their measured values. This requires even more Monte Carlo runs.
[**2.1 Mean-field methods** ]{}
It is possible to avoid long Monte Carlo runs using mean field methods. Two such methods have been proposed. One is based on heuristic mean-field equations first applied to magnetic systems by Weiss more than a century ago. The other is based on improved equations first proposed by Onsager. They were applied to spin glasses by Thouless, Anderson and Palmer (Thouless et al, 1977), so nowadays in statistical mechanics they are called TAP equations.
Both the original mean-field and TAP equations are approximate partial solutions to the forward problem, i.e., calculating the magnetizations $m_i$. From them, we will derive corresponding solutions to the inverse problem. These solutions are approximate, but become very good in the limits of weak coupling or, for dense connections, large $N$.
The idea of mean field theory is very simple. The exact $m_i$, conditional on a set of $S_j$ connected to $i$ by the coupling matrix $J_{ij}$, is the difference of Boltzmann probabilities p(S\_i=1| {S\_j}) - p(S\_i=-1| {S\_j}) &=&\
&=& ( h\_i + \_jJ\_[ij]{}S\_j). The approximation consists of replacing the $S_j$ inside the tanh by their averages $m_i$: m\_i = ( h\_i + \_jJ\_[ij]{}m\_j). This is apparently a good approximation when the sum on $j$ has many terms (a loose kind of central-limit argument). This is called the mean field limit.
Onsager argued that the contribution to the neighbor magnetizations $m_j$ from the central unit $S_i$ itself should not be counted in estimating the field acting on $S_i$. This leads to corrected mean-field equations, m\_i = . TAP showed that these equations, rather than (\[MFT\]), should be used in spin glasses, where the $J_{ij}$ are random, with a zero or very small mean, because then the Onsager correction term is of the same order as the naive mean field. In this sense, these equations are the correct mean field equations for spin glasses. It has become conventional to call them “TAP equations”. We will use the abbreviation “nMF” for the naive mean field equations (\[MFT\]).
Plefka showed that (\[MFT\]) and (\[TAP\]) are the first two in a sequence of better and better approximations that can be derived systematically (Plefka, 1982), but we will only consider these first two here.
It is convenient to write the TAP equations in the form h\_i = \^[-1]{}m\_i - \_jJ\_[ij]{}m\_j + m\_i\_jJ\_[ij]{}\^2(1-m\_j\^2) The matrix \^[-1]{}\_[ij]{} = = -J\_[ij]{} -2J\_[ij]{}\^2m\_im\_j is the inverse susceptibility matrix. In equilibrium statistical mechanics there is a theorem that the susceptibility matrix is (up to a factor of the temperature, which we are setting equal to 1) equal to the correlation matrix C\_[ij]{} = (S\_i - m\_i)(S\_j - m\_j) Thus, if we know the correlation matrix, we simply need to invert it and solve (for $i\neq j$) ([C]{}\^[-1]{})\_[ij]{} = -J\_[ij]{} -2J\_[ij]{}\^2m\_im\_j for $J_{ij}$ (Kappen and Rodriguez, 1998; Tanaka, 1998; Roudi et al, 2009). This is the TAP algorithm. There are $n(n-1)/2$ quadratic equations to be solved, but they are decoupled. For the original, “naive” mean field approximation, it is even simpler: The last term on the right hand side of (\[inversionformula\]) is absent, and we just have $J_{ij} = -({\sf
C}^{-1})_{ij}$.
[**3 Kinetic Ising Models** ]{}
The Ising model as described so far has no dynamics. It is defined solely by the Gibbs equilibrium distribution (\[Ising\]). Since the system we are trying to understand is a noisy dynamical one, we would like to fit its data – the spike trains – by a stochastic dynamical model. And the Ising model can be given a dynamics in a natural way, as follows (Glauber, 1963). At each time, every neuron has a chance $\propto dt$ of being updated in the infinitesimal interval $[t,t+dt)$ (i.e., neuron updates are independent Poisson processes). For a neuron that is updated, we compute the total “field” H\_i(t) = h\_i(t) +\_jJ\_[ij]{}S\_j(t). (We can allow the external field $h_i$ to depend on time.) We then let $S_i$ take its next value $S_i(t+\Delta t)$ according to the probability P(S\_i(t+t) | { S\_j(t) }) = = \[ 1 - S\_i(t+t) H\_i(t)\]. It is then possible to show that if $h_i$ is independent of $t$ and the matrix $\sf
J$ is symmetric, this model has a stationary distribution given by (\[Ising\]).
Since we would like to identify the $J_{ij}$ with synaptic strengths, which are not symmetric, we relax the symmetry condition. Furthermore, here we will update all the neurons simultaneously instead of randomly asynchronous, because it makes the model easier to apply to time-binned data, such as we are assuming we have here. (We set $\Delta t = 1$ from now on.) With either of these changes, the Gibbs equilibrium distribution (\[Ising\]) is no longer a stationary solution of the dynamics, even if the $h_i$ are time-independent. In the case that they are and $\sf J$ is symmetric, there is a stationary distribution, P\[[**S**]{}\] = (Peretto, 1984). When the $h_i$ are time-independent and $\sf J$ is not symmetric, a stationary distribution may still exist, but it cannot be written down in closed from. Since such a state is not described by the Gibbs equilibrium distribution (\[Ising\]), we call it a non-equilibrium state, even though it is stationary.
[**3.1 Stationary case: exact algorithm** ]{}
For simplicity, we consider first the case where the $h_i$ are time-independent and we assume a stationary distribution of the $S_i$. If we are given a set of spike trains in the form ${\bf S} =\{ S_i(t)
\}$, $1 \le i \le N$, we can derive an algorithm for finding the model parameters $J_{ij}$ and $h_i$ by performing gradient ascent on the log-likelihood of the data under the model, which is L\[[**S**]{}, [J]{}, [**h**]{}\] = \_[it]{} \[S\_i(t+1)H\_i(t) - 2 H\_i(t)\], with $H_i(t)$ given by (\[fieldatt\]). This gives learning rules h\_i &=&\
J\_[ij]{} &=& (Roudi and Hertz, 2011a). Equations (\[dh\]) and (\[dJ\]) have a form analogous to (\[Boltzh\]) and (\[BoltzJ\]) for the equilibrium case: the right hand sides are differences between averages over the data and averages under the current model. However, here the latter can be evaluated directly and quickly from the data, so this algorithm is generally much faster than the equilibrium one.
It is practical in implementing this algorithm to express the neuronal firing variables in terms of the differences $\delta S_i(t) = S_i(t) -
m_i$, with $m_i = \langle S_i(t) \rangle_t$ and to write $H_i(t)$ in the form H\_i(t) = b\_i + \_j J\_[ij]{} S\_j(t) with b\_i = h\_i + \_j J\_[ij]{} m\_j. Then we have b\_i &=&\
J\_[ij]{} &=& { D\_[ij]{} - S\_j(t)\_t }, with D\_[ij]{} = S\_i(t+1) S\_j(t) , the one-step-delayed correlation matrix. The factor in brackets in the time average in (\[dJ2\]) can be recognized as the expectation value of $\delta S_i(t+1)$, given the measured $S_j(t)$, so the second term is the one-time-step delayed correlation matrix for the model, conditional on the data at $t$.
This algorithm is exact, in the sense that for data generated by a kinetic Ising model of this form, it will recover the $J_{ij}$ and $h_i$ exactly, after infinite iterations, for infinite data. (By “infinite data” we mean an infinite number of time steps.)
[**3.2 Stationary case: mean-field methods** ]{}
Like the equilibrium model, this model allows systematic and rapid mean-field and TAP approximations (Roudi and Hertz, 2011a). To derive the mean-field algorithm, we start by assuming that the $m_i$ (of the data) satisfy the mean-field equations (\[MFT\]). Then, on the right-hand side of (\[dJ\]), we write $S_i(t) = m_i + \delta S_i(t)$ and expand the tanh to first order in the $J_{ik}$. First-order terms in $\delta S$ cancel, and we are left with J\_[ij]{} = S\_i(t+1) S\_j(t) - (1-m\_i\^2)\_k J\_[ik]{} S\_k(t) S\_j(t) . If the $J_{ik}$ are the true ones, then we should have $\delta J_{ij}=0$. Thus we obtain a simple matrix equation = [A]{}[J]{}[C]{} where C\_[ij]{}&=& S\_i(t) S\_j(t)\
A\_[ij]{} &=& (1-m\_i\^2)\_[ij]{} The $J$ matrix can be found simply by matrix inversion and multiplication in terms of first- and second-order statistics of the data: = [A]{}\^[-1]{}[D]{}[C]{}\^[-1]{}. Knowing the $J_{ij}$, the inversion task is then completed by solving the MF equations (\[MFT\]) for $h_i$: h\_i = \^[-1]{}m\_i - \_j J\_[ij]{}m\_j.
Likewise, we can get a TAP approximation by assuming that the $m_i$ of the data satisfy the TAP equations (\[TAP\]) and expanding the tanh to third order. After a little algebra, and ignoring some terms which are small for networks with weak, dense connections, we again find the formula (\[MF2\]), but where now A\_[ij]{} = (1-m\_i\^2)(1-F\_i)\_[ij]{}, with F\_i = (1-m\_i\^2)\_k J\_[ik]{}\^2 (1-m\_k\^2). We cannot solve for the $J_{ik}$ directly in this case, since now $A_{ij}$ depends on them through the $F_i$. One way to do it is by iteration, using, say, the mean-field result (\[MFJ\]) in (\[Fdef\]) to get an initial estimate of $F_i$, using this in (\[ATAP\]) to get better $J_{ij}$s, and so on. However in the stationary case this iteration can be circumvented by a trick: Since we can write $J_{ik}$ as $J_{ik}^{nMF}/(1-F_i)$, we can use (\[Fdef\]) to obtain F\_i = , which is a cubic equation that we can solve for $F_i$, given the mean-field $\sf J$. The relevant root is the one that goes to zero as $J_{ij} \to 0$. This root cannot exceed $1/3$, which restricts this method to fairly weak couplings. Still, it is an improvement on the mean-field approximation.
In the same way as for the MF case, once the $J_{ij}$ are found, we can go back to the TAP equations in the form (\[TAPhofm\]) to obtain the $h_i$.
[**3.3 Nonstationary case: exact algorithm** ]{}
Most of the above carries over to the nonstationary case, where the $h_i$ depend on $t$. However, now different time bins are not statistically equivalent, so we require data from (many) runs of the system. Thus, now we will write the spike trains as $S_i(t,r)$, where $r$ labels the different runs and $t$ is the time index within a run. The “magnetizations” are now $t$–dependent averages over runs, m\_i(t) = S\_i(t,r) \_r, and we define $\delta S_i(t,r) = S_i(t,r) - m_i(t)$.
In (\[LL\]) $S_i(t+1)$ and $H_i(t)$ acquire a run argument $r$. The learning rules analogous to (\[db\]) and (\[dJ2\]) are again straightforward to derive by differentiating with respect to $b_i(t)$ and $J_{ij}$: b\_i(t) &=&\
J\_[ij]{} &=& T , where $T$ is the run length and D\_[ij]{} = S\_i(t+1,r) S\_j(t,r) \_[r,t]{}. Note that it is not the same as the $D_{ij}$ (\[Ddef\]) because now the $\delta S_i$ are defined relative to the time-dependent run averages $m_i(t)$, not the time-independent $m_i$.
When we use the word “exact" to describe this algorithm, we mean it in the same sense as for the stationary algorithm, but “infinite data” now means an infinite number of runs to average over in computing the statistics.
[**3.4 Nonstationary case: mean-field methods** ]{}
First one needs time-dependent mean-field and TAP equations. These have been derived recently (Roudi and Hertz, 2011b), using a systematic scheme analogous to that employed by Plefka for the equilibrium case. The dynamic TAP equations for the time-dependent magnetizations take the form m\_i(t+1) = For (naive) mean field, the equations are the same except that the last term inside the brackets is absent.
Now the same procedure that led to (\[MF2\]) gives (Roudi and Hertz, 2011a) D\_[ij]{} = \_k J\_[ik]{}(1-m\_i\^2(t+1))C\_[kj]{}(t)\_t, where C\_[kj]{}(t) = S\_k(t,r) S\_j(t,r)\_r. If we define a set of matrices ${\sf B}^{(i)}$ by B\^[(i)]{}\_[kj]{} = (1-m\_i\^2(t+1))C\_[kj]{}(t)\_t, we can again solve for $\sf J$ by simple matrix algebra: J\_[ij]{} = \_k D\_[ik]{} \[([B]{}\^[(i)]{})\^[-1]{}\]\_[kj]{}. The calculation is a little more time-consuming than the stationary one, because now we have to invert a different matrix ${\sf B}^{(i)}$ for each $i$. Nevertheless, it is still much faster than using the exact algorithm.
For TAP, in the same way as in the stationary calculation, a correction factor $1-F_i$ appears. Now $F_i$ is time-dependent, F\_i(t) = (1-m\_i\^2(t+1))\_k J\_[ik]{}\^2 (1-m\_k\^2(t)), and the $(1-F_i)$-factor enters the time average that defines the matrices ${\sf B}^{(i)}$: B\^[(i)]{}\_[kj]{} = (1-m\_i\^2(t+1))(1-F\_i(t))C\_[kj]{}(t)\_t, Now there is no simple equation one can solve for $F_i(t)$, as there was in the stationary case, so solving for the $\sf
J$-matrix has to be done iteratively.
However, we have found that qualitatively good reconstruction can be done if we make the approximation of factoring the average in (\[defBnonstat\]): (1-m\_i\^2(t+1))(1-F\_i(t))C\_[kj]{}(t)\_t (1-F\_i)(1-m\_i\^2(t+1))C\_[kj]{}(t)\_t, where $F_i = \langle F_i(t)\rangle_t$. Now $F_i$ solves a cubic equation, as in the stationary case: F\_i (1-F\_i\^2)= \_k (J\_[ik]{}\^[nMF]{})\^2(1-m\_i\^2(t+1)) (1-m\_k\^2(t))\_t, and, given $F_i$, one has $J_{ij} = J_{ij}^{nMF}/(1-F_i)$, again as in the stationary case.
[**4 Errors** ]{}
In a Bayesian way, we can interpret the likelihood $\exp L({\sf J},{\bf
b})$ as a probability density for the $J$s and $b$s with a flat prior. This density will peak around the parameters found by the algorithm, and if the data are generated by the kinetic Ising model itself, the peak will lie at their true values in the limit of infinite data. The spread of the probability density around the peak gives a measure of how the parameters obtained from a single data set will vary from one data set to another. Because the log likelihood is proportional to the data set size, it is sufficient, for large $T$ (stationary case) or many runs (nonstationary case) to expand it to second order in the deviations $\delta J_{ij}$ from the values ${\sf J}_0, {\bf b}_0$ at the maximum, so the resulting density is a Gaussian. Suppressing, for simplicity, the dependence on the $b$s and the overall normalization, we find, for the nonstationary case, P([J]{}) \_i ( L([J]{}\_0,[**b**]{}\_0)- RT \_[jk]{} B\_[jk]{}\^[(i)]{}J\_[ij]{}J\_[ik]{}), where $R$ is the number of runs and the $B_{jk}^{(i)}$ are the matrix elements that occurred above (\[defBnonstat\]) in the mean-field algorithm. Here we can see that $RTB_{jk}^{(i)}$ are the elements of the Fisher information matrix of the distribution of $J$s. The factor of $RT$ means that the fluctuations of the $J$s around their central values go to zero like $1/\sqrt{RT}$ for large data sets. This makes more precise our statement that the exact algorithms described above recover the correct model parameters correctly in the limit of infinite data, if those data are generated by our kinetic Ising model.
To get a little more insight, consider the stationary case, where $B_{jk}^{(i)} = T(1-m_i^2)C_{jk}$. The one can see (1) that $J_{ij}$s with different first (postsynaptic) indices are uncorrelated, and (2) $J_{ij}$s with different second (presynaptic) indices will have strongly correlated errors if they have large projections along eigenvectors of $\sf C$ that have small eigenvalues. Thus, elements of $\sf J$ identified as large by the mean field approximation (\[MFJ\]) are exactly the ones for which the estimate is most uncertain. Finally, if correlations in the data are very weak, $\sf C$ is almost diagonal, and J\_[ij]{}J\_[ik]{} = . These considerations show that reconstruction is harder when the data are strongly correlated and when rates are low ($m_i$ near -1).
For the mean-field or TAP algorithms, the distribution of $J$s will have a form like (\[GaussianP\]), but it will not be centered at the true maximum. The mean square errors will be J\_[ij]{}J\_[ik]{} = + (J\_[ij]{}\^0 - J\_[ij]{}\^[true]{}) (J\_[ik]{}\^0 - J\_[ik]{}\^[true]{}) . In cases (such as weak coupling) where the approximations are good, the mean square errors will show the $1/RT$-dependence for small data sets, but for large $RT$ they will level off at a residual value given by the second term in (\[approxerr\]).
If the data come from a different model than our kinetic Ising one, we can never obtain its true parameters, even if we use the “exact” algorithms. In this case we say that the model that generated the data was “unrealizable”. The only “realizable” cases are ones where the model used to make the fit is the same as the one used to make the data. When the two models are in some sense close to each other and the $J$s have a meaning in the true model (for example, if we added a term with weak delayed interactions $\sum_j K_{ij}S_j(t-1)$ to (\[fieldatt\])), we can have a situation like that for the approximate algorithms, with errors that first decrease with data set size and then level off at residual values. But in general all we can say is that we have found the best kinetic Ising model, i.e., the kinetic Ising model most likely to have generated the data.
The real “true” model, i.e., nature itself, is far from our kinetic Ising model. Neurons have much more complex dynamics than our model endows them with, and synapses have dynamics of their own that we have ignored entirely in the present formulation. Thus, in applying our methods to neural data, we have to be rather cautious about interpreting the connection strengths $J_{ij}$ that we obtain.
Furthermore, even if we could replace our simple binary units by models with realistic dynamics and include suitable models for the synapses, we could not expect to recover their parameters correctly from an inference procedure like this, because most of the neurons in the true network are not recorded. A typical experiment records spikes of about 100 neurons in a region of cortex of area about 10 ${\rm mm}^2$, which contains of the order of a million neurons. Hence, even ignoring the fact that neurons in this region are coupled to many outside it, we have data from only about 0.01% of the local network. It is clearly an audacious project to hope to learn about the network connections and dynamics from such limited data (hence the resort to terms like “functional connectivity” to describe the results).
On the other hand, we have to start somewhere, and the simple description of a neuron as a stochastic element that fires at a rate that is a sigmoidal function of its net input is at least qualitatively consistent with neurophysiology. While we cannot expect to associate all the $J_{ij}$s we find with direct synaptic connections, we would be surprised if the strongest $J_{ij}$s we found were not indicative of synapses between recorded neurons (assuming we had chosen the time bin size for our data sensibly, i.e., around the sum of typical synaptic and membrane time constants). Our findings (Roudi and Hertz, 2011a; Sect 10.6.2 below), as well as results obtained applying this approach to genetic regulatory network networks (Lezon et al, 2006), support this optimism.
[**5 Regularization** ]{}
The algorithms above can be given a simple little extra twist that allows us to eliminate the smallest (presumably, least significant) connections in the network in a controlled way. The idea (Tibshirani, 1996; Ravikumar et al, 2010) is to try to minimize a cost function E = -L\[[**S**]{}, [J]{}, [**h**]{}\] +\_[ij]{} | J\_[ij]{} |. In a Bayesian perspective, the addition of this term is the imposition of a prior $\exp(-\lambda \sum |J_{ij}|)$ on the $J_{ij}$. It leads to a new term in the update rules (\[dJ\]), (\[dJ2\]) and (\[dJnonstat\]) for $J_{ij}$ proportional to $-\lambda
{\rm sgn} J_{ij}$. To see its effect, think of the space of $J$s, with an axis for each pair $ij$. Under learning, the extra term pushes the $J$ vector toward the edges of a hyper-octahedron where many of its components vanish. The regularization parameter $\lambda$ is chosen small, so this is a weak effect for large $J_{ij}$. But it is a big effect for $J_{ij} \sim \lambda$ or smaller, and it tends to drive these $J$s to zero. Thus this modification of the learning performs a kind of automatic pruning of the network, the degree of which we can control by adjusting $\lambda$.
A possibility for another kind of regularization arises in using the nonstationary algorithms described above. The source of the time dependence of the driving fields $h_i(t)$ or $b_i(t)$ is the time dependence of the measured $m_i(t)$, i.e., of the firing rates. For realistic amounts of data (perhaps 100 repetitions of the stimulus in an experiment), one cannot expect to estimate a rate within a time bin to better than 10% accuracy. The apparent rates will fluctuate this much from one time bin to the next, and the inferred $h_i(t)$ will inherit these fluctuations. If we believe that the true rates vary more smoothly than this, we can add a further term to the cost function, E\_[smooth]{} = \_[it]{}\[h\_i(t) - h\_i(t+1)\]\^2. The resulting extra term in $\delta h_i(t)$ is proportional to $\kappa [h_i(t-1)-2h_i(t)+h(t+1)]$. This smoothes $h_i$ by averaging $h_i(t)$ a little with $h_i(t\pm 1)$.
[**6 Testing the Models**]{}
We have tested these models on three kinds of data:
\(1) Data generated by the models themselves, where we can measure directly how well they recover the (known) network parameters.
\(2) Data generated by a semi-realistic computational model of a cortical network. This model is much more complex than the kinetic Ising model we try to fit its data with, but we still know the true connectivity of the network.
\(3) Data recorded from a real biological network, in this case of salamander retinal ganglion cells.
[**6.1 Testing the models on themselves**]{}
Testing the algorithms described above on how well they reconstruct the couplings of a kinetic Ising model has two purposes: (1) For the exact algorithms, to verify that they work according to the theoretical picture set out above, in particular, checking that the mean square errors fall off like $1/T$ (stationary case) or $(RT)^{-1}$ (nonstationary case). (2) For the approximate algorithms, evaluating the average residual error in (\[approxerr\]) and studying how it depends on the strength of the couplings in the model. We summarize the findings (Roudi and Hertz, 2011a) here.
Fig. 1 shows mean square errors in the $J$s for a model in which the $J_{ij}$ are chosen independently from a Gaussian distribution of mean zero and standard deviation $g/\sqrt{N}$, where $N$ is the number of neurons in the network. Here we have taken $N = 20$ and all $h_i = 0$. The nMF and TAP errors follow the exact-algorithm result $1/T$ for small $T$ and then flatten out at minimum values (higher for nMF that TAP).
The asymptotic nMF and TAP errors are systematic: For infinite data, nMF systematically underestimates the magnitude of the $J_{ij}$ and TAP overestimates them (though less so). In fact, the factor $1-F_i$ relating the TAP and nMF $J_{ij}$s is the lowest-order correction of the nMF underestimation. We can see this simply, for zero field, by expanding the tanh in (\[dJ\]) or (\[dJ2\]) at $\delta J_{ij} = 0$ to third order in the $J$s (instead of just to linear order as we did in (\[MF1\])). We get D\_[ij]{} = \_k J\_[ik]{} C\_[kj]{} - \_[klm]{}J\_[ik]{}J\_[il]{}J\_[im]{}S\_kS\_lS\_mS\_j+ . (All correlations here are equal-time.) In the sum over $k$, $l$ and $m$, terms with $k=l$, $l=m$ and $m=k$ dominate. Multiplying on the right by the inverse of ${\sf C}$ and using the nMF result (\[MFJ\]), we find J\_[ij]{}\^[nMF]{} = J\_[ij]{} - ( \_k J\_[ik]{}\^2 ) J\_[ij]{} (plus corrections of relative order $1/N$). Now we can use the fact that for our SK-like model, the quantity $\sum_k J_{ik}^2 = g^2$, again ignoring corrections of higher order in $1/N$, so we obtain J\_[ij]{} = . The denominator can be recognized as the TAP correction factor $1-F_i$ when the $h_i=0$. This implies an asymptotic mean square error \_[nMF]{} = (J\_[ij]{}-J\_[ij]{}\^[nMF]{})\^2= . The solid curves in Fig 1a are $1/T
+ \epsilon_{\rm nMF}$, evidently a quite good fit for the values of $g$ from 0.1 to 0.16 for which they are plotted. The black dots in Fig 2 show the nMF $J$s plotted against the true ones. When the size of the data set is increased from $10^4$ to $10^6$, the scatter is reduced and the systematic underestimation becomes clear. In the limit of infinite data, we expect the scatter to disappear, leaving a clean linear behavior near $J_{ij}=0$ with a slope $1-g^2$.
The TAP errors shown in Fig 1b are much smaller that the nMF ones, because they include the lowest-order correction (\[JnMFcorrected\]). We can find the leading error in the TAP approximation by expanding the tanh in (\[dJ\]) to fifth order. A little algebra like that leading to (\[enMF\]) leads to an asymptotic error estimate \_[TAP]{} = . Furthermore, one can show that TAP systematically overestimates the magnitudes of the $J$s, in the same way that nMF underestimated them. For a large network, we expect this to describe the asymptotic TAP error. However, there are finite-size corrections, of order $g^6/N^3$. These would be negligible for $N \gg 1/g^2$, but they are not for the small network ($N=20$) at the values of $g$ studied here. However, taking both (\[eTAP\]) and the finite-size corrections into account, we are able to account for the measured errors (Fig. 1b).
In Fig 2 one can see that the TAP $J$s (square markers) lie much closer to the true ones than the nMF $J$s (\*) do, but a careful look at Fig 2b reveals the systematic overestimation.
Finally, we give an example of nonstationary inference. We generated $R
= 100$ runs, each of length $T= 10^5$ steps for a model with $h_i = 0.5
\cos (2\pi t/100)$. The couplings are small enough ($g = 0.05$) that nMF is quite accurate, as Fig 3a shows. Furthermore, the fields are also quite well recovered (dots, Fig 3c) by solving (\[dynTAP\]) (without the TAP term) for $h_i(t)$.
The $J$s obtained if one does the calculation assuming stationarity are systematically too large (Fig 3b). This is because of the apparent correlations induced in the data by the common external field. In calculating correlations we should actually compute $\langle \delta
S_i(t) \delta S_j(t') \rangle_r$ using $\delta S_i(t) = S_i(t) -
m_i(t)$, with time-dependent $m_i(t)$. However, if we assume stationarity and use $\delta S_i(t) = S_i(t) - \langle m_i(t)
\rangle_t$, instead, we will overestimate the correlations. (This is sometimes called “stimulus-induced correlations”.) These spurious correlations (all positive, in this case) then lead to a spurious increase in the inferred $J_{ij}$s. One can use these spurious $J$s in (\[dynTAP\]) (again, here with the TAP term absent) to infer $h_i(t)$, but because the $J$s are too big, the resulting $h_i(t)$ do not have a modulation amplitude as large as they should (Fig 3c, crosses).
[**6.2 Application To Data From Model Cortical Network**]{}
For a first nontrivial test of the approach, we try it on data generated by a realistic model cortical network (Hertz, 2010). This is a network of 1000 spiking neurons, 80% excitatory and 20% inhibitory, with Hodgkin-Huxley-like intrinsic dynamics. They are driven by an additional external population of 800 Poisson neurons firing tonically at 10 Hz. All neurons, both internal and external, are connected by conductance-based synapses in a random fashion, with a connection probability of 10% (except that the external neurons are not connected to each other). The strengths of these synaptic conductances were chosen so that the network was in a high-conductance balanced state (Amit and Brunel, 1997; van Vreeswijk and Sompolinsky, 1998) with average conductances in the range measured experimentally (Destexhe et al, 2003).
The magnitudes of the synaptic conductances did not vary within a class (excitatory-to-excitatory, excitatory-to-inhibitory, etc.), except for the fact that 90% of them were zero because of the random dilution, though they did vary somewhat in their temporal characteristics. For such a system, a natural and important question to ask is how well we can identify the connections that are actually present.
We choose for analysis the spike trains of the inhibitory neurons with rates over 10 Hz. There were 95 of them. Their mean rate was 32 Hz, with a maximum of 83 Hz. Higher firing rates give better statistics, and the inhibitory synapses are the strongest ones in this model, so this choice makes the task of fitting the model and identifying the connections present easier than it might otherwise be. However, it is useful to test the method on an easier problem before trying to solve a more difficult one. The data were put into the form to be used by the algorithms by binning the spikes with 10-ms bins.
To study how much data is necessary for the fit, we did the analysis using the exact stationary algorithm for data sets of size $T$ from 1000 up to 64000 time bins. For each such $T$, we calculated the log likelihood of the data. For comparison, we also calculated these numbers for independent-neuron models based on the same partial data sets. The results were adjusted with Akaike corrections equal to the number of parameters in the models (Akaike, 1974).
Fig 4 shows the log likelihoods (per time step, per neuron), with and without Akaike corrections. They appear to converge to a value of $-0.306 \pm 0.001$, and evidently there is little point in using more than around 50000 time steps of data. Also shown are the log likelihoods for an independent-neuron model (all $J_{ij} = 0$). For these, the Akaike correction is so small that the corrected and uncorrected curves would fall almost on top of each other, so only the adjusted values are actually plotted. The curve is nearly flat at at value of about -0.53. It is evident that the model with $J$s is much better than the one without them.
![Cortical model data (tonic firing): Log likelihoods as a function of data set size $T$ for fits with kinetic Ising and independent-neuron models. Dashed line: log likelihood. Solid line: Akaike-adjusted log likelihood. Dotted line: independent-neuron model ($J_{ij}=0$). Despite its greater number of parameters, the model with couplings is clearly better, and the difference is evident for all the data set lengths in the range shown.[]{data-label="Fig4"}](Fig4.eps){height="7" width="7cm"}
The Akaike-adjusted log likelihood is a suitable statistic for comparing model quality, but one does not have an clear idea of what a value of -0.306 means for the quality of the network reconstruction, in particular for the problem of identifying the connections present in this diluted network. To do this, we use another statistic, defined as follows.
![Distributions of connection strengths found for cortical model network data for inhibitory connections present in the network (solid line) and connections not present in the network (dashed line). Averages of results of performing nMF inference on 30 sets of 50 neurons chosen randomly from the 95 inhibitory neurons with rates $> 10$ Hz. []{data-label="Fig5"}](Fig5.eps){height="7" width="7cm"}
Consider the values of $J_{ij}$ that the algorithm finds for the connections that are actually present in the network. These estimates will have some spread around their mean value, because of the limited data and the mismatch between the kinetic Ising model and the real network. For the same reason, the values assigned to the $J_{ij}$s for which the connections $\{ ij\}$ are not present, which should be zero, will also be spread around their (small) mean value (Fig 5). If the spreads of these two distribution are small compared to the difference between their means, we can easily identify the true connections, even if we do not know which ones they are [*a priori*]{}. A measure of the difficulty of the task which we can compute (knowing the true connectivity) is the noise/signal ratio d , where the standard deviations and means are over the $J$s found by the algorithm. Fig 6 shows $d$ as a function of data set size $T$, calculated for the $J$s obtained by both the exact algorithm and the nMF and TAP approximate formulae. The TAP and nMF results are too close to each other to distinguish, so only the latter is plotted. The plot in Fig 5 was made using TAP results for $T=200000$ and shows the kind of errors one makes for $d \approx 0.47$: The false positive rate is 5.6% and the false negative rate is 7.2%. For $d \approx 0.3$, which one achieves with the exact algorithm, there are almost never any errors.
![Fitting cortical network data: Noise/signal ratio $d$, as function of data set length $T$ for nMF (dashed) and exact (solid) learning algorithms.[]{data-label="Fig6"}](Fig6.eps){height="7" width="7cm"}
Note, furthermore, that this reconstruction was achieved for a strongly undersampled network – There were 1000 neurons in the network but the reconstructions were done using only 50 of them at a time. This lends some support to our hope that reconstruction, at least for strong synapses, might be possible even though we cannot record from all the neurons in the network.
In this example, we used the stationary versions of the algorithms because the rate of the external driving population was constant. We have also studied a case where the external rate varies in time, performing the corresponding nonstationary analysis, which gives similar results. Results will be published elsewhere.
[**6.3 Application To Data From Salamander Retina**]{}
We have also analyzed a data set provided by Michael Berry of recordings from 40 ganglion cells in a salamander retina. The retina was stimulated 120 times by a 26.5-s movie clip (a total recording time of 3180 s). A nonstationary treatment is natural, in view of the time-dependent stimulus. We expect a significant contribution to the connections identified in a stationary analysis from stimulus-induced correlations, as in Fig 3b.
Such data have been analyzed in the past using the Gibbs equilibrium analysis described in Section 2 (Schneidman et al 2006), which assumes stationarity. It is interesting to see whether the $J_{ij}$s obtained that way are related to those obtained using a stationary kinetic Ising model. They are not expected to be the exactly the same, since the Gibbs $J$s are determined solely by the equal-time correlations and the kinetic Ising ones depend on the one-time-step delayed correlations. Nevertheless, we can hope that if we have chosen the time bin size sensibly, the connections identified by the two methods will be similar.
To find out, we fit both the Gibbs model and the stationary version of our kinetic Ising model to these data. Fig. 7a shows a scatter plot of the resulting two sets of $J_{ij}$s. While they are not so similar as to be concentrated along a straight line in the plot, they do tend to have the same sign (most of the points are in the first or third quadrants).
The kinetic Ising model allows for an asymmetric $\sf J$ matrix, while symmetry is enforced in the Gibbs model. To see how asymmetric the kinetic Ising $\sf J$ matrix is here, we made a scatter plot of the elements of $\sf J$ against those of its transpose (Fig 7b). Evidently, the magnitudes of $J_{ij}$ and $J_{ji}$ can be different, but they do not differ wildly, and they almost always agree in sign. This is consistent with the fact that they agree qualitatively with the fully symmetric $J$s found for the Gibbs model and with our expectation that they may be the result, to a strong degree, of stimulus-induced correlations, which are symmetric by definition.
To find out to what degree, we then performed a nonstationary analysis, with time-dependent $b_i(t)$ in our kinetic Ising model. In addition, we made a fit for a model (also nonstationary) with independent neurons, in which all time dependence of firing rates is explained in terms of the time-dependent $b$s.
We compared these three models on the basis of the log likelihoods of the data under them, with Akaike corrections for the different numbers of model parameters. We found that the stationary model (Akaike-adjusted log likelihood per neuron per time bin $-0.128$ bits) was significantly worse than the nonstationary ones ($-0.0927$ bits for the independent model and $-0.0906$ bits for the full model). The difference between the fits given by the nonstationary models with and without $J_{ij}$s is quite small (0.0021 bits, only 2.4% of the total log likelihood). A full description of this analysis will be published elsewhere.
We conclude that essentially all the connections found for the stationary and Gibbs models come from stimulus-induced correlations. Once we include time-dependent $b$s in the model, adding $J$s gives almost no improvement in the fit. Although it is disappointing not to be able to identify any important intrinsic connections in this network, the result illustrates how the nonstationary analysis can uncover features of the system that stationary models (including the Gibbs model) can not.
[**7 Further Developments** ]{}
In all the above, we restricted the treatment to the simplest kind of kinetic Ising model, for pedagogical purposes. It is not hard to extend the model in various ways to bring it closer to neurobiological reality. The most obvious way is to modify the memory-less dynamics (\[fieldatt\]-\[KIdef\]) by letting the firing probability at $t+1$ depend on spikes earlier than $t$: H\_i(t) = h\_i(t) +\_[j,s = 1]{}\^J\_[ij]{}(s)S\_j(t-s+1). Thus, each connection in the model is characterized by a temporal kernel $J_{ij}(s)$. It describes synaptic and membrane potential dynamics. It is straightforward to derive exact and mean-field learning algorithms for such a model.
Models of this kind, called generalized linear models (GLMs), have been studied in recent years (Truccolo et al, 2004; Okatan et al, 2006). They were used (Pillow et al, 2008) in an extensive study of the signaling by a population of 27 monkey retinal ganglion cells. They have also been used (Rebesco et al, 2010) to track changes in connection strengths in a network in which synaptic plasticity was induced by microstimulation. Thus, the feasibility and utility of such solutions of the inverse problem for real neural networks have been demonstrated. The simple cases described in this chapter can be thought of as “poor man’s” GLMs. Since they have fewer parameters, they may be particularly useful models when data are limited. It may also be useful to extend the mean-field methods discussed here to GLMs.
Another couple of directions in which further development of these models would be useful are the following:
\(1) As mentioned above, one never records from all the neurons in a network. A systematic approach to the inference problem in the presence of “hidden units” is needed. This problem could usefully be explored first for the simplest model (\[fieldatt\]-\[KIdef\]) in a realizable case, to gain some knowledge about how to model the hidden part of the population.
\(2) Even the model with temporal synaptic kernels assumes “current-based” synapses, i.e., that a presynaptic spike causes a particular current to flow in or out of the postsynaptic neuron, independent of the state of that neuron. Actually, what the presynaptic spike causes is the stochastic opening and closing of a channel, selective for particular ions, in the postsynaptic membrane. The synaptic current (or, more precisely, its average over many presynaptic transmitter release events) is then the product of the conductance of the channel if it were open, the time-dependent probability that it is open, conditional on the presynaptic spike, and the difference between the instantaneous postsynaptic membrane potential and the reversal potential for that channel. The difference between such “conductance based” synapses and current-based ones may be quantitatively small for excitatory synapses, since their reversal potentials are far above the firing threshold, but they are not for inhibitory ones, whose reversal potentials are not much lower than typical subthreshold membrane potentials. It would be desirable and interesting to extend the kind of modeling here to include a little of this potentially relevant biophysics, starting perhaps with the limiting case of “shunting inhibition”.
\(3) Finally, we mention a new mean-field inversion algorithm (M[é]{}zard and Sakallariou, 2011; Sakallariou et al, 2011) which reconstructs kinetic Ising models exactly when the $\sf J$ matrix is completely asymmetric. It would be interesting to see how well it works on data from more realistic models or experiments.
Ackley D, Hinton GE, Sejnowski TJ (1985) A learning algorithm for Boltzmann machines. Cognitive Science 9:147-169.\
Akaike H (1974) A new look at the statistical model identification. IEEE Trans Aut Control 19:716Ð723.\
Amit DJ, Brunel N (1997) Model of global spontaneous activity and local structured activity during delay periods in the cerebral cortex. Cerebral Cortex 7:237-252.\
Destexhe A, Rudolph M, Par[é]{} D (2003) The high-conductances state of cortical neurons [*in vivo*]{}. Nature Reviews Neurosci 4:739-751.\
Glauber RJ (1963) Time-dependent statistics of the Ising model. J Math Phys 4:294-307.\
Hertz, J (2010) Cross-correlations in high-conductance states of a model cortical network. Neural Comp 22:427-447.\
Kappen HJ, Rodriguez FB (1998) Efficient learning in Boltzmann machines using linear response theory. Neural Comp 10:1137-1156.\
Lezon TR, Banavar JR, Cieplak M, Maritan A (2006) Using the principle of entropy maximization to infer genetic interaction networks from gene expression patterns. Proc Nat Acad Sci USA 103:19033Ð19038.\
McCulloch WS, Pitts W (1943) A logical calculus of ideas immanent in nervous activity. Bull Math Biophys 5:115-133.\
M[é]{}zard M, Sakallariou J (2011) Exact mean field inference in asymmetric kinetic Ising systems. arXiv:1103.3433.\
Okatan M, Wilson MA, Brown EN (2005) Analyzing functional connectivity using a network likelihood model of ensemble neural spiking activity. Neural Comp 17:1927-1961.\
Peretto P (1984) Collective properties of neural networks: a statistical physics approach. Biol Cybern 50:51-62.\
Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, Simoncelli EP (2008) Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature 454:995-999.\
Plefka T (1982) Convergence conditions of the TAP equation for the infinite-ranged Ising spin glass model. J Phys A 15:1971-1978.\
Ravikumar P, Wainwright M, Lafferty JD (2010). High-dimensional Ising model selection using $\ell_1$-regularised logistic regression. Ann Stat 38:1287-1319.\
Rebesco JM, Stevenson IH, K[ö]{}rding KP, Solla SA (2010) Rewiring neural interactions by micro-stimulation. Frontiers System Neurosci 4:39.\
Roudi Y, Hertz J (2011a) Mean-field theory for nonequilibrium network reconstruction. Phys Rev Lett 106:048702.\
Roudi Y, Hertz J (2011b) Dynamical TAP equations for non-equilibrium spin glasses. J Stat Mech P03031.\
Roudi Y, Tyrcha J, Hertz J (2009) The Ising model for neural data: model quality and approximate methods for extracting functional connectivity. Phys Rev E 79:051915.\
Sakellariou J, Roudi Y, M[é]{}zard M, Hertz J (2011) Effect of coupling asymmetry on mean-field solutions of direct and inverse Sherrington-Kirkpatrick model. arXiv:1106.0452.\
Schneidman E, Berry MJ II, Segev R, Bialek W (2006) Weak pairwise correlations imply strongly correlated network states in a neural population. Nature 440:1007-1012.\
Tanaka T (1998) Mean-field theory of Boltzmann learning. Phys Rev E 58:2302-2310.\
Thouless DJ, Anderson PW, Palmer RG (1977) Solution of ‘solvable model of a spin glass’. Phil Mag 35:593-601.\
Tibshirani R (1996) Regression shrinkage and selection via the lasso. J Roy Stat Soc B 58:267-288.\
Truccolo W, Eden UT, Fellows MR, Donoghue JP, Brown EN (2004) A point process framework for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects. J Neurophysiol 93:1074-1089.\
Van Vreeswijk C, Sompolinsky H (1998) Chaotic balanced state in a model of cortical circuits. Neural Comp 10:1321-1371.
[^1]: Statisticians would call these covariances, but here we follow the convention in statistical physics and call them correlations.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Our scientific knowledge is increasingly built on software output. User code which defines data analysis pipelines and computational models is essential for research in the natural and social sciences, but little is known about how to ensure its correctness. The structure of this code and the development process used to build it limit the utility of traditional testing methodology. Formal methods for software verification have seen great success in ensuring code correctness but generally require more specialized training, development time, and funding than is available in the natural and social sciences. Here, we present a Python library which uses lightweight formal methods to provide correctness guarantees without the need for specialized knowledge or substantial time investment. Our package provides runtime verification of function entry and exit condition contracts using refinement types. It allows checking hyperproperties within contracts and offers automated test case generation to supplement online checking. We co-developed our tool with a medium-sized ($\approx$3000 LOC) software package which simulates decision-making in cognitive neuroscience. In addition to helping us locate trivial bugs earlier on in the development cycle, our tool was able to locate four bugs which may have been difficult to find using traditional testing methods. It was also able to find bugs in user code which did not contain contracts or refinement type annotations. This demonstrates how formal methods can be used to verify the correctness of scientific software which is difficult to test with mainstream approaches.'
author:
- Maxwell Shinn
bibliography:
- 'paranoid-submitted-refs.bib'
title: Refinement type contracts for verification of scientific investigative software
---
Introduction
============
Over the last several decades, software engineering has made great strides in developing tools and processes for verifying the correctness of computer software. Although verification has been strikingly successful across many different domains, it has not been widely applied to scientific software. It has been estimated that 5%-100% of scientific software output is incorrect due to undetected software bugs [@Soergel2015], and this is evident from the many retractions caused by undetected bugs[^1]. Some of these retractions were covered in depth by the popular press. In one case, a researcher discovered that the results defining his career were built on a software bug, forcing him to retract five of his most important papers from top journals [@Miller2006]. In another case, a software bug reversed the results of highly influential economics research which was widely cited in public policy decisions [@Herndon2013]. Section \[examples\] will provide concrete examples which illustrate common sources of such bugs.
One ubiquitous class of scientific software has received little to no attention from the verification community: this software is user code which is characterized by (a) small amounts of code (usually 50-500 LOC) (b) written by domain experts with little to no formal training in software engineering [@Hannay2009; @Chilana2009; @Saltelli2014], which is (c) run a limited number of times, (d) has specifications that change unpredictably on a daily or hourly basis, (e) is used almost exclusively by the original developer, and (f) has no testing oracle because the output of the software is the object of investigation [@Hatton1994; @Kanewala2013; @Soergel2015]. We refer to this class of software as “investigative software”. Investigative software is written on a daily basis by countless researchers across the natural and social sciences. Some common examples of investigative software include: scripts to load experimental data and perform statistical tests using statistical libraries; simulations of a computational model; a pipeline which performs complicated preprocessing operations on input data; or a script used internally to make business decisions. This paper presents a tool which is used to verify investigative software.
Testing and verifying correctness of investigative software is difficult due to many technical and cultural factors including unclean data, the lack of a testing oracle, and the insufficiency of standard testing procedures for the structure and goals of scientific programs [@Kanewala2014; @Hook2009; @Johanson2018; @Joppa2013; @Kamali2015; @Sanders2008]. There already exist methods in the scientific community for checking program correctness, but they have serious limitations which limit their utility as discussed in Section \[otherwork\]; as a result, investigative software is most commonly validated by determining whether the program’s output matches the expectations of the scientist, posing a fundamental violation of the scientific method [@Sanders2008]. A more convenient and effective method for ensuring the quality of scientific software is needed.
Formal methods are able to check the correctness of investigative software. However, state of the art formal verification tools are time-consuming to implement and require substantial formal methods expertise, which make their use impractical for investigative software. A survey estimated that scientists spend about 30% of their time writing software [@Hannay2009], so the need to formally verify this code is a substantial hurdle to productivity.
By contrast, runtime verification is easy to use and especially well-suited to investigative software. It requires little training to use, yet is able to achieve many of the same goals; because investigative software is usually developed by its sole user, it doesn’t matter whether bugs are caught during development or detected at runtime through verification condition violations. This reflects the fact that investigative software development often cannot be meaningfully separated from data analysis [@Paine2017; @Hook2009; @Kelly2015; @Sanders2008]. Researchers need access to lightweight formal methods to improve the correctness of their software without slowing down the research process.
We created Paranoid Scientist[^2], a Python library for verifying the correctness of investigative software. Paranoid Scientist employs runtime verification to check software correctness. Developers specify function behavior through entry and exit conditions, thereby creating function contracts which must be satisfied by each function execution [@Barnett2003; @Hatcliff2012]. Contracts are specified in two parts. First, each function argument and the return value are specified modularly using refinement types [@Freeman1994; @Vazou2014], which are types defined by predicates. Second, additional constraints such as a dependence of the return type on the function arguments or the function argument types on each other are specified using predicates written in pure Python. These constraints may also depend on previous calls to the function in order to check hyperproperties. In addition to runtime checking, Paranoid Scientist may further use the refinement type contracts to automatically generate test cases. Critically, Paranoid Scientist uses simple syntax and is intuitive to those without a background in formal methods or software engineering. We ran our tool on real-world investigative software and found that it was able to catch four undetected bugs while imposing a 1.05–6.41 factor performance penalty.
Motivating examples {#examples}
===================
We provide two motivating examples from real-world investigative software written for research in the biomedical sciences.
### Incorrect function usage
Figure \[fig:matmult\] shows an example bug in a function designed to find the reverse complement of a DNA sequence. Briefly, DNA is the primary medium for long-term information storage in biological organisms. Each strand is composed of long sequences of four “base” molecules—adenine (), guanine (), cytosine (), and thymine (). For chemical stability, each strand of DNA within an organism is usually bound to a “reverse complement” strand whereby the and bases and the and bases are swapped and the resulting sequence is reversed. By contrast, RNA is a biological medium for short-term information storage. It shares a nearly identical structure with DNA, but replaces with uracil (). Consequently, when forming the RNA reverse complement, is swapped with instead of .
The function in Figure \[fig:matmult\] uses a simple mechanism to compute the reverse complement DNA sequence. It accepts a list of characters and, using Python’s equivalent of the Unix program “tr”, it converts all to , to , to , and to , and then reverses the result. While this function behaves as intended, an erroneous usage of the function occurs when the user tries to find the reverse complement of an RNA sequence instead of a DNA sequence. The function accepts the valid RNA sequence and returns another valid RNA sequence without error. However, is not the reverse complement of , because the should be replaced with an . The function is only designed to operate on DNA sequences, and thus cannot properly deal with the RNA base .
[0.52]{}
``` {basicstyle="\scriptsize\ttfamily"}
# Find the reverse complement of
# DNA. Assumes `seq` is a list of
# chars "A", "G", "C", and "T".
def complement_sequence(seq):
# Convert `seq` to string and
# use regex translate
c_str = ''.join(seq).translate(
str.maketrans('AGCT', 'TCGA'))
return list(reversed(c_str))
# Correct usage, return ['G','A','T']
complement_sequence(['A','T','C'])
# BUG: Called for RNA instead of DNA,
# Return ['C','G','U'], however U
# is not its own complement.
complement_sequence(['U','C','G'])
```
[0.51]{}
``` {basicstyle="\scriptsize\ttfamily"}
# Perform CPU-intensive preprocessing
timeseries = preprocess_ts(timeseries)
# Find timeseries pairwise correlations
corr_matrix = corrcoef(timeseries)
# Fisher z-transform (arctanh).
# Assumes abs(correlation) <= 1
norm = fisher_transform(corr_matrix)
save_csv("matrix.csv", norm)
corr_matrix = load_csv("matrix.csv")
# BUG: normalized before saving,
# these values violate assumptions
# of the fisher_transform function.
norm = fisher_transform(corr_matrix)
# Convert matrix to undirected graph
G = matrix_to_graph(norm)
```
Paranoid Scientist is able to ensure correctness in this example. An annotated version of the function is shown in Figure \[fig:matmultfixes\]. Most importantly for this example, the decorator checks that the argument value is a list consisting only of values , , , and . This means that passing an RNA sequence to the function in Figure \[fig:matmult\] will raise an error, because the “U” element in the list is invalid for DNA. Even if entry conditions were not specified, the “U” would have been caught as an invalid output type by the return type specification in the decorator, and by the exit condition in the decorator which specifies that all values of the input must be different than the corresponding values of the reversed output. These properties are checked at runtime to ensure the function is receiving correct input and producing correct output.
[0.56]{}
``` {basicstyle="\scriptsize\ttfamily"}
@accepts(List(Set('AGCT')))
@returns(List(Set('AGCT')))
@ensures("all(seq[i] != return[::-1][i] \
for i in range(0, len(seq)))")
def complement_sequence(seq):
...
```
[0.45]{}
``` {basicstyle="\scriptsize\ttfamily"}
@accepts(NDArray(t=Range(-1, 1)))
@returns(NDArray(t=Number))
@ensures("return.shape == \
corr_values.shape")
def fisher_transform(corr_values):
...
```
### NaN propagation
Another example bug is shown in Figure \[fig:fmri\], which is a condensed version of a bug from real-world investigative software. This code is the final step in a pipeline which converts a functional magnetic resonance imaging (fMRI) scan—a type of brain scan which allows researchers to look at brain activity over time—to an undirected graph where edges represent strong correlations in brain activity [@Bullmore2009]. First, several computationally-intensive preprocessing steps are applied to the timeseries, and then the pairwise Pearson correlation of each region is computed. Pearson correlation is a value from -1 to 1 inclusive, which is converted to a more statistically-informative value from negative infinity to infinity using the Fisher z-transform, or equivalently hyperbolic arctangent. This is then saved to a file so that the computationally intensive steps do not need to be repeated.
After saving, the timeseries can be reloaded and turned into an undirected graph. However, before doing so, the function is erroneously applied a second time. The second time it is called, the inputs hold values from negative infinity to infinity. This is outside of the domain of the hyperbolic arctangent function, causing it to return NaN for inputs less than -1 or greater than 1. In practice, this gives reasonable values for small- and medium-valued correlations but NaN for large correlations. These NaN values are masked by the function , returning a graph which does not show evidence of the NaN values; this creates a graph which appears to be correct for all but the largest correlations. This bug therefore created subtle changes in the resulting graph topology which were not noticed immediately, causing several weeks of work analyzing the resulting graphs to be lost.
Paranoid Scientist is able to detect this bug. Figure \[fig:fmrifixes\] shows annotations for the function. This function takes any $N$-dimensional array () with values ranging from -1 to 1, and returns an $N$-dimensional array of the same shape with elements which are numbers. NaN is not a valid number. These annotations would have been sufficient to catch the bug in Figure \[fig:fmri\]. Alternatively, annotations which specified the valid input of would have also been able to detect this bug.
If either of these two bugs had appeared in software which had an oracle, there is a high probability that the difference in behavior would manifest as an observable bug and the behavior could be corrected [@Kanewala2013]. Investigative software is written because the result is unknown, so in these cases the bugs may never have been found.
Package summary
===============
Refinement types
----------------
Function entry and exit conditions are specified in part by refinement types. In our tool, refinement types are defined by a predicate which checks whether an input is an element of the type. Predicates are constructed by Python functions using purely Python code, and thus may reach arbitrary levels of complexity without depending on a domain specific language. This allows types to be defined in terms of their scientific purpose and conceptual properties instead of as datatypes [@Vazou2014]. These refinement types are akin to what one would write when documenting the function. For example, one type defined by default is “Number”, which can be a float or an int but not NaN or $\pm$inf. Types can also represent more complex properties. For example, a discrete probability distribution is a list with non-negative elements which sum to 1, and a correlation matrix is a symmetric positive semidefinite matrix with 1 on the diagonal. A list of all types included by default is included in the Appendix. Any class can be used in place of a type by checking whether the passed value is a subclass of the given class. Alternatively, classes can define a method to determine whether the passed value is an instance of the class, described in further detail in Section \[syntax\].
Entry and exit conditions
-------------------------
In addition to refinement types, additional entry and exit conditions may be specified for conditions involving multiple function arguments or involving function arguments and the return value. For example, there may be a constraint that the first argument is greater than or equal to the second argument, or that the function returns a matrix with the number of rows and columns specified by the input arguments. Conditions are specified as a string which is evaluated as Python code.
Function properties may depend on more than a single execution of the function. For example, function concavity and function monotonicity are hyperproperties which cannot be determined at runtime when considering a single function execution. Paranoid Scientist saves in memory a list of arguments and return values from previous function executions. All future function executions are compared against these past values. For functions which are executed many times, a naive implementation would cause serious performance and memory penalties, limiting the practicality of this feature. We address this problem by saving only a subset of function calls and using reservoir sampling [@Vitter1985] to test against a uniform distribution across all function calls; as a result, verification of hyperproperties is not performed across all previous calls but rather checked across a sample of previous calls which is uniformly-distributed across time. An example of one such hyperproperty is shown in Figure \[fig:hyper\].
``` {basicstyle="\scriptsize\ttfamily"}
@accepts(Number)
@returns(Number)
@ensures("t >= t` --> return >= return`") # Monotonic function
def cube(t):
return t**3
```
Syntax
------
Refinement types for function arguments and return values are specified using the and function decorators, respectively. Further entry and exit conditions are passed as strings of Python code to the and function decorators, respectively. The strings are evaluated using a namespace which includes the function arguments and additional user-specified libraries. In , the special value represents the return value of the function. For testing hyperproperties, function arguments and return values from previous executions can be accessed by appending one or more backtick characters as a suffix, a notation which is reminiscent of the “prime” symbol from mathematics. Additionally, syntactic sugar is available for the two common idioms “implies” with the syntax and “if and only if” with the syntax.
In addition to the default types, refinement types may be defined manually. Types are classes which define two methods: one to test values for adherence to the type, and a second to generate values of the type for use in automated testing, as described below. The type definition may optionally accept arguments to specify parameterizable behavior or generics. Any existing class can also be used as a type by using the class name in place of a refinement type. In these cases, Paranoid Scientist only tests whether the element is an instance of the class, where adherence to the Liskov substitution principle is assumed by default, i.e. subclasses are also considered to be elements of their parent class. For more precise control over the checking, class methods may be defined analogous to the methods for testing and generating values in stand-alone refinement types, allowing a single class to serve both as a normal class and also as a refinement type.
Automated testing
-----------------
The use of entry and exit conditions for each function makes it possible to perform unit tests automatically. A stand-alone command line utility takes the program to be checked as input and individually tests each function in the program with values generated using a specialized method from the refinement type specifications, similar to fuzz testing [@Duran1984]. These generated values are passed as arguments to each function as long as the values satisfy the function’s entry conditions. Because investigative software very seldom includes tests, this increases robustness.
Not all functions can be tested. Those with unspecified types, strict entry conditions, or arguments which cannot be automatically generated will not produce any test cases. Likewise, some tests run for a very long time under certain parameterizations; these are killed after some designated time duration to balance correctness with the practical constraints of testing. Paranoid Scientist will report to the user a list of the functions which could not be tested so that these may be targeted for further testing.
More detailed information about syntax and automated testing is available in the package documentation.
Performance evaluation
======================
We evaluate the runtime performance of Paranoid Scientist on several examples programs drawn from investigative software in cognitive neuroscience.
Design matrix construction (`design`)
: Construct a design matrix for a generalized linear model, similar to the analysis performed in [@Park2014].
Nodal versatility (`versatility`)
: Compute the versatility [@Shinn2017] of a node in an undirected graph with respect to a community detection algorithm.
Decision-making simulation (`pyddm_sim`)
: Simulate decision-making using the PyDDM software package (see Section \[sec:pyddm\]).
Decision-making fitting (`pyddm_fit`)
: Fit a decision-making model to simulated data using the PyDDM software package (see Section \[sec:pyddm\]).
Performance benchmarks are shown in Table \[tab:performance\]. Overall, annotations comprised between 25-30% of the lines of code. There is a performance penalty for runtime verification, and this penalty varies depending on the details of the code. This penalty falls within the ranges suggested for “deliverable” (a 3x slowdown) or “usable” (10x slowdown or less) in runtime checks [@Takikawa2016].
`design` `versatility` `pyddm_sim` `pyddm_fit`
-------------------------- ------------------- -------------------- -------------------- -------------------
Total LOC 162 86 5 20
Program LOC 117 68 5 20
Annotation LOC 45 18 (user code) (user code)
% Code annotations 28% 26% 0% 0%
Runtime w/ checking (s) $8.452 \pm 0.052$ $29.047 \pm 0.158$ $13.685 \pm 0.057$ $4.726 \pm 0.160$
Runtime w/o checking (s) $3.239 \pm 0.022$ $27.606 \pm 0.205$ $2.136 \pm 0.013$ $3.066 \pm 0.121$
Slowdown factor 2.61 1.05 6.41 1.54
: **Performance benchmarks.** Runtime for each example program is shown with standard error of the mean over 10 runs.
\[tab:performance\]
These examples were derived from real-world investigative software. Notably, a previously undetected bug was found in the `design` example when Paranoid Scientist annotations were added to it, which caused incorrect binning of data before performing the regression analysis.
Case study {#sec:pyddm}
==========
We used our tool while developing PyDDM[^3], a decision-making simulator for cognitive neuroscience. PyDDM’s development was initially intended for a specific series of studies and was later released to other research groups [@Gewaltig2014]. Overall, Paranoid Scientist annotations comprised about 10% of the codebase. Over 95% of these annotations require refinement types which cannot be checked using existing static type checkers for Python. Hyperproperties were specified for less than 5% of function exit conditions.
We briefly describe the motivation for PyDDM. PyDDM aids in the study of a simple form of decision-making whereby two options are presented and the subject must choose one of the two based on either preference or matching a given stimulus. This type of decision-making is often studied using the drift-diffusion model (DDM), which posits that all decisions consist of some underlying evidence signal plus noise in continuous time [@Ratcliff2008]. The process of making a decision relies on integrating evidence for each option over time and coming to the final decision when the total integrated evidence surpasses some confidence criterion. Decision-making is usually studied for simple decisions over a short duration of time (<5 sec), such as for determining whether a sock is black or dark blue, but the model can also be used for decisions which may span days or weeks, such as deciding between two job offers. In either case, there is a trade-off between the time it takes to make the decision and the accuracy expected.
The DDM represents evidence integration as a diffusion process governed by the first passage time of a stochastic differential equation across a boundary. Analytical solutions for the DDM are fast to compute, but only specific versions of the model can be solved analytically. Recent experiments have found that humans and animals exhibit behavior which differs from these specific versions of the model. In order to explain experimental data, it is necessary to use a more general version of the model which can only be solved numerically. PyDDM provides a consistent interface to a collection of analytical and numerical algorithms for solving the generalized DDM, selecting the best routine for each model.
As a result, PyDDM contains a mix of optimized routines for finding solutions to stochastic differential equations, and a set of object-oriented interfaces to make these routines convenient to use. This includes predefined models for the most common use cases. Consequently, PyDDM must be used in conjunction with user code written by neuroscience researchers who may have limited experience in software engineering. Thus, the goals of verification are two-fold: detecting errors in PyDDM itself, and detecting errors in PyDDM user code.
Detecting errors in PyDDM
-------------------------
We found four non-trivial bugs in PyDDM using Paranoid Scientist. Briefly, the bugs were:
1. Under certain circumstances, a small function which was assumed to return a positive number would return a negative number due to a typo in a mathematical equation. This was caught by a return type of “Positive”.
2. An algorithm assumed that the output of a previous step in the processing pipeline yielded a vector with 0 as the first element. It was intuitive that this should be the case. When the numerical algorithm was upgraded, this no longer held true. Only one distant branch in the pipeline relied on this value being zero. This was caught by a precondition specifying the first element of the array should be 0.
3. When a model of a particular form was fit to data, a discretization approximation in a non-central portion of the code would exploit the limited numerical resolution of the simulation in order to select unnatural parameters which in turn artificially inflated the model’s performance metrics. This was caught by a postcondition which checked that the distribution would integrate to 1.
4. Particular inputs caused numerical instabilities in one of the three simulation methods and made the probability distribution contain values slightly less than 0. This was caught by a precondition on a different function which required all elements of the input array to be greater than 0.
Bugs (1) and (2) would have been very difficult to detect without our tool, and detecting bug (3) would have required manually examining a large amount of intermediate output. These three bugs would have slightly impacted scientific results. Bug (4) would have likely been noticed eventually but would have caused a substantial time investment to locate. In addition to these bugs, Paranoid Scientist was able to detect an internal inconsistency in how data were stored. Though this did not manifest in a bug which affected results, it had the potential to do so in the future.
Detecting errors using traditional methodology
----------------------------------------------
In addition to Paranoid Scientist annotations, unit tests and manual code review were used to catch bugs in PyDDM. One non-trivial bug was found in unit testing which was not detected by Paranoid Scientist, but this bug did not impact results:
1. When the core simulator’s representation of a probability distribution was extended to include support for storing the distribution of incomplete trials, this unexpectedly modified the behavior of a distant piece of the code which utilized that representation.
Additionally, two non-trivial bugs were detected through code review which neither Paranoid Scientist nor unit/integration tests were able to catch; these bugs also did not impact results:
1. When constructing a diffusion matrix as a part of the core simulation routine, certain rare but important cases utilized the previous timestep instead the next timestep.
2. A simulation algorithm is automatically chosen for each model, but this choice was suboptimal for a small number of models.
Detecting errors in user code
-----------------------------
In addition to PyDDM’s core library code, a key feature is its extensibility with user code to define new models. Due to the complexity of models which can be defined by users, it is important to catch errors in user code even if the users do not use Paranoid Scientist annotations.
Paranoid Scientist was able to find three bugs in user code, even though this code did not have Paranoid Scientist annotations. All three of these would have impacted results:
1. Two subjects completed an experimental task with different task parameters, but parameters were mixed because the expression should have been . This caused a parameterization which was valid on its own but not within the context of the data. This was caught by a precondition which checked that one parameter was less than or equal to all elements of a data array.
2. A user-defined function to generate discrete probability distributions sometimes produced an invalid distribution. This was caught by a precondition checking that the distribution summed to 1.
3. Boundaries were initialized randomly according to a normal distribution. However, sometimes these bounds would be erroneously initialized such that some mass of the initial probability distribution had already crossed the bounds. This was caught by a precondition which checked that two input vectors were the same size.
Limitations
===========
Runtime checking imposes penalties on the program’s speed. Paranoid Scientist has not been optimized for speed, though such optimization is possible in the future. Previous work has demonstrated improved performance of runtime checks through a client–server architecture [@Dimopoulos2015] and through optional contracts [@Dimoulas2013]. Performance could also be improved by producing a certificate during runtime which can be checked after execution.
Paranoid Scientist is compatible with all Python features and does not require the programmer to limit herself to a Python subset or to use a wrapper of the Python executable. Nevertheless, some less-commonly used Python features may cause problems if incorporated into contracts due to the present implementation of runtime checking, especially if these features are stateful. For example, Python objects are allowed to change their value when accessed, but this violates the assumptions of runtime checking. Likewise, contracts cannot yet be specified for generators. If these features are needed in contracts, Paranoid Scientist includes the type as an alternative.
During runtime, Paranoid Scientist is able to implicitly deal with side effects relevant to the results of the computation such as modifications to global state and file IO. However, automated tests are unable to deal with these side effects. Paranoid Scientist does not have an explicit model of these or other side effects such as exceptions and printing, because a clear specification of these is seldom critical for investigative software.
Python’s syntax for type annotations[^4] provides a convenient way to specify types. Paranoid Scientist uses function decorators instead of type annotations. Type annotations would be suitable for the and decorators, but not for the or decorators, so using decorators for all of these improves syntax consistency. The use of decorators allows type annotations to be used for other purposes in the same codebase, and may avoid confusion among less-experienced Python programmers who are not used to this new syntax, or among users who run older versions of Python.
Related work {#otherwork}
============
### Formal methods for scientific software
The present focus on investigative software differs from previous work on verifying scientific software, which focuses on floating point operations [@Boldo2007; @Goubault2011] or high performance computing [@Gunnels2001]. These tools are effective for specific types of scientific software, but the methodology they impose does not reflect the environment in which most investigative software is written and used [@Sanders2008]. Besides formal methods, prior work on testing scientific software does not focus on investigative software; instead, it focuses on large or collaborative software projects [@Gewaltig2014; @Carver2007; @Sarma2016], software written by seasoned software engineers instead of researchers with limited formal training [@Lundgren2016; @Kanewala2013; @Heaton2015], software without an oracle [@Kanewala2013], or computationally- or numerically-intensive software [@Weyuker1982; @Hochstein2008; @Clune2011].
### Testing scientific software
Several recognized methods exist for software testing in the scientific community, but these methods have serious limitations. One method involves rewriting a piece of software one or more times by independent parties and comparing the output for identical input [@Weyuker1982; @Patel2017]. In practice this is not feasible for most investigative software, due to the fact that it is written by a single individual for a limited number of executions. Another method is running the software with simplified parameters or artificial data for which the result is known [@Weyuker1982; @Patel2017]. This leaves the most scientifically important pieces of the code untested, and it is often difficult to determine equivalence of the software’s output with the known result due to stochasticity or floating point arithmetic [@Weyuker1982; @Kanewala2014]. Meta-morphic testing has been proposed as an alternative means of testing software without an oracle, which involves testing specific properties which are required to hold [@Patel2017; @Chen2009; @Giannoulatou2014; @Chen1998]. This requires a deep knowledge of testing methodology and a code structure which facilitates such tests.
### Python type systems
In recent years there has been a proliferation of static type checkers within the Python ecosystem, starting with MyPy[^5] and continued by Facebook’s Pyre[^6], Google’s PyType[^7], and Microsoft’s PyRight[^8]. This has been further advanced by PEP 484[^9], which standardized a syntax for type annotations for functions in the Python language. These type checkers introduce neither overhead nor speedups, as the types are checked before and not during runtime. Thus, the type information is not enforced during program execution.
Reticulated Python [@Vitousek2014] includes runtime checks for annotated types using three different methods. The two bugs described in [@Vitousek2014] which were caught using Reticulated Python would have occurred as exceptions without the runtime checks; with the present work, we are more interested in bugs which might have otherwise gone undetected. One of Reticulated Python’s modes of operation can insert undetected bugs into the code by not preserving object identity, demonstrating the different objectives between Reticulated Python and the present work. Additional packages for runtime checking of static data types in Python, such as the “enforce”[^10] or “typeguard”[^11] packages, share many similarities to the “transient” method in Reticulated Python.
### Python contracts libraries
While contracts were first conceived for Python in the language’s infancy [@Ploscha1997], contracts are still rare in Python code. The most popular contract library for Python is “PyContracts”[^12], which embeds a domain specific language into Python for specifying properties that each argument must satisfy. It is difficult to specify complex properties or to create properties which rely on more than one argument, such as “argument 1 is greater than argument 2”. PyBlame [@Arai2016] provides a sophisticated contract library for Python which integrates with the debugger, but a detailed comparison could not be performed due to the lack of availability of the PyBlame source code. Data validation libraries, such as “cerberus”[^13] and “voluptuous”[^14], ensure that datasets satisfy particular conditions, and thus may be used in conjunction with Paranoid Scientist.
Nagini is a package which provides full static verification for a Python subset [@Eilers2018]. Python code is converted to an intermediate language and conditions are specified using contracts. In addition to arbitrary assertions, it can reason about exceptions, memory safety, data-race conditions, and input–output. However, approximately half of the lines of code must be devoted to the specification, and it has difficulty inferring properties about non-Python code such as C libraries.
Conclusions and future directions
=================================
It is difficult to overestimate the importance of investigative software in scientific research, but few studies have examined effective techniques for ensuring its correctness. Paranoid Scientist uses lightweight formal methods to provide correctness guarantees about this difficult-to-test class of software. It does so through a combination of contracts and refinement types in a way which is easy to use for those without explicit training in formal methods. We demonstrated that Paranoid Scientist can be used to find bugs in scientific software which would have impacted results and would have been otherwise difficult to detect.
Investigative software is built in an environment which poses two unique challenges for Hoare-style verification and static checking of preconditions and postconditions. First, the verification technique must be usable by scientists with little to no training in computer science. The state of the art techniques require a deep knowledge of formal methods to use them effectively [@Eilers2018]. Second, the amount of time spent verifying the software must be small compared to the amount of time spent writing the code to be verified. Current techniques are time consuming to implement. By contrast, our tool requires approximately as much time to implement as does writing function documentation.
Technical constraints of investigative software raise further challenges for formal methods. Investigative software in Python relies heavily on non-Python code such as C libraries and integrated shell commands, and thus static verification would require scaffolding for usability in practice. This scaffolding mandates more effort for verification and strong familiarity of the user with formal methods, exacerbating the previously discussed environmental challenges. Additionally, techniques such as type inference are conceptually incompatible with the present approach because the types in Paranoid Scientist often depend on the purpose of the code; for example, it may be “valid” to accept a probability less than 0 but this does not make sense on a scientific level. A gradual static verification approach, analogous to gradual typing [@Sieka2007], may be useful.
Lightweight formal methods may also be applied to investigative software in other programming languages. The present work targets Python due to its ubiquity in investigative software and powerful metaprogramming capabilities to simplify implementation. Besides Python, common languages for investigative software include Matlab, Julia, and R. While the bug shown in Figure \[fig:fmri\] may not have occurred in other languages, different programming languages have different advantages and disadvantages for the correctness of investigative software. For example, Matlab by default defines the constants and to be $\sqrt{-1}$, but allows these to be assigned other values by users. As a result, mistaken variable initialization or names can cause undetected bugs. Additionally, unexpected files saved to locations in Matlab’s `PATH` can cause erroneous versions of scripts or data to be loaded, or even cause built-in functions to change their behavior. An implementation of lightweight formal methods as described here would be able to catch these and other bugs in Matlab.
It is critical to verify the correctness of investigative software, but technical and cultural constraints limit the effectiveness of conventional techniques. Lightweight formal methods as implemented in Paranoid Scientist provide a convenient and effective way to check the correctness of investigative software.
Acknowledgments {#acknowledgments .unnumbered}
===============
Thank you to Ruzica Piskac and Anastasia Ershova for a critical review of the manuscript; Clarence Lehman, Daeyeol Lee, and John Murray for helpful discussions; Michael Scudder for PyDDM code reviews; and Norman Lam for PyDDM development and code reviews. Funding was provided by the Gruber Foundation.
Appendix: Default types {#appendix-default-types .unnumbered}
=======================
[lX]{}\
Numeric & A floating point or integer\
ExtendedReal & A floating point or integer, excluding NaN\
Number & A floating point or integer, excluding NaN and $\pm$inf\
Integer & An integer\
Natural0 & An integer greater than or equal to zero\
Natural1 & An integer greater than zero\
Range & A number with a value between two specified numbers, inclusive\
RangeClosedOpen & A number with a value between two specified numbers, inclusive on the bottom and exclusive on the top\
RangeOpenClosed & A number with a value between two specified numbers, exclusive on the bottom and inclusive on the top\
RangeOpen & A number with a value between two specified numbers, exclusive\
Positive0 & A number greater than or equal to zero\
Positive & A number greater than zero\
NDArray & A Numpy , optionally with a given dimensionality or elements which satisfy a given type\
\
String & A Python string\
Identifier & A non-empty alphanumeric string with underscores and hyphens\
Alphanumeric & A non-empty alphanumeric string\
Latin & A non-empty string with Latin characters only\
\
Tuple & A Python tuple, with elements which satisfy given types\
List & A Python list, with elements which satisfy a given type\
Dict & A Python dictionary, keys and values which satisfy given types\
Set & A Python set, with elements which satisfy a given type\
ParametersDict & A dictionary which may include only a subset of keys, with values which satisfy given types\
\
And & Logical AND of two or more types\
Or & Logical OR of two or more types\
Not & Logical NOT of a type\
\
Boolean & Either or\
Function & A Python function\
Constant & A single specified value is accepted\
Nothing & Only , equivalent to\
Unchecked & Any value (always succeeds)\
Void & No value is accepted (always fails)\
Maybe & Either a value of the specified type or else\
Self & The argument to a method\
PositionalArguments & Optional positional arguments to functions\
KeywordArguments & Optional keyword arguments to functions\
[^1]: See <http://retractionwatch.com/>
[^2]: <https://github.com/mwshinn/paranoidscientist>
[^3]: <https://github.com/mwshinn/PyDDM>
[^4]: <https://www.python.org/dev/peps/pep-0484/>
[^5]: <http://mypy-lang.org/>
[^6]: <https://pyre-check.org/>
[^7]: <https://google.github.io/pytype/>
[^8]: <https://github.com/Microsoft/pyright>
[^9]: <https://www.python.org/dev/peps/pep-0484/>
[^10]: <https://github.com/RussBaz/enforce/>
[^11]: <https://github.com/agronholm/typeguard>
[^12]: <https://andreacensi.github.io/contracts/>
[^13]: <https://docs.python-cerberus.org/en/stable/>
[^14]: <https://github.com/alecthomas/voluptuous>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we establish a result regarding the connection between continuous maximal regularity and generation of analytic semigroups on a pair of densely embedded Banach spaces. More precisely, we show that continuous maximal regularity for a closed operator $A: E_1 \rightarrow E_0$ implies that $A$ generates a strongly continuous analytic semigroup on $E_0$ with domain equal $E_1$.'
title: Continuous Maximal Regularity and Analytic Semigroups
---
<span style="font-variant:small-caps;">Jeremy LeCrone</span>
<span style="font-variant:small-caps;">Gieri Simonett</span>
Introduction
============
In this paper we consider the following abstract inhomogeneous equation $$\label{Problem}
\begin{cases}
\frac{d}{dt}u(t) - Au(t) = f(t), \quad\text{$t \in J$},\\
u(0)=u_0 .
\end{cases}$$ Here we are assuming that $J:=[0,T]$ for a fixed $T>0$, or $J:=\mathbb{R}_+=[0,\infty)$, and that $A: D(A) \subset E_0 \rightarrow E_0$ is a closed, densely defined operator on the complex Banach space $(E_0, \| \cdot \|_0)$. It then follows that $(E_1, \| \cdot \|_1) := (D(A), \| \cdot \|_{D(A)})$, equipped with the graph norm $\| \cdot \|_{D(A)}$, is also a complex Banach space and $E_1$ is densely (continuously) embedded in $E_0$. We assume that $f : J \rightarrow E_0$ and $u_0\in E_0$ are known, so that the unknown quantity in is the function $u : J \rightarrow E_0$. We say that $u$ is a *classical solution* to on $J$ if $u \in C(J,E_0) \cap C^1( \dot{J}, E_0) \cap C( \dot{J}, E_1) $ and $u$ satisfies for $t \in \dot{J} := J \setminus \{ 0 \}$.
In approaching this problem, it would be beneficial to know from the outset that the operator $A$ generates a strongly continuous semigroup $\{ e^{tA}: t \geq 0 \}$ on $E_0$. Then one can derive an explicit representation for solutions to . More precisely, if a solution $u$ of exists then it is represented by the variation of parameters formula as $$\label{Eqn:VarOfParams}
u(t)= e^{tA}u_0 + \int_0^t {e^{(t - \tau)A} f(\tau) \, d \tau} \qquad \text{for} \quad t \in J,$$ under the assumption that $A$ generates a C$_0$-semigroup. However, as we will show in our main result, one can do away with this assumption if the operator is known to satisfy the conditions of continuous maximal regularity.
In case $A$ generates an analytic semigroup, sufficient conditions for continuous maximal regularity were first obtained by Da Prato and Grisvard [@DPG79]. The results contained in this paper show that the assumption of $A$ generating an analytic semigroup is also necessary for the Da Prato-Grisvard result.
The theory of maximal regularity has proved itself to be an indispensable tool in the analysis of nonlinear parabolic problems, and it has been used by many authors. In the case of continuous maximal regularity we just mention the references [@DaLu88; @ES97; @ES98; @EMS98; @Si95].
Maximal Regularity {#sec:MaximalRegularity}
==================
In the remainder of this paper we will use the following notation: if $E_0$ and $E_1$ are Banach spaces we say that $(E_0,E_1)$ is a pair of embedded Banach spaces if $E_1\subset E_0$ and the natural injection $i:E_1\to E_0$ is continuous. If, in addition, $E_1$ is a dense subset of $E_0$ then we call $(E_0,E_1)$ a pair of densely embedded spaces. Suppose $(E_0,E_1)$ is a pair of embedded Banach spaces and $A:E_1\to E_0$ is a linear operator. Then we can also interpret $A:E_1\subset E_0\to E_0$ as an unbounded linear operator on $E_0$ with domain $D(A)=E_1$. It is then meaningful to say that $A$ is a closed operator on $E_0$. For a given pair $(E_0,E_1)$ of embedded Banach spaces we shall often consider linear operators $A:E_1\to E_0$ with the property that $A\in\mathcal{B}(E_1,E_0)$ (i.e. $A:E_1\to E_0$ is a bounded operator) and the additional property that $A$ is closed on $E_0$. These combined properties turn out to be equivalent to the fact that $E_1$ and $(D(A),\|\cdot\|_{D(A)})$ carry equivalent norms, see [@Amann95 Lemma I.1.1.2].
Next we give a general definition of maximal regularity, and then move on to focus on the more specific continuous maximal regularity.
Let $J := [0,T]$ for $T > 0$, or $J := \mathbb{R}_+ = [0,\infty)$, and assume that $(E_0,E_1)$ is a pair of densely embedded Banach spaces. Moreover, suppose that $(\mathbb{E}_0(J),\mathbb{E}_1(J))$ is a pair of Banach spaces such that $$\label{eqn:BlockESpaces}
\mathbb{E}_0(J) \subset L_{1,\text{loc}}(J,E_0) \qquad
\mathbb{E}_1(J) \subset W^{1}_{1,\text{loc}}(J,E_0) \cap L_{1,\text{loc}}(J,E_1).$$ Define the *trace operator* $\gamma : \mathbb{E}_1(J) \rightarrow E_0$ by $\gamma u := u(0)$ for $u \in \mathbb{E}_1(J)$ and the *trace space* $\gamma \mathbb{E}_1$ by $$\begin{split}
\gamma \mathbb{E}_1 &:= \{ x \in E_0 : x = \gamma u \text{ for some } u \in \mathbb{E}_1(J) \},\\
\| x \|_{\gamma} &:= \| x \|_{\gamma \mathbb{E}_1} := \inf \{ \| u \|_{\mathbb{E}_1(J)} : u \in \mathbb{E}_1(J) \text{ and } \gamma u = x \}.
\end{split}$$ For $A\in\mathcal{B}(E_1,E_0)$ we say that $(\mathbb{E}_0(J),\mathbb{E}_1(J))$ is a *pair of maximal regularity for $A$* if $$\label{eqn:MaximalRegularity}
\left( \frac{d}{dt} - A, \gamma \right) \in \text{Isom} \big( \mathbb{E}_1(J), \mathbb{E}_0(J) \times \gamma \mathbb{E}_1 \big).$$ That is, $(\mathbb{E}_0(J),\mathbb{E}_1(J))$ enjoys the property of maximal regularity if, for every $(f,x) \in \mathbb{E}_0(J) \times \gamma \mathbb{E}_1$, there exists a unique function $u \in \mathbb{E}_1(J)$ such that $u$ satisfies $$\label{eqn:MaxRegEqn}
\begin{cases}
\frac{d}{dt}u(t) - A u(t) = f(t) & \text{for $t \in J$}\\
\gamma u = u(0) = x
\end{cases}$$ and the mapping $[(f,x) \mapsto u]$ is continuous. For a pair $(\mathbb{E}_0(J),\mathbb{E}_1(J))$ of maximal regularity for $A$, we define the *solution operator* $$K_A: \mathbb{E}_0(J) \rightarrow \mathbb{E}_1(J) \quad \text{by}
\quad K_Af := \left( \frac{d}{dt} - A, \gamma \right)^{-1} (f,0).$$
We say that the pair enjoys the property of *continuous* maximal regularity if $$(\mathbb{E}_0(J), \, \mathbb{E}_1(J)) = (BU\!C(J,E_0), BU\!C^1(J,E_0) \cap BU\!C(J,E_1)).$$ Here $BU\!C(J,E)$ denotes the space of all bounded uniformly continuous functions $u:J\to E$, and $BU\!C^1(J, E)$ stands for all functions in $BU\!C(J,E)$ whose derivative also shares this property. In case $J=[0,T]$ is a compact interval we will use the shorter notation $C$ and $C^1$ in place of $BU\!C$ and $BU\!C^1$, respectively. We refer to [@Amann95] for more background information on maximal regularity.
Notice that $(\mathbb{E}_0(J), \mathbb{E}_1(J))$ a pair of continuous maximal regularity for $A$ implies existence of a unique classical solution $u$ to for every $f \in \mathbb{E}_0(J)$ and $x \in E_1$, since $E_1$ and $\gamma \mathbb{E}_1$ coincide in this case, and there exists a positive constant M, independent of $f$ and $x$, such that the estimate $$\label{MaxReg}
\sup_{t \in J} \big( \| \dot{u}(t) \|_{0} + \| u(t) \|_{1} \big) \leq M \left( \sup_{t \in J} \| f(t) \|_{0} + \| x \|_{1} \right)$$ holds. Furthermore, it follows from Theorem \[MainResult\] below and that the solution operator produces the function $$(K_Af)(t) = \int_0^t {e^{(t - \tau)A} f(\tau) \, d \tau} \qquad \text{for} \quad t \in J.$$ The following theorem is a modification of a result by Dore [@Dore93] and Prüss [@Pruss02], who demonstrated a proof in the case of $L_p$ maximal regularity.
\[MainResult\] Fix $T > 0$ (or $T = \infty$) and $J := [0,T]$ (or $J := \mathbb{R}_+)$ and let $(E_0, E_1)$ be a pair of densely embedded Banach spaces. If $A \in \mathcal{B}(E_1, E_0)$ is a closed operator on $E_0$ and $$\big( \mathbb{E}_0(J),\mathbb{E}_1(J) \big) = \left( BU \! C(J, E_0), BU \! C^1(J, E_0) \cap BU \! C(J, E_1) \right)$$ is a pair of continuous maximal regularity for $A$, then $A$ generates a strongly continuous analytic semigroup on $E_0$. Moreover, in the case $J = \mathbb{R}_+$, it follows that $s(A) < 0$ where $s(A) := \sup \{ {\rm Re}\, \mu: \mu \in \sigma (A) \}$ is the spectral bound of $A$.
We begin by demonstrating the result for $T < \infty$ and then we extend to the unbounded interval $\mathbb{R}_+$. In particular, we show that there exist constants $\omega \in \mathbb{R}$ and $N \geq 1$ so that $$\{ \mu \in \mathbb{C}: \text{Re} \, \mu > \omega \} \subset \rho(A) \quad \text{and}
\quad \|(\mu - A)^{-1} \|_{\mathcal{B}(E_0)} \leq \frac{N}{1 + | \mu |} \quad \text{for} \quad \text{Re} \, \mu > \omega \, .$$\
***Claim 1:*** There exists $\omega_1 \geq 0$ sufficiently large so that the a priori estimate $$\label{a-priori}
\|x\|_1+|\mu| \|x\|_0\le C\|(\mu-A)x\|_0,\quad x\in E_1,\quad \text{Re} \, \mu > \omega_1,$$ holds. In particular, $(\mu - A) : E_1 \rightarrow E_0$ is injective for $\text{Re} \, \mu > \omega_1 \, .$
- Let $x \in E_1, \, \mu \in \mathbb{C}$ and set $v_{\mu}(t) := v_{\mu}(t,x) = e^{\mu t} x$. Then $v_{\mu} \in {\mathbb E}_1(J)$ and satisfies $$\dot{v}_{\mu}(t) - A v_{\mu}(t) = g_{\mu}(t), \quad v_{\mu}(0) = x,$$ where $g_{\mu}(t) = e^{\mu t}(\mu - A) x \in \mathbb{E}_0(J)$. By assumption, we see that inequality (\[MaxReg\]) holds and implies $$\label{I1}
\sup_{t \in J} e^{\text{Re}\,\mu t} \big(\|x\|_1 + |\mu|\|x\|_0 \big)
\leq M \left( \sup_{t \in J} e^{\text{Re}\,\mu t} \| (\mu - A)x \|_0 + \| x \|_1 \right).$$ Now, choosing $\omega_1 \geq 0$ large enough so that $\displaystyle 2M \leq \sup_{t \in J} e^{\omega_1 t}$ we obtain $$\label{I2}
\| x \|_1 + | \mu | \| x \|_0 \leq 2M \| (\mu - A)x \|_0 , \quad \text{Re} \, \mu > \omega_1,$$ thus establishing the claim.
Now let $x \in E_0$, $\mu \in \mathbb{C}_+ := \{ z \in \mathbb{C}: \text{Re} \, \mu > 0 \}$ and define the functions $$f_{\mu}(t) := e^{-\mu t}x \in BU \!C(\mathbb{R}_+,E_0), \qquad u_{\mu} := u_{\mu}(\cdot,x) := K_A f_{\mu} \in \mathbb{E}_1(J) \, .$$ Notice that continuity of the embedding $E_1 \hookrightarrow E_0$ and continuity of $K_A$ imply existence of positive constants $c_1$ and $c_2$ so that $$\label{embedding}
\sup_{t \in J} \| u_{\mu}(t) \|_0 \leq c_1 \sup_{t \in J} \| u_{\mu} (t) \|_1 \leq c_1 \| u_{\mu} \|_{\mathbb{E}_1} \leq c_2 \| f_{\mu} \|_{\mathbb{E}_0} = c_2 \sup_{t \in J} \| f_{\mu}(t) \|_0 \leq c_2 \| x \|_0 \, .
$$ In particular, we see that $$\label{uMuBound}
\| u_{\mu} (t) \|_0 \leq c_2 \| x \|_0 \qquad \mu \in \mathbb{C}_+, \; t \in J.$$ Also, for $\mu \in \mathbb{C}_+$ we note that $\bar{\mu} \in \mathbb{C}_+$ and define the functions $U_{\mu}: E_0 \rightarrow E_1$ and $V_{\mu}: E_0 \rightarrow E_0$ as $$\label{eqn:BigUmu}
\begin{aligned}
U_{\mu}x :&= 2 {\rm Re}\, \mu \int_0^T e^{- \mu t} u_{\bar{\mu}}(t,x) \, dt,\\
V_{\mu}x :&= \frac{2 {\rm Re} \, \mu \, e^{- \mu T}}{ \left(1 - e^{- 2 {\rm Re} \, \mu T} \right)} \, u_{\bar{\mu}} (T,x) \, .
\end{aligned}$$\
***Claim 2:*** There exists $\omega_2 \geq 0$ sufficiently large so that $(\mu - A) : E_1 \rightarrow E_0$ is surjective for $\text{Re} \, \mu > \omega_2 \, .$
- First notice that $$\| V_{\mu} x \|_{0} = \frac{2 {\rm Re} \, \mu \, e^{- {\rm Re} \, \mu T}}
{ \left(1 - e^{- 2 {\rm Re} \, \mu T} \right)} \, \| u_{\bar{\mu}} (T,x) \|_0
\leq c_2\frac{2 {\rm Re} \, \mu \, e^{- {\rm Re} \, \mu T}}{ \left(1 - e^{- 2 {\rm Re} \, \mu T} \right)} \, \| x \|_0 \, ,$$ converges to 0 as Re$\, \mu \rightarrow \infty$. So we fix $\omega_2 \geq 0$ large enough that $\| V_{\mu} \|_{\mathcal{B}(E_0)}
< \frac{1}{2}$ for $\text{Re} \, \mu > \omega_2 \, .$ Multiplying the relation $$\frac{d}{dt} u_{\bar{\mu}}(t,x) - A u_{\bar{\mu}}(t,x) = e^{- \bar{\mu} t} x$$ by $2 ({\rm Re} \, \mu) e^{- \mu t}$ and integrating over the interval $[0,T]$ yields $$(\mu - A)U_{\mu} x = \big(1 - e^{- 2 {\rm Re} \, \mu T} \big)
\, \left( I - V_{\mu} \right) \, x, \qquad x \in E_0, \, \mu \in \mathbb{C}_+ \, .$$ Here we use the fact that $A$ is a closed operator and so we are free to interchange $A$ with integration. Therefore, choosing $\text{Re} \, \mu > \omega_2$ we conclude that $(\mu - A)U_{\mu}$ is invertible by the Neumann series. However, invertibility of $(\mu - A)U_{\mu}$ implies surjectivity of $(\mu - A)$ and so the claim is established.
With claims 1 and 2 established, we see that choosing $\omega \geq \omega_1 \vee \omega_2$ implies $\{ \mu \in \mathbb{C}: {\rm Re} \mu > \omega \} \subset \rho(A)$ and the inequality $$\| ( \mu - A )^{-1} \|_{\mathcal{B}(E_0)} \leq \frac{2M(1 \vee c_1)}{1 + | \mu |}, \qquad {\rm Re}\, \mu > \omega \, ,$$ follows from (\[I2\]). It is well known that these properties are sufficient for $A$ to generate an analytic semigroup on $E_0$, c.f. [@Amann95 Theorem 1.2.2], which concludes the proof for the finite interval $J = [0,T]$.\
Now we consider the case of continuous maximal regularity on $\mathbb{R}_+$. More precisely, assume that $\big( \mathbb{E}_0 (\mathbb{R}_+), \mathbb{E}_1 (\mathbb{R}_+) \big) =
\big( BU \!C( \mathbb{R}_+, E_0), BU \!C^1( \mathbb{R}_+, E_0) \cap BU \!C(\mathbb{R}_+, E_1) \big)$ is a pair of maximal regularity for $A$.\
***Claim 3:*** For any $T > 0, \, \big( \mathbb{E}_0 (J), \mathbb{E}_1 (J) \big)$ is a pair of maximal regularity for $A$, where $J = [0,T]$.
- Note that any function $f \in C(J, E_0)$ can be extended to $Ef \in BU \!C( \mathbb{R}_+, E_0)$ by setting $(Ef)(t) = f(T)$ for $t \geq T$. Moreover, denoting by $R$ the restriction operator to the interval $J$, we see that $u := R \big( \frac{d}{dt} - A, \gamma \big)^{-1} (Ef, x) \in \mathbb{E}_1(J) \,$ is a solution to problem (\[eqn:MaxRegEqn\]) on $J$, for $f \in C(J,E_0)$ and $x \in E_1$. It remains to show that this solution $u$ is unique in $\mathbb{E}_1(J)$. In fact, it suffices to show that $u \equiv 0$ is the unique solution in $\mathbb{E}_1(J)$ to problem (\[eqn:MaxRegEqn\]) with $(f,x) = (0,0)$. By way of contradiction, suppose $\tilde{u} \in \mathbb{E}_1(J)$ solves (\[eqn:MaxRegEqn\]) with $(f,x) = (0,0)$ and there exists $t_1 \in (0,T)$ such that $\tilde{u}(t_1) \not= 0$. Then, define $v := \big( \frac{d}{dt} - A, \gamma \big)^{-1} (0, \tilde{u}(t_1))$, so that $v \in \mathbb{E}_1(\mathbb{R}_+)$ and $v$ satisfies $\dot{v} - Av = 0$ on $\mathbb{R}_+$ and $v(0) = \tilde{u}(t_1)$. Define $$w(t) :=
\begin{cases}
\tilde u(t), & \text{$t \in [0, t_1],$}\\
v(t - t_1), & \text{$ t \in [t_1, \infty),$}
\end{cases}$$ and it follows that $w \in \mathbb{E}_1 (\mathbb{R}_+)$ and $w$ satisfies (\[eqn:MaxRegEqn\]) with $(f,x)=(0,0)$, on $\mathbb{R}_+$. However, note that $w(t_1) = \tilde u(t_1) \not= 0$, so that $w \not\equiv 0$, a contradiction.
Thus, by Claims 1 and 2 above, we see that $A$ generates an analytic semigroup on $E_0$. It remains to consider the spectral bound $s(A)$.
***Claim 4:*** The spectral bound of $A$ is strictly negative, i.e. $s(A) < 0$.
- By assumption of continuous maximal regularity on $\mathbb{R}_+$, and by Claim 3, there exists a positive constant $M$ (independent of $T > 0$) such that $$\label{T-indepenent}
\| u \|_{{\mathbb E}_1(J)} \leq M \big( \| f \|_{{\mathbb E}_0(J)} + \| x \|_1 \big), \quad
(f, x) \in {\mathbb E}_0(J) \times E_1,$$ for the solution $u = R \big( \frac{d}{dt} - A, \gamma \big)^{-1} (Ef, x)$ of (\[eqn:MaxRegEqn\]) on $J = [0,T]$. From this estimate, we conclude that inequality (\[I1\]) holds for any $T > 0$ and so Claim 1 is true for $\omega_1 = 0$ in this case. Furthermore, setting $$U_{\mu}x := 2 {\rm Re}\, \mu \int_0^\infty e^{- \mu t} u_{\bar{\mu}}(t,x) \, dt
\qquad \text{and} \qquad V_\mu :=0,$$ we see that Claim 2 holds for $\omega_2 = 0$. Therefore, it follows that $${\mathbb C}_+\subset\rho(A)\quad\text{and}\quad \| (\mu-A)^{-1} \|_{\mathcal{B}(E_0)}
\leq \frac{c}{1 + | \mu |},\quad {\rm Re}\,\mu >0\,.$$ Since this estimate is uniform in ${\rm Re}\,\mu >0$, we conclude that the imaginary axis must be contained in the resolvent set of $A$ and the estimate continues to hold for ${\rm Re}\,\mu\ge 0$. This implies that $\sigma(A)\subset[{\rm Re}\,\mu<0]$ and $s(A)<0$ as claimed.
[**(a)**]{} If the underlying Banach spaces $E_0$ and $E_1$ are real, we can apply Theorem \[MainResult\] to the complexification of $E_0$, $E_1$ and $A$. Then, by restriction back to the real case we derive results in the original, real-valued, setting. Hence, we lose no generality here by focusing on only the complex case.\
[**(b)**]{} Suppose $(C(J,E_0), C^1(J,E_0) \cap C(J,E_1))$ is a pair of maximal regularity for some $A \in \mathcal{B}(E_1, E_0)$. Then by a result of Baillon [@Ba80] (see also [@EG92]) either $E_{1}=E_{0}$, or $E_{0}$ contains a closed subspace which is isomorphic to $c_{0}$, the space of null sequences. In particular, if $E_{0}$ is reflexive or weakly sequentially compact then continuous maximal regularity cannot occur.\
[**(c)**]{} If one already knows that $A$ generates a $C_0$-semigroup on $E_0$, then DaPrato and Grisvard proved that continuous maximal regularity implies that $A$ is the generator of an analytic semigroup, see also [@Amann95 Proposition III.3.1.1].\
[**(d)**]{} For the reader’s convenience we briefly describe a situation which shows that many interesting operators $A$ give rise to continuous maximal regularity. Assume that $A\in\mathcal{H}(E_{1},E_{0})$ (i.e. $-A$ generates a strongly continuous analytic semigroup on $E_0$ with $D(A) = E_1$) and define $$\begin{split}
&E_{2}:=E_{2}(A):=( D(A^{2}),\|\cdot\, \|_{E_{2}}), \\
&\|\cdot \|_{E_{2}}:= \|\cdot \|_{E_{2}(A)}:=\|A\cdot \|_{E_{1}} +
\|\cdot \|_{E_{1}}.
\end{split}$$ Then $(E_{2},\|\cdot\|_{E_{2}})$ is a Banach space, and $E_2$ is densely embedded in $E_1$. We set $$\begin{split}
&E_{\theta}:=(E_{0},E_{1})_{\theta}, \\
&E_{1+\theta}:=E_{1+\theta}(A):=(E_{1},E_{2}(A))_{\theta},\qquad
\theta \in (0,1), \\ &A_{\theta}:=\text{the maximal
$E_{\theta}$-realization of $A$},
\end{split}$$ where $(\cdot,\cdot)_\theta$ denotes the continuous interpolation method of Da Prato and Grisvard [@DPG79], see also [@Amann95; @Ang90; @CleSim01; @Lun95]. It is then well-known that $A_{\theta}\in
\mathcal{H}(E_{\theta},E_{1+\theta})$, and it turns out that $A_{\theta}$ gives rise to continuous maximal regularity.\
[**Theorem.**]{} [(Da Prato-Grisvard)]{} Suppose that $\theta\in (0,1)$, $T > 0$. Let $J = [0,T]$. Then $ \displaystyle
(\mathbb{E}_0(J),\mathbb{E}_1(J)):=
(C(J,E_{\theta}),
C^{1}(J,E_{\theta})\cap C(J,E_{1+\theta}))
$ is a pair of maximal regularity for $A_{\theta}$.
[**(a) Maximal Regularity With Prescribed Singularity**]{} We demonstrate that Theorem \[MainResult\] continues to hold in a more general setting, namely a space of continuous functions which allow for singularities at zero. These spaces have for instance been studied in [@Amann95; @Ang90; @CleSim01]. This setting is well adapted for the study of quasilinear parabolic equations, see for instance [@CleSim01; @EMS98]. Assume that $\sigma \in (0,1)$, $E$ is a Banach space over $\mathbb{K} \,\big( = \mathbb{R} \; \text{or} \; \mathbb{C} \big)$, $J=[0,T]$ for some $T>0$, and we set $\dot J:=J\setminus\{0\}$. Then define $$\begin{aligned}
BU\!C_{1- \sigma}(J,E):=\!\big\{u\in C(\dot J,E):
[t\mapsto t^{1- \sigma}u(t)]\in BU\!C(\dot J,E),
\lim_{t\to 0^{+}}t^{1- \sigma}\|u(t)\|=0\big\},
\end{aligned}$$ which is a Banach space when equipped with the norm $$\|u\|_{C_{1-\sigma}}:=\sup_{t\in \dot J}t^{1-\sigma}\|u(t)\|_{E},
\quad u\in BU\!C_{1-\sigma}(J,E), \quad \sigma\in (0,1).$$ Further we introduce the subspace $$BU\!C^{1}_{1-\sigma}(J,E):=\{u\in C^{1}(\dot J,E)
: u,\dot{u} \in BU\!C_{1-\sigma}(J,E)\}.$$ Now, given a pair of densely embedded Banach spaces $(E_0,E_1)$ and $\sigma \in (0,1)$, we consider the pair of associated function spaces $$\label{2.1}
\begin{aligned}
\mathbb{E}_{0}(J):&=BU\!C_{1-\sigma}(J,E_{0}),\\
\mathbb{E}_{1}(J):&=BU\!C^{1}_{1-\sigma}(J,E_{0})\cap BU\!C_{1-\sigma}(J,E_{1}),
\end{aligned}$$ where $\mathbb{E}_{1}(J)$ is a Banach space with the norm $$\|u\|_{\mathbb{E}_{1}(J)}:=
\sup_{t\in\dot J}t^{1-\sigma}(\| \dot{u}(t)\|_{E_{0}}+\|u(t)\|_{E_{1}}).$$ It is well-known that the trace space is then given by $\gamma{\mathbb E}_1=(E_0,E_1)_\sigma$, $\sigma\in (0,1),$ where $(\cdot,\cdot)_\sigma$ denotes the continuous interpolation method, see for instance.
Let $(E_0, E_1)$ be a pair of densely embedded Banach spaces, $J = [0,T]$ with $T > 0$, and $A\in\mathcal{B}(E_1,E_0)$ a closed operator on $E_0$. If $$\big( \mathbb{E}_0(J),\mathbb{E}_1(J) \big) = \left( BU \! C_{1 - \sigma}(J, E_0), BU \! C_{1 - \sigma}^1(J, E_0) \cap BU \! C_{1 - \sigma}(J, E_1) \right)$$ is a pair of maximal regularity for $A$, for some $\sigma \in (0,1]$, then $A$ generates a strongly continuous analytic semigroup on $E_0$.
The methods presented in the proof of Theorem \[MainResult\] apply in this setting with minor modifications. We provide an outline of the proof in this case in order to illuminate the necessary modifications. For Claim 1, fix $x \in E_1$ and define $v_{\mu} \in \mathbb{E}_1(J)$, $g_{\mu}\in \mathbb{E}_0(J)$ as before. Using the continuous embedding $E_1\hookrightarrow \gamma{\mathbb E}_1$ we derive as above the inequality $$\sup_{t \in J} t^{1 - \sigma} e^{{\rm Re} \, \mu t} \big( |\mu| \| x \|_0 + \| x \|_1 \big)
\le M \big( \sup_{t \in J} t^{1 - \sigma}e^{{\rm Re} \, \mu t} \| (\mu - A) x \|_0 + \| x \|_1 \big),$$ which is the analog of (\[I1\]) in this setting. Choosing $\omega_1 $ so that $ 2M \leq \sup_{t \in J} t^{1 - \sigma} e^{\omega_1 t},$ we see that (\[I2\]) holds as before.
Meanwhile, for Claim 2 we consider $f_{\mu}, \, u_{\mu}, \, U_{\mu}$ and $V_{\mu}$ defined as above, for $\mu \in \mathbb{C}_+$. Recalling the continuity and embedding constants $c_1, \, c_2$, we see, for $t \in \dot{J}$, $$t^{1 - \sigma}\| u_{\bar{\mu}}(t) \|_0
\le \| u_{\bar{\mu}} \|_{\mathbb{E}_0(J)}
\le c_1 \| u_{\bar{\mu}} \|_{\mathbb{E}_1(J)}
\le c_2 \| f_{\bar{\mu}} \|_{\mathbb{E}_0(J)}
\le c_2 T^{1-\sigma}\| x \|_0.$$ This shows that (\[uMuBound\]) holds for $t = T$ and the remainder of the proof follows exactly as in Theorem \[MainResult\].
[**(b) $L_p$ Maximal Regularity**]{}
For a given pair of densely embedded Banach spaces $(E_0,E_1)$ one defines the pair of functions spaces $$\begin{split}
{\mathbb E}_1(J):&=W^1_p(J,E_0)\cap L_p(J,E_1), \\
{\mathbb E}_0(J):&=L_p(J,E_0).
\end{split}$$ It is well-known that the trace space is then given by $\gamma{\mathbb E}_1=(E_0,E_1)_{1-1/p,p}$.
An inspection of the proof of Theorem \[MainResult\] shows that the same methods apply to this $L_p$-maximal regularity setting, with minor modifications. Moreover, the methods presented herein considerably simplify the arguments in [@Dore93; @Pruss02].
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous reviewer for valuable suggestions. We would also like to thank Jan Prüss and Mathias Wilke for helpful conversations.
[99]{}
(MR1345385) H. Amann, “Linear and Quasilinear Parabolic Problems. Vol. I," Monographs in Mathematics, 89, Birkh$\ddot{\text{a}}$user, Boston, 1995.
(MR1059647) S. Angenent, *Nonlinear analytic semiflows*, Proc. Roy. Soc. Edinburgh Sect. A, **115** (1990), 91–107.
(MR0577152) J.B. Baillon, *Charactère borné de certains genérateurs de semigroupes linéaires dans les espaces de Banach*, C.R. Acad. Sc. Paris, **290** (1980), 757–760.
(MR1838320) P. Clément and G. Simonett, *Maximal regularity in continuous interpolation spaces and quasilinear parabolic equations*, J. Evol. Equ., **1** (2001), 39–67.
(MR0551075) G. Da Prato and P. Grisvard, *Équations d’évolution abstraites nonlinéaires de type parabolique*, Ann. Mat. Pura Appl.(4), **120** (1979), 329–396.
(MR0921935) G. Da Prato and A. Lunardi, *Stability, instability and center manifold theorem for fully nonlinear autonomous parabolic equations in Banach space*, Arch. Rational Mech. Anal., **101** (1988), no. 2, 115Ð-141.
(MR1225809) G. Dore, *L$^p$ regularity for abstract differential equations*, in “Functional Analysis and Related Topics, 1991 (Kyoto)" (eds. H. Komatsu), Lecture Notes in Math., 1540, Springer, Berlin (1993), 25–38.
(93i:47053) B. Eberhardt and G. Greiner, *Baillon’s theorem on maximal regularity*, Acta Appl. Math., **27** (1992), 47Ð-54.
(MR1466667) J. Escher and G. Simonett, *Classical solutions of multidimensional Hele-Shaw models*, SIAM J. Math. Anal., **28** (1997), 1028–1047.
(MR1607952) J. Escher and G. Simonett, *A center manifold analysis for the Mullins-Sekerka model*, J. Differential Equations, **143** (1998), 267–292.
(MR1638074) J. Escher, U.F. Mayer and G. Simonett, *The surface diffusion flow for immersed hypersurfaces*, SIAM J. Math. Anal., **29** (1998), 1419–1433.
(MR1329547) A. Lunardi, “Analytic Semigroups and Optimal Regularity in Parabolic Problems", Progress in Nonlinear Differential Equations and their Applications, 16, Birkhäuser Verlag, Basel, 1995.
(MR1988408) J. Prüss, *Maximal regularity for evolution equations in $L_p$-spaces*, Conf. Semin. Mat. Univ. Bari, No. 285 (2002), 1–39.
(MR1306591) G. Simonett, *Center manifolds for quasilinear reaction-diffusion systems*, Differential Integral Equations, **8** (1995), 753–796.
Received July 2010; revised March 2011.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have performed Monte Carlo studies of the 3D $XY$ model with random uniaxial anisotropy, which is a model for randomly pinned spin-density waves. We study $L \times L \times L$ simple cubic lattices, using $L$ values in the range 16 to 64, and with random anisotropy strengths of $D / 2 J$ = 1, 2, 3, 6 and $\infty$. There is a well-defined finite temperature critical point, $T_c$, for each these values of $D / 2 J$. We present results for the angle-averaged magnetic structure factor, $S ( k )$ at $T_c$ for $L = 64$. We also use finite-size scaling analysis to study scaling functions for the critical behavior of the specific heat, the magnetization and the longitudinal magnetic susceptibility. Good data collapse of the scaling functions over a wide range of $T$ is seen for $D / 2 J$ = 6 and $\infty$. For our finite values of $D / 2 J$ the scaled magnetization function increases with $L$ below $T_c$, and appears to approach an $L$-independent limit for large $L$. This suggests that the system is ferromagnetic below $T_c$.'
author:
- Ronald Fisch
title: 'Finite-Size Scaling Critical Behavior of Randomly Pinned Spin-Density Waves'
---
Introduction
============
The Harris-Plischke-Zuckermann model[@HPZ73] has long been used to study the effects of random uniaxial anisotropy on ferromagnetism. The Hamiltonian of this random anisotropy model (RAM) is $$H_{RAM} ~=~ - J \sum_{\langle ij \rangle} \vec{\bf S}_{i} \cdot \vec{\bf S}_{j}
~-~ D \sum_{i} ( \hat{\bf n}_{i} \cdot \vec{\bf S}_{i} )^2 \, ,$$ where each $\vec{\bf S}_{i}$, the dynamical on site $i$, is usually taken to be a classical three-component spin of unit length. Each $\hat{\bf n}_{i}$ is a time-independent unit vector. The $\hat{\bf n}$ on different sites are assumed to be uncorrelated random variables. $\sum_{\langle ij \rangle}$ is a sum over nearest neighbors on some lattice. In this work we will use a simple cubic lattice with periodic boundary conditions, and we will study the case of two-component ($n = 2$) spins.
As was discussed in some detail in an earlier paper,[@Fis95] if one chooses the $\vec{\bf S}_{i}$ and the $\hat{\bf n}_{i}$ to be two-component vectors, then the Hamiltonian can be mapped onto a model of a spin-density wave (SDW) in an anisotropic material with an easy axis. For $XY$ spins, [*i.e.*]{} $n = 2$, the Hamiltonian of the model may be rewritten as $$H ~=~ - J \sum_{\langle ij \rangle} \cos ( \phi_{i} - \phi_{j} )
~-~ {D \over 2} \sum_{i} [ \cos ( 2 ( \phi_{i} - \theta_{i} ) )
~-~ 1 ] \, .$$ Each $\phi_{i}$ is a dynamical variable which takes on values between 0 and $2 \pi$. The $\langle ij \rangle$ indicates here a sum over nearest neighbors on a simple cubic lattice of size $L \times L \times L$. We choose each $\theta_{i}$ to be an independent identically distributed quenched random variable, with the probability distribution $$P ( \theta_i ) ~=~ 1 / 2 \pi \,$$ for $\theta_i$ between 0 and $2 \pi$. A constant term has also been added to the anisotropy, to make the Hamiltonian well-behaved in the limit $ D / J \to \infty $.
In this work we will study Eqn. 2 on the simple cubic lattice over a range of $D / J$, using Monte Carlo simulations. The large increase in available computing resources over the last fifteen years makes possible significant improvements over the earlier results.[@Fis95] By studying a range of $L$, we will be able to learn about the stability of long-range order against random pinning which respects the Kramers degeneracy, such as alloy disorder, and the critical behavior of a SDW in an easy-axis material with this type of pinning.
Random pinning effects
======================
In the limit $ D / J \to \infty $, often called the Ising limit, both analytical[@DV80; @HCB87; @FH90] and numerical[@JK80; @Fis95; @TPV06; @LLMPV07] calculations become substantially simplified. This is due to the fact that in this Ising limit the random anisotropy term in the Hamiltonian becomes a projection operator, and each spin has only two allowed states. It has been argued that for large $D / J$ the behavior is close to the $D / J = \infty$ limit as long as $T \ll D$.[@HCB87] It has also been found, however, that for $n = 2$ at low temperatures and moderately large values of $D / J$ the magnetization per spin on $L \times L \times L$ simple cubic lattices, $|\vec{M} ( L )|$, decreases[@Fis95] as the temperature, $T$, is lowered. This effect was not seen for $D / J = \infty$.
A similar effect is seen in the case of the random bond Ising model (RBIM), where the Nishimori gauge symmetry causes the magnetization to have a maximum at a finite $T$ on the Nishimori line.[@Nis81; @HTPV07] The RBIM is the natural extension[@CL77; @Fis90] of the RAM to the case of Ising spins, $n = 1$. Thus it should be expected that the phase diagram of the $n = 2$ RAM has a close relation to that of the RBIM. However, there are aspects of the phase diagrams which remain somewhat mysterious. For example, Chen and Lubensky[@CL77] found that the critical exponents which describe the stability of the ferromagnet-spin glass-paramagnet multicritical point for the random bond model in $6 - \epsilon$ dimensions are well-behaved for $n = 1$, but become complex for $n = 2$ and $n = 3$. One interpretation of this puzzling result is that the multicritical point itself becomes unstable in $6 - \epsilon$ dimensions, so that it becomes a region of the phase diagram, rather than a single point. In this expanded multicritical region one might expect to find quasi-long-range order (QLRO). Although an explicit calculation has not been done, a similar result would not be surprising for the RAM. The existence of QLRO in the RAM was first suggested by Aharony and Pytte[@AP80] in 1980. They later[@AP83] pointed out that higher order terms might make the correlation length, $\xi$, finite below $T_c$. Feldman[@Fel01] has argued that QLRO should be common in disordered magnets and similar systems.
Thus there are a number of possibilities available for the topology of the phase diagram. In a Cayley-tree mean-field theory, where QLRO does not occur, it is known[@HCB87] that in the limit $D / J \to \infty$ the phase diagram depends on the parameter $z / n$, where $z$ is the number of nearest neighbor spins. Thus it is to be expected that the phase diagram in three dimensions will also depend on the lattice type, $n$ and the range of the exchange interactions, as well as on $T / J$ and $D / J$. For the simple cubic lattice (which has $z = 6$) it has been shown[@LLMPV07] that in the limit $D / J \to \infty$ the ground state is an Ising spin glass when $n \ge 3$. For small $D / J$, however, where one does not expect the qualitative behavior to depend on $z$, Feldman[@Fel01] predicts QLRO in $d = 3$ for $n \le 4$. In the $n = 3$ case, this appears to be confirmed by Monte Carlo calculations.[@Fis98]
The presence of a reentrant phase is difficult to demonstrate conclusively using the type of numerical calculations we have performed here. It was only relatively recently that reentrance was demonstrated convincingly in the $d = 2$ RBIM.[@WHP03; @HTPV08] There may also be a range of $D / J$ for which the three-dimensional $n = 2$ model has a reentrant ferromagnetic phase. One motivation for believing this is that reentrance is frequently observed in laboratory experiments. Another is the work of Pelcovits, Pytte and Rudnick[@PPR78; @Pel79] who argue that ferromagnetism should be unstable in the RAM for low $T$ and small $D / J$. Since magnetization can increase with increasing $T$ at low $T$, (which was not known at the time of their work,) it is not correct to claim that the absence of ferromagnetism near $T = 0$ precludes the existence of a ferromagnetic phase in the RAM at a somewhat higher $T$.
Larkin[@Lar70] studied a model for a vortex lattice in a type-II superconductor. His model replaces the spin-exchange term of the Hamiltonian with a harmonic potential, so that each $\phi_{i}$ is no longer restricted to lie in a compact interval. He argued that for any non-zero value of a random field this model has no long-range order on a lattice whose dimension $d$ is less than or equal to four. This argument, using the harmonic potential instead of the spin-exchange, is only rigorously correct in the limit $n \to \infty$.
A more intuitive derivation of the result was given by Imry and Ma,[@IM75] who assumed that the increase in the energy of an $L^d$ lattice when the order parameter is twisted at a boundary scales as $L^{d - 2}$, just as it does in the nonrandom ferromagnet. As argued by Imry and Ma,[@IM75] and later justified more carefully,[@AIM76; @PS82] within an $\epsilon$-expansion one finds the phenomenon of “dimensional reduction". Within this perturbation theory the critical exponents of any $d$-dimensional $O(n)$ random-field model (RFM) (for which the Kramers degeneracy is broken by the randomness) appear to be identical to those of an ordinary $O(n)$ model of dimension $d - 2$. For the Ising ($n = 1$) case, this dimensional reduction was shown rigorously to be incorrect.[@Imb84; @BK87] Another interesting development was the calculation of Mezard and Young,[@MY92] who showed that random fields caused breaking of replica symmetry below $T_c$ for any finite value of $n$. Thus there is no good reason to expect that dimensional reduction should be correct near $T_c$ for any finite value of $n$.
Although there is certainly a family resemblance between the RFM and the RAM, the difference between breaking the Kramers degeneracy at the level of Hamiltonian and breaking it spontaneously has profound consequences. One such consequence is Theorem 4.4 of Aizenman and Wehr,[@AW90] which applies to the RFM, but not to the RAM. A naive but not entirely misleading analogy may be drawn between the relationship of the RFM to the RAM and the relationship between applying a uniform magnetic field to a ferromagnet or to an antiferromagnet. The field which couples linearly to the order parameter has a qualitatively stronger effect than the field which couples quadratically to the order parameter.
Translation invariance of $H_{RAM}$ is broken for any non-zero value of $D$, since the vectors $\hat{\bf n}_{i}$ are random. Within a high-temperature perturbation theory, performing a configuration average over the ensemble of random lattices appears to restore translation invariance above $T_c$. However, the radius of convergence of this perturbation theory cannot be greater for $D \ne 0$ than it is for $D = 0$. For models described by Eqn. (1), the $T_c$ predicted by extrapolating the low orders of perturbation theory is always maximal at $D = 0$. This implies that for $D \ne 0$ the high-temperature perturbation theory does not converge near $T_c$. The inadequacy of perturbation theory to describe $XY$ models in $d = 3$, because of the effects of vortex lines, has been discussed by Halperin.[@Hal81] While it is not clear than Halperin’s argument is valid for the RFM, where the Kramers degeneracy is broken by the Hamiltonian, it should be valid for the RAM. Thus it seems quite implausible that for $d < 4$ the twist energy for Eqn. (2) really scales as $L^{d - 2}$ when $D \ne 0$, even though this is correct to all orders in the configuration-averaged perturbation theory.
The argument of Pelcovits[@Pel79] for the $n = 2$ RAM, which is a prototype for much subsequent work,[@Fish85] assumes that if one goes to small enough $D / J$ and low enough $T$ then the effects of vortex lines can be ignored. In essence, what is done is to replace the spin variables by a noncompact “elastic manifold”. These authors then claim that this does not effect the behavior one is studying. However, this cannot be true when one considers behavior on scales larger than the Imry-Ma length.[@Cop91]
The basic point is that Imry-Ma-type arguments for continuous O([*n*]{}) spins ([*i.e.*]{} $n \ge 2$) are not self-consistent. One begins by assuming that the random field is weak so that the twist energy scales as $L^{d - 2}$, as in the absence of the randomness. Then one shows that, if $d \le 4$ and $T < T_c$, the effective coupling to the random field increases as $L$ increases. If the effective random coupling is strong, however, then assuming that the twist energy is uniformly distributed throughout the volume is not reasonable. The conclusion which should be drawn from this is that a deeper analysis is needed when $d \le 4$.
In order to understand whether the problems with perturbation theory are actually due to vortex lines, and thus restricted to the $n = 2$ case, or if similar problems can also be expected for $n > 2$, it may be helpful to reconsider the analysis of Pelcovits, Pytte and Rudnick.[@PPR78] These authors show that within their perturbation theory the pure O([*n*]{}) mean-field theory critical fixed point remains stable against random anisotropy for $d > 4$. This contrasts to the random-field case, where mean-field theory is only stable for $d > 6$. Then they argue that for $d \le 4$ and $n \ge 2$ there is no stable critical fixed point for the RAM, because under rescaling transformations the random anisotropy coupling constant runs off to $\infty$. However, they did not (and within their formulation could not) examine the possibility that there could exist another ferromagnetic critical fixed point at a large value of $D / J$. The reason why such an object may exist is that there exist alternative formulations of mean-field theory[@DV80; @HCB87] for the RAM in the limit $D / J \to \infty$.
It is useful to consider the generalization of Eqn. (2) to $p$-fold random fields.[@Aha81] In the $d = 2$ case[@HKY81; @CO82] it has been shown that there continues to be a Kosterlitz-Thouless phase as long as $p^2 > 8$, [*i.e.*]{} for $p^2 > 8$ a weak $p$-fold random field does not destroy the Kosterlitz-Thouless phase. It was claimed by Aharony[@Aha81] that ferromagnetism should be unstable for any value of $p$ when $d = 3$. However, a computer simulation study[@Fis92] for $p = 3$ is not consistent with this claim, which is based on the weak randomness perturbation theory around the $D = 0$ model. This $d = 3$ computer simulation finds that there is a mass gap at $T = 0$, an effect which cannot be reproduced within the perturbation theory. The interpretation of this is that for $p \ge 3$ the thickness of a domain wall remains finite in the limit $L \to \infty$, [*i.e.*]{} the domain wall becomes localized by random pinning.
Removing vortex lines from the pure $XY$ model by letting the vortex fugacity become large forces the system into a ferromagnetic state at any temperature.[@KSW86; @Fis95b] This result is true even in the presence of a strong $p = 2$ random anisotropy,[@Fis00] but the $p = 2$ case is more complicated than $p \ge 3$. For $p = 2$, as we shall see, the domain walls probably have a fractal structure, rather than becoming completely localized.
Numerical results
=================
In this work, we will present results obtained from heat bath Monte Carlo calculations. The data were obtained from $L \times L \times L$ simple cubic lattices with $16 \le L \le 64$ using periodic boundary conditions. The calculations were done for a 12-state clock model, [*i.e.*]{} a ${\bf Z}_{12}$ approximation[@Fis97] to the $XY$ model of Eqn. (2). The computer program was an adaptation of the code used recently for the $XY$ model in a random field,[@Fis07] modified to replace the random field term with the random 2-fold anisotropy term of Eqn. (2). For any integer value of the quantity $D / 2 J$ one can use a lookup table for the Boltzmann factors, because all the energies in the problem are then expressible as sums of integers and integer multiples of $\sqrt{3}$. The values of $D / 2 J$ for which data were obtained are 1, 2, 3, 6, and $\infty$.
The discretization of the phase space of the model has significant effects at very low $T$, but the effects at the temperatures we study here are expected to be negligible compared to our statistical errors. The probability distributions for the local magnetization of equilibrium states which are calculated for the ${\bf Z}_{12}$ model are found to have very small contributions from the third and higher harmonics of $\cos ( \phi )$ and $\sin ( \phi )$. This is strong evidence that the 12-state clock model is an accurate approximation to the $XY$ model within our range of parameters. The ${\bf Z}_{12}$ model shows equivalent behavior for $D$ and $- D$, unlike the ${\bf Z}_{6}$ model used earlier.[@Fis95]
The program uses two independent linear congruential pseudorandom number generators, one for choosing the values of the $\theta_i$, and a different one for the Monte Carlo spin flips, which are performed by a single-spin-flip heat-bath algorithm. The code was checked by setting $D = 0$, and seeing that the known behavior of the pure ferromagnetic system was reproduced correctly.
Each sample was started off in a random spin state at a temperature significantly above the $T_c$ for the pure model, and cooled slowly. Thermal averages for $S (\vec{\bf k})$ were obtained at a set of temperatures spanning the critical region.
![\[Fig.1\] Angle-averaged structure factor for $64 \times 64 \times 64$ lattices with $D / 2 J = 1$ at $T = 2.203125$. The axes are scaled logarithmically.](RAD2J1SkofTc.EPS){width="3.4in"}
![\[Fig.2\] Angle-averaged structure factor for $64 \times 64 \times 64$ lattices with $D / 2 J = 2$ at $T = 2.1875$. The axes are scaled logarithmically.](RAD2J2SkofTc.EPS){width="3.4in"}
![\[Fig.3\] Angle-averaged structure factor for $64 \times 64 \times 64$ lattices with $D / 2 J = 3$ at $T = 2.171875$. The axes are scaled logarithmically.](RAD2J3SkofTc.EPS){width="3.4in"}
![\[Fig.4\] Angle-averaged structure factor for $64 \times 64 \times 64$ lattices with $D / 2 J = 6$ at $T = 2.078125$. The axes are scaled logarithmically.](RAD2J6SkofTc.EPS){width="3.4in"}
![\[Fig.5\] Angle-averaged structure factor for $64 \times 64 \times 64$ lattices with $D = \infty$ at $T = 1.921875$. The axes are scaled logarithmically.](RADINFSkofTc.EPS){width="3.4in"}
The magnetic structure factor, $S (\vec{\bf k}) = \langle | \vec{\bf M}(\vec{\bf
k}) |^2 \rangle $, for $n = 2$ spins is $$S (\vec{\bf k}) ~=~ L^{-3} \sum_{ i,j } \cos ( \vec{\bf k} \cdot
\vec{\bf r}_{ij}) \langle \cos ( \phi_{i} - \phi_{j}) \rangle \, ,$$ where $\vec{\bf r}_{ij}$ is the vector on the lattice which starts at site $i$ and ends at site $j$. Here the angle brackets denote a thermal average. For a RAM with $n > 1$, unlike the RBIM, the longitudinal part of the magnetic susceptibility, $\chi_l$, which is given by $$T \chi_l (\vec{\bf k}) ~=~ 1 - M^2 ~+~ L^{-3} \sum_{ i \ne j } \cos (
\vec{\bf k} \cdot \vec{\bf r}_{ij}) (\langle \cos ( \phi_{i} - \phi_{j}
) \rangle ~-~ Q_{ij} ) \, ,$$ where $M^2 ~=~ \langle |\vec{\bf M}| \rangle^2$, and $Q_{ij} ~=~ \langle
\vec{\bf S}_{i} \rangle \cdot \langle \vec{\bf S}_{j} \rangle$. For O(2) spins $$M^2 ~=~ L^{-6} [ \langle |\sum_{i} \cos ( \phi_{i} )| \rangle^2
~+~ \langle |\sum_{i} \sin ( \phi_{i} )| \rangle^2 ] \, ,$$ and $$Q_{ij} ~=~ \langle \cos ( \phi_{i} ) \rangle \langle \cos ( \phi_{j} )
\rangle ~+~ \langle \sin ( \phi_{i} ) \rangle \langle \sin ( \phi_{j} )
\rangle \, .$$
Thus $M^2$ is not the same as $S$, even above $T_c$. The scalar quantity $M^2$ is a well-behaved function of the lattice size $L$ for finite lattices, which approaches its large $L$ limit smoothly as $L$ increases, except possibly at a phase transition. The vector $\vec{\bf M}$, on the other hand, may not be a well-behaved function of $L$ for an $XY$ model in a two-fold random field. Knowing the local direction in which $\vec{\bf M}$ is pointing, averaged over some small part of the lattice, may not give us a strong constraint on what $\vec{\bf M}$ for the entire lattice will be.
The critical exponent $\eta$, is defined at $T = T_c$ by the small $\vec{k}$ behavior $$S (\vec{\bf k}) ~\approx~ c |\vec{\bf k}|^{- (2 - \eta)} \, ,$$ where $c$ is some constant. For each value of $D / J$, results for four different $L = 64$ configurations of the random anisotropy $\theta_i$ were averaged. The same four samples of random $\theta_i$ were used for all values of $T$, and all values of $D / J$, in order to facilitate the comparison of results for different values of $T$ and $D$.
All of the data shown in these figures were obtained from Monte Carlo runs which used hot start initial conditions, starting at at temperature well above $T_c$. The value of $T$ was then lowered in steps. The initial part of the run at each $T$ was discarded to allow the system to equilibrate. For these $L = 64$ runs with $D / 2 J$ = 1, at each $T$ a sequence of spin states obtained at intervals of 20,480 Monte Carlo steps per spin (MCS) was Fourier transformed and averaged. For the larger values of $D$, where the relaxation times are longer, this interval was chosen to be 102,400 MCS. The number of these selected spin states was chosen to be 16 for each of the finite values of $D / J$, and 32 for $D / J = \infty$. The Fourier transformed spin state data were then binned according to the values of $k$ = $|\vec{\bf k}|$, to give the angle-averaged $S ( k )$. Finally, a configuration average over the four random samples was performed. Both equally weighted and logarithmically weighted averages were tried. No significant differences were found between these two types of weighting, and only the equally weighted averages will be displayed here.
The results for $D / 2 J$ = 1, 2, 3, 6, and $\infty$ are shown in Figs. 1, 2, 3, 4, and 5, respectively. The values of $T$ which are used in these figures are convenient binary fraction approximations to the values of $T_c$ at these values of $D / 2 J$. The best estimates of $T_c$ were determined later, by the analysis of the data over a range of $L$ and $T$. We see from these figures that $S ( k )$ is only a weakly varying function of $D / J$, at least for $L = 64$.
The values of $2 - \eta$, as displayed on the figures, were found by least-squares fits to the data points for $0 < k < \pi/8$, where the data are well-approximated by Eqn. 8. Note that $\eta$ appears to be a slowly varying monotonic function of $D / J$, and that the extrapolation of $\eta$ down toward $D$ = 0 appears to be significantly different from the value of $\eta$ found for the nonrandom $n = 2$ ferromagnet.[@LGZJ80; @LT89] It is also interesting to note that the value of $\eta$ found for $D = \infty$ appears to be identical to the value of $\eta$ for the nonrandom system, but the significance of this is unclear.
The fact that $\eta$ appears to vary with $D / J$ is an indication that the claim of Reed[@Reed91] is too simplistic. He did not calculate a numerical value for $\eta$, but he argued that the finite-size scaling (FSS) behavior at $D / 2 J$ = 1 was indistinguishable from that of the nonrandom system.
One should not conclude from these data that $\eta$ is varying continuously with $D$, so that there is a line of critical points. Another explanation of the data is that for any $D$ we have a function $D_{eff} ( D / J, L )$, which increases very slowly as $L$ increases, up to a value $D_{eff} = D^*$ Then we will only find $\eta_{eff} = \eta^*$ when $L$ becomes large enough so that $D_{eff}
\approx D^*$. In the Cayley tree mean-field approximation,[@HCB87] whether $D^*$ is finite or infinite depends on the value of $z / n$.
 Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 1$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J1avmagfss.EPS "fig:"){width="3.1in"} Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 1$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J1chi_lfss.EPS "fig:"){width="3.1in"}
 Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 2$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J2avmagfss.EPS "fig:"){width="3.1in"} Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 2$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J2chi_lfss.EPS "fig:"){width="3.1in"}
 Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 3$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J3avmagfss.EPS "fig:"){width="3.1in"} Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 3$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J3chi_lfss.EPS "fig:"){width="3.1in"}
 Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 6$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J6avmagfss.EPS "fig:"){width="3.1in"} Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 6$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RAD2J6chi_lfss.EPS "fig:"){width="3.1in"}
 Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / J = \infty$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RADINFavmagfss.EPS "fig:"){width="3.1in"} Finite-size scaling near $T_c$ for $L \times L \times L$ lattices with $D / J = \infty$. (a) Configuration-averaged magnetization vs. temperature. (b) $\chi_l$ vs. temperature. The $y$-axis is scaled logarithmically.](RADINFchi_lfss.EPS "fig:"){width="3.1in"}
If we make the assumption that the usual critical exponent scaling laws for translation invariant models remain valid for the RAM, we can easily obtain values of the exponent combinations $\beta / \nu$ and $\gamma / \nu$ from our computed values of $2 - \eta$. These combinations are exactly what we need for FSS of the magnetization $| \vec{\bf M} ( L, T ) |$ and the magnetic susceptibility[@FB72] $\chi_l ( L, T )$. Thus, by making standard FSS plots,[@NB99] we can test the validity of these scaling laws for the RAM.
In Fig. 6(a) we show a FSS plot of the configuration average of $| \vec{\bf M}
( L, T ) |$ on $L \times L \times L$ lattices, for $L$ between 16 and 64. The number of sample configurations used for each $L < 64$ was 8 for $D / 2 J$ = 1, 2, 3 and 6, and 16 for $D = \infty$. For $L$ = 64, the number of samples was 4 for all $D$. Fig. 6(b) shows a similar plot for $\chi_l$. Figs. 7, 8, 9 and 10 show the corresponding plots for $D / 2 J$ = 2, 3, 6 and $\infty$, respectively. Since the values of $\eta$ used here were taken from the fits to the small k behavior of $S ( k )$, the only two adjustable fitting parameters used in these figures were the values of $\nu$ and $T_c$, which were required to be identical for parts (a) and (b) of each figure.
In these FSS plots, the temperature coordinate scales as $(T - T_c)L^{1 /
\nu}$. The reader should note that the range of $T$ which we cover in these plots is about an order of magnitude larger than the range which one would typically use for a problem where one is already confident about the nature of the phase transition, and one is trying to obtain high precision estimates of $T_c$ and the critical exponents by concentrating on the range of $T$ where $\xi \approx L$. As a consequence of this, the spacings between the values of $T$ for which we have taken data are rather large. Thus we are unable to use histogram reweighting[@FLS95] to obtain essentially continuous values for the thermodynamic functions.
From the results given in these figures, we see that the estimates of $\nu$ increase monotonically and the estimates of $T_c$ decrease monotonically as $D / 2 J$ increases. We also see that the peak in $\chi_l$ is slightly above $T_c$ for finite $L$, which is typical for ferromagnetic critical behavior. The data collapse is good near this peak, which is the range of $T$ for which $\xi > L$. We do not give estimates of statistical errors for $\nu$, because we believe that the variation in $\nu$ in the range $D / 2 J$ = 1 to 6 is due to variation in the value of $D_{eff}$. We will discuss this further in the next section. The errors in the values of $T_c$ are estimated to be less than $\pm 0.01$.
 Finite-size scaling of the difference between $c_H ( T_c )$= 3.00 and the configuration-averaged $c_H ( L, T )$ vs. temperature near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 1$.](RAD2J1c_Hfss.EPS){width="3.4in"}
 Finite-size scaling of the difference between $c_H ( T_c )$= 2.65 and the configuration-averaged $c_H ( L, T )$ vs. temperature near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 2$.](RAD2J2c_Hfss.EPS){width="3.4in"}
 Finite-size scaling of the difference between $c_H ( T_c )$= 2.40 and the configuration-averaged $c_H ( L, T )$ vs. temperature near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 3$.](RAD2J3c_Hfss.EPS){width="3.4in"}
 Finite-size scaling of the difference between $c_H ( T_c )$= 1.94 and the configuration-averaged $c_H ( L, T )$ vs. temperature near $T_c$ for $L \times L \times L$ lattices with $D / 2 J = 6$.](RAD2J6c_Hfss.EPS){width="3.4in"}
 Finite-size scaling of the difference between $c_H ( T_c )$= 0.81 and the configuration-averaged $c_H ( L, T )$ vs. temperature near $T_c$ for $L \times L \times L$ lattices with $D = \infty$.](RADINFc_Hfss.EPS){width="3.4in"}
Fig. 11 shows the difference between an estimate of the specific heat at $T_c$ for an infinite system, $c_H ( T_c )$, and the calculated specific heat of a finite system at temperature $T$, $c_H ( L, T )$ for $D / 2 J$ = 1. The only new adjustable fitting parameter here is $c_H ( T_c )$. Figs. 12, 13, 14 and 15 show the corresponding plots for $D / 2 J$ = 2, 3, 6 and $\infty$, respectively. The values of $c_H ( T_c )$ decrease monotonically as $D / 2 J$ increases. In all cases, the values of $c_H ( T_c )$ given in the figures are estimated to be accurate to about 1%. As we also saw for $|\vec{M}|$ and $\chi_l$, the FSS data collapse is not good below $T_c$ for $D / 2 J \le
3$. The results for $D = \infty$ are in very good agreement with the earlier results[@Fis95] obtained with the $Z_6$ approximation.
Discussion
==========
According to Imry and Ma[@IM75] and Pelcovits, Pytte and Rudnick[@PPR78], for small $D / J$ this model should appear ferromagnetic when $L$ is smaller than the “Imry-Ma length”, which is determined by balancing the domain wall energy against the energy of random pinning. If this length exists, when $L$ is larger than the Imry-Ma length, the system will break up into domains, without long-range order. We have argued here, however, that in the presence of random pinning one should not believe that the domain wall energy scales as $L^{d - 2}$. One can try to patch up this picture by assuming that the domain wall energy scales as $L^{d - \sigma_{dw}}$, with $3/2 < \sigma_{dw} < 2$. If this were the case, then it would still be possible for $d = 3$ to find a length scale where the domain wall energy balanced the random pinning energy. Then it would continue to be true in $d = 3$ that the system would break up into Imry-Ma-like domains when $L$ became very large.
What we see in our FSS plots for $D / 2 J \le 3$, however, is that when $T < T_c$ the leading correction to finite-size scaling increases the magnetization as $L$ increases. Therefore this model appears to be stable against domain formation for $D / 2 J \le 3$, at least for some range of $T$ below $T_c$. The natural interpretation of this result is that $\sigma_{dw}$ must be less than 3/2 in $d = 3$ for the $n = 2$ case.
Fig. 12, the FSS magnetization plot for $D / 2 J$ = 3, shows that as $L$ increases the data for $T < T_c$ seem to be converging to a scaling function which is independent of $L$. For the data in Fig. 15 for $D / 2 J$ = 6, the data appear to be in this $L$-independent limit. If we were able to do the Monte Carlo calculations at substantially larger values of $L$, we would expect to see the same type of convergence for $D / 2 J$ = 1 and 2.
If we had data at such large values of $L$, so that the magnetization scaling function had converged to an $L$-independent limit, then our estimates of the critical exponents would be expected to shift somewhat. Therefore, it is likely that $\eta^*$, the true value of $\eta$ in the range of $D / 2 J$ from 1 to 6, is actually independent of $D / J$.
The reader must also remember that the ferromagnetic phase is allowed to be reentrant. Therefore, we do not claim that the ferromagnetic behavior which we see below $T_c$ must be stable down to $T$ = 0 over the entire range of $D / J$. Also, we do not claim stable ferromagnetism for very large values of $D / J$. It must be stated, however, that this only applies to the simple cubic lattice with nearest neighbor interactions. We expect that it would be possible to stabilize a ferromagnetic phase at $D = \infty$ by adding further neighbor finite-range exchange interactions.
We point out that our earlier claim[@Fis95] of infinite magnetic susceptibility without ferromagnetism when $D = \infty$ was based on results at $T = 0$.[@Fis91] Since the magnetization of finite simple cubic lattices with $D = \infty$ seems to be a monotonically decreasing function of $T$,[@Fis95] however, we consider the existence of true ferromagnetism on this lattice to be unlikely at any $T$ for $D = \infty$.
The author sees no reason to believe that the exponent $\sigma_{dw}$ should be independent of $n$ for $d = 3$. Thus, while we claim the existence of a ferromagnetic phase for $n = 2$, we are not making any claim here about the behavior for $n = 3$. We do expect that $\sigma_{dw}$ must converge to 2 in the limit $n \to \infty$, in agreement with the result of Larkin.[@Lar70] The reason for this is that for $n \to \infty$ the “elastic membrane” approximation becomes valid.
Clearly, it would be desirable to obtain a direct estimate of $\sigma_{dw}$, by, for example, calculating the change in energy of a sample between periodic and antiperiodic boundary conditions along one direction. Since the energies involved are subextensive and the domain wall energy goes to zero at $T_c$, it is difficult to do such a calculation.
Summary
=======
In this work we have presented Monte Carlo results for the $d = 3$ $XY$ random anisotropy model, Eqn. (2), for several values of the anisotropy strength $D / J$. By studying the finite-size scaling behavior of $L
\times L \times L$ simple cubic lattices over the range $16 \le L \le 64$, we find that, for values of $D / J$ which are not very large, there appears to be a finite-temperature critical point at which the model undergoes a transition into a ferromagnetic phase. For this lattice at very large $D / J$, the transition appears to be into a phase with QLRO, but no true magnetization.
The author thanks the Physics Department of Princeton University for providing use of the computers on which the data were obtained.
[49]{}
R. Harris, M. Plischke and M. J. Zuckermann, Phys. Rev. Lett. [**31**]{}, 160 (1973). R. Fisch, Phys. Rev. B [**51**]{}, 11507 (1995). B. Derrida and J. Vannimenus, J. Phys. C [**13**]{}, 3261 (1980). R. Fisch and A. B. Harris, Phys. Rev. B [**41**]{}, 11305 (1990). A. B. Harris, R. G. Caflisch and J. R. Banavar, Phys. Rev. B, [**35**]{}, 4929 (1987). C. Jayaprakash and S. Kirkpatrick, Phys. Rev. B [**21**]{}, 4072 (1980). F. P. Toldin, A. Pelissetto and E. Vicari, J. Stat. Mech. P06002 (2006). F. Liers, J. Lukic, E. Marinari, A. Pelissetto and E. Vicari, Phys. Rev. B [**76**]{}, 174423 (2007). H. Nishimori, Prog. Theor. Phys. [**66**]{}, 1169 (1981). M. Hasenbusch, F. P. Toldin, A. Pelissetto and E. Vicari, Phys. Rev. B [**76**]{}, 184202 (2007). J. H. Chen and T. C. Lubensky, Phys. Rev. B [**16**]{}, 2106 (1977). R. Fisch, Phys. Rev. B [**41**]{}, 11705 (1990). A. Aharony and E. Pytte, Phys. Rev. Lett. [**45**]{}, 1583 (1980). A. Aharony and E. Pytte, Phys. Rev. B [**27**]{}, 5872 (1983). D. E. Feldman, Int. J. Mod. Phys. B [**15**]{}, 2945 (2001). R. Fisch, Phys. Rev. B [**58**]{}, 5684 (1998). C. Wang, J. Harrington and J. Preskill, Ann. Phys. (N.Y.) [**303**]{}, 31 (2003). M. Hasenbusch, F. P. Toldin, A. Pelissetto and E. Vicari, Phys. Rev. E [**77**]{}, 051115 (2008). R. A. Pelcovits, E. Pytte and J. Rudnick, Phys. Rev. Lett. [**40**]{}, 476 (1978). R. A. Pelcovits, Phys. Rev. B [**19**]{}, 465 (1979) A. I. Larkin, Zh. Eksp. Teor. Fiz. [**58**]{}, 1466 (1970) \[Sov. Phys. JETP [**31**]{}, 784 (1970)\]. Y. Imry and S.-K. Ma, Phys. Rev. Lett. [**35**]{}, 1399 (1975). A. Aharony, Y. Imry and S.-K. Ma, Phys. Rev. Lett. [**36**]{}, 1364 (1976). G. Parisi and N. Sourlas, Nucl. Phys. [**B206**]{}, 321 (1982). J. Z. Imbrie, Phys. Rev. Lett. [**53**]{}, 1747 (1984). J. Bricmont and A. Kupiainen, Phys. Rev. Lett. [**59**]{}, 1829 (1987). M. Mezard and A. P. Young, Europhys. Lett. [**18**]{}, 653, (1992). M. Aizenman and J. Wehr, Commun. Math. Phys. [**130**]{}, 489 (1990). B. I. Halperin, in [*Physics of Defects, Les Houches 1980*]{}, edited by R. Balian, M. Kleman and J. P. Poirier (North Holland, Amsterdam, 1981) pp. 837-844. e.g. D. S. Fisher, Phys. Rev. B [**31**]{}, 7233 (1985). S. N. Coppersmith, Phys. Rev. B [**44**]{}, 2887 (1991). A Aharony, J. Phys. C [**14**]{}, L841 (1981). A. Houghton, R. D. Kenway and S. C. Ying, Phys. Rev. B [**23**]{}, 298 (1981). J. L. Cardy and S. Ostlund, Phys. Rev. B [**25**]{}, 6899 (1982). R. Fisch, Phys. Rev. B [**46**]{}, 242 (1992). G. Kohring, R. E. Shrock and P. Wills, Phys. Rev. Lett. [**57**]{}, 1358 (1986). R. Fisch, Phys. Rev. B [**52**]{}, 12512 (1995). R. Fisch, Phys. Rev. B [**62**]{}, 361 (2000). R. Fisch, Phys. Rev. B [**55**]{}, 8211 (1997). R. Fisch, Phys. Rev. B [**76**]{}, 214435 (2007). J. C. Le Guillou and J. Zinn-Justin, Phys. Rev. B [**21**]{}, 3976 (1980). Y.-H. Li and S. Teitel, Phys. Rev. B [**40**]{}, 9122 (1989). P. Reed, J. Phys. A [**24**]{}, L117 (1991). M. E. Fisher and M. N. Barber, Phys. Rev. Lett. [**28**]{}, 1516 (1972). M. E. J. Newman and G. T. Barkema, [*Monte Carlo Methods in Statistical Physics*]{}, (Oxford University Press, Oxford, 1999), pp. 232-239. A. M. Ferrenberg, D. P. Landau and R. H. Swendsen, Phys. Rev. E [**51**]{}, 5092 (1995). R. Fisch, Phys. Rev. Lett. [**66**]{}, 2041 (1991).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce an adelic Cartier divisor over a trivially valued field and discuss the bigness of it. For bigness, we give the integral representation of the arithmetic volume and prove the existence of limit of it. Moreover, we show that the arithmetic volume is continuous and log concave.'
author:
- 'Tomoya Ohnishi[^1]'
title: Volume function over a trivially valued field
---
Introduction
============
Arakelov geometry is a kind of arithmetic geometry. Beyond scheme theory, it has been developed to treat a system of equations with integer coefficients such by adding infinite points. In some sense, it is an extension of Diophantine geometry. It is started by Arakelov [@arakelov1974intersection], who tried to define the intersection theory on an arithmetic surface. His result was not complete, but it was done by Faltings [@faltings1984calculus]. He gave the complete intersection theory on a arithmetic surface, such as the arithmetic Riemann-Roch theorem, the Noether formula and so on. Before a higher dimensional case, we would see the 1-dimensional case, that is, the arithmetic curves.
An arithmetic curve is the spectrum of the integer ring $O_K$ of a number field $K$. For example, ${\mathrm{Spec}\,}{\mathbb{Z}}$ is an arithmetic curve. It is not a proper scheme, so it is difficult to treat ${\mathrm{Spec}\,}{\mathbb{Z}}$ like an algebraic curve. For instance, the degree of principal divisors might not be zero, and the set of Cartier divisors modulo principal divisors is trivial. This problem is solved by considering the valuation theory. The scheme ${\mathrm{Spec}\,}{\mathbb{Z}}$ is the set of prime numbers in ${\mathbb{Q}}$, which corresponds to the finite places of ${\mathbb{Q}}$. But the field ${\mathbb{Q}}$ has not only finite places but also the infinite place. Hence by adding the point which corresponds to the infinite place to ${\mathrm{Spec}\,}{\mathbb{Z}}$, we can make ${\mathrm{Spec}\,}{\mathbb{Z}}$ “compact”. This idea is very successful. Now we can define an arithmetic divisor on ${\mathrm{Spec}\,}{\mathbb{Z}}$ as a pair of a Cartier divisor and a real number, which is a counterpart of the infinite place (more precisely, it is a continuous ${\mathbb{R}}$-valued function on a single point set). The degree of principal divisors is not zero in general, but the arithmetic degree of arithmetic principal divisors is always zero because of the product formula: $$|f|_\infty \cdot \prod_{p : \text{primes}} |f|_p = 1, \quad \text{for}\ \forall f \in {\mathbb{Q}}^\times,$$ where ${|{\raisebox{.4ex}{.}}|}_\infty$ is the usual absolute value and ${|{\raisebox{.4ex}{.}}|}_p$ is the $p$-adic absolute value on ${\mathbb{Q}}$. Moreover, the set of arithmetic divisors modulo arithmetic principal divisors on ${\mathrm{Spec}\,}{\mathbb{Z}}$ is isomorphic to ${\mathbb{R}}$. This result corresponds to the fact that the Picard group of ${\mathbb{P}}^1$ is isomorphic to ${\mathbb{Z}}$. In this way like an algebraic curve, we can study the geometry on ${\mathrm{Spec}\,}{\mathbb{Z}}$ equipped with the infinite point. Above all, the product formula plays an important role.
Next, we briefly recall the higher dimensional case. An arithmetic variety $X$ is an integral scheme flat and quasi-projective over ${\mathrm{Spec}\,}{\mathbb{Z}}$. As we equip ${\mathrm{Spec}\,}{\mathbb{Z}}$ with the infinite point, we consider not only $X$ but also the counterpart of the infinite point, which is the analytic space $X({\mathbb{C}})$. The main tools of Arakelov geometry are the intersection theory and an arithmetic divisor. The intersection theory on $X$ was given by Gillet-Soul[é]{} [@gillet1990arithmetic] by using Green currents on $X({\mathbb{C}})$. They also proved the general arithmetic Riemann-Roch theorem on $X$ [@gillet1992arithmetic]. One of the most famous application of the arithmetic Riemann-Roch theorem is the solution of the Mordell conjecture due to Voijta [@vojta1991siegel]. As related to arithmetic divisors, it was used to show the existence of small sections of $H^0(X,\overline{D})$, where $\overline{D}$ is an arithmetic divisor. A small section is a section whose norm is less than or equal 1. In Arakelov geometry, it plays the same role as a global section in algebraic geometry. Therefore it is also important to study the asymptotic behavior of the amount of small sections of $H^0(X,n\overline{D})$ as $n \rightarrow \infty$. To study it, Moriwaki [@moriwaki2009continuity] introduced the arithmetic volume $${\widehat{{\mathrm{vol}}}}(\overline{D}) = \limsup_{m \rightarrow +\infty} \frac{\hat{h}^0(X,m\overline{D})}{m^d/d!},$$ where $d = \dim X$ and $\hat{h}^0(X,\overline{D}) = \log \# \{ \text{small sections of}\ H^0(X,\overline{D})\}$. Moriwaki [@moriwaki2009continuity] proved the continuity of the arithmetic volume, Chen and Boucksom [@boucksom2011okounkov] proved the concavity. For general theory, we refer to [@moriwaki2014arakelov].
Recently, by considering the analytic space associated with not only infinite places but also finite places, the study of the adelic version of Arakelov geometry [@moriwaki2016adelic] has also developed. In this theory, the analytic space associated with a non-Archimedean absolute value in the sense of Berkovich plays the main role. Moreover, as a further generalization, Chen and Moriwaki [@chen2019arakelov] introduced the notion of adelic curves. It is a field equipped with a measure space which is consist of absolute values, and the “product formula”. This notion contains several classical settings. For example, we can treat algebraic curves and arithmetic curves as adelic curves. However, since the notion of adelic curves is very general, things that cannot be considered as a curve may be an adelic curve. One of them is a trivially valued field. It is a field $K$ equipped with the trivial product formula, that is, $$|f|_0 = 1, \quad \text{for}\ \forall f \in K^\times.$$ Hence Arakelov geometry over a trivially valued field is the geometry of schemes over an adelic curve ${\mathrm{Spec}\,}K$.
As the classical Arakelov geometry, there is the arithmetic volume function of an adelic Cartier divisor $\overline{D}$ on a projective variety $X$, which was introduced by Moriwaki and Chen [@chen2017sufficient]: $${\widehat{{\mathrm{vol}}}}(\overline{D}) = \limsup_{n \rightarrow \infty} \frac{{\widehat{\deg}}_+(n\overline{D})}{n^{d+1}/(d+1)!},$$ where $d = \dim X$. In the classical setting, the invariants $\hat{h}^0(X,\overline{D})$ and ${\widehat{\deg}}_+(\overline{D})$ behave in a similar way. For detail, we refer to [@chen2010arithmetic], [@chen2015majorations] and [@chen2019arakelov]. We say that an adelic Cartier divisor $\overline{D}$ is big if ${\widehat{{\mathrm{vol}}}}(\overline{D}) > 0$. In this paper, we will show several properties of the arithmetic volume ${\widehat{{\mathrm{vol}}}}({\raisebox{.4ex}{.}})$.
Let $\overline{D},\overline{E}$ be adelic Cartier divisors. The arithmetic volume has the following properties:
1. (integral formula). $${\widehat{{\mathrm{vol}}}}(\overline{D}) = (d+1)\int_0^\infty F_{\overline{D}}(t)\ \mathrm{d}t,$$ where $F_{\overline{D}}$ is a function given by $\overline{D}$. (c.f. Theorem \[vol1\]).
2. (limit existence). $${\widehat{{\mathrm{vol}}}}(\overline{D}) = \lim_{n \rightarrow \infty} \frac{{\widehat{\deg}}_+(n\overline{D})}{n^{d+1}/(d+1)!}.$$ (c.f. Theorem \[vol1\]).
3. (continuity). If $D$ is big, we have $$\lim_{\epsilon \rightarrow 0} {\widehat{{\mathrm{vol}}}}(\overline{D} + \epsilon \overline{E}) = {\widehat{{\mathrm{vol}}}}(\overline{D}).$$ (c.f. Theorem \[cvol\]).
4. (homogeneity). For $a \in {\mathbb{R}}_{>0}$, $${\widehat{{\mathrm{vol}}}}(a\overline{D}) = a^{d+1}{\widehat{{\mathrm{vol}}}}(\overline{D}).$$ (c.f. Corollary \[vol2\]).
5. (log concavity). If $\overline{D},\overline{E}$ are big. we have $${\widehat{{\mathrm{vol}}}}(\overline{D} + \overline{E})^{\frac{1}{d+1}} \geq {\widehat{{\mathrm{vol}}}}(\overline{D})^{\frac{1}{d+1}} + {\widehat{{\mathrm{vol}}}}(\overline{E})^{\frac{1}{d+1}}.$$ (c.f. Theorem \[cvol2\]).
Section 2 is devoted to preliminary such as algebraic geometry, normed vector spaces and the Berkovich spaces. In Section 3, we see the fundamental result of Arakelov geometry over a trivially absolute values. In Section 4, we discuss the properties of the arithmetic volume. For example, we will prove the continuity and the concavity of the arithmetic volume.
Preliminary
===========
${\mathbb{Q}}$- and ${\mathbb{R}}$-divisors
-------------------------------------------
Let $X$ be a variety over a field $K$ and $K(X)$ be a function field of $X$. By abuse of notation, we also denote the (constant) sheaf of rational functions on $X$ by $K(X)$. Firstly, we recall the definitions of Cartier divisors and Weil divisors (for detail, see [@hartshorne2013algebraic] and [@liu2006algebraic]).
Let ${\mathrm{Div}}(X) := H^0(X,K(X)^\times/{\mathcal{O}}_X^\times)$, whose element is called a *Cartier divisor*. A non-zero rational function $f \in K(X)^\times$ naturally gives rise to a Cartier divisor, which is called a *principal Cartier divisor* (or simply a *principal divisor*) and denoted by $(f)$. We denote the group law on ${\mathrm{Div}}(X)$ additive way. We say that two Cartier divisors $D_1,D_2 \in {\mathrm{Div}}(X)$ are *linearly equivalent* if $D_1-D_2$ is principal, which is denoted by $D_1 \sim D_2$. We set ${\mathrm{Pic}}(X) := {\mathrm{Div}}(X)/\sim$, which is called the *Picard group of $X$*. We say that a Cartier divisor $D \in {\mathrm{Div}}(X)$ is *effective* if it is contained in the image of the canonical map $H^0(X,{\mathcal{O}}_X \cap K(X)^\times) \rightarrow H^0(X,K(X)^\times/{\mathcal{O}}_X^\times)$. For two Cartier divisors $D_1,D_2$, we write $D_1 \geq D_2$ if $D_1-D_2$ is effective. In particular, we write $D \geq 0$ if $D$ is effective. For an open subset $U$ of $X$, let $D|_U$ be the image of $D$ by the canonical restriction $H^0(X,K(X)^\times/{\mathcal{O}}_X) \rightarrow H^0(U,K(X)^\times/{\mathcal{O}}_X^\times)$, which gives a Cartier divisor on $U$.
By definition, for $D \in {\mathrm{Div}}(X)$, there is an open covering $\{U_i\}$ of $X$ such that $D$ is given by some non-zero rational function $f_i \in K(X)^\times$ on $U_i$ and $f_i/f_j \in {\mathcal{O}}_X(U_i \cap U_j)^\times$ for $i \neq j$. In the above setting, $D$ is effective if and only if $f_i$ is regular on $U_i$, that is, $f_i \in {\mathcal{O}}_X(U_i)$ for all $i$.
We can associate any Cartier divisor $D = \{(U_i,f_i)\} \in {\mathrm{Div}}(X)$ with a subsheaf ${\mathcal{O}}_X(D) \subset K(X)$, which is given by ${\mathcal{O}}_X(D)|_{U_i} := f_i^{-1}{\mathcal{O}}_X|_{U_i}$. It is well-known that this construction is independent of the choice of a representation $\{(U_i,f_i)\}$ of $D$ and ${\mathcal{O}}_X(D)$ is an invertible ${\mathcal{O}}_X$-module on $X$.
Let $D_1,D_2$ be Cartier divisors.
1. ${\mathcal{O}}_X(D_1) \simeq {\mathcal{O}}_X(D_2)$ if $D_1 \sim D_2$.
2. ${\mathcal{O}}_X(D_1 + D_2) \simeq {\mathcal{O}}_X(D_1) \otimes_{{\mathcal{O}}_X} {\mathcal{O}}_X(D_2)$.
We denote $\Gamma(U,{\mathcal{O}}_X(D))$ by $\Gamma(U,D)$ for an open subset $U$ of $X$. For any open subset $U$ of $X$, we have $$\label{sec1}
\Gamma(U,D) = \{f \in K(X)^\times \,|\, (D + (f))|_U \geq 0 \} \cup \{0\}$$ by definition.
Conversely, we can associate any invertible ${\mathcal{O}}_X$-module ${\mathcal{L}}$ with a Cartier divisor $D$ such that ${\mathcal{L}}\simeq {\mathcal{O}}_X(D)$. Let $s$ be a non-zero rational section of ${\mathcal{L}}$, that is, $s \in {\mathcal{L}}_\eta \setminus \{0\}$ where $\eta$ is the generic point of $X$. Let $\{U_i\}$ be an open covering of $X$ which trivializes ${\mathcal{L}}$, and $\omega_i \in {\mathcal{L}}(U_i)$ be a local basis of ${\mathcal{L}}$ for each $i$. Then $s$ is denoted by $f_i\omega_i$ on $U_i$ for some $f_i \in K(X)$. The date $\{(U_i, f_i)\}$ gives the required Cartier divisor $\mathrm{div}(s)$. For example, if we choose $1$ as a rational section of ${\mathcal{O}}_X(D)$, then we have $\mathrm{div}(1) = D$ by its construction.
Next, we assume that $X$ is normal. Let $X^{(1)} = \{x \in X \,|\, \mathrm{codim}_X\overline{\{x\}} = 1\}$. For $x \in X^{(1)}$, let $[x] := \overline{\{x\}}$, which is an irreducible closed subset of $X$ and called a *prime divisor*.
Let ${\mathrm{WDiv}}(X) := \bigoplus_{x \in X^{(1)}} {\mathbb{Z}}[x]$, whose element is called a *Weil divisor*. If we write $$D=\sum_{x \in X^{(1)}}n_x[x],$$ we set ${\mathrm{ord}}_x(D) := n_x$. We say that a Weil divisor $D \in {\mathrm{WDiv}}(X)$ is *effective* if ${\mathrm{ord}}_x(D) \geq 0$ for all $x \in X^{(1)}$. For two Weil divisors $D_1,D_2$, we write $D_1 \geq D_2$ if $D_1-D_2$ is effective. In particular, we write $D \geq 0$ if $D$ is effective. For a non-empty open subset $U$ of $X$, let $$D|_U := \sum_{x \in X^{(1)} \cap U} {\mathrm{ord}}_x(D)[x],$$ which is called the restriction of a Weil divisor $D$ on $U$.
Let $x \in X^{(1)}.$ Since $X$ is normal, ${\mathcal{O}}_{X,x}$ is a discrete valuation ring. Hence we have the normalized discrete valuation ${\mathrm{ord}}_x$ on $K(X)$ associated with ${\mathcal{O}}_{X,x}$. For a non-zero rational function $f \in K(X)^\times$, let $$(f) := \sum_{x \in X^{(1)}} {\mathrm{ord}}_x(f) [x].$$ This is a Weil divisor and such a divisor is called a *principal Weil divisor* (or simply a *principal divisor*). We say that two Weil divisors $D_1,D_2 \in {\mathrm{WDiv}}(X)$ are *linearly equivalent* if $D_1-D_2$ is principal. Then we write $D_1 \sim D_2$.
We can associate any Cartier divisor $D \in {\mathrm{Div}}(X)$ with a Weil divisor as follows: For any $x \in X^{(1)}$, let $f \in K(X)$ be a local equation around $x$ of $D$. Then we set ${\mathrm{ord}}_x(D) := {\mathrm{ord}}_x(f)$. It is independent of the choice of a local equation. Hence we can define that $$D := \sum_{x \in X^{(1)}} {\mathrm{ord}}_x(D)[x].$$ This construction gives a homomorphism $\varphi:{\mathrm{Div}}(X) \rightarrow {\mathrm{WDiv}}(X)$.
1. The homomorphism $\varphi$ is injective. Moreover, $\varphi$ is an isomorphism if $X$ is regular. Hence we sometimes identify a Cartier divisor with a Weil divisor.
2. For any $D_1, D_2 \in {\mathrm{Div}}(X)$, $D_1 \sim D_2$ as Cartier divisors if and only if $D_1 \sim D_2$ as Weil divisors.
3. For any $D \in {\mathrm{Div}}(X)$, $D \geq 0$ as Cartier divisors if and only if $D \geq 0$ as Weil divisors.
We can associate any Weil divisor $D$ with a subsheaf ${\mathcal{O}}_X(D) \subset K(X)$, which is defined by $${\mathcal{O}}_X(D)|_U := \{f \in K(X)^\times \,|\, (D + (f))|_U \geq 0 \} \cup \{0\}$$ for any open subset $U$ of $X$. By (\[sec1\]), if $D$ is Cartier, the above construction gives the same invertible ${\mathcal{O}}_X$-module ${\mathcal{O}}_X(D)$. However ${\mathcal{O}}_X(D)$ is not invertible if $D$ is not Cartier.
Let ${\mathbb{K}}= {\mathbb{Q}}$ or ${\mathbb{R}}$. Let us introduce the definition of ${\mathbb{K}}$-divisors.
Let ${\mathrm{Div}}(X)_{\mathbb{K}}:= {\mathrm{Div}}(X) \otimes_{\mathbb{Z}}{\mathbb{K}}$, ${\mathrm{WDiv}}(X)_{\mathbb{K}}:= {\mathrm{WDiv}}(X) \otimes_{\mathbb{Z}}{\mathbb{K}}$ and $K(X)_{\mathbb{K}}^\times := K(X)^\times \otimes_{\mathbb{Z}}{\mathbb{K}}$. An element of ${\mathrm{Div}}(X)_{\mathbb{K}}$ (resp. ${\mathrm{WDiv}}(X)_{\mathbb{K}}$, $K(X)_{\mathbb{K}}^\times$) is called a *${\mathbb{K}}$-Cartier divisor* (resp. a *${\mathbb{K}}$-Weil divisor*, a *${\mathbb{K}}$-rational function*) on $X$. Clearly, Cartier divisors and ${\mathbb{Q}}$-Cartier divisors (resp. Weil divisors and ${\mathbb{Q}}$-Weil divisors) are ${\mathbb{R}}$-Cartier divisors (resp. ${\mathbb{R}}$-Weil divisors). A non-zero ${\mathbb{K}}$-rational function $f \in K(X)_{\mathbb{K}}^\times$ naturally gives rise to a ${\mathbb{K}}$-Cartier divisor (or equivalently a ${\mathbb{K}}$-Weil divisor), which is called a *${\mathbb{K}}$-principal divisor* and denoted by $(f)$. We say that two ${\mathbb{R}}$-Cartier divisors (resp. ${\mathbb{R}}$-Weil divisors) $D_1,D_2$ are *${\mathbb{K}}$-linearly equivalent* if $D_1-D_2$ is ${\mathbb{K}}$-principal, which is denoted by $D_1 \sim_{\mathbb{K}}D_2$. We say that a ${\mathbb{K}}$-Cartier divisor (resp. a ${\mathbb{K}}$-Weil divisor) $D$ is *effective* if $D$ is a linear combination of effective divisors with positive coefficients in ${\mathbb{K}}$. We write $D_1 \geq D_2$ if $D_1-D_2$ is effective. In particular, we write $D \geq 0$ if $D$ is effective.
Similarly to Cartier divisors, for $D \in {\mathrm{Div}}(X)_{\mathbb{K}}$, there is an open covering $\{U_i\}$ of $X$ such that $D$ is given by some non-zero ${\mathbb{K}}$-rational function $f_i \in K(X)_{\mathbb{K}}^\times$ on $U_i$ and $f_i/f_j \in ({\mathcal{O}}_X(U_i \cap U_j) \otimes_{\mathbb{Z}}{\mathbb{K}})^\times$ for $i \neq j$.
Let $D \in {\mathrm{WDiv}}(X)_{\mathbb{K}}$. By definition, we can write $D = \sum_{x \in X^{(1)}} k_x[x]$, where $k_x \in {\mathbb{K}}$ and $k_x = 0$ for all but finitely many $x \in X^{(1)}$. Then we define the round down of $D$ as follows: $$\lfloor D \rfloor := \sum_{x \in X^{(1)}} \lfloor k_x \rfloor [x].$$ This is a Weil divisor and $\lfloor D \rfloor = D$ if and only if $D \in {\mathrm{WDiv}}(X)$.
For $D \in {\mathrm{WDiv}}(X)_{\mathbb{K}}$, the *associated ${\mathcal{O}}_X$-module* ${\mathcal{O}}_X(D)$ is defined by ${\mathcal{O}}_X(\lfloor D \rfloor)$. Then we have $H^0(X,D) = \{f \in K(X)^\times \,|\, D + (f) \geq 0 \} \cup \{0\}$. We remark that $D + (f) \geq 0 \Leftrightarrow \lfloor D \rfloor + (f) \geq 0$ for any $f \in K(X)^\times$ and ${\mathcal{O}}_X(2D)$ is not isomorphic to ${\mathcal{O}}_X(D) \otimes_{{\mathcal{O}}_X} {\mathcal{O}}_X(D)$ in general.
Let $D \in {\mathrm{WDiv}}(X)_{\mathbb{K}}$. Then $H^0(X,D)$ is a finite-dimensional vector space over $K$.
Big divisors
------------
We simply recall the definitions of bigness of Cartier divisors. Let $X$ be a variety over a field $K$.
\[big\] Let $D$ be a Cartier divisor on $X$. Let $h^0(D) := \dim_K H^0(X,D)$ and $d = \dim X$. We define the *volume* ${\mathrm{vol}}(D)$ of $D$ as follows: $${\mathrm{vol}}(D) := \limsup_{n \rightarrow +\infty} \frac{h^0(nD)}{n^d/d!}.$$ We say that $D$ is *big* if ${\mathrm{vol}}(D) > 0$.
Later we will consider the volume of an ${\mathbb{R}}$-Weil divisor. Hence we extends the above definition.
Let $D$ be an ${\mathbb{R}}$-Weil divisor on a normal variety $X$. We define a function $\mathfrak{h}_D:{\mathbb{R}}_+ \rightarrow {\mathbb{Z}}$ by $\mathfrak{h}_D(t) := \dim_K H^0(tD) = \dim_K H^0(\lfloor tD \rfloor)$. The *volume* of $D$ is defined by $${\mathrm{vol}}(D) := \limsup_{t \rightarrow +\infty} \frac{\mathfrak{h}_D(t)}{t^d/d!},$$ where $d = \dim X$. We say that $D$ is *big* if ${\mathrm{vol}}(D) > 0$.
By Fulger, Koll[á]{}r and Lehmann [@fulger2016volume], the above definition agrees with one in Definition \[big\] if $X$ is proper and $D$ is Cartier.
Finally we recall the well-known properties of the volume function ${\mathrm{vol}}({\raisebox{.4ex}{.}})$ without a proof (for detail, see [@lazarsfeld2017positivity]).
Let $X$ be a proper normal variety and $d = \dim X$. Let $D,E$ be ${\mathbb{R}}$-Cartier (or ${\mathbb{R}}$-Weil) divisors on $X$.
1. $\displaystyle {\mathrm{vol}}(D) = \lim_{n \rightarrow +\infty} \frac{h^0(nD)}{n^d/d!} \left(= \lim_{t \rightarrow +\infty} \frac{\mathfrak{h}_D(t)}{t^d/d!} \right)$.
2. For $a \in {\mathbb{R}}_{>0}$, ${\mathrm{vol}}(aD) = a^d{\mathrm{vol}}(D)$.
3. The volume function ${\mathrm{vol}}({\raisebox{.4ex}{.}})$ is continuous, that is, ${\mathrm{vol}}(E) \rightarrow {\mathrm{vol}}(D)$ as $E \rightarrow D$ (which means that each coefficients of $E$ converge coefficients of $D$ as an ${\mathbb{R}}$-Weil divisor).
4. The volume function ${\mathrm{vol}}({\raisebox{.4ex}{.}})$ is $d$-concave on big divisors, that is, if $D,E$ are big, then $${\mathrm{vol}}(D+E)^{1/d} \geq {\mathrm{vol}}(D)^{1/d} + {\mathrm{vol}}(E)^{1/d}.$$
Normed vector space
-------------------
In this section, we study fundamental properties of a normed vector space over a field equipped with an absolute value. But we mainly consider a trivially valued field.
Let $K$ be a field.
We say that a map ${|{\raisebox{.4ex}{.}}|}:K \rightarrow {\mathbb{R}}_+$ is an *absolute value on $K$* if it satisfies the following conditions:
1. $\forall a \in K$, $|a|=0 \Leftrightarrow a=0$.
2. $\forall a,b \in K$, $|a|{\raisebox{.4ex}{.}}|b|=|ab|$.
3. (*triangle inequality*) $\forall a,b \in K$, $|a+b| \leq |a|+|b|$.
If an absolute value ${|{\raisebox{.4ex}{.}}|}$ also satisfies the following inequality $$\forall a,b \in K,\ |a+b| \leq \max\{|a|,|b|\},$$ we say that ${|{\raisebox{.4ex}{.}}|}$ is *non-Archimedean*. Otherwise, ${|{\raisebox{.4ex}{.}}|}$ is called *Archimedean*.
We say that an absolute value ${|{\raisebox{.4ex}{.}}|}$ on $K$ is *trivial* if it satisfies that $|a|=1$ for any $a \in K \setminus \{0\}$. A field $K$ equipped with the trivial absolute value ${|{\raisebox{.4ex}{.}}|}$ is called a *trivially valued field*. Clearly, the trivial absolute value is non-Archimedean and a trivially valued field is complete.
Let $V$ be a vector space over $K$.
We say that a map ${\|{\raisebox{.4ex}{.}}\|}:V \rightarrow {\mathbb{R}}_+$ is a *(multiplicative) norm over $(K,{|{\raisebox{.4ex}{.}}|})$* if it satisfies the following conditions:
1. $\forall v \in V$, $\|v\|=0 \Leftrightarrow v=0$.
2. $\forall a \in K$ and $v \in V$, $\|av\|=|a|{\raisebox{.4ex}{.}}\|v\|$.
3. (*triangle inequality*) $\forall v,w \in V$, $\|v+w\| \leq \|v\|+\|w\|$.
If a norm ${\|{\raisebox{.4ex}{.}}\|}$ also satisfies the following inequality $$\forall v,w \in V,\ \|v+w\| \leq \max\{\|v\|,\|w\|\},$$ we say that ${\|{\raisebox{.4ex}{.}}\|}$ is *ultrametric*. A pair $(V,{\|{\raisebox{.4ex}{.}}\|})$ is called a *normed vector space*.
Let $V_\bullet = \bigoplus_{n=0}^{\infty} V_n$ be a graded ring over $K$ such that $V_n$ is a vector space over $K$ for all $n$ and $V_0 = K$. Let ${|{\raisebox{.4ex}{.}}|}$ be an absolute value on $K$ and ${\|{\raisebox{.4ex}{.}}\|}_n$ be a norm of $V_n$ over $(K,{|{\raisebox{.4ex}{.}}|})$ for $n \in {\mathbb{Z}}_{\geq 0}$ such that ${\|{\raisebox{.4ex}{.}}\|}_0={|{\raisebox{.4ex}{.}}|}$ on $V_0=K$.
We say that $$(V_\bullet,{\|{\raisebox{.4ex}{.}}\|}_\bullet) := \bigoplus_{n=0}^\infty (V_n,{\|{\raisebox{.4ex}{.}}\|}_n)$$ is a *normed graded ring over $(K,{|{\raisebox{.4ex}{.}}|})$* if $\|v_m {\raisebox{.4ex}{.}}v_n\|_{m+n} \leq \|v_m\|_m{\raisebox{.4ex}{.}}\|v_n\|_n$ for all $v_m \in V_m$ and $v_n \in V_n$.
Let $W_\bullet = \bigoplus_{n=0}^\infty W_n$ be a $V_\bullet$-module such that $W_n$ is a vector space over $K$ for all $n$. Let $h \in {\mathbb{Z}}_{>0}$. We say that $W_\bullet$ is a *$h$-graded $V_\bullet$-module* if $v_m{\raisebox{.4ex}{.}}w_n \in W_{hm+n}$ for all $v_m \in V_m$ and $w_n \in W_n$. If $h=1$, $W_\bullet$ is simply called a *graded $V_\bullet$-module*.
Let ${\|{\raisebox{.4ex}{.}}\|}_{W_n}$ be a norm on $W_n$ over $(K,{|{\raisebox{.4ex}{.}}|})$ for $n \in {\mathbb{Z}}_{\geq0}$.
We say that $$(W_\bullet,{\|{\raisebox{.4ex}{.}}\|}_{W_\bullet}) := \bigoplus_{n=0}^\infty (W_n,{\|{\raisebox{.4ex}{.}}\|}_{W_n})$$ is a *normed $h$-graded $(V_\bullet,{\|{\raisebox{.4ex}{.}}\|}_\bullet)$-module* if $\|v_m{\raisebox{.4ex}{.}}w_n\|_{W_{hm+n}} \leq \|v_m\|_m{\raisebox{.4ex}{.}}\|w_n\|_{W_n}$ for all $v_m \in V_m$ and $w_n \in W_n$. If $h=1$, $(W_\bullet,{\|{\raisebox{.4ex}{.}}\|}_{W_\bullet})$ is simply called a *normed graded $(V_\bullet,{\|{\raisebox{.4ex}{.}}\|}_\bullet)$-module*.
In the following, let $(V,{\|{\raisebox{.4ex}{.}}\|})$ be an ultrametrically normed vector space over a trivially valued field $(K,{|{\raisebox{.4ex}{.}}|})$ and $\dim_K(V) < +\infty$.
\[norm1\]
1. Let $v_1,\dots, v_n \in V$. If $\|v_1\|,\dots, \|v_n\|$ are all distinct, then we have $\|v_1+\cdots+v_n\|=\max\{\|v_1\|,\dots,\|v_n\|\}$.
2. $\#\{\|v\| \,|\, v \in V \} \leq \dim_K(V) + 1$.
\(1) By induction of $n$, it is sufficient to show in the case of $n=2$. Let $\|x_1\| > \|x_2\|$. By definition, we have $\|x_1+x_2\| \leq \|x_1\|$. On the other hand, $\|x_1\| = \|(x_1 + x_2) - x_2\| \leq \max\{\|x_1+x_2\|,\|x_2\|\}$. Since $\|x_1\| > \|x_2\|$, we have $\|x_1\| \leq \|x_1+x_2\|$. Hence we get a conclusion.
\(2) It suffices to show that $v_1,\dots, v_n \in V$ are linearly independent if $\|v_1\|,\dots, \|v_n\|$ are all distinct. We assume that $v_1,\dots,v_n \in V \setminus \{0\}$ are not linearly independent, that is, $a_1 v_1+\cdots+a_n v_n = 0$ for some $a_1,\dots,a_n \in K$. We can assume that $a_i \neq 0$ for all $i$. Since $K$ is trivially valued, we have $\|av\|=\|v\|$ for any $a \in K$ and $v \in V$. Hence by (1), we have $0 = \|a_1 v_1 +\cdots+ a_n v_n\| = \max\{\|v_1\|,\dots,\|v_n\|\}$, which is a contradiction.
We set $${\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|}) := \{ v \in V \,|\, \|v\| \leq e^{-t} \} \quad \text{for} \ t \in {\mathbb{R}}.$$ Remark that ${\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|})$ is a vector space over $K$ for any $t \in {\mathbb{R}}$ because ${|{\raisebox{.4ex}{.}}|}$ is trivial. Then $\{{\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|})\}_{t \in {\mathbb{R}}}$ satisfies the following conditions:
\[rfil\]
1. For sufficiently positive $t \in {\mathbb{R}}$, ${\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|}) = \{0\}$.
2. For sufficiently negative $t \in {\mathbb{R}}$, ${\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|}) = V$.
3. For any $t \geq s$, ${\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|}) \subseteq {\mathcal{F}}^s(V,{\|{\raisebox{.4ex}{.}}\|})$.
4. The function ${\mathbb{R}}\ni t \mapsto \dim_K {\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|})$ is left-continuous.
\(1) and (2) follow from Lemma \[norm1\] and (3) and (4) follow from by definition.
We set $${\lambda_{\max}}(V,{\|{\raisebox{.4ex}{.}}\|}) := \sup\{t \in {\mathbb{R}}\,|\, {\mathcal{F}}^t(V,{\|{\raisebox{.4ex}{.}}\|}) \neq \{0\} \}.$$ By convention, ${\lambda_{\max}}(V,{\|{\raisebox{.4ex}{.}}\|}) = -\infty$ if $V = \{0\}$. By Proposition \[rfil\], we have ${\lambda_{\max}}(V,{\|{\raisebox{.4ex}{.}}\|}) < +\infty$ and by Lemma \[norm1\], we can replace “$\sup$” by “$\max$” in the above definition.
Berkovich space {#BS}
---------------
Let $K$ be a field equipped with an absolute value ${|{\raisebox{.4ex}{.}}|}$. We assume that $K$ is complete with respect to ${|{\raisebox{.4ex}{.}}|}$. Let $X$ be a scheme over ${\mathrm{Spec}\,}K$. We define the analytification of $X$ in the sence of Berkovich (for detail, see [@berkovich2012spectral]).
The *analytification of $X$ in the sense of Berkovich*, or *Berkovich space associated to $X$* is the set of pairs $x = (p,{|{\raisebox{.4ex}{.}}|}_x)$ where $p \in X$ and ${|{\raisebox{.4ex}{.}}|}_x$ is an absolute value on the residue field $\kappa(x) := \kappa(p)$ which is an extension of ${|{\raisebox{.4ex}{.}}|}$, denoted by $X^{\mathrm{an}}$. The map $j:X^{\mathrm{an}}\rightarrow X, (p,{|{\raisebox{.4ex}{.}}|}_x) \mapsto p$ is called the *specification map*.
Let $U$ be a non-empty Zariski open subset of $X$. The subset $U^{\mathrm{an}}:= j^{-1}(U)$ of $X^{\mathrm{an}}$ is called a *Zariski open subset of $X^{\mathrm{an}}$*. A regular function $f \in {\mathcal{O}}_X(U)$ on $U$ define a function $|f|$ on $U^{\mathrm{an}}$ as follows: $$|f|(x) := |f(j(x))|_x \quad \text{for} \ x \in U^{\mathrm{an}}.$$ We also denote $|f|(x)$ by $|f|_x$.
We define a topology on $X^{\mathrm{an}}$ as the most coarse topology which makes $j$ and $|f|$ continuous for any Zariski open subset $U$ of $X$ and $f \in {\mathcal{O}}_X(U)$. This is called the *Berkovich topology*. Remark that $X^{\mathrm{an}}$ is Hausdorff (resp. compact) if $X$ is separated (resp. proper) over ${\mathrm{Spec}\,}K$.
Let $f:X \rightarrow Y$ be a morphism of schemes over ${\mathrm{Spec}\,}K$. There is a continuous map $f^{\mathrm{an}}:X^{\mathrm{an}}\rightarrow Y^{\mathrm{an}}$ such that the following diagram is commutative: $$\begin{tikzcd}
X \arrow{r}{f} \arrow{d}{j} & Y \arrow{d}{j} \\
X^{\mathrm{an}}\arrow{r}{f^{\mathrm{an}}} & Y^{\mathrm{an}}\end{tikzcd}$$
Concretely, $f^{\mathrm{an}}$ is constructed as follows: Let $x = (p,{|{\raisebox{.4ex}{.}}|}_x) \in X^{\mathrm{an}}$ and $q = f(p) \in Y$. We remark that $\kappa(y)=\kappa(q)$ is a subfield of $\kappa(x)=\kappa(p)$. Then $y = f^{\mathrm{an}}(x)$ is given by $q=f(p)$ and the absolute value ${|{\raisebox{.4ex}{.}}|}_y$ on $\kappa(q)$ which is the restriction of ${|{\raisebox{.4ex}{.}}|}_x$.
In the following, $(K,{|{\raisebox{.4ex}{.}}|})$ is a trivially valued field. For $x \in X$, let $x^{\mathrm{an}}= (x,{|{\raisebox{.4ex}{.}}|}_0) \in X^{\mathrm{an}}$ where ${|{\raisebox{.4ex}{.}}|}_0$ is the trivial absolute value on $\kappa(x)$. This correspondence gives a section of $j$, which is denoted by $\sigma:X \rightarrow X^{\mathrm{an}}$.
Now we introduce an important subset of $X^{\mathrm{an}}$. We assume that $X$ is normal projective variety over ${\mathrm{Spec}\,}K$. Let $\eta \in X$ be the generic point of $X$ and $X^{(1)} = \{x \in X \,|\, \mathrm{codim}_X\overline{\{x\}} = 1\}$. Let $K(X)$ be the function field of $X$. Firstly, for $x \in X^{(1)}$, we set $$(\eta^{\mathrm{an}},x^{\mathrm{an}}) := \left\{ \xi \in X^{\mathrm{an}}\,\middle|\, j(\xi)=\eta, {|{\raisebox{.4ex}{.}}|}_\xi = e^{-t(\xi){\mathrm{ord}}_x({\raisebox{.4ex}{.}})}\ \text{on}\ K(X), t(\xi) \in (0,+\infty) \right\}$$ and $$[\eta^{\mathrm{an}},x^{\mathrm{an}}] := \{\eta^{\mathrm{an}}\} \cup (\eta^{\mathrm{an}},x^{\mathrm{an}}) \cup \{x^{\mathrm{an}}\}.$$ Then the correspondence $\xi \mapsto t(\xi)$, $\eta^{\mathrm{an}}\mapsto 0$ and $x^{\mathrm{an}}\mapsto +\infty$ gives a homeomorphism from $(\eta^{\mathrm{an}},x^{\mathrm{an}})$ (resp. $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$) to $(0,+\infty)$ (resp. $[0,+\infty]$). Hence we sometimes identify $(\eta^{\mathrm{an}},x^{\mathrm{an}})$ (resp. $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$) with $(0,+\infty)$ (resp. $[0,+\infty]$).
We set $X_{\mathrm{div}}^{\mathrm{an}}:= \bigcup_{x \in X^{(1)}} [\eta^{\mathrm{an}},x^{\mathrm{an}}]$. Then we can illustrate $X_{\mathrm{div}}^{\mathrm{an}}$ by an infinite tree as follows:
(0,1) node\[above\][$\eta^{\mathrm{an}}$]{} – (0,0) node\[below\][$x^{\mathrm{an}}$]{}; (0,1)–(-3,0); (0,1)–(-2,0); (0,1)–(1,0); (0,1)–(3,0); (-1,0) node[$\cdots$]{}; (2,0) node[$\cdots$]{}; (0,1) circle \[radius=1.5pt\]; (0,0) circle \[radius=1.5pt\]; (-3,0) circle \[radius=1.5pt\]; (-2,0) circle \[radius=1.5pt\]; (1,0) circle \[radius=1.5pt\]; (3,0) circle \[radius=1.5pt\];
We remark that $X_{\mathrm{div}}^{\mathrm{an}}= X^{\mathrm{an}}$ if $\dim X = 1$.
\[den\] $X_{\mathrm{div}}^{\mathrm{an}}$ is dense in $X^{\mathrm{an}}$.
For the proof, it is sufficient to show that, for any regular function $f$ on a Zariski open set $U$ in $X$ and any $x \in U^{\mathrm{an}}$, the value $|f|(x)$ is belonged to the closure $W$ of $\{ |f|(z) \,|\, z \in X_\mathrm{div}^{\mathrm{an}}\cap U^{\mathrm{an}}\} \subset {\mathbb{R}}_+$. If $f$ has no pole on $X$, then $f$ is regular on the whole $X$, so $f$ is a constant function and algebraic over $k$ because $X$ is normal and projective. Therefore $|f|(z) = 1$ on $X^{\mathrm{an}}$, so it is clear that $|f|(x) \in W$.
We next assume that $f$ has poles on $X \setminus U$. In this case, there are $y,y' \in X^{(1)}$ such that $f(y)=0$ and $f$ has a pole at $y'$ because $X$ is normal. Then, $|f|(t) = e^{-at}$ for $t \in (\eta^{\mathrm{an}},y^{\mathrm{an}})$, $|f|(t') = e^{a't'}$ for $t' \in (\eta^{\mathrm{an}},y'^{\mathrm{an}})$ for some $a,a'>0$ and $|f|(\eta^{\mathrm{an}}) = 1$, which implies that $W={\mathbb{R}}_+$ and we complete the proof.
Let ${\mathbb{R}}_{>0}$ be the multiplicative group of positive real numbers. There is an action of ${\mathbb{R}}_{>0}$ to $X^{\mathrm{an}}$. For $r \in {\mathbb{R}}_{>0}$ and $x=(p,{|{\raisebox{.4ex}{.}}|}_x) \in X^{\mathrm{an}}$, we define $$r^*x := (p,{|{\raisebox{.4ex}{.}}|}_x^r).$$ We also denote $r^*x$ by $x^r$. This action is called the *scaling action* in [@boucksom2018singular]. The scaling action is free faithful and preserve the subset $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$ for all $x \in X^{(1)}$.
Finally, we introduce the reduction map ${\mathrm{red}}:X^{\mathrm{an}}\rightarrow X$. For $x \in X^{\mathrm{an}}$, let $\widehat{\kappa}(x)$ be the completion of $\kappa(x)$ with respect to ${|{\raisebox{.4ex}{.}}|}_x$ and we also denote the absolute value on $\widehat{\kappa}(x)$ by ${|{\raisebox{.4ex}{.}}|}_x$. We set $o_x := \{f \in \widehat{\kappa}(x) \,|\, |f|_x \leq 1 \}$ and $m_x := \{f \in \widehat{\kappa}(x) \,|\, |f|_x < 1 \}$. Then $o_x$ is a local ring and $m_x$ is the maximal ideal of $o_x$. If ${|{\raisebox{.4ex}{.}}|}_x$ is trivial on $\kappa(x)$, then $o_x = \kappa(x)$ and $m_x = \{0\}$. Let $p_x:{\mathrm{Spec}\,}\widehat{\kappa}(x) \rightarrow X$ be a $K$-morphism of schemes defined by $j(x)$ and $\iota_x:{\mathrm{Spec}\,}\widehat{\kappa}(x) \rightarrow {\mathrm{Spec}\,}o_x$ be a $K$-morphism defined by the inclusion $o_x \hookrightarrow \widehat{\kappa}(x)$. By the valuation criterion of properness (for instance, see [@hartshorne2013algebraic]), there is a unique $K$-morphism $\phi_x:{\mathrm{Spec}\,}o_x \rightarrow X$ such that $p_x = \phi_x \circ \iota_x$. $$\begin{tikzcd}
{\mathrm{Spec}\,}\widehat{\kappa}(x) \arrow{r}{p_x} \arrow{d}{\iota_x} & X \arrow{d} \\
{\mathrm{Spec}\,}o_x \arrow[dashed]{ru}{\exists ! \phi_x} \arrow{r} & {\mathrm{Spec}\,}K
\end{tikzcd}$$ Then we define ${\mathrm{red}}(x) \in X$ to be the image of $m_x$ by $\phi_x$. The map ${\mathrm{red}}:X^{\mathrm{an}}\rightarrow X$ defined by the above correspondence is called the *reduction map*. The morphism $\phi_x$ induces a homomorphism ${\mathcal{O}}_{X,{\mathrm{red}}(x)} \rightarrow o_x$. Hence we have $$\label{red}
\forall f \in {\mathcal{O}}_{X,{\mathrm{red}}(x)}, |f|_x \leq 1.$$ We remark that $j \neq {\mathrm{red}}$. For example, for any $x \in X$, ${\mathrm{red}}(x^{\mathrm{an}}) = x$ and for any $\xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}})$, ${\mathrm{red}}(\xi) = x$. It is known that ${\mathrm{red}}:X^{\mathrm{an}}\rightarrow X$ is anti-continuous, that is, for any open set $U$ of $X$, ${\mathrm{red}}^{-1}(U)$ is closed in $X^{\mathrm{an}}$.
Adelic ${\mathbb{R}}$-Cartier divisors over a trivially valued field
====================================================================
In this section, we study fundamental properties of Arakelov geometry over a trivially valued field. Throughout this section, let $K$ be a trivially valued field, $X$ be a normal projective variety over ${\mathrm{Spec}\,}K$ and $X^{\mathrm{an}}$ be the analytification of $X$ in the sense of Berkovich. Let $K(X)$ be the function field of $X$.
Green functions
---------------
Let $U^{\mathrm{an}}$ be a non-empty Zariski open subset of $X^{\mathrm{an}}$. We denote by $C^0(U^{\mathrm{an}})$ the set of continuous functions on $U^{\mathrm{an}}$. We define $$\widehat{C}^0(X^{\mathrm{an}}) := \varinjlim_{\substack{\text{non-empty} \\ \text{Zariski open} \\ \text{subset of} \ X^{\mathrm{an}}}} C^0(U^{\mathrm{an}}).$$ Then $C^0(U^{\mathrm{an}})$ and $\widehat{C}^0(X^{\mathrm{an}})$ are ${\mathbb{R}}$-algebras and we have a canonical homomorphism $C^0(U^{\mathrm{an}}) \rightarrow \widehat{C}^0(X^{\mathrm{an}})$. Since $U^{\mathrm{an}}$ is dense in $X^{\mathrm{an}}$, this homomorphism is injective. Hence we sometimes identify a function in $C^0(U^{\mathrm{an}})$ with a function in $\widehat{C}^0(X^{\mathrm{an}})$. We say that a function in $\widehat{C}^0(X^{\mathrm{an}})$ *extends to a continuous function on $U^{\mathrm{an}}$* if it is in the image of the canonical injection $C^0(U^{\mathrm{an}}) \rightarrow \widehat{C}^0(X^{\mathrm{an}})$.
Let $D$ be an ${\mathbb{R}}$-Cartier divisor on $X$. We say that a function $g \in \widehat{C}^0(X^{\mathrm{an}})$ is a *$D$-Green function of $C^0$-type* (or simply a *Green function of $D$*) if for any non-empty Zariski open subset $U$ of $X$ and any local equation $f \in K(X)_{\mathbb{R}}^\times$ of $D$ on $U$, the function $g + \log|f|$ extends to a continuous function on $U^{\mathrm{an}}$.
Let ${\mathbb{P}}_K^n = \mathrm{Proj}\,K[T_0,\dots,T_n]$ be the $n$-dimensional projective space. We set $z_i = T_i/T_0$ for $i=0,\dots,n$ and $D = \{ T_0=0 \}$. Then $g=\log\max\{a_0,a_1|z_1|,\dots,a_n|z_n|\} \in \widehat{C}^0({\mathbb{P}}_K^{n,{\mathrm{an}}})$, where $a_0,\dots,a_n \in {\mathbb{R}}_{>0}$, is a $D$-Green function of $C^0$-type.
\[pg\] Let $D,D'$ be ${\mathbb{R}}$-Cartier divisors on $X$ and $g,g' \in \widehat{C}^0(X^{\mathrm{an}})$ be Green function of $D,D'$ respectively.
1. For any $s \in K(X)_{\mathbb{R}}^\times$, $-\log|s| \in \widehat{C}^0(X^{\mathrm{an}})$ is a Green function of $(s)$.
2. For any $a,a' \in {\mathbb{R}}$, $ag + a'g'$ is a Green function of $aD+a'D'$.
3. If $D$ is the zero divisor, a $D$-Green function of $C^0$-type coincides to a continuous function on $X^{\mathrm{an}}$.
4. Let $\pi :Y \rightarrow X$ be a morphism of projevtive varieties over $K$ such that $\pi(Y) \nsubseteq \mathrm{Supp}D$. Then $\pi^*g = g \circ \pi^{\mathrm{an}}$ is a Green function of $\pi^*D$.
\(1) It follows from that $s$ is a local equation of $(s)$ on any Zariski open subset.
\(2) Let $U$ be a non-empty Zariski open subset of $X$, $f,f'$ be local equations of $D,D'$ on $U$ respectively. Then $f^a{f'}^{a'}$ is a local equation of $aD+a'D'$ on $U$ and $(ag+a'g')+(a\log|f|+a'\log|f'|)=a(g+\log|f|)+a'(g'+\log|f'|)$ extends a continuous function on $U^{\mathrm{an}}$.
\(3) It follows from (2).
\(4) Let $U$ be a non-empty Zariski open subset of $X$, $f$ be local equations of $D$ on $U$. Then $\pi^*f = f \circ \pi$ is a local equation of $\pi^*D$ on $\pi^{-1}(U)$ and $\pi^*g + \log|\pi^*f| = (g + \log|f|) \circ \pi^{\mathrm{an}}$ extends a continuous function on $(\pi^{\mathrm{an}})^{-1}(U^{\mathrm{an}})$ because $\pi^{\mathrm{an}}$ is continuous.
\[eg\] For any ${\mathbb{R}}$-Cartier divisor $D$, there exists a $D$-Green function of $C^0$-type.
Firstly, we assume that $D$ is an ample Cartier divisor. Let $m$ be a positive integer such that $mD$ is very ample. Then we have an closed immersion $\pi:X \hookrightarrow {\mathbb{P}}_K^n = \mathrm{Proj}K[T_0,\dots,T_n]$ such that ${\mathcal{O}}_X(mD) = \pi^*{\mathcal{O}}(1)$. We set $z_i = T_i/T_0$ for $i=0,\dots,n$ and $D_0 = \{ T_0=0 \}$. Then $g_0 = \log\max\{1,|z_1|,\dots,|z_n|\}$ is a $D_0$-Green function of $C^0$-type. By Proposition \[pg\] (4), $\pi^*g_0$ is a $\pi^*D_0$-Green function of $C^0$-type. Since $mD$ and $\pi^*D_0$ are linearly equivalent, there is a non-zero rational function $s \in K(X)^\times$ such that $mD = \pi^*D_0 + (s)$. Then $(\pi^*g_0 - \log|s|)/m$ gives a $D$-Green function of $C^0$-type.
Next, we assume that $D$ is a Cartier divisor. Then we can write $D$ as $A-A'$ where $A,A'$ are ample divisors. From the previous discussion, there are Green functions $g_A,g_{A'}$ of $A,A'$ respectively. Then $g_A-g_{A'}$ gives a $D$-Green function of $C^0$-type.
In general, there are Cartier divisors $D_1,\dots,D_n$ and $a_1,\dots,a_n \in {\mathbb{R}}$ such that $D=a_1D_1+\cdots+a_nD_n$. Let $g_i$ be a $D_i$-Green function of $C^0$-type for $1,\dots,n$. Then $a_1g_1+\cdots+a_ng_n$ gives a $D$-Green function of $C^0$-type.
\[eff\] Let $D$ be an effective ${\mathbb{R}}$-Cartier divisor on $X$ and $g$ be a $D$-Green function of $C^0$-type. Then the function $e^{-g} \in \widehat{C}^0(X^{\mathrm{an}})$ extends to a non-negative continuous function on $X^{\mathrm{an}}$.
Let $U$ be a non-empty Zariski open subset of $X$ and $f$ be a local equation of $D$ on $U$. Since $g+\log|f|$ extends a continuous function on $U^{\mathrm{an}}$, $e^{-g} = |f| {\raisebox{.4ex}{.}}e^{-(g+\log|f|)}$ extends a non-negative continuous function on $U^{\mathrm{an}}$. We remark that $|f| \in C^0(U^{\mathrm{an}})$ because $D$ is effective. By gluing continuous functions, $e^{-g}$ extends a non-negative continuous function on $X^{\mathrm{an}}$.
By Proposition \[eff\], we sometimes consider a Green function of an effective ${\mathbb{R}}$-Cartier divisor as a map $X^{\mathrm{an}}\rightarrow {\mathbb{R}}\cup \{+\infty\}$.
Continuous metrics on an invertible ${\mathcal{O}}_X$-module
------------------------------------------------------------
Let ${\mathcal{L}}$ be an invertible ${\mathcal{O}}_X$-module. We say that a family $\varphi = \{{|{\raisebox{.4ex}{.}}|}_\varphi(x)\}_{x \in X^{\mathrm{an}}}$ is a *metric on ${\mathcal{L}}$* if ${|{\raisebox{.4ex}{.}}|}_\varphi(x)$ is a norm on ${\mathcal{L}}(x) := {\mathcal{L}}\otimes_{{\mathcal{O}}_X} \widehat{\kappa}(x)$ for all $x \in X^{\mathrm{an}}$. A metric $\varphi = \{{|{\raisebox{.4ex}{.}}|}_\varphi(x)\}_{x \in X^{\mathrm{an}}}$ on ${\mathcal{L}}$ is *continuous* if for any Zariski open subset $U$ of $X$ and non-zero section $s \in H^0(U,{\mathcal{L}})\setminus\{0\}$, $|s|_\varphi(x)$ is a continuous function on $U^{\mathrm{an}}$.
Let ${\mathcal{L}}$ be an invertible ${\mathcal{O}}_X$-module on $X$ and $\varphi$ be a continuous metric on ${\mathcal{L}}$. Then we define a norm ${\|{\raisebox{.4ex}{.}}\|}_\varphi$ on $H^0(X,{\mathcal{L}})$ by $$\|s\|_\varphi := \sup_{x \in X^{\mathrm{an}}} |s|_\varphi(x) \quad \text{for} \ s \in H^0(X,{\mathcal{L}}).$$
Let ${\mathcal{L}}, {\mathcal{L}}'$ be invertible ${\mathcal{O}}_X$-modules. Let $\varphi = \{{|{\raisebox{.4ex}{.}}|}_\varphi(x)\}_{x \in X^{\mathrm{an}}}$ and $\varphi' = \{{|{\raisebox{.4ex}{.}}|}_{\varphi'}(x)\}_{x \in X^{\mathrm{an}}}$ be metrics of ${\mathcal{L}},{\mathcal{L}}'$ respectively. We define the metric $\varphi + \varphi'$ of ${\mathcal{L}}\otimes {\mathcal{L}}'$ by $$|s \otimes s'|_{\varphi+\varphi'}(x) = |s|_\varphi(x) {\raisebox{.4ex}{.}}|s|_{\varphi'}(x)$$ for $x \in X^{\mathrm{an}}$, $s \in {\mathcal{L}}(x)$ and $s' \in {\mathcal{L}}'(x)$. We set the *dual metric $-\varphi$ of $\varphi$ on ${\mathcal{L}}^\vee$* as $$|\alpha(s)|_x = |\alpha|_{-\varphi}(x) {\raisebox{.4ex}{.}}|s|_\varphi(x)$$ for $x \in X^{\mathrm{an}}$, $s \in {\mathcal{L}}(x)$ and $\alpha \in {\mathcal{L}}^\vee(x) = ({\mathcal{L}}(x))^\vee$. If $\varphi$ and $\varphi'$ are continuous, $\varphi + \varphi'$ and $-\varphi$ are also continuous by definition.
We see the relation between Green functions and continuous metrics. Let ${\mathcal{L}}$ be an invertible ${\mathcal{O}}_X$-module and $\varphi = \{{|{\raisebox{.4ex}{.}}|}_\varphi(x)\}_{x \in X^{\mathrm{an}}}$ be a continuous metric on ${\mathcal{L}}$. Let $s$ be a non-zero rational section of ${\mathcal{L}}$. Then $(\mathrm{div}(s), -\log|s|_\varphi)$ is a $\mathrm{div}(s)$-Green function of $C^0$-type. In fact, let $U$ be a non-empty Zariski open subset of $X$ which trivialize ${\mathcal{L}}$ and $\omega$ be a local basis of ${\mathcal{L}}$ on $U$. Then $s$ is denoted by $f\omega$ for some $f \in K(X)$. Since $\mathrm{div}(s)$ is defined by $f$ on $U$, we have $$-\log|s|_\varphi + \log|f| = -\log|f| - \log|\omega|_\varphi + \log|f| = -\log|\omega|_\varphi,$$ which is a continuous function on $U^{\mathrm{an}}$.
Conversely, let $D$ be a Cartier divisor and $g$ be a $D$-Green function of $C^0$-type. Then we equip ${\mathcal{O}}_X(D)$ with a continuous metric $\varphi_g = \{{|{\raisebox{.4ex}{.}}|}_g(x)\}_{x \in X^{\mathrm{an}}}$ as follows: Let $U$ be a non-empty Zariski open subset of $X$ and $f$ be a local equation of $D$ on $U$. Since $1/f$ is a local basis of ${\mathcal{O}}_X(D)$ on $U$, we can denote any section $s \in {\mathcal{O}}_X(D)(U)$ by $a/f$ for some $a \in {\mathcal{O}}_X(U)$. Then we define $|s|_g(x) := |a|_x {\raisebox{.4ex}{.}}e^{-g(x) - \log|f|_x}$ for $x \in U^{\mathrm{an}}$.
By Proposition \[pg\], a continuous metric $\varphi$ on ${\mathcal{O}}_X$ corresponds to a continuous function $g_\varphi$ on $X^{\mathrm{an}}$. So for continuous metrics $\varphi$ and $\psi$ on an invertibel ${\mathcal{O}}_X$-module, we write $\varphi \geq \psi$ if the continuous function $g_{\varphi-\psi}$, which corresponds to $\varphi-\psi$, is non-negative on $X^{\mathrm{an}}$.
Adelic ${\mathbb{R}}$-Cartier divisors
--------------------------------------
Let ${\mathbb{K}}= {\mathbb{Q}},{\mathbb{R}}$ or a blank symbol.
We say that a pair $\overline{D} = (D,g)$ is an *adelic ${\mathbb{K}}$-Cartier divisor* if $D$ is an ${\mathbb{K}}$-Cartier divisor and $g$ is a $D$-Green function of $C^0$-type. We denote by $\widehat{{\mathrm{Div}}}(X)_{\mathbb{K}}$ the set of ${\mathbb{K}}$-Cartier divisors. Remark that $\widehat{{\mathrm{Div}}}(X)_{{\mathbb{R}}} \ncong \widehat{{\mathrm{Div}}}(X) \otimes_{{\mathbb{Z}}} {\mathbb{R}}$. A non-zero ${\mathbb{K}}$-rational function $f \in K(X)_{\mathbb{K}}^\times$ naturally gives an adelic ${\mathbb{K}}$-Cartier divisor $((f),-\log|f|)$, which is called a *${\mathbb{K}}$-principal* and denoted by $\widehat{(f)}$. We say that two adelic ${\mathbb{R}}$-Cartier divisors $\overline{D}_1,\overline{D}_2$ are *${\mathbb{K}}$-linearly equivalent* if $\overline{D}_1-\overline{D}_2$ is ${\mathbb{K}}$-principal. Let $\widehat{{\mathrm{Pic}}}(X)$ be $\widehat{{\mathrm{Div}}}(X)$ modulo linearly equivalence and it is called the *arithmetic Picard group*. An adelic ${\mathbb{K}}$-Cartier divisor $(D,g)$ is *effective* if $D$ is effective and $g$ is a non-negative. Then we write $(D,g) \geq 0$.
Let $\overline{D} = (D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. Then the set of “global sections” $H^0(D)$ is given by $$H^0(D) = \{f \in K(X)^\times \,|\, D + (f) \geq 0 \} \cup \{0\}.$$ Let $s \in H^0(D) \setminus \{0\}$. By Proposition \[eff\], the function $|s| e^{-g} = e^{-g+\log|s|}$ extends to a non-negative function on $X^{\mathrm{an}}$. We denote this function by $|s|_g:X^{\mathrm{an}}\rightarrow {\mathbb{R}}_+$. Then we define $$\|s\|_g := \sup_{x \in X^{\mathrm{an}}} |s|_g(x).$$ The map ${\|{\raisebox{.4ex}{.}}\|}_g :H^0(D) \rightarrow {\mathbb{R}}_+$ gives an ultrametric norm on $H^0(D)$ over $K$ and it coincides with the supremum norm induced by the continuous metric on ${\mathcal{O}}_X(D)$ corresponding to $g$. Moreover, $\bigoplus_{n=0}^\infty (H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng})$ is a normed graded ring over $K$ by definition.
We set $${\lambda_{\max}}(D,g) := {\lambda_{\max}}(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_g),$$ and $${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) := \limsup_{n \rightarrow +\infty} \frac{1}{n}{\lambda_{\max}}(nD,ng).$$ Since $\bigoplus_{n=0}^\infty (H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng})$ is a normed graded ring, the sequence $\{{\lambda_{\max}}(nD,ng)\}_n$ is super-additive, that is, $${\lambda_{\max}}((m+n)D,(m+n)g) \geq {\lambda_{\max}}(mD,mg) + {\lambda_{\max}}(nD,ng) \quad \text{for}\ \forall m,n \in {\mathbb{Z}}_+.$$ Hence by Fekete’s lemma, we have $${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) = \lim_{n \rightarrow +\infty} \frac{1}{n}{\lambda_{\max}}(nD,ng) = \sup_{n \geq 1} \frac{1}{n}{\lambda_{\max}}(nD,ng).$$ Later, we will show that ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) < +\infty$.
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor. We say that a non-zero global section $s \in H^0(D) \setminus \{0\}$ is a *small section* if $\|s\|_g \leq 1$ or equivalently $s \in {\mathcal{F}}^0(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_g)$. Moreover, if $\|s\| < 1$, it is called a *strictly small section*.
Let $\overline{D}=(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. Then we have $${\mathcal{F}}^0(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_g) = \left\{ s \in K(X)^\times \, \middle| \, \overline{D} + \widehat{(s)} \geq 0 \right\} \cup \{0\}.$$
Let $s \in H^0(D) \setminus \{0\}$. By definition, $$\begin{aligned}
\|s\|_g \leq 1 &\Leftrightarrow e^{-g+\log|s|} \leq 1\ \text{on}\ X^{\mathrm{an}}\\
&\Leftrightarrow g-\log|s| \geq 0 \ \text{on}\ X^{\mathrm{an}}.
\end{aligned}$$
Small sections play the similar role as global sections in algebraic geometry. Therefore we are interested in the asymptotic behavior of ${\mathcal{F}}^0(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_{ng})$ as $n \rightarrow +\infty$.
Associated ${\mathbb{R}}$-Weil divisors
---------------------------------------
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. For any $x \in X^{(1)}$, $$\mu_x(g) := \inf_{\xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}})} \frac{g(\xi)}{t(\xi)} \in {\mathbb{R}}\cup \{-\infty\}.$$ Clearly $\mu_x(g) \geq 0$ if and only if $g \geq 0$ on $(\eta^{\mathrm{an}},x^{\mathrm{an}})$. Moreover $\mu_x(g) = -\infty$ if and only if $g(\eta^{\mathrm{an}}) < 0$, which implies that if $\mu_x(g) = -\infty$ for some $x \in X^{(1)}$, then $\mu_x(g) = -\infty$ for every $x \in X^{(1)}$.
The above invariant $\mu_x(g)$ has following properties:
\[mu0\] Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. For all but finitely many $x \in X^{(1)}$, we have $\mu_x(g) \leq 0$.
Let $U$ be a non-empty Zariski open subset of $X$ such that $g$ is a continuous function on $U^{\mathrm{an}}$. Then $g$ is continuous on $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$ for all $x \in U \cap X^{(1)}$. Since $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$ is compact, $g|_{[\eta^{\mathrm{an}},x^{\mathrm{an}}]}$ is bounded above. Hence we have $\mu_x(g) \leq 0$ for for all $x \in U \cap X^{(1)}$, which implies the assertion because $X^{(1)} \setminus U$ is a finite set.
\[mu1\] Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$ and $x \in X^{(1)}$.
1. For any $s \in K(X)_{\mathbb{R}}^\times$, we have $$\mu_x(g - \log |s|) = \mu_x(g) + \mathrm{ord}_x(s).$$
2. We have $\mu_x(g) \leq \mathrm{ord}_x(D)$.
\(1) By definition of $X_{\mathrm{div}}^{\mathrm{an}}$, for any $s \in K(X)_{\mathbb{R}}^\times$, we have $$-\log|s|(\xi) = t(\xi){\mathrm{ord}}_x(s), \quad \xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}}).$$ Hence we obtain that $$\mu_x(g-\log|s|) = \inf_{\xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}})} \frac{g(\xi)-\log|s|(\xi)}{t(\xi)} = \mu_x(g) + {\mathrm{ord}}_x(s).$$
\(2) Let $f \in K(X)_{\mathbb{R}}^\times$ be a local equation of $D$ around $x$. Then $g + \log|f|$ extends to a continuous function on $[\eta^{\mathrm{an}},x^{\mathrm{an}}]$. Since $(g + \log|f|)|_{[\eta^{\mathrm{an}},x^{\mathrm{an}}]}$ is bounded above, we have $\mu_x(g + \log|f|) \leq 0$. By (1), we get $\mu_x(g) \leq {\mathrm{ord}}_x(D)$.
Now we introduce an important divisor.
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. We say that $(D,g)$ is *$\mu$-finite* if $\mu_x(g) = 0$ for all but finitely many $x \in X^{(1)}$, which is equivalent to $\mu_x(g) \geq 0$ for all but finitely many $x \in X^{(1)}$ by Proposition \[mu0\]. If $(D,g)$ is $\mu$-finite, we can define an ${\mathbb{R}}$-Weil divisor on $X$ as follows: $$D_{\mu(g)} := \sum_{x \in X^{(1)}} \mu_x(g) [x].$$ It is called an *${\mathbb{R}}$-Weil divisor associated with $(D,g)$*. Remark that $D_{\mu(g)}$ may not be an ${\mathbb{R}}$-Cartier divisor.
For example, if $(D,g)$ has a Dirichlet property (which means that $(D,g)$ is ${\mathbb{R}}$-linearly equivalent to an effective adelic ${\mathbb{R}}$-Cartier divisor), then $(D,g)$ is $\mu$-finite.
By Proposition \[mu1\], we have $D_{\mu(g)} \leq D$ and $$\label{mulin}
{(D + (s))}_{\mu(g - \log |s|)} = D_{\mu(g)} + (s).$$
Let $(D,g)$ be a $\mu$-finite adelic ${\mathbb{R}}$-Cartier divisor on $X$. Then $(D,g)$ is effective if and only if $D_{\mu(g)}$ is effective.
We first assume that $(D,g)$ is effective. Then $g$ is non-negative on $X^{{\mathrm{an}}}$, so $\mu_x(g) \geq 0$ for any $x \in X^{(1)}$, which implies $D_{\mu(g)}$ is effective.
Conversely we assume that $D_{\mu(g)}$ is effective. Then $g$ is non-negative on $X_{\mathrm{div}}^{{\mathrm{an}}}$, but $X_{\mathrm{div}}^{{\mathrm{an}}}$ is dense in $X^{{\mathrm{an}}}$ by Lemma \[den\], so it follows that $g$ is non-negative on the whole $X^{{\mathrm{an}}}$. Moreover by Proposition \[mu1\], we have $$\mathrm{ord}_x(D) \geq \mu_x(g) \geq 0$$ for any $x \in X^{(1)}$, which completes the proof.
By the above proposition and the equation (\[mulin\]), we have the following corollary:
\[as1\] $$H^0(D_{\mu(g)}) = \mathcal{F}^0(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_g) = \{ s \in H^0(D) \,|\, \|s\|_g \leq 1 \}.$$
Canonical Green function
------------------------
For any ${\mathbb{R}}$-Cartier divisor $D$, we can naturally give a $D$-Green function of $C^0$-type as follows: For any $x \in X^{\mathrm{an}}$, let $f \in K(X)_{\mathbb{R}}^\times$ be a local equation of $D$ around ${\mathrm{red}}(x) \in X$. Then we define $$g_D^\mathrm{c}(x) := -\log|f|_x.$$ This definition is independent of the choice of a local equation. In fact, let $f' \in K(X)_{\mathbb{R}}^\times$ be another local equation. Then there is an element $a \in ({\mathcal{O}}_{X,{\mathrm{red}}(x)})_{\mathbb{R}}^\times$ such that $f' = af$. Since $|a|_x = 1$ by (\[red\]), we have $-\log|f'|_x = -\log|f|_x$.
\[cg\] The function $g_D^\mathrm{c}$ is a $D$-Green function of $C^0$-type.
It is enough to show that for any non-empty Zariski open subset $U$ of $X$ and local equation $f$ of $D$ on $U$, $g_D^\mathrm{c} + \log|f|$ extends to a continuous function on $U^{\mathrm{an}}$. Let $x \in U^{\mathrm{an}}$. If ${\mathrm{red}}(x) \in U$, then $g_D^\mathrm{c}(x) = -\log|f|_x$. Hence we have $g_D^\mathrm{c}(x) + \log|f|_x = 0$. Next, we assume that ${\mathrm{red}}(x) \notin U$. Let $U'$ be a non-empty Zariski open neighborhood of ${\mathrm{red}}(x)$ and $f'$ be a local equation of $D$ on $U'$. Then we have $g_D^\mathrm{c}(x) = -\log|f'|_x$ We remark that $j(x) \in U'$, hence $U \cap U' \neq \emptyset$. There is a non-zero regular function $u \in ({\mathcal{O}}_X(U \cap U'))_{\mathbb{R}}^\times$ such that $f' = uf$ on $U \cap U'$. Therefore we obtain that $g_D^\mathrm{c}(x) + \log|f|_x = -\log|u|_x$, which is continuous on $U^{\mathrm{an}}\cap {U'}^{\mathrm{an}}$. Finally, let $y \in U^{\mathrm{an}}\cap {U'}^{\mathrm{an}}$ such that ${\mathrm{red}}(y) \in U$. Since $u \in ({\mathcal{O}}_{X,{\mathrm{red}}(y)})_{\mathbb{R}}^\times$, we have $|u|_y = 1$ by (\[red\]). Hence $g_D^\mathrm{c}(y) + \log|f|_y = - \log|u|_y = 0$, which completes the proof.
Proposition \[cg\] gives the another proof of Proposition \[eg\].
The function $g_D^\mathrm{c}$ is called the *canonical Green function of $D$*.
\[cgp\]
1. For any $s \in K(X)_{\mathbb{R}}^\times$, $g_{(s)}^\mathrm{c} = -\log|s|$.
2. For any $D,D' \in {\mathrm{Div}}(X)_{\mathbb{R}}$ and $a,a' \in {\mathbb{R}}$, $g_{aD+a'D'}^\mathrm{c} = ag_D^\mathrm{c} + a'g_{D'}^\mathrm{c}$.
\(1) Since $(s)$ is globally defined by $s$, it follows by definition of the canonical Green function.
\(2) Let $x \in X^{\mathrm{an}}$ and $f,f'$ be local equations of $D,D'$ around ${\mathrm{red}}(x)$ respectively. Then $f^a{f'}^{a'}$ is a local equation of $aD+a'D'$ around ${\mathrm{red}}(x)$. Hence we have $$g_{aD+a'D'}^\mathrm{c}(x) = -\log|f^a{f'}^{a'}|_x = -a\log|f|_x-a'\log|f'|_x = ag_D^\mathrm{c}(x)+a'g_{D'}^\mathrm{c}(x).$$
Using the canonical Green function, we can define the following injective homomorphism: $$\phi:{\mathrm{Div}}(X) \rightarrow \widehat{{\mathrm{Div}}}(X), \quad D \mapsto (D,g_D^\mathrm{c}).$$ By Proposition \[cgp\], it induces an injective homomorphism $\overline{\phi}:{\mathrm{Pic}}(X) \rightarrow \widehat{{\mathrm{Pic}}}(X)$ such that the following diagram is commutative: $$\begin{tikzcd}
{\mathrm{Div}}(X) \arrow{r}{\phi} \arrow{d} & \widehat{{\mathrm{Div}}}(X) \arrow{d} \\
{\mathrm{Pic}}(X) \arrow{r}{\overline{\phi}} & \widehat{{\mathrm{Pic}}}(X).
\end{tikzcd}$$
Height function
---------------
Here we see the hight function on $X^{\mathrm{an}}$ associated with an adelic ${\mathbb{R}}$-Cartier divisor, which is introduced by Chen and Moriwaki [@chen2017sufficient].
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. We set $h_{(D,g)}^{\mathrm{an}}:= g - g_D^\mathrm{c}$, which is called *hight function on $X^{\mathrm{an}}$ associated with $(D,g)$*.
\[hp\] Let $\overline{D},\overline{D}'$ be adelic ${\mathbb{R}}$-Cartier divisors on $X$.
1. For any $s \in K(X)_{\mathbb{R}}^\times$, $h_{\widehat{(s)}}^{\mathrm{an}}= 0$ on $X^{\mathrm{an}}$.
2. For any $a,a' \in {\mathbb{R}}$, $h_{a\overline{D}+a'\overline{D}'}^{\mathrm{an}}= ah_{\overline{D}}^{\mathrm{an}}+a'h_{\overline{D}'}^{\mathrm{an}}$ on $X^{\mathrm{an}}$.
It immediately follow from Proposition \[cgp\].
For any adelic ${\mathbb{R}}$-Cartier divisor $(D,g)$ on $X$, $h_{(D,g)}^{\mathrm{an}}$ is a continuous function on $X^{\mathrm{an}}$. Hence we have the following homomorphism: $$\psi:\widehat{{\mathrm{Div}}}(X) \rightarrow C^0(X^{\mathrm{an}}), \quad (D,g) \mapsto h_{(D,g)}^{\mathrm{an}}.$$ This homomorphism is surjective. By Proposition \[hp\], it induces a surjective homomorphism $\overline{\psi}:\widehat{{\mathrm{Pic}}}(X) \rightarrow C^0(X^{\mathrm{an}})$ such that the following diagram is commutative: $$\begin{tikzcd}
\widehat{{\mathrm{Div}}}(X) \arrow{r}{\psi} \arrow{d} & C^0(X^{\mathrm{an}}) \\
\widehat{{\mathrm{Pic}}}(X) \arrow{ru}{\overline{\psi}}
\end{tikzcd}$$
The following sequence is exact: $$\begin{tikzcd}
0 \arrow{r} & {\mathrm{Pic}}(X) \arrow{r}{\overline{\phi}} & \widehat{{\mathrm{Pic}}}(X) \arrow{r}{\overline{\psi}} & C^0(X^{\mathrm{an}}) \arrow{r} & 0.
\end{tikzcd}$$ In particular, $\widehat{{\mathrm{Pic}}}(X) \simeq {\mathrm{Pic}}(X) \oplus C^0(X^{\mathrm{an}})$.
Since $\psi \circ \phi = 0$ by definition, we have $\overline{\psi} \circ \overline{\phi} = 0$. Let $(D,g) \in \widehat{{\mathrm{Div}}}(X)$ such that $\overline{\psi}(D,g) = 0$. Then there are $H \in {\mathrm{Div}}(X)$ and $s \in K(X)$ such that $(D,g) = (H,g_H^\mathrm{c}) + (s,-\log|s|)$. By Proposition \[cgp\], we have $g = g_H^\mathrm{c} - \log|s| = g_D^\mathrm{c}$, which implies that $(D,g) = \phi(D)$. Hence we obtain that $\mathrm{Im}\, \overline{\phi} = \mathrm{Ker}\, \overline{\psi}$.
For $X = {\mathrm{Spec}\,}K$, the Berkovich space $X^{\mathrm{an}}$ associated with $X$ is a single point. Hence we have $C^0(X^{\mathrm{an}}) = {\mathbb{R}}$, which implies that $$\widehat{{\mathrm{Pic}}}(X) \simeq {\mathbb{R}}.$$ This result corresponds to the fact that the Picard group of ${\mathbb{P}}^1$ is isomorphic to ${\mathbb{Z}}$ and the arithmetic Picard group of ${\mathrm{Spec}\,}{\mathbb{Z}}$ is isomorphic to ${\mathbb{R}}$.
Scaling action for Green functions
----------------------------------
We saw that the multiplicative group ${\mathbb{R}}_{>0}$ acts $X^{\mathrm{an}}$ (see Section \[BS\]). Here we see that it also acts the set of $D$-Green functions of $C^0$-type. Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. For $r \in {\mathbb{R}}_{>0}$, we define $$r^*g(x) := rg(x^{1/r}).$$ The function $r^*g$ is also $D$-Green function of $C^0$-type. In fact, let $U$ be a non-empty Zariski open subset of $X$ and $f$ be a local equation of $D$ on $U$. Then we have $$r^*g(x) + \log|f|_x = rg(x^{1/r}) + r\log|f|_x^{1/r} = r(g(x^{1/r}) + \log|f|_{x^{1/r}})$$ on $U^{\mathrm{an}}$, which is continuous. This action is also called the *sacling action*.
\[sag\] Let $D$ be an ${\mathbb{R}}$-Cartier divisor on $X$, $g,g'$ be $D$-Green functions of $C^0$-type.
1. The scaling action is linear, that is, for $r \in {\mathbb{R}}_{>0}$, $$r^*(g + g') = r^*g + r^*g'.$$
2. The scaling action preserves the canonical Green function $g_D^\mathrm{c}$.
\(1) It is clear by definition.
\(2) Let $x \in X^{\mathrm{an}}$ and $f$ be a local equation of $D$ around ${\mathrm{red}}(x)$. Then $g_D^\mathrm{c}(x) = -\log|f|_x$. Hence we have $$r^*g_D^\mathrm{c}(x) = -r\log|f|_{x^{1/r}} = -r\log|f|_x^{1/r} = -\log|f|_x = g_D^\mathrm{c}(x)$$ for any $r \in {\mathbb{R}}_{>0}$.
Arithmetic volume
=================
In this section, we introduce the arithmetic volume function and the bigness of adelic ${\mathbb{R}}$-Cartier divisors. And we study some properties of the arithmetic volume function. Throughout this section, let $K$ be a trivially valued field and $X$ be a normal projective variety over ${\mathrm{Spec}\,}K$.
Big adelic ${\mathbb{R}}$-Cartier divisors
------------------------------------------
We introduce the counterparts of $h^0(D)$ and ${\mathrm{vol}}({\raisebox{.4ex}{.}})$ in Arakelov geometry, which is given by Chen and Moriwaki [@chen2017sufficient]. We set $${\widehat{\deg}}_+(D,g) := \int_0^{+\infty} \dim_K {\mathcal{F}}^t(H^0(D),{\|{\raisebox{.4ex}{.}}\|}_g) \ \mathrm{d}t,$$ and $${\widehat{{\mathrm{vol}}}}(D,g) := \limsup_{n \rightarrow +\infty} \frac{{\widehat{\deg}}_+(nD,ng)}{n^{d+1}/(d+1)!},$$ where $d = \dim X$.
We say that an adelic ${\mathbb{R}}$-Cartier divisor $(D,g)$ is *big* if ${\widehat{{\mathrm{vol}}}}(D,g) > 0$.
\[mbig\] Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. If $(D,g)$ is big, then $(D,g)$ is $\mu$-finite, $D_{\mu(g)}$ is big and ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) > 0$. In particular, $D$ is big.
If $(D,g)$ is big, $(D,g)$ is ${\mathbb{Q}}$-linearly equivalent to an effective adelic ${\mathbb{R}}$-Cartier divisor, which implies that $(D,g)$ is $\mu$-finite. By definition, for any integer $n > 0$, we have $${\widehat{\deg}}_+(nD,ng) \leq \dim_k{\mathcal{F}}^0(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng}){\max} \{{\lambda_{\max}}(nD,ng),0 \}.$$ Therefore we have $${\widehat{{\mathrm{vol}}}}(D,g) \leq (d+1){\mathrm{vol}}(D_{\mu(g)}){\max} \{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g),0 \},$$ by corollary \[as1\]. Since ${\widehat{{\mathrm{vol}}}}(D,g) > 0$, we have ${\mathrm{vol}}(D_{\mu(g)}) > 0$ and ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) > 0$.
Existence of limit of the arithmetic volume
-------------------------------------------
Firstly, we define $$\nu_{\max}(D,g) := \sup \{ t \in {\mathbb{R}}\,|\, (D,g-t) \ \text{is $\mu$-finite} \}.$$
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. We have $${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \leq \nu_{\max}(D,g) \leq g(\eta^{\mathrm{an}}).$$
Clearly we can assume ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \in {\mathbb{R}}$. For a sufficiently large integer $n > 0$, there is a non-zero element $s \in H^0(nD) \setminus \{0\}$ such that $\|s\|_{ng} \leq e^{-{\lambda_{\max}}(nD,ng)}$, which is equivalent to $\|s\|_{ng-{\lambda_{\max}}(nD,ng)} \leq 1$. Therefore $(D,g-{\lambda_{\max}}(nD,ng)/n)$ is effecive, which implies $$\frac{1}{n}{\lambda_{\max}}(nD,ng) \leq \nu_{\max}(D,g).$$ Taking a supremum, we get ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \leq \nu_{\max}(D,g)$.
Next we show $\nu_{\max}(D,g) \leq g(\eta^{\mathrm{an}})$. For any $\epsilon > 0$, $g(\zeta) - (g(\eta^{\mathrm{an}}) + \epsilon)$ is negative around $\zeta = \eta^{\mathrm{an}}$. So we have $\mu_x(g-(g(\eta^{\mathrm{an}}) + \epsilon)) = -\infty$ for any $x \in X^{(1)}$, which implies that $(D,g-(g(\eta^{\mathrm{an}}) + \epsilon))$ is not $\mu$-finite and $\nu_{\max}(D,g) \leq g(\eta^{\mathrm{an}}) + \epsilon$. Since $\epsilon$ is arbitrary, we conclude that $\nu_{\max}(D,g) \leq g(\eta^{\mathrm{an}})$.
The above inequality is sometimes strict. For example, let $X = {\mathbb{P}}_K^1 = \mathrm{Proj}K[T_0,T_1]$, $z=T_1/T_0$, $D = \{T_0=0\}$ and $x_\infty = (0:1)$. Let $g_1 = 2\log\max\{2,|z|\} - \log\max\{1,|z|\}$. Then $g_1(\xi) = 2\log2$ for $\xi \in [\eta^{\mathrm{an}},x^{\mathrm{an}}]$ for $x \neq x_\infty$ and $$g_1(\xi) =
\begin{cases}
2\log2 - \xi & (0 \leq \xi \leq \log2) \\
\xi & (\log2 \leq \xi),
\end{cases}$$ on $[\eta^{\mathrm{an}},x_\infty^{\mathrm{an}}]$. Hence we have $$\mu_x(g_1-t) =
\begin{cases}
0 & (x \neq x_\infty) \\
\log2 - t & (x = x_\infty),
\end{cases}$$ for $t \leq 2\log2$ and $\mu_x(g_1-t) = -\infty$ for $t > 2\log2$ and all closed point $x$ of $X$. Therefore we obtain that ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) = \log2$ and $\nu_{\max}(D,g) = 2\log2$.
Next, we set $$h(\xi) =
\begin{cases}
-\xi & (0 \leq \xi \leq 1) \\
-1 & (1 \leq \xi),
\end{cases}$$ for $\xi \in [\eta^{\mathrm{an}},x^{\mathrm{an}}]$ and all closed point $x$ of $X$, which is a continuous function on $X^{\mathrm{an}}$. We define a $D$-Green function $g_2$ as $\log\max\{1,|z|\} + h$. Then we have $$\mu_x(g_2-t)=
\begin{cases}
0 & (x \neq x_\infty) \\
1 & (x = x_\infty)
\end{cases}$$ for $t \leq -1$, $$\mu_x(g_2-t)=
\begin{cases}
-1 - t & (x \neq x_\infty) \\
-t & (x = x_\infty)
\end{cases}$$ for $-1 \leq t \leq 0$ and $\mu_x(g_2-t) = -\infty$ for $t >0$ and all closed point $x$ of $X$. Hence we obtain that $\nu_{\max}(D,g_2) = -1$ and $g(\eta^{\mathrm{an}}) = 0$.
For any integer $n > 0$, let $$P_n^{(D,g)}(t) := \frac{\dim_K{\mathcal{F}}^{nt}(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng})}{n^d/d!}$$ where $d = \dim X$. If there is no confusion, we write it simply $P_n(t)$. By definition, if $(D,g)$ is $\mu$-finite, $$P_n(0) = \frac{\dim_K{\mathcal{F}}^0(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng})}{n^d/d!} = \frac{\dim_KH^0(nD_{\mu(g)})}{n^d/d!},$$ so we have $$\label{p1}
\lim_{n \rightarrow +\infty}P_n(0) = {\mathrm{vol}}(D_{\mu(g)}).$$
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. For any $\epsilon \in {\mathbb{R}}$, we have $$P_n^{(D,g-\epsilon)}(t) = P_n^{(D,g)}(t+\epsilon).$$
For any $s \in H^0(nD)$, we have $$\|s\|_{n(g -\epsilon)} \leq e^{-nt} \Leftrightarrow \|s\|_{ng} \leq e^{-n(t+\epsilon)}.$$ Hence we get $$\dim_K{\mathcal{F}}^{nt}(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{n(g-\epsilon)}) = \dim_K{\mathcal{F}}^{n(t+\epsilon)}(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng}),$$ which implies that $P_n^{(D,g-\epsilon)}(t) = P_n^{(D,g)}(t+\epsilon)$.
In particular, $$P_n^{(D,g)}(t) = P_n^{(D,g-t)}(0).$$ So by the equation (\[p1\]), we have $$\label{p2}
\lim_{n \rightarrow +\infty}P_n(t) = {\mathrm{vol}}(D_{\mu(g-t)})$$ for any $t < \nu_{\max}(D,g)$.
If we define $$\begin{aligned}
F_{(D,g)}(t) := \left \{
\begin{array}{ll}
{\mathrm{vol}}(D_{\mu(g-t)}) & (t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)) \\
0 & (t > {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)),
\end{array}
\right .\end{aligned}$$ we get the following theorem by the equation (\[p2\]):
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. The sequence $\{ P_n(t) \}_{n \geq 1}$ converges pointwise to $F_{(D,g)}(t)$ on ${\mathbb{R}}\setminus\{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)\}$.
The sequence $\{ P_n(t) \}_{n \geq 1}$ is uniformly bounded on $(0,{{\lambda_{\max}}^{\mathrm{asy}}}(D,g))$, so we get the main theorem in this section by using bounded convergence theorem:
\[vol1\] Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. We have $$\begin{aligned}
{\widehat{{\mathrm{vol}}}}(D,g) &= \lim_{n \rightarrow +\infty} \frac{{\widehat{\deg}}_+(nD,ng)}{n^{d+1}/(d+1)!} \\
&= (d+1)\int_0^{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)} F_{(D,g)}(t) \ \mathrm{d}t.\end{aligned}$$
By definition, $${\widehat{\deg}}_+(nD,ng) = \int_0^{{\lambda_{\max}}(nD,ng)} \dim_K{\mathcal{F}}^t(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng}) \ \mathrm{d}t.$$ Substituting $t$ for $nt$, we have $${\widehat{\deg}}_+(nD,ng) = n\int_0^{\frac{1}{n}{\lambda_{\max}}(nD,ng)} \dim_K{\mathcal{F}}^{nt}(H^0(nD),{\|{\raisebox{.4ex}{.}}\|}_{ng}) \ \mathrm{d}t.$$ Therefore we get $$\frac{{\widehat{\deg}}_+(nD,ng)}{n^{d+1}/(d+1)!} = (d+1)\int_0^{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)} P_n(t) \ \mathrm{d}t.$$ We remark that ${\lambda_{\max}}(nD,ng)/n \leq {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $P_n(t) = 0$ if $t > {\lambda_{\max}}(nD,ng)/n$. So by using bounded convergence theorem, we get the conclusion.
\[vol2\] The arithmetic volume ${\widehat{{\mathrm{vol}}}}({\raisebox{.4ex}{.}})$ is $(d+1)$-homogeneous. Namely, for any adelic ${\mathbb{R}}$-Cartier divisor $(D,g)$ and $a \in {\mathbb{R}}_{>0}$, we have $${\widehat{{\mathrm{vol}}}}(aD,ag) = a^{d+1} {\widehat{{\mathrm{vol}}}}(D,g).$$
We have ${{\lambda_{\max}}^{\mathrm{asy}}}(aD,ag) = a{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $F_{(aD,ag)}(at) = a^d F_{(D,g)}(t)$ because the algebraic volume is $d$-homogeneous. Therefore by Theorem \[vol1\], we have $$\begin{aligned}
{\widehat{{\mathrm{vol}}}}(aD,ag) &= (d+1) \int_0^{{{\lambda_{\max}}^{\mathrm{asy}}}(aD,ag)} F_{(aD,ag)}(t) \ \mathrm{d}t \\
&= a(d+1) \int_0^{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)} a^d F_{(D,g)}(t) \ \mathrm{d}t \\
&= a^{d+1} {\widehat{{\mathrm{vol}}}}(D,g).\end{aligned}$$
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor and $r \in {\mathbb{R}}_{>0}$.
1. ${{\lambda_{\max}}^{\mathrm{asy}}}(D,r^*g) = r{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$.
2. ${\widehat{{\mathrm{vol}}}}(D,r^*g) = r{\widehat{{\mathrm{vol}}}}(D,g)$.
\(1) For a non-zero element $f \in H^0(nD) \setminus \{0\}$, we have $$|f|_{nr^*g}(x) = \exp(-nrg(x^\frac{1}{r})+\log|f|_x) = \exp(r(-ng(x^\frac{1}{r})+\log|f|_{x^\frac{1}{r}})) = |f|_{ng}(x^\frac{1}{r}).$$ Hence we get ${\lambda_{\max}}(nD,nr^*g) = r{\lambda_{\max}}(nD,ng)$ for any positive integer $n$, which implies that ${{\lambda_{\max}}^{\mathrm{asy}}}(D,r^*g) = r{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$.
\(2) If $D$ is not big, ${\widehat{{\mathrm{vol}}}}(D,g) = 0$ for any $D$-Green function $g$. So we can assume that $D$ is big. For $t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $x \in X^{(1)}$, we have $$\begin{aligned}
\mu_x(r^*g-t) &= \inf_{\xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}})} \frac{rg(\xi^\frac{1}{r})-t}{t(\xi)} \\
&= \inf_{\xi \in (\eta^{\mathrm{an}},x^{\mathrm{an}})} \frac{g(\xi^\frac{1}{r})-t/r}{t(\xi^\frac{1}{r})} = \mu_x(g-t/r),\end{aligned}$$ which implies that $D_{\mu(r^*g-t)} = D_{\mu(g-t/r)}$ and hence $F_{(D,r^*g)}(t) = F_{(D,g)}(t/r)$. Therefore, by Theorem \[vol1\] and (1), we get $$\begin{aligned}
{\widehat{{\mathrm{vol}}}}(D,r^*g) &= (d+1) \int_0^{r{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)} F_{(D,g)}(t/r)\ \mathrm{d}t \\
&= (d+1)r \int_0^{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)} F_{(D,g)}(t')\ \mathrm{d}t' \quad \left(t' = \frac{t}{r}\right) \\
&= r{\widehat{{\mathrm{vol}}}}(D,g).\end{aligned}$$
Finally, we prove a simple criterion of the bigness of an adelic ${\mathbb{R}}$-Cartier divisor.
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. We assume that $D$ is big. Then the following conditions are equivalent:
1. $(D,g)$ is big.
2. ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) > 0$.
3. For $\forall n \gg 0$, there is a strictly small section of $H^0(nD)$.
\(1) $\Rightarrow$ (2) It follows from Proposition \[mbig\].
\(2) $\Rightarrow$ (1) It is sufficient to show that $D_{\mu(g)}$ is big. In fact, if $D_{\mu(g)}$ is big, $D_{\mu(g-t)}$ is also big for $t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ because ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g-t) = {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) - t$ for $t \in {\mathbb{R}}$. Then we have $${\widehat{{\mathrm{vol}}}}(D,g) = (d+1)\int_0^\infty {\mathrm{vol}}(D_{\mu(g-t)})\ \mathrm{d}t > 0$$ by Theorem \[vol1\]. Now we prove that $D_{\mu(g)}$ is big. Since $D$ is big, there is an ample divisor $A$ such that $mD-A$ is effective for some $m \in {\mathbb{Z}}_{>0}$. Let $s \in H^0(mD-A) \setminus \{0\}$ be a non-zero section such that the map $H^0(kA) \rightarrow H^0(kmD)$ is given by multiplication by $s^{\otimes k}$ for all $k > 0$. We denote the image of the map $H^0(kA) \rightarrow H^0(kmD)$ by $V_k$ and $V_0 = K$. Since the graded ring $\bigoplus_{k=0}^\infty V_k$ is finitely generated, there is $a \in {\mathbb{R}}$ such that $\|v\|_{kmg} \leq e^{-akm}$ for all $v \in V_k$ and a sufficiently large $k>0$. Let $\epsilon$ be a real number such that $0 < \epsilon < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$. Then we can find $p \in {\mathbb{Z}}_{>0}$ such that there is a non-zero element $s_p \in H^0(pD) \setminus \{0\}$ with $\|v_p\|_{pg} \leq e^{-p\epsilon}$ and $p > -am/\epsilon$ because $\epsilon < {\lambda_{\max}}(pD,pg)/p$ for a sufficiently large $p>0$. The image $W_k$ of the composition of the map $H^0(kA) \rightarrow H^0(kmD) \rightarrow H^0(k(m+p)D)$ is given by multiplication by $(ss_p)^{\otimes k}$ for all $k>0$. Hence for any $w \in W_k$, we can write $w = v \otimes (s_p)^{\otimes k}$ with $v \in V_k$ and we have $$\|w\|_{k(m+p)g} \leq \|v\|_{kmg} {\raisebox{.4ex}{.}}\|s_p\|_{pg}^k \leq e^{-akm}e^{-kp\epsilon} = e^{-k(am+p\epsilon)} \leq 1,$$ which implies that $W_k \subset H^0(k(m+p)D_{\mu(g)})$ for a sufficiently large $k > 0$. Therefore we obtain that ${\mathrm{vol}}((m+p)D_{\mu(g)}) \geq {\mathrm{vol}}(A) > 0$, which is required.
\(2) $\Rightarrow$ (3) Since ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) > 0$, we have ${\lambda_{\max}}(nD,ng) > 0$ for a sufficiently large $n > 0$. Hence there is a non-zero section $s \in H^0(nD) \setminus \{0\}$ such that $\|s\|_{ng} \leq e^{-{\lambda_{\max}}(nD,ng)} < 1$.
\(3) $\Rightarrow$ (2) Let $s$ be a strictly small section of $H^0(nD)$. Then we have ${\lambda_{\max}}(nD,ng) \geq -\log\|s\|_{ng} > 0$. Therefore we obtain that ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \geq {\lambda_{\max}}(nD,ng)/n > 0$.
Continuity of $F_{(D,g)}(t)$
----------------------------
Firstly, we will prove a very useful lemma:
\[con\] Let $V$ be a convex cone and let $f:V \rightarrow {\mathbb{R}}$ be a concave function. Namely, for any $v,v' \in V$ and $a,a' \geq 0$, $$f(av+a'v') \geq af(v) + a'f(v').$$ If $g(t) := v+tv'$ is a map from some open interval $(a,b) \subset {\mathbb{R}}$ to $V$ for fixed elements $v,v' \in V$, then $f \circ g$ is a concave function on $(a,b)$. In particular, $f \circ g$ is continuous on $(a,b)$.
For any $t,t' \in (a,b)$ and $0 \leq \epsilon \leq 1$, we have $$\begin{aligned}
f \circ g(\epsilon t + (1-\epsilon)t') &= f(v+(\epsilon t + (1-\epsilon)t')v') \\
&= f(\epsilon (v+tv') + (1-\epsilon)(v+t'v')) \\
& \geq \epsilon f(v+tv') + (1-\epsilon) f(v+t'v') \\
&= \epsilon f \circ g(t) + (1-\epsilon)f \circ g(t').\end{aligned}$$
For an adelic ${\mathbb{R}}$-Cartier divisor $(D,g)$, it follows immediately from the above lemma that $\mu_x(g-t)$ is a continuous concave function on $(-\infty,{{\lambda_{\max}}^{\mathrm{asy}}}(D,g))$ for every $x \in X^{(1)}$. Hence we get the following proposition:
\[D1\] Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$. For any $t,t' < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $0 \leq \epsilon \leq 1$, we have $$D_{\mu(g-(\epsilon t + (1-\epsilon)t'))} \geq \epsilon D_{\mu(g-t)} + (1-\epsilon)D_{\mu(g-t')}.$$
Let $(D,g)$ be an adelic ${\mathbb{R}}$-Cartier divisor on $X$ and $d = \dim X$. Then $F_{(D,g)}(t)$ is a $d$-concave function on $(-\infty,{{\lambda_{\max}}^{\mathrm{asy}}}(D,g))$, that is, $F_{(D,g)}(t)^{1/d}$ is concave on $(-\infty,{{\lambda_{\max}}^{\mathrm{asy}}}(D,g))$. In particular, $F_{(D,g)}(t)$ is continuous on ${\mathbb{R}}\setminus \{{{\lambda_{\max}}^{\mathrm{asy}}}(D,g)\}$.
By definition, $$F_{(D,g)}(t) = {\mathrm{vol}}(D_{\mu(g-t)})$$ for $t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$. Since the algebraic volume is $d$-concave on a big cone, for any $t,t' < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $0 \leq \epsilon \leq 1$, we have $$\begin{aligned}
F_{(D,g)}(\epsilon t + (1-\epsilon)t')^{\frac{1}{d}} &= {\mathrm{vol}}(D_{\mu(g - (\epsilon t + (1-\epsilon)t'))})^{\frac{1}{d}} \\
&\geq {\mathrm{vol}}(\epsilon D_{\mu(g-t)} + (1-\epsilon)D_{\mu(g-t')})^{\frac{1}{d}} \quad (\because \text{Proposition \ref{D1}}) \\
&\geq \epsilon \ {\mathrm{vol}}(D_{\mu(g-t)})^{\frac{1}{d}} + (1-\epsilon){\mathrm{vol}}(D_{\mu(g-t')})^{\frac{1}{d}} \\
&= \epsilon F_{(D,g)}(t)^{\frac{1}{d}} + (1-\epsilon)F_{(D,g)}(t')^{\frac{1}{d}}.\end{aligned}$$
In general, we cannot extend $F_{(D,g)}$ to a continuous function on the whole ${\mathbb{R}}$. For example, let $X = {\mathbb{P}}_K^1 = \mathrm{Proj}K[T_0,T_1]$, $z=T_1/T_0$, $D = \{T_0=0\}$ and $x_\infty = (0:1)$ Let $g = \log\max\{1,|z|\}$. Then we have $$\mu_x(g-t)=
\begin{cases}
0 & (x \neq x_\infty) \\
1 & (x = x_\infty)
\end{cases}$$ for $t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) = 0$. Hence we obtain that $$F_{(D,g)}(t) =
\begin{cases}
1 & (t < 0) \\
0 & (t > 0).
\end{cases}$$
Continuity of the arithmetic volume
-----------------------------------
Firstly, we will prove the continuity of ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ for an adelic ${\mathbb{R}}$-Cartier divisor $(D,g)$.
\[lam1\] Let $(D,g),(D',g')$ be adelic ${\mathbb{R}}$-Cartier divisors on $X$. We have $${{\lambda_{\max}}^{\mathrm{asy}}}(D+D',g+g') \geq {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) + {{\lambda_{\max}}^{\mathrm{asy}}}(D',g').$$
For any integers $n,n' > 0$, there are non-zero elements $s \in H^0(nD) \setminus \{0\}$ and $s' \in H^0(n'D') \setminus \{0\}$ such that $$\|s\|_{ng} \leq e^{-{\lambda_{\max}}(nD,ng)} \ \text{and} \ \|s'\|_{n'g'} \leq e^{-{\lambda_{\max}}(n'D',n'g')}.$$ Since $s^{\otimes n'} \otimes s'^{\otimes n} \in H^0(nn'(D+D')) \setminus \{0\}$, we have $$\|s^{\otimes n'} \otimes s'^{\otimes n}\|_{nn'(g+g')} \leq (\|s\|_{ng})^{n'}(\|s'\|_{n'g'})^n \leq e^{-n'{\lambda_{\max}}(nD,ng)-n{\lambda_{\max}}(n'D',n'g')},$$ which implies $$\frac{1}{nn'}{\lambda_{\max}}(nn'(D+D'),nn'(g+g')) \geq \frac{1}{n}{\lambda_{\max}}(nD,ng) + \frac{1}{n'}{\lambda_{\max}}(n'D',n'g').$$ Since ${{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \geq {\lambda_{\max}}(nD,ng)/n$, we have $${{\lambda_{\max}}^{\mathrm{asy}}}(D+D',g+g') \geq \frac{1}{n}{\lambda_{\max}}(nD,ng) + \frac{1}{n'}{\lambda_{\max}}(n'D',n'g').$$ Taking a supremum with respect to $n,n'$, we complete the proof.
\[lam2\] Let $\overline{D}=(D,g),\overline{D}'=(D',g')$ be adelic ${\mathbb{R}}$-Cartier divisors on $X$. We assume $D$ is big. Then $\lambda(t) := {{\lambda_{\max}}^{\mathrm{asy}}}(\overline{D}+t\overline{D}')$ is a real-valued function on some open interval $(a,b) \subset {\mathbb{R}}$ containing 0, and concave on $(a,b)$. In particular $\lambda(t)$ is continuous on $(a,b)$.
Since $D$ is big, $D+tD'$ is big for $|t| \ll 1$, which implies that $\lambda(t)$ is definable on a sufficiently small open neighborhood of 0. Moreover, using Lemma \[lam1\], we can prove the concavity of $\lambda(t)$ by Lemma \[con\].
Next, we prove the continuity of the arithmetic volume ${\widehat{{\mathrm{vol}}}}({\raisebox{.4ex}{.}})$. Let $(D,g),(D',g')$ be adelic ${\mathbb{R}}$-Cartier divisors on $X$ and we assume $D$ is big. We set $$(D_\epsilon,g_\epsilon) := (D,g) + \epsilon(D',g'),$$ and $$\begin{aligned}
F_\epsilon(t) := \left \{
\begin{array}{ll}
{\mathrm{vol}}((D_\epsilon)_{\mu(g_\epsilon-t)}) & (t < {{\lambda_{\max}}^{\mathrm{asy}}}(D_\epsilon,g_\epsilon)) \\
0 & (t > \lambda_{{\max}}^{{\mathrm{asy}}}(D_\epsilon,g_\epsilon)).
\end{array}
\right .\end{aligned}$$ We remark that this function is well-defined if $|\epsilon| \ll 1$ by Proposition \[lam2\].
The function $F_\epsilon(t)$ converges pointwise to $F_{(D,g)}(t)$ on ${\mathbb{R}}\setminus \{ {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \}$ as $|\epsilon| \rightarrow 0$. More precisely, for any $t \in {\mathbb{R}}\setminus \{ {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \}$, $F_\epsilon(t)$ is continuous with respect to $\epsilon$ on a sufficiently small open neighborhood of $\epsilon = 0$.
We first assume $t > {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$. By Proposition \[lam2\], there is $\delta > 0$ such that ${{\lambda_{\max}}^{\mathrm{asy}}}(D_\epsilon,g_\epsilon) < t$ if $|\epsilon| < \delta$. Then $F_\epsilon(t) = F_{(D,g)}(t) = 0$, which is required.
Next we assume $t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$. Similarly, there is $\delta > 0$ such that ${{\lambda_{\max}}^{\mathrm{asy}}}(D_\epsilon,g_\epsilon) > t$ if $|\epsilon| < \delta$. Then $F_\epsilon(t)$ is $d$-concave with respect to $\epsilon$ on $(-\delta,\delta)$, where $d = \dim X$. In fact, by Lemma \[con\], for any $\epsilon,\epsilon' \in (-\delta,\delta)$ and $0 \leq \zeta \leq 1$, we have $$(D_{\zeta\epsilon+(1-\zeta)\epsilon'})_{\mu(g_{\zeta\epsilon+(1-\zeta)\epsilon'}-t)} \geq \zeta(D_\epsilon)_{\mu(g_\epsilon-t)} + (1-\zeta)(D_{\epsilon'})_{\mu(g_{\epsilon'}-t)}.$$ Therefore $F_\epsilon(t)$ is $d$-concave with respect to $\epsilon$ on $(-\delta,\delta)$ because $F_\epsilon(t) = {\mathrm{vol}}((D_\epsilon)_{\mu(g_\epsilon-t)})$ and the algebraic volume is $d$-concave. In particular, $F_\epsilon(t)$ is continuous with respect to $\epsilon$ on $(-\delta,\delta)$.
Since $F_\epsilon(t)$ is uniformly bounded with respect to $\epsilon$ and $${\widehat{{\mathrm{vol}}}}(D_\epsilon,g_\epsilon) = (d+1) \int_0^{+\infty} F_\epsilon(t) \ \mathrm{d}t$$ by Theorem \[vol1\], we get the continuity of the arithmetic volume by bounded convergence theorem:
\[cvol\] Let $\overline{D}=(D,g),\overline{D}'=(D',g')$ be adelic ${\mathbb{R}}$-Cartier divisors on $X$. We assume $D$ is big. Then ${\widehat{{\mathrm{vol}}}}(\overline{D}+\epsilon \overline{D}')$ converges to ${\widehat{{\mathrm{vol}}}}(\overline{D})$ as $|\epsilon| \rightarrow 0$.
Log concavity of the arithmetic volume
--------------------------------------
Firstly, we will prove some inequalities:
\[inq1\] Let $a,b,p$ and $\epsilon$ be real numbers such that $a,b \geq 0$, $p > 0$ and $0 < \epsilon < 1$. Then we have the following inequality: $$(\epsilon a^p + (1-\epsilon)b^p)^{\frac{1}{p}} \geq a^\epsilon b^{1-\epsilon} \geq \min \{a,b\}.$$
If $ab = 0$, the assertion is clear, so we assume that $a,b > 0$. Moreover, the inequality $a^\epsilon b^{1-\epsilon} \geq \min \{a,b\}$ is also clear. Now, we will show the first inequality. Since $\log x$ is concave on $(0,+\infty)$, we have $$\log (\epsilon x + (1-\epsilon)y) \geq \epsilon \log x + (1-\epsilon) \log y$$ for any $x,y > 0$. Substituting $x$ for $a^p$ and $y$ for $b^p$,
$$\begin{aligned}
\log (\epsilon a^p + (1-\epsilon)b^p) \geq \epsilon \log a^p + (1-\epsilon) \log b^p
\Longleftrightarrow &\log (\epsilon a^p + (1-\epsilon)b^p)^\frac{1}{p} \geq \log a^\epsilon b^{1-\epsilon} \\
\Longleftrightarrow & (\epsilon a^p + (1-\epsilon)b^p)^{\frac{1}{p}} \geq a^\epsilon b^{1-\epsilon}\end{aligned}$$
\[inq2\] Let $V$ be a convex cone. Let $f : V \rightarrow (0,+\infty)$ be a non-negative $d$-homogeneous function for some $d > 0$, that is, $$f(av) = a^df(v)$$ for any $a > 0$ and $v \in V$. Then the following conditions are equivalent:
1. $f$ is $d$-concave, that is, $$f(\epsilon v + (1-\epsilon)v')^\frac{1}{d} \geq \epsilon f(v)^\frac{1}{d} + (1-\epsilon) f(v')^\frac{1}{d}$$ for every $v,v' \in V$ and $0 \leq \epsilon \leq 1$.
2. $f(\epsilon v + (1-\epsilon)v') \geq \min \{ f(v), f(v') \}$ for every $v,v' \in V$ and $0 \leq \epsilon \leq 1$.
Firstly, we assume $(1)$ and we can assume $\min \{f(v),f(v')\} = f(v)$. Then we have $$f(\epsilon v + (1-\epsilon)v')^\frac{1}{d} \geq \epsilon f(v)^\frac{1}{d} + (1-\epsilon) f(v')^\frac{1}{d} \geq f(v)^\frac{1}{d}.$$ Raising both sides to $d$-th power, we have $$f(\epsilon v + (1-\epsilon)v') \geq f(v) = \min \{ f(v), f(v') \}.$$
Next we assume $(2)$. If we set $$w = f(v)^{-\frac{1}{d}}v,\ w'= f(v')^{-\frac{1}{d}}v',\ \epsilon = \frac{f(v)^\frac{1}{d}}{f(v)^\frac{1}{d} + f(v')^\frac{1}{d}},$$ we have $$\begin{aligned}
\epsilon w + (1-\epsilon) w' &= \frac{1}{f(v)^\frac{1}{d} + f(v')^\frac{1}{d}}(v+v'), \\
\min \{f(w),f(w')\} &= 1.\end{aligned}$$ By the inequality $(2)$ for $w,w'$ and $\epsilon$, we have $$\begin{aligned}
(f(v)^\frac{1}{d} + f(v')^\frac{1}{d})^{-d}f(v+v') \geq 1
\Longleftrightarrow & f(v+v') \geq (f(v)^\frac{1}{d} + f(v')^\frac{1}{d})^d \\
\Longleftrightarrow & f(v+v')^\frac{1}{d} \geq f(v)^\frac{1}{d} + f(v')^\frac{1}{d},\end{aligned}$$ which implies the inequality $(1)$ because $f$ is $d$-homogeneous.
Moreover, we will use the following inequality so called “Pr[é]{}kopa-Leindler inequality”. It was proved by Pr[é]{}kopa [@1971logarithmic] [@prekopa1973logarithmic] and Leindler [@leindler1972certain] (for detail, see [@gardner2002brunn]).
\[inq3\] Let $0<\epsilon<1$ and $f,g,h:{\mathbb{R}}^n \rightarrow [0,+\infty)$ be measurable functions. We assume $$h(\epsilon x + (1-\epsilon)y) \geq f(x)^\epsilon g(y)^{1-\epsilon}$$ for any $x,y \in {\mathbb{R}}^n$. Then we have $||h||_1 \geq ||f||_1^\epsilon ||g||_1^{1-\epsilon}$, that is, $$\int_{{\mathbb{R}}^n} h \ \mathrm{d}\nu \geq \left( \int_{{\mathbb{R}}^n} f \ \mathrm{d}\nu \right)^\epsilon \left( \int_{{\mathbb{R}}^n} g \ \mathrm{d}\nu \right)^{1-\epsilon}$$ where $\nu$ is the Lubesgue measure on ${\mathbb{R}}^n$.
Now, we start to prove the log concavity of ${\widehat{{\mathrm{vol}}}}({\raisebox{.4ex}{.}})$.
\[cvol2\] The arithmetic volume ${\widehat{{\mathrm{vol}}}}({\raisebox{.4ex}{.}})$ is $(d+1)$-concave for $d = \dim X$. More precisely, for any big adelic ${\mathbb{R}}$-Cartier divisors $(D,g),(D',g')$, we have $${\widehat{{\mathrm{vol}}}}(D+D',g+g')^\frac{1}{d+1} \geq {\widehat{{\mathrm{vol}}}}(D,g)^\frac{1}{d+1} + {\widehat{{\mathrm{vol}}}}(D',g')^\frac{1}{d+1}.$$
For $0 < \epsilon < 1$, we set $$(D_\epsilon,g_\epsilon) := \epsilon(D,g) + (1-\epsilon)(D',g'),$$ and $$\begin{aligned}
\Theta_{(D,g)}(t) := \left \{
\begin{array}{ll}
(d+1){\mathrm{vol}}(D_{\mu(g-t)}) & (0 \leq t < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)) \\
0 & (\text{otherwise}).
\end{array}
\right .\end{aligned}$$ Then, we have $$\label{ve1}
{\widehat{{\mathrm{vol}}}}(D,g) = ||\Theta_{(D,g)}||_1,\ {\widehat{{\mathrm{vol}}}}(D',g') = ||\Theta_{(D',g')}||_1,\ {\widehat{{\mathrm{vol}}}}(D_\epsilon,g_\epsilon) = ||\Theta_{(D_\epsilon,g_\epsilon)}||_1$$ by Theorem \[vol1\]. We claim that $$\label{ve2}
\Theta_{(D_\epsilon,g_\epsilon)}(\epsilon x + (1-\epsilon)y) \geq \Theta_{(D,g)}(x)^\epsilon \Theta_{(D',g')}(y)^{1-\epsilon} \quad \text{for any} \ x,y \in {\mathbb{R}}.$$ In fact, if $x<0, {{\lambda_{\max}}^{\mathrm{asy}}}(D,g) \leq x, y<0$ or ${{\lambda_{\max}}^{\mathrm{asy}}}(D',g') \leq y$, we have $\Theta_{(D,g)}(x)=0$ or $\Theta_{(D',g')}(y)=0$, so the inequality (\[ve2\]) is clear in this case. And if $0 \leq x < {{\lambda_{\max}}^{\mathrm{asy}}}(D,g)$ and $0 \leq y < {{\lambda_{\max}}^{\mathrm{asy}}}(D',g')$, we have $$\mu_z(g_\epsilon -(\epsilon x + (1-\epsilon)y)) \geq \epsilon \mu_z(g-x) + (1-\epsilon)\mu_z(g'-y)$$ for any $z \in X^{(1)}$, which implies that $$(D_\epsilon)_{\mu(g_\epsilon - (\epsilon x + (1-\epsilon)y))} \geq \epsilon D_{\mu(g-x)} + (1-\epsilon)D'_{\mu(g'-y)}.$$ Since the algebraic volume is $d$-concave, we obtain $${\mathrm{vol}}((D_\epsilon)_{\mu(g_\epsilon - (\epsilon x + (1-\epsilon)y))})^\frac{1}{d} \geq \epsilon \ {\mathrm{vol}}(D_{\mu(g-x)})^\frac{1}{d} + (1-\epsilon){\mathrm{vol}}(D'_{\mu(g'-y)})^\frac{1}{d}.$$ By Lemma \[inq1\], we get $${\mathrm{vol}}((D_\epsilon)_{\mu(g_\epsilon - (\epsilon x + (1-\epsilon)y))}) \geq {\mathrm{vol}}(D_{\mu(g-x)})^\epsilon {\mathrm{vol}}(D'_{\mu(g'-y)})^{1-\epsilon},$$ which is equivalent to the inequality (\[ve2\]). Therefore by Pr[é]{}kopa-Leindler inequality, we have $||\Theta_{(D_\epsilon,g_\epsilon)}||_1 \geq ||\Theta_{(D,g)}||_1^\epsilon ||\Theta_{(D',g')}||_1^{1-\epsilon}$. By Lemma \[inq1\] again, we have $||\Theta_{(D_\epsilon,g_\epsilon)}||_1 \geq \min \{ ||\Theta_{(D,g)}||_1, ||\Theta_{(D',g')}||_1 \}$, which is the inequality $${\widehat{{\mathrm{vol}}}}(D_\epsilon,g_\epsilon) \geq \min \{ {\widehat{{\mathrm{vol}}}}(D,g), {\widehat{{\mathrm{vol}}}}(D',g') \}$$ by (\[ve1\]). Since the arithmetic volume is $(d+1)$-homogeneous by Corollary \[vol2\], we have $${\widehat{{\mathrm{vol}}}}(D_\epsilon,g_\epsilon)^\frac{1}{d+1} \geq \epsilon \ {\widehat{{\mathrm{vol}}}}(D,g)^\frac{1}{d+1} + (1-\epsilon) {\widehat{{\mathrm{vol}}}}(D',g')^\frac{1}{d+1},$$ by Lemma \[inq2\], which completes the proof.
[99]{}
S. J. Arakelov. Intersection theory of divisors on an arithmetic surface. , Vol. 8, No. 6, pp. 1167–1180, 1974.
V. G. Berkovich. . No. 33. American Mathematical Soc., 2012.
S. Boucksom and H. Chen. Okounkov bodies of filtered linear series. , Vol. 147, No. 4, pp. 1205–1229, 2011.
S. Boucksom and M. Jonsson. Singular semipositive metrics on line bundles on varieties over trivially valued fields. , 2018.
H. Chen. Arithmetic [F]{}ujita approximation. , Vol. 43, No. 4, pp. 555–578, 2010.
H. Chen. Majorations explicites des fonctions de [H]{}ilbert-[S]{}amuel g[é]{}om[é]{}trique et arithm[é]{}tique. , Vol. 1, No. 279, pp. 99–137, 2015.
H. Chen and A. Moriwaki. Sufficient conditions for the [D]{}irichlet property. , 2017.
H. Chen and A. Moriwaki. Arakelov geometry over adelic curves. , 2019.
G. Faltings. Calculus on arithmetic surfaces. , pp. 387–424, 1984.
M. Fulger, J. Koll[á]{}r, and B. Lehmann. Volume and [H]{}ilbert functions of $\mathbb{R}$-divisors. , Vol. 65, No. 2, pp. 371–387, 2016.
R. Gardner. The [B]{}runn-[M]{}inkowski inequality. , Vol. 39, No. 3, pp. 355–405, 2002.
H. Gillet and C. Soul[é]{}. Arithmetic intersection theory. , Vol. 72, No. 1, pp. 94–174, 1990.
H. Gillet and C. Soul[é]{}. An arithmetic [R]{}iemann-[R]{}och theorem. , Vol. 110, No. 1, pp. 473–543, 1992.
R. Hartshorne. , Vol. 52. Springer Science & Business Media, 2013.
R. K. Lazarsfeld. , Vol. 48. Springer, 2017.
L. Leindler. On a certain converse of [H]{}[ö]{}lder’s inequality. 2. , Vol. 33, pp. 217–223, 1972.
Q. Liu and R. Erne. , Vol. 6. Oxford University Press, 2006.
A. Moriwaki. Continuity of volumes on arithmetic varieties. , Vol. 18, No. 3, pp. 407–457, 2009.
A. Moriwaki. , Vol. 244. American Mathematical Soc., 2014.
A. Moriwaki. , Vol. 242. American Mathematical Society, 2016.
A. Pr[é]{}kopa. Logarithmic concave measures with application to stochastic programming. , Vol. 32, pp. 301–316, 1971.
A. Pr[é]{}kopa. On logarithmic concave measures and functions. , Vol. 34, pp. 335–343, 1973.
P. Vojta. Siegel’s theorem in the compact case. , Vol. 133, No. 3, pp. 509–548, 1991.
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a theory of the subgap tails of the density of states in a diffusive superconductor containing magnetic impurities. We show that the subgap tails have two contributions: one arising from mesoscopic gap fluctuations, previously discussed by Lamacraft and Simons, and the other associated to the long-wave fluctuations of the concentration of magnetic impurities. We study the latter both in small superconducting grains and in bulk systems \[$d=1,2,3$\], and establish the dimensionless parameter that controls which of the two contributions dominates the subgap tails. We observe that these contributions are related to each other by dimensional reduction. We apply the theory to estimate the effects of a weak concentration of magnetic impurities \[$\approx 1 {\rm p.p.m}$\] on the phase coherence of Josephson qubits. We find that at these typical concentrations, magnetic impurities are relevant for the dephasing in large qubits, designed around a $10\;{\rm \mu m}$ scale, where they limit the quality factor to be $Q<10^4-10^5$.'
author:
- 'Alessandro Silva and Lev B. Ioffe'
title: Subgap states in dirty superconductors and their effect on dephasing in Josephson qubits
---
Introduction. {#sec0}
=============
Recent experimental studies of disordered conductors demonstrated that even a weak concentration of magnetic impurities may lead to important effects, especially in the context of dephasing and energy relaxation. In particular, the experiments in mesoscopic Au and Cu wires have shown that the frequently observed saturation of the dephasing time $\tau _{\varphi }$ at low temperatures [@Mohanty; @Pierre; @Pierre2] is likely due to a small unavoidable presence of magnetic impurities with the concentration as low as $0.1-1\;{ppm}$. On a theoretical side it was shown [@Glazman] that magnetic impurities at these concentrations can explain the puzzling anomalous energy dependence of the relaxation rate observed in diffusive Au and Cu wires [@Pothier; @Pierre3].
It is well known that a significant concentration of magnetic impurities strongly affects conventional superconductors [@Abrikosov]. Further, it was shown recently that even a small concentration of magnetic impurities leads to exponentially small tails in the density of states within the gap [@Balatsky; @Aleiner; @Shytov; @Brouwer; @Lamacraft]. Because such subgap states can trap and release the quasiparticles, they can become an important for a number of physical applications, e.g. for Josephson-junction qubits. For example, since the resulting subgap states are localized, the quasiparticles that are generated during qubit manipulations or readout [@Martinis1], can get trapped in these localized subgap states. These trapped quasiparticles provide extra degrees of freedom that contribute to the dephasing and dissipation in subsequent qubit operations. Although the typical materials used for the fabrication of Josephson-junction qubits \[ e.g., ${Al-Al_{2}O_{3}-Al}$\], are not intentionally doped with magnetic impurities, some concentration of them is unavoidable, especially in small devices where they are due to the surface effects. Therefore, it is important to estimate the effects of diluted magnetic impurities on the density of states in these systems and the resulting dephasing.
With this goal in mind we present here a detailed theory of the subgap states in a diffusive superconductors containing low concentrations of magnetic impurities. In particular, we study the contribution to the subgap density of states due to spatial fluctuations of the concentration of magnetic impurities. Combining the results of our analysis with those recently obtained by Lamacraft and Simons [@Lamacraft], who considered the subgap tails resulting from universal mesoscopic gap fluctuations, we provide a full picture of the subgap states, and establish in which regime one or the other type of the fluctuation dominates the physics. Finally, the results of our theoretical analysis are applied to estimate the subgap density of states generated by a concentration of $\approx 1$ ppm of magnetic impurities in aluminum-based qubits. The result of our analysis is that the smallest superconducting islands used in modern experiments might contain a few subgap states while larger qubits can contain about $10^{3}$ of them. Finally, we give the estimates the qubit decoherence resulting from trapped quasiparticles and conclude that they provide an important source of the decoherence in larger devices but negligible in the smallest ones.
The rest of paper is organized as follows. In Sec. \[sec00\] we present the qualitative arguments and the results of our analysis. The technical details of the density of states computation are presented in Sec. [sec01]{}: there we first discuss the effect of fluctuations on the disorder averaged density of states in a large collection of superconducting grains \[ Sec. \[sec1\]\], and then extend the analysis to the bulk systems ($d=1,2,3$) in Sec. \[sec2\]. In Sec. \[sec2b\] we present the details of our estimates of the subgap density of states in superconducting qubits and its effect on the decoherence. Finally, Sec. \[sec3\] gives the conclusions.
Overview and Main Results.
==========================
\[sec00\]
The effects of impurities on conventional superconductivity were extensively studied in the last four decades. It is well known that dilute non-magnetic impurities preserving the time reversal symmetry do not affect the superconducting gap [@Anderson] while a finite concentration of magnetic impurities suppresses both the critical temperature $T_{c}$ and the gap in the local density of states, $E_{g}$ [@Abrikosov]. The effect of weak magnetic impurities is controlled by the dimensionless parameter $$\zeta =\frac{1}{\tau _{s}\Delta },$$where $\tau _{s}$ is the spin-scattering time and $\Delta $ the bare superconducting gap. At the critical concentration of magnetic impurities corresponding to $\zeta =1$ the gap in the density of states closes but $%
T_{c}$ remains finite; at larger concentrations one the superconductor becomes gapless [@Abrikosov]. Subsequent extensions of this theory have included the effect of strong magnetic impurities. While a single strong magnetic impurity generates a bound intra-gap quasi-particle state and a local suppression of the BCS gap [@Liu; @Rusinov; @Balatsky3], a finite concentration of them leads to the formation of an intra-gap impurity band [@Liu; @Shiba; @Rusinov2]. As the concentration is further increased, the impurity band broadens, merging with the BCS continuum, and eventually resulting in a complete gap destruction.
For small impurity concentration corresponding to $\zeta <1$, the Abrikosov-Gorkov theory [@Abrikosov] predicts a vanishing density of states for $\mid E\mid <E_{g}=\Delta (1-\zeta ^{2/3})^{3/2}$. This conclusion, however, neglects the effects of the impurity distribution fluctuations [@Balatsky]. One expects that such fluctuations smear the gap edge and lead to the exponential tails in the density of states extending down into the superconducting gap, similarly to the Lifshits tails [@Lifshits] in the density of states below the band edge of a disordered conductor [@Halperin; @Lax; @Zittarz].
As in the case of Lifshits tails one can visualize the the subgap states in a bulk sample as appearing locally in places where the impurity potential decreases the energy of the quasiparticle. Further, the main contribution to the states deep in the tail is dominated by the specific impurity potential that is called the optimal fluctuation. Historically, the first example of such optimal fluctuations in the context of superconductivity was found by Larkin and Ovchnikov [@Larkin] who considered the effects of the interaction constant inhomogeneity on the density of the subgap states. Recently, various mechanisms leading to such sub-gap states in superconductors containing magnetic impurities have been considered, both in lightly magnetically doped superconductors [Shytov]{}, in small superconducting grains [@Aleiner; @Brouwer; @Lamacraft], and in diffusive superconductors [@Balatsky; @Lamacraft]. We will concentrate on the latter case: dirty superconductors containing both magnetic and non-magnetic impurities, under the conditions $\zeta \ll 1$ and $l\ll \xi $, where $l$ is the mean free path relative to momentum scattering and $\xi $ is the coherence length.
The subgap tails in this regime were recently considered by Lamacraft and Simons [@Lamacraft]. Slightly above the gap the density of states predicted by Abrikosov and Gorkov is $$\begin{aligned}
&& \nonumber \\
\nu &=&\frac{1}{\pi L^{d}}\sqrt{\frac{E-E_{g}(\zeta )}{\Delta _{g}^{3}}}
\label{AG_DOS} \\
\Delta _{g}^{-3/2} &\simeq & \pi \nu _{0}L^{d}\sqrt{\frac{2}{3\Delta }}\frac{%
1}{\zeta ^{2/3}}\end{aligned}$$where $\nu _{0}$ is the density of states per unit volume in the normal state, $L$ is the sample size, and $d$ the dimensionality of the system. Formulating the problem in terms of a supersymmetric Sigma model, Lamacraft and Simons obtained a density of states close but below the gap $$\frac{\langle \nu \rangle }{\nu _{0}}\propto \exp \left[ -\tilde{a}%
_{d}\left( \frac{\lambda _{0}}{L}\right) ^{d}\left( \frac{E_{g}-E}{\Delta
_{g}}\right) ^{3/2}\right] \label{Lama}$$where $\tilde{a}_{d}\sim 1$ and $$\lambda _{0}=\xi \left( \frac{\Delta }{E_{g}-E}\right) ^{1/4} \label{size}$$is the linear size of the optimal fluctuation.
This result has a clear physical interpretation. In zero dimensions, the action is a universal function of the rescaled energy $(E_{g}-E)/\Delta_g$, where $\Delta_g$ can be interpreted as the effective level spacing in the right above the AG gap. This indicates that the gap fluctuations can be seen as Random Matrix-like fluctuations of the edge of a Wigner semicircle [Tracy]{}. Indeed, the zero dimensional result was previously conjectured [Brouwer]{} on the basis of the universality hypothesis of Random Matrix Theory.
In a bulk system \[$d\neq 0$\], the subgap localized states giving the density of states Eq.(\[Lama\]) originate from particular configurations of normal and magnetic impurities favoring mesoscopic fluctuations of the gap edge. On the other hand, it is intuitively clear that another type of inhomogeneous fluctuations should also contribute to the sub-gap tails of the DOS: local fluctuations of the concentration $n_{imp}$ of magnetic impurities dictated by their Poissonian statistics . A local increase of $n_{imp}$, and therefore of the spin-scattering rate $\zeta $, above its average value implies a local suppression of the AG gap, and as a result the formation of sub-gap localized quasi-particle states [@Balatsky1].
Below, we will show that the mechanism generating subgap states described above is complementary to universal mesoscopic gap fluctuations. This complementarity is particularly transparent in the peculiar zero dimensional limit, where the explicit form of the subgap tails depends on the way the statistics over disorder realizations is acquired . Indeed, if gap fluctuations are studied in a single sample, e.g varying the boundary conditions through gate voltages, the rate $\zeta $ is constant, and the universal result of Lamacraft and Simons [@Lamacraft] always applies. On the other hand, if one considers a large collection of the superconducting grains $\zeta $ is going to display sample to sample fluctuations. As shown below, depending on the system parameters, either the former or the latter effect is bigger.
While in the zero dimensional case one has the possibility to single out mesoscopic gap fluctuations by considering a single mesoscopic grain, in finite dimensional systems this is not possible. Since the analysis of Ref.\[\] neglects large scale fluctuations of $\zeta $, limiting the scattering rate to be a constant independent on position, only a direct comparison of the resulting subgap tails can establish which of the two mechanisms gives the dominant contribution to the asymptotic subgap density of states.
In the following Sections, we shall study in detail the physics of the subgap states due to the fluctuations of the concentration of magnetic impurities, or, equivalently, of $\zeta $. We describe the dirty superconductor by the quasiclassical Usadel equations [@Usadel; @Kopnin] in which the spin-scattering rate $\zeta $ becomes a position dependent statistical quantity. The variable $\zeta $ inherits its statistics from the Poissonian distribution of magnetic impurities. In this framework, we find the optimal fluctuation of $\zeta $ by solving the associated variational problem and the resulting instanton equations, and discuss the deep analogies between the problem at hand and the Lifshits tails in disordered conductors. We then calculate the expression of the asymptotic subgap tails of the DOS, including gaussian prefactors, as a function of the distance from the gap edge $$\delta \epsilon =\frac{(E_{g}-E)}{\Delta} ,$$and of the average scattering rate $\zeta $, impurity concentration $n_{imp}$, and dimensionality \[$d=0,1,2,3$\]. For every dimensionality, we compare our results with the subgap tails due to mesoscopic gap fluctuations and establish the parameter that controls which of the two types of fluctuations dominated the physics of subgap states.
The results of our analysis may be summarized as follows. While the typical size of the optimal fluctuations associated to a local increase of $\zeta $ is given by Eq.(\[size\]), the subgap density of states originating from the $n_{imp}$ fluctuations is $$\begin{aligned}
\label{DOS0}
\frac{\langle \nu \rangle }{\nu _{0}} &\propto &\exp \left[ -a_{d}\frac{%
n_{imp}\xi ^{d}}{\zeta ^{4/3}}(\delta \epsilon )^{2-d/4}\right] \nonumber
\\
&=&\exp \left[ -a_{d}\left( \frac{\lambda _{0}}{L}\right) ^{d}\left( \frac{%
E_{g}-E}{\delta E_{g}}\right) ^{2}\right]\end{aligned}$$where $a_{d}$ is a constant, calculated below, $\delta E_{g}=\Delta \zeta
^{2/3}/\sqrt{N}$, and $N=n_{imp}\;L^{d}$. A more detailed formula for the DOS is given by Eq.(\[finalDOS\]). The effective dimensionality of the system is determined by comparing the linear size of the sample to the typical size of an optimal fluctuation $\lambda $, Eq.(\[size\]). In particular, for $d=0$ the parameter that determines whether the physics of tail states is dominated by mesoscopic fluctuations \[Eq. (\[Lama\])\] or fluctuations of $%
\zeta $ \[Eq. (\[DOS0\])\] is the ratio $$\beta =\frac{\Delta _{g}}{\delta E_{g}}$$For $\beta \gg 1$ the asymptotics is dictated by Eq.(\[Lama\]), while for $%
\beta \ll 1$ the result of Eq.(\[DOS0\]) applies. In higher dimensions, a detailed comparison is more involved. However it is interesting to notice that, apart from a numerical constant, the ratio of the actions relative to the two optimal fluctuations \[see Eq. (\[Lama\]), and Eq. (\[DOS0\])\] is independent of dimensionality.
Subgap DOS associated to fluctuations of $\protect\zeta$.
=========================================================
\[sec01\]
In this Section, we present the theory of the subgap density of states associated to fluctuations of the spin-scattering rate $\zeta $, starting with the simple zero dimensional case, and subsequently extending it to the case of finite dimensional systems.
Fluctuations of $\protect\zeta$ in small grains.
------------------------------------------------
\[sec1\]
We begin with the simpler zero dimensional case, the results in this subsection apply to the samples which are smaller than the size of the optimal fluctuation.
As explained in the previous section, the zero dimensional case is peculiar because the form of the subgap DOS depends on the whether statistics is acquired measuring the DOS in a single sample, e.g. varying the boundary conditions, or in a large collection of small grains. In the first case, the subgap tails are always dominated by mesoscopic Random Matrix-like gap fluctuations. In the following, we will analyze the case of a large collection of small grains, where both mesoscopic gap fluctuations and fluctuations of $\zeta $ contribute to the subgap tails.
For every value of the dimensionless spin-scattering rate $\tilde{\zeta}$, and slightly above the gap edge, the DOS in the Abrikosov-Gorkov approximation is given by $$\begin{aligned}
\nu &=&\frac{1}{\pi L^{3}}\sqrt{\frac{E-E_{g}(\tilde{\zeta})}{\Delta _{g}^{3}%
}}, \\
\Delta _{g}^{-3/2} &\simeq & \pi \nu _{0}L^{3}\sqrt{\frac{2}{3\Delta }}\frac{%
1}{\tilde{\zeta}^{2/3}}. \label{levspacing}\end{aligned}$$where $L$ is the linear size of each grain. We can estimate the DOS tail resulting from sample to sample fluctuations of the concentration of impurities by promoting $\tilde{\zeta}$ to be a statistical variable, with average $\zeta $ and variance $\zeta ^{2}/N$, where $N$ is the average number of paramagnetic impurities per sample. The average DOS is $$\frac{\langle \nu \rangle }{\nu _{0}}=\frac{4}{\zeta ^{2/3}}\sqrt{\frac{2}{3}%
}\int \frac{d(\delta \zeta )}{\sqrt{2\pi \eta _{0}}}\sqrt{\frac{%
E-E_{g}(\zeta +\delta \zeta )}{\Delta }}e^{-\frac{(\delta \zeta )^{2}}{2\eta
_{0}}},$$where $\eta _{0}=\zeta ^{2}/N$. Since $$\frac{E_{g}(\zeta +\delta \zeta )}{\Delta }\simeq \frac{E_{g}(\zeta )}{%
\Delta }-\frac{\delta \zeta }{\zeta ^{1/3}},$$introducing $\delta \epsilon =(E_{g}(\zeta )-E)/\Delta $, we obtain $$\begin{aligned}
\frac{\langle \nu \rangle }{\nu _{0}} &\propto &\int_{\zeta ^{1/3}\delta
\epsilon }^{+\infty }d(\delta \zeta )\sqrt{\delta \zeta -\zeta ^{1/3}\delta
\epsilon }\;e^{-\frac{(\delta \zeta )^{2}}{2\eta _{0}}} \nonumber
\label{finalzeroDOS} \\
&\propto &\left[ \frac{1}{2S}\right] ^{3/4}e^{-S}.\end{aligned}$$Here $$S=\frac{1}{2}\frac{N}{\zeta ^{4/3}}(\delta \epsilon )^{2}=\frac{1}{2}\left(
\frac{E_{g}-E}{\delta E_{g}}\right) ^{2}, \label{actionme}$$is the action, assumed to be $S>>1$, and $$\delta E_{g}=\Delta \frac{\zeta ^{2/3}}{\sqrt{N}},$$is the typical scale of gap fluctuations. The final result is $$\frac{\langle \nu \rangle }{\nu _{0}}\propto \frac{1}{(\delta \epsilon
)^{3/2}}\exp \left[ -\frac{1}{2}\frac{N}{\zeta ^{4/3}}(\delta \epsilon )^{2}%
\right] .$$ Let us compare now these results to the asymptotic DOS resulting from mesoscopic fluctuations of the gap edge [@Lamacraft]. These fluctuations are characterized by the action $$S_{SUSY}=\frac{4}{3}\left( \frac{E_{g}-E}{\Delta _{g}}\right) ^{3/2}.$$As consequence of the universality of this result, this action is of order $%
1 $ at energies below the gap corresponding to the effective level spacing right above the AG gap edge, i.e. $$E_{g}-\omega \approx \Delta _{g}.$$At the same time, the action of Eq.(\[actionme\]) is of order one the distance from the gap becomes of the order of the typical gap fluctuations $$E_{g}-\omega \approx \delta E_{g}.$$It follows that the parameter that roughly determines which one of the two mechanisms dominates is $$\begin{aligned}
\label{beta0}
\beta =\frac{\Delta _{g}}{\delta E_{g}}. \end{aligned}$$ For $\beta \gg 1$ the asymptotic tails are dominated by mesoscopic fluctuations and the result obtained in Ref. \[\] applies. In contrast, when $\beta \ll 1$ fluctuations of the impurity concentration dominate the physics, and Eq.(\[finalzeroDOS\]) describe the asymptotic tails.
In more detail, if $\beta \ll 1$, as the energy $E$ is decreased from $E_{g}$ towards the Fermi level, beyond $E_{g}-E\approx \Delta _{g}$ the mesoscopic gap fluctuations become exponentially rare. However, for $\Delta
_{g}<E_{g}-E<\delta E_{g}$ the system is still in the range of the typical gaussian fluctuations of $E_{g}$ associated to fluctuations of $\zeta $, which dominate the physics. As the energy is decreased further, both asymptotic result are applicable, and to determine which wins, one has to compare the actions. The direct comparison shows that in the range $\delta
E_{g}\ll E_{g}-E\ll \delta E_{g}/\beta ^{3}$ the density of states is dominated by the action associated to optimal fluctuations of $\zeta $. At the crossover point $E_{g}-E\simeq \delta E_{g}/\beta ^{3}$, both actions are of the same order $S\simeq S_{SUSY}\simeq 1/\beta ^{6}\gg 1$ but typically at this point the density of states is negligible. For example, for $\beta =0.3$, at the crossover $S\simeq 10^{3}$, and $exp[-S]$ is practically zero. Therefore, we conclude that the simple rule to determine which mechanism dominates is to compare the typical gap fluctuations to the effective level spacing: the largest wins.
Bulk Optimal fluctuations. {#sec2}
--------------------------
In this section, we study the effect of spatial fluctuations of the spin-scattering rate in relation to subgap localized states. In particular, we will consider a dirty superconductor containing both normal and magnetic impurities, under the following conditions $$\begin{aligned}
\label{conditions}
l \ll \xi \;\;\;\;\;\;\zeta \ll 1\;\;\;\mathrm{and}\;\;\;n_{imp}\xi^d \gg 1,\end{aligned}$$ where $l$ is the mean free paths relative to non-magnetic scattering, $\xi$ is the coherence length, $d$ is the dimensionality of the problem, and $%
n_{imp}$ is the average concentration of magnetic impurities in the system. This set of conditions describe a disordered superconductor containing a relatively large number of weak magnetic impurities.
The condition $l\ll \xi $ implies that the problem can be studied in the framework of the semiclassical approximation [@Kopnin]. In particular, parametrizing both the semiclassical Green’s function $g(r,\epsilon )$ and anomalous Green’s function $f(r,\epsilon )$ in terms of a phase $\theta $ by $$\begin{aligned}
g(\mathbf{r},\epsilon ) &=&\cos [\theta (\mathbf{r},\epsilon )], \\
f(\mathbf{r},\epsilon ) &=&i\sin [\theta (\mathbf{r},\epsilon )],\end{aligned}$$one gets the Usadel equation [@Usadel; @Kopnin] $$\begin{aligned}
\label{Usadel}
&& \nabla ^{2}\theta (\mathbf{r})+i\epsilon \sin [\theta (\mathbf{r})]-\cos
[\theta (\mathbf{r})]-\nonumber \\
&&\zeta \sin [\theta (\mathbf{r})]\cos [\theta (\mathbf{r})] =0. \end{aligned}$$Here the unit length is $\xi =\sqrt{D/2\Delta }$, and $D$ is the diffusion constant, and $\epsilon =E/\Delta $.
In this section, we will study small deviations from the solutions of the * uniform* Usadel equation $$(i\epsilon -\frac{\zeta }{2}\cos (\theta ))\sin (\theta )-(1+\frac{\zeta }{2}%
\sin (\theta ))\cos (\theta )=0. \label{uniformUsadel}$$Let us derive a few well known properties of such solutions, corresponding to the Abrikosov-Gorkov mean field theory [@Abrikosov]. Setting $\theta
=-\pi /2-i\;\mathrm{arctanh}(u)$, we have $$\begin{aligned}
\sin (\theta ) &=&-\frac{1}{\sqrt{1-u^{2}}}, \\
\cos (\theta ) &=&-i\frac{u}{\sqrt{1-u^{2}}}.\end{aligned}$$Therefore, the solution of Eq. (\[uniformUsadel\]) are all those $u$ such that $$\epsilon =u\left( 1-\zeta \frac{1}{\sqrt{1-u^{2}}}\right) .$$From this equation we obtain that the gap edge is at $\epsilon _{0}=(1-\zeta
^{2/3})^{3/2}$. In particular, the density of states per unit volume $$\nu =\nu _{0}\mathrm{Re}[\cos (\theta )],$$is uniform and equal to zero for $\mid \epsilon \mid <\epsilon _{0}$. At the gap edge the parameter $u$ is equal to $u_{0}=(1-\zeta ^{2/3})^{1/2}$. Correspondingly, one has the solution $\theta _{edge}=-\pi /2-i~\;\mathrm{%
arctanh}(u_{0})$.
The Usadel equation is obtained neglecting large scale fluctuations of the impurity concentration. This assumption limits the scattering to be homogeneous in the appropriate long wavelength limit. On the other hand, in a realistic system the distribution of impurities is poissonian. This implies that the number of impurities in any finite volume element fluctuates, with a variance equal to the average.
Following this observation, we introduce an effective theory describing the fluctuations of the concentration of magnetic impurities on length scales of the order, or larger, than the coherence length. In terms of the Usadel equation Eq.(\[Usadel\]) $\zeta $ becomes a position dependent random variable. As a consequence of the Poissonian statistics of the impurities, we have $$\begin{aligned}
\zeta (\mathbf{r}) &=&\zeta +\delta \zeta (\mathbf{r}), \\
\langle \delta \zeta \rangle &=&0, \\
\langle \delta \zeta (\mathbf{r})\delta \zeta (\mathbf{r^{\prime }})\rangle
&=&\left( \frac{\zeta ^{2}}{n_{imp}\;\xi ^{d}}\right) \;\delta (\mathbf{r}-%
\mathbf{r^{\prime }}),\end{aligned}$$where $\zeta $ is the average, uniform, dimensionless, spin-scattering rate, and $d$ is the dimensionality.
Let us look for solutions of Eq.(\[Usadel\]) at energies slightly below the gap ($\epsilon =\epsilon _{0}-\delta \epsilon $) in the form $$\theta (\mathbf{r})=\theta _{edge}-i\phi (\mathbf{r}).$$Expanding Eq.(\[Usadel\]) in $\phi ,\delta \epsilon $, and $\delta \zeta $, in the limit $\zeta \ll 1$ one obtains $$\begin{aligned}
\label{expUsadel1}
\nabla ^{2}\phi +\frac{3}{2}\zeta ^{1/3}\phi ^{2}=\frac{1}{\zeta ^{1/3}}%
\left( \delta \epsilon -\frac{1}{\zeta ^{1/3}}\delta \zeta \right) . \end{aligned}$$ Rescaling again the length in units $\lambda =\xi \;(2/3)^{1/4}$, and defining $\psi =(3/2\;\;\zeta ^{2/3})^{1/2}\phi $, we recast Eq. ([expUsadel1]{}) in the simpler form $$\begin{aligned}
\label{expUsadel}
\nabla ^{2}\psi +\psi ^{2}=\delta \epsilon -f(\mathbf{r}), \end{aligned}$$ where $$f(\mathbf{r})=\frac{\delta \zeta }{\zeta ^{1/3}}.$$In particular, $$\begin{aligned}
\langle f(\mathbf{r})f(\mathbf{r^{\prime }})\rangle &=&\eta \;\delta (%
\mathbf{r}-\mathbf{r^{\prime }}), \nonumber \\
\eta &\equiv &\left( \frac{\zeta ^{4/3}}{n_{imp}\;\xi ^{d}}\left( \frac{3}{2}%
\right) ^{d/4}\right) . \label{eta}\end{aligned}$$
Let us split $\psi $ as $\psi =-x+iy$. Then we have the system $$\begin{aligned}
-\nabla ^{2}x+x^{2}-y^{2} &=&\delta \epsilon -f(\mathbf{r}),
\label{saddlepoint1} \\
-\frac{1}{2}\nabla ^{2}y+x\;y &=&0. \label{saddlepoint2}\end{aligned}$$Interestingly, this set of equations is analogous to the equations obtained by Larkin and Ovchinikov [@Larkin], in the context of the study of gap smearing in inhomogeneous superconductors.
Notice that the analytical properties of the quasi-classical Green’s functions impose some constraints on the solutions of these equations. Indeed, the DOS is $$\begin{aligned}
\label{DOS1}
\nu(\mathbf{r})/\nu_0=\mathrm{Re }[\cos(\theta)] \simeq \sqrt{\frac{2}{%
3~\zeta^{4/3}}}\;y(\mathbf{r}),\end{aligned}$$ where $\nu_0$ is the bare DOS per unit volume at the Fermi level. Therefore, we must have $y(\mathbf{r})>0$.
Our aim is to evaluate the average DOS $\langle \nu (\mathbf{r})\rangle /\nu
_{0}=\sqrt{2/3~\zeta ^{4/3}}\langle y(\mathbf{r})\rangle $ at a distance $%
\delta \epsilon $ below the average gap. In particular, $$\langle y(\mathbf{r})\rangle =\frac{\int \;\mathcal{D}[f(\mathbf{r})]\;y(%
\mathbf{r},\delta \epsilon \mid f(\mathbf{r}))\;e^{-\frac{1}{2\eta }\;\int d%
\mathbf{r}(f(\mathbf{r}))^{2}}}{\int \;\mathcal{D}[f(\mathbf{r})]\;e^{-\frac{%
1}{2\eta }\;\int d\mathbf{r}(f(\mathbf{r}))^{2}}}. \label{<y(r)>}$$
For energies below, but not too close, to the average gap edge, the leading contribution to the DOS comes from exponentially rare fluctuations of $f(%
\mathbf{r})$, describing a local increase of the impurity concentration, and therefore a local suppression of the gap below its average value. In particular, we have to find the *optimal* fluctuation of $f(\mathbf{r}%
)$, i.e. the configuration of $f(\mathbf{r})$ associated to the dominant contribution to $\langle \nu(\mathbf{r}) \rangle$.
In order to be characterized by a finite action, an optimal fluctuation has to be such that $f(\mathbf{r})\rightarrow 0$ as $r=~\mid ~\mathbf{r}%
~\mid \rightarrow \infty $. The asymptotic behavior of any solution of Eqs.(\[saddlepoint1\]-\[saddlepoint2\]) with $f(r)$ describing an optimal fluctuation is therefore $$\begin{aligned}
y(r) &\rightarrow &0\;\;\;\mathrm{as}\;\;\;r\rightarrow +\infty ,
\label{BC1} \\
x(r) &\rightarrow &\sqrt{\delta \epsilon }\;\;\;\mathrm{as}%
\;\;\;r\rightarrow +\infty . \label{BC2}\end{aligned}$$Using Eq.(\[saddlepoint2\]), the second condition implies $$\frac{\nabla ^{2}y}{2y}\rightarrow \sqrt{\delta \epsilon }\;\;\;\mathrm{as}%
\;\;\;r\rightarrow +\infty . \label{BCy}$$Since the system is diffusive, and scattering is isotropic, it is natural to assume a spherically symmetric optimal fluctuation, its scale being $\lambda
$, i.e. $f(r)=f~(~r~/~\lambda ~)$. Then $y(r)=y(r/\lambda )$ which implies $$\frac{\nabla _{r}^{2}y}{2y}=1/\lambda ^{2}\frac{\nabla _{(r/\lambda )}^{2}y}{%
2y}.$$Thus Eq.(\[BCy\]) implies $\lambda \propto (\delta \epsilon )^{-1/4}$.
The problem above has many analogies to the problem of localized states in the Lifshits tails of a disordered system [@Halperin; @Lax; @Zittarz] \[see Appendix A\]. The analogy becomes clear if we substitute $V(r)=x-\sqrt{\delta
\epsilon }$ and rewrite Eq.(\[saddlepoint2\]) as $$\begin{aligned}
\left[ -\frac{1}{2}\nabla ^{2}+V(r)\right] y(r) &=&-\sqrt{\delta \epsilon }%
\;y(r), \label{Sch1} \\
V(r) &\rightarrow &0\;\;\;\;\mathrm{as}\;\;\;\;\;r\rightarrow +\infty .
\nonumber\end{aligned}$$As in the Lifshits tails problem, we are essentially looking for a potential well $V(r)$ that admits a ground state at energy $-\sqrt{\delta \epsilon }$. A ground state condition is important because the nodes in the “wave function” $y(r)$ are not consistent with the analytical properties of the quasi-classical Green’s functions. Despite the close analogy, it is important to notice two differences with the Lifshits tails problem: *(i)* our wave function $y(r)$ is a component of the Usadel phase directly proportional to the DOS, and * (ii)* the statistical weight of the potential in the standard Lifshits tails problem is $S\propto \int
dr\;(V(r))^{2}$ while here it is $$\begin{aligned}
&& \nonumber \\
S &=&\frac{1}{2\eta }\int d\mathbf{r}\;(f(r))^{2} \nonumber \\
&=&\int \frac{1}{2\eta }d\mathbf{r}\;(y^{2}+\nabla ^{2}V-V^{2}-2\sqrt{\delta
\epsilon }V)^{2}, \label{S}\end{aligned}$$where we used Eq.(\[saddlepoint1\]) to express $f(r)$ in terms of $V(r)$ and $y$.
With the exponential accuracy one can neglect the $y^{2}$ term in Eq.(\[S\]) because the integration over the fluctuations near the saddle point characterized by Eq.(\[Sch1\]) leads to a non-zero contribution to the density of states Eq.(\[<y(r)>\]) \[see Appendix B\]. To find the parametric dependence of the action, we note that the Schrödinger equation Eq.(\[Sch1\]) implies that the optimal fluctuation has a scale $\lambda \propto (\delta
\epsilon )^{-1/4}$ and an amplitude $V_{0}(r)\propto -\sqrt{\delta \epsilon }$ . Writing $V=\sqrt{\delta \epsilon }\;v(r/\lambda )$ we get $$\begin{aligned}
S &\simeq &\frac{1}{2\eta }\int d\mathbf{r}\left( \nabla ^{2}V-V^{2}-2\sqrt{%
\delta \epsilon }V\right) ^{2} \nonumber \\
&=&\frac{1}{2\eta }(\delta \epsilon )^{2-d/4}\int d(\mathbf{r}/\lambda
)\left( \nabla ^{2}v-v^{2}-2v\right) ^{2} \nonumber \\
&=&\mathrm{const}\frac{n_{imp}\xi ^{d}}{\zeta ^{4/3}}(\delta \epsilon
)^{2-d/4}, \label{S_scale}\end{aligned}$$where $\eta $ is given by Eq.(\[eta\]).
We now find the exact shape of the optimal fluctuation and the resulting numerical constant in Eq.(\[S\_scale\]) from the Euler-Langrange equations. We have to look for a the most probable $f(r)$ which when inserted in Eq.([saddlepoint1]{}-\[saddlepoint2\]), produces a solution $y(r)$, positive everywhere, and such that $\nabla ^{2}y/(2y)\rightarrow \sqrt{\delta
\epsilon }$ at infinity. This general strategy can be applied to the classical problem of Lifshits tails, as explained in Appendix.A.
Expressing $f(r)$ in terms of $x$ and $y$, we get the dimensionless action $$\begin{aligned}
S &=&\frac{1}{2\eta }\int d\mathbf{r}\left[ \nabla ^{2}x-x^{2}+\delta
\epsilon \right] ^{2}, \label{actionmin} \\
x &=&\frac{\nabla ^{2}y}{2y},\end{aligned}$$where $d$ is the dimensionality. This action admits a trivial stationary point corresponding to zero action or $f(r)=0$ in Eqs (\[saddlepoint1\]-\[saddlepoint2\]) characterized by$$x^{\prime \prime}+\frac{d-1}{r}\;x^{\prime}-x^{2}+\delta \epsilon =0,
\label{trivialsaddle}$$where the prime indicates the derivative with respect to $r$. It is possible to show that this equation admits instanton solutions. However, these solutions have either $y(r)=0$ at every point in space, or $%
y(r)<0$ somewhere. The analytic properties of quasi-classical Green’s functions indicate that the latter solutions are not allowed. On the other hand, the solutions with $y(r)=0$ do not contribute to the DOS in the present approximation scheme. However, these solutions become meaningful in the Sigma model approach to this problem; there they describe the saddle point corresponding to mesoscopic fluctuations generating subgap states [Lamacraft, Ostrovsky, Dima]{}.
Let us now find the equations associated with the nontrivial saddle points. In spherical coordinates, the Euler-Lagrange equation corresponding to the action, Eq.(\[actionmin\]), is $$\begin{aligned}
&&\partial _{r}^{2}\left( r^{d-1}\;\mathcal{A}\right) -(d-1)\;\partial
_{r}\left( r^{d-2}\;\mathcal{A}\right) \nonumber \\
&&-2\;x\;r^{d-1}\mathcal{A}=0, \label{NonTrivialSaddle} \\
&&\mathcal{A}[x,x^{\prime },x^{\prime \prime },r]=x^{\prime \prime }+\frac{%
d-1}{r}x^{\prime }-x^{2}+\delta \epsilon .\end{aligned}$$The problem of finding the instanton solutions of this equation describing the optimal fluctuations is considerably simplified by the observation that, in all dimensions \[$d=1,2,3$\], the solutions of $$\mathcal{A}[x,x^{\prime },x^{\prime \prime },r]=\frac{2\;x^{\prime }}{r},
\label{Ansatz}$$are also solutions of Eq.(\[NonTrivialSaddle\]). Therefore, the nontrivial saddle points [@foot] are described by the system $$\begin{aligned}
x^{\prime \prime }+\frac{d-3}{r}\;x^{\prime }-x^{2}+\delta \epsilon &=&0,
\label{EL1} \\
y^{\prime \prime }+\frac{d-1}{r}\;y^{\prime } &=&x\;y. \label{EL2}\end{aligned}$$It is interesting to notice that the equation above for the variable $x(r)$ coincides with the equation describing the trivial saddle point, Eq.([trivialsaddle]{}), in $d-2$ dimensions. In three dimensions we get $$\begin{aligned}
x_{0} &=&\sqrt{\delta \epsilon }\left( 1-3\;(\mathrm{sech}\left[ r/\lambda %
\right] )^{2}\right) , \\
y_{0} &\propto &\;\frac{1}{r/\lambda }\tanh \left[ r/\lambda \right] (%
\mathrm{sech}\left[ r/\lambda \right] )^{2}, \\
\lambda &=&\left( \frac{\delta \epsilon }{4}\right) ^{-1/4}.\end{aligned}$$The optimal fluctuation of the dimensionless spin-flip scattering rate can be then evaluated using $$f_{0}(r)=\delta \epsilon -\left( \frac{\nabla ^{2}y_{0}}{2y_{0}}\right)
^{2}+\nabla ^{2}\left( \frac{\nabla ^{2}y_{0}}{2y_{0}}\right)$$Thus, in three dimensions, the action at the nontrivial saddle point is $$\begin{aligned}
S &=&\frac{4\pi }{2\eta }\left( \frac{\delta \epsilon }{4}\right) ^{5/4}%
\frac{384}{5} \nonumber \label{3d} \\
&\simeq &63\left[ \frac{n_{imp}\;\xi ^{3}}{\zeta ^{4/3}}\right] \delta
\epsilon ^{5/4}.\end{aligned}$$
In one and two dimensions the solution of Eq.(\[EL1\]-\[EL2\]) is not so straightforward. In order to estimate the action of the optimal fluctuation, we had to minimize the action $S$ in terms of a variational ansatz $%
y_{a,b}(r)$ depending on two free parameters. The asymptotic form of any variational ansatz $y_{a,b}(r)$ must be $$\begin{aligned}
y(r) &\rightarrow &\frac{e^{2r/\lambda }}{\sqrt{r/\lambda }}\;\;\;\;\;\;(d=2)
\\
y(r) &\rightarrow &e^{2r/\lambda }\;\;\;\;\;\;(d=1)\end{aligned}$$in order to obtain to a finite action. The result of this calculation is $$\begin{aligned}
S &\simeq &14\left[ \frac{n_{imp}\;\xi ^{2}}{\zeta ^{4/3}}\right] \delta
\epsilon ^{3/2}\;\;\;\;(d=2) \label{2d} \\
S &\simeq &2.8\left[ \frac{n_{imp}\;\xi }{\zeta ^{4/3}}\right] \delta
\epsilon ^{7/4}\;\;\;\;(d=1) \label{1d}\end{aligned}$$
Clearly $\langle y(r)\rangle \propto \exp [-S]$. This equality must be supplemented by the calculation of the prefactor, i.e. gaussian fluctuations, using the standard technique due to Zittarz and Langer [Zittarz]{}. This calculation is reported in some detail in Appendix.B. The result is $$\frac{\langle \nu \rangle }{\nu _{0}}\simeq \tilde{b}_{d}\frac{\sqrt{%
n_{imp}\xi ^{d}}}{\zeta ^{4/3}}\delta \epsilon ^{\alpha _{d}}\exp \left[ -%
\tilde{a}_{d}\frac{n_{imp}\xi ^{d}}{\zeta ^{4/3}}\delta \epsilon ^{2-d/4}%
\right] \label{finalDOS}$$where the exponent is $\alpha _{d}=1/8(d(10-d)-12)$. The values of the numerical coefficient of the action $\tilde{a}_{d}$, as well as the estimated value of the numerical constants $\tilde{b}_{d}$ according to the calculation reported in Appendix.B, are summarized in the following table
$$\begin{aligned}
\begin{array}{|c||c|c|}
\hline
\mathrm{dimensions} & \;\;\;\;\;\;\;\tilde{a}_d\;\;\;\;\;\;\; &
\;\;\;\;\;\;\;\tilde{b}_d\;\;\;\;\;\;\; \\ \hline\hline
1 & 2.8 & 0.0007 \\ \hline
2 & 14 & 0.006 \\ \hline
3 & 63 & 0.2 \\ \hline
\end{array}
\nonumber\end{aligned}$$
First of all, the result of Eq.(\[finalDOS\]) is asymptotic, and requires $%
S \gg 1$. This condition translates into $\delta\epsilon \gg
\delta\epsilon_0 $, where $$\begin{aligned}
\label{limit}
\delta\epsilon_d \equiv \left[\frac{1}{\tilde{a}_{d}} \frac{\zeta^{4/3}}{
n_{imp}\xi^d} \right] ^{\frac{4}{8-d}}.\end{aligned}$$ Since we assumed $\zeta \ll 1$, $n_{imp}\xi^d \gg 1$ , one has $%
\delta\epsilon_0 \ll 1$.
As expected, the final result given by (\[finalDOS\]) crosses over to the result of the zero dimensional calculation performed in Sec. \[sec1\] in the limit $d\rightarrow 0$. The relation between the $d\neq 0$ and $d=0$ results becomes simpler when we rewrite Eq.(\[finalDOS\]) as $$\frac{\langle \nu \rangle }{\nu _{0}}\propto \exp \left[ -\tilde{a}%
_{d}\left( \frac{\lambda _{0}}{L}\right) ^{d}\left( \frac{E_{g}-E}{\delta
E_{g}}\right) ^{2}\right] , \label{compactDOS}$$where $\lambda _{0}=\xi /(\delta \epsilon )^{1/4}$ \[see Eq.(\[size\])\]. Though this way of writing the result makes the $d\rightarrow 0$ limit transparent, it should be kept in mind that the action in finite dimensional systems does not depend on the sample size $L$ but only on intensive quantities.
Let us now compare the asymptotic tails resulting from fluctuations of $%
\zeta $, Eq.(\[finalDOS\]) with those associated to mesoscopic gap fluctuations [@Lamacraft]. The latter give an asymptotic DOS $$\begin{aligned}
\frac{\langle \nu \rangle }{\nu _{0}} &\propto &\exp \left[ -S_{SUSY}\right]
, \nonumber \label{finalDOSSUSY} \\
S_{SUSY} &=&a_{d}\left( \frac{4\pi \nu _{0}\Delta \xi ^{d}}{\zeta ^{2/3}}%
\right) (\delta \epsilon )^{3/2-d/4},\end{aligned}$$where $a_{d}$ is a numerical constant. In one dimension it is given by $%
a_{1}\approx 8^{4}\sqrt{24}/5\approx 4000$. Introducing $\lambda _{0}=\xi
/(\delta \epsilon )^{1/4}$, one may write $$S_{SUSY}\simeq a_{d}\left( \frac{\lambda _{0}}{L}\right) ^{d}\left( \frac{%
E_{g}-E}{\Delta _{g}}\right) ^{3/2},$$with $\Delta _{g}$ given by Eq.(\[levspacing\]). From this equation we conclude that the ratio between this action and the action of the optimal fluctuations of $\zeta $ \[Eq.(\[compactDOS\])\] is independent on dimensionality (apart from a numerical prefactor). This is a direct consequence of the fact that the typical linear sizes of the optimal fluctuations associated to the two mechanisms are identical.
In order to establish which fluctuation dominates the physics of the subgap tails, we first notice that $S_{SUSY}$ is of order of unity at $\delta
\epsilon \simeq \delta \epsilon _{d}^{\prime }$, where $$\delta \epsilon _{d}^{\prime }\equiv \left( \frac{1}{a_{d}}\frac{\zeta ^{2/3}%
}{4\pi \nu _{0}\Delta \xi ^{d}}\right) ^{\frac{4}{6-d}}.$$Therefore, in a finite dimensional system, the parameter that determines which one of the two mechanisms dominates is $$\beta _{d}\equiv \frac{\delta \epsilon _{d}^{\prime }}{\delta \epsilon _{d}}.
\label{betad}$$
As in the zero dimensional case, when $\beta _{d}\gg 1$ the subgap tails are dominated by mesoscopic gap fluctuations, and the asymptotics of the subgap tails is described by Eq.(\[finalDOSSUSY\]). In contrast, when $\beta
_{d}\ll 1$ the physics is dominated by the fluctuations of the concentration of magnetic impurities, and Eq.(\[finalDOS\]) applies. More precisely, as the energy is lowered from the gap edge, i.e. as $\delta \epsilon $ increases, first the asymptotic result relative to mesoscopic gap fluctuations starts being applicable beyond $\delta \epsilon \simeq \delta
\epsilon _{d}^{\prime }$. However, for $\delta \epsilon _{d}^{\prime
}<\delta \epsilon <\delta \epsilon _{d}$ the system is within the range of typical fluctuations of $\zeta $, which dominate over the exponential tails associated to mesoscopic gap fluctuations. Increasing $\delta \epsilon $ beyond $\delta \epsilon _{d}$ both asymptotic results are applicable and the two actions should be compared. In particular, $$\frac{S}{S_{SUSY}}=\frac{(\delta \epsilon _{d}^{\prime })^{\frac{6-d}{4}}}{%
(\delta \epsilon _{d})^{\frac{8-d}{4}}}\;\delta \epsilon ^{1/2},$$implying that for $$\delta \epsilon _{d}\ll \delta \epsilon \ll \delta \epsilon \left( \frac{1}{%
\beta _{d}}\right) ^{\frac{6-d}{2}},$$the asymptotic tails are dominated by fluctuations of $\zeta $. At the crossover, i.e. $\delta \epsilon \simeq (\delta \epsilon _{d})^{\frac{8-d}{2}%
}/(\delta \epsilon _{d}^{\prime })^{\frac{6-d}{2}}$, both actions are $$S\simeq S_{SUSY}\simeq \left( \frac{1}{\beta _{d}}\right) ^{\frac{(6-d)(8-d)%
}{8}}>>>1.$$This implies that, at the crossover, the density of states is already negligible. For example, for $d=2$ and at the crossover point, $S\simeq
(1/\beta _{d})^{3}$. Taking, $\beta _{d}=0.1$, one obtains at the crossover $%
S\approx 10^{3}$, implying that $\exp (-S)$ is practically zero.
Shallow quasiparticle traps in superconducting Qubits.
======================================================
\[sec2b\]
In the previous sections we found that a low concentration of the paramagnetic impurities lead to a small but finite density of the localized subgap quasiparticle states. In this section, we discuss the effect of these states on small superconducting devices, especially the ones proposed for quantum computation. Because the density of these localized states is always very low, they do not affect much the macroscopic properties such as specific heat or Josephson current of big junctions. However, even a small number of quasiparticles trapped in these states might affect the long time phase coherence in the Cooper pair boxes [@Nakamura],[Devoret,Martinis2]{} or flux qubits[@Mooji], so we shall focus on such devices here.
The building blocks of these circuits are typically $\mathrm{%
Al-Al_{2}O_{3}-Al}$ or $\mathrm{Nb-Al_{2}O_{3}-Nb}$ Josephson junctions. When quasiparticles are generated, say, during the readout process, most of them quickly recombine and disappear in the condensate. In order to eliminate the remaining quasiparticles, superconducting circuits incorporate normal metal leads acting as quasiparticle traps [@Devoret; @Martinis1]. At low temperatures, these leads can be viewed as the sinks for all the quasiparticles that may diffuse to them. However, a quasiparticle localized within a superconducting electrode cannot diffuse to the quasiparticle traps; instead it remains localized inside the superconducting circuit and represents an active degree of freedom that may contribute to decoherence in subsequent operations of the qubit \[see Fig. \[fig:fig1\]\].
We begin with the estimate of the number of localized states in a typical experimental setup. We focus on aluminum based qubits. The typical parameters for aluminum are $\Delta =200\;\mathrm{\mu eV}$, $D=50\;\mathrm{%
cm^{2}/s}$, $v_{F}=2 \times 10^{8}\;\mathrm{cm/s}$, and $\nu _{0}\approx
1.5 \times 10^{22}\;\mathrm{cm^{-3}\;eV^{-1}}$. This implies that the coherence length is $\xi \approx 0.1\;\mathrm{\mu m}$, while the mean free path is $%
l\approx 0.01\;\mathrm{\mu m}\ll \xi $. Therefore, the system is diffusive, and the first condition for the applicability of the results obtained in previous sections is satisfied \[see Eq.(\[conditions\])\]. Further, because the thickness of the samples is about or less than the coherence length, they are effectively two dimensional.
The strength of the spin-flip scattering and the number of paramagnetic impurities is more difficult to estimate reliably. The problem is that for a typical sample of thickness $l\approx 0.1\mu m$ surface scattering is rather important and it is very likely that there it contains a significant spin-flip contribution. Indeed, the surface of the $Al$ devices is covered by a thin layer of glassy $Al_{2}O_{3}$. A typical glass has a concentration of $n_{TLS}\sim 10^{7}-10^{8}\mu m^{-3}$ of two level systems which are usually attributed to the trapped electrons [@Burin1995]. This translates into a surface density of $10^{3}-10^{4}\mu m^{-2}$ of free spins in the boundary layer of $Al_{2}O_{3}$ that interact with the electrons in the metal or into the effective density $10^{4}-10^{5}\mu
m^{-3} $ per unit volume of the device. Because this estimate counts only the spins associated with the two level systems in a glass, we expect that the actual number is somewhat bigger. On the other hand, assuming that the strength of the spin-spin interaction $J\approx 0.1\;eV$ and its range $%
r\approx 4\;\mathring{A}$ we get that the volume density $n_{sf}$**$%
\simeq $**$10^{5}\mu m^{-3}$ leads to a spin-scattering time of **$\tau
_{s}\approx $**$10^{-8}\;s~$ which is the lower bound given by the weak localization magnetoresistance data [@Gershenson]. Thus, we believe that the realistic estimate for the effective density of the paramagnetic impurities in the wires of the $0.1\mu m$ thickness is about $%
n_{sf}$**$\simeq $**$10^{5}\mu m^{-3}$.
For this value of the the paramater $\beta $ \[see Eq. (\[betad\])\] is very small: $\beta _{2}\approx 10^{-5}\ll 1$. Therefore, the asymptotics described by Eq.(\[finalDOS\]) applies. Using now Eq.([finalDOS]{}) for the density of states we estimate the total number of the localized states per unit area in these conditions: $\rho \sim 1-10\;\mu
m^{-3}$. This means that for the large junctions of the area $A\approx
10\times 10\;\mu m^{2}$, one gets about $N_{sub}\approx 10^{2}-10^{3}$ localized states on each superconducting electrode. The length scale associated to these states is about $1\mu m$. With the parameters above, we expect these localized states to be very shallow, typically $0.01\;\mu eV$ below the gap edge. This energy scale is two orders of magnitude smaller than the typical temperature in qubit experiments, $T\simeq 10\;mK\simeq
1\;\mu eV/k_{B}$. This implies that if a non-equilibrium quasiparticle generated in the duty cycle of the qubit gets trapped into one of these localized states, it will eventually absorb a phonon, get excited above the mobility edge, and diffuse out of the sample. However, at such low temperatures, the time scale $\tau _{tr}$ associated to these phonon-assisted detrapping processes is very long, $\tau _{tr}\approx
100s\;\;$ [@tau; @Lawrence] . At the same time, in a superconducting qubit composed of large junctions of area $10\times 10\;\mu m^{2}$ and operating at a frequency $f=10\;GHz$ the qubit relaxation rate associated to the coupling between the phase degrees of freedom and a single localized quasiparticle is $\Gamma _{qp} \approx 10^{3}\;Hz$ \[see Appendix C\]. Therefore, the presence of $10^{2}-10^{3}$ localized quasiparticles interacting with the phase degrees of freedom across the junction is expected to limit the quality factor to be $Q=\nu _{0}/\Gamma _{qp}\approx 10^{4}-10^{5}$ . It is important to mention that for smaller systems, e.g. of area $A\approx
1\times 1\;\mu m^{2}$, the probability of having a single localized subgap state of area considerably smaller than $A$ is negligible. The subgap tails are effectively zero dimensional and the tail states extend to the whole system. So, an electron trapped in such state can always escape to the lead through the Josephson contact.
Conclusions.
============
\[sec3\]
We have shown that in a diffusive superconductor containing magnetic impurities at small concentrations \[$\zeta <1$\], the subgap tails of the DOS have two distinct contributions, one due to the universal mesoscopic gap fluctuations [@Lamacraft], the second resulting from inhomogeneous fluctuations of the spin scattering time $\zeta $.
We calculated the contribution to the density of states due to the fluctuations of $\zeta $ \[see Eq.(\[finalzeroDOS\]) for $d=0$ and Eq.([finalDOS]{}) for $d>0$\] and established, through a direct comparison of the results, the parameter that controls which of the two mechanisms dominates the physics of subgap states \[see Eq.(\[beta0\])-(\[betad\])\]. The two mechanism are related by dimensional reduction, both at the level of instanton equations and in the dependence of the actions on $\delta\epsilon$. The deep reason behind the occurrence of dimensional reduction in this context is presently unclear, and requires further study.
On the theoretical side, we believe that the mechanism generating subgap localized states studied in this work is not limited to the dirty superconductors containing paramagnetic impurities. For example, our study is clearly related to the study of inhomogeneous superconductors performed in a seminal paper by Larkin and Ovchinikov, Ref.\[\]. In particular, in three dimensions they obtained result similar to ([finalDOS]{}). It is interesting to notice that a recent solution of the same problem by the replica Sigma model approach [@Meyer] gave the same asymptotic behavior as the mesoscopic gap fluctuations \[Eq.(\[Lama\])\]. The source of this discrepancy was previously attributed to the impossibility to construct in the present context a Lifshits argument [Meyer]{} that was used in the estimate of the action of the optimal fluctuation in Ref.\[\]. In contrast, we have argued that both results correctly describe two different physical effects that work in parallel: mesoscopic Random-Matrix like gap fluctuations [Meyer]{} and long wavelength fluctuations of the coarse grained gap [Larkin]{}. In analogy with the present study, a direct comparison of the two resulting actions determines which mechanism dominates the asymptotics of the density of states.
Finally, we applied the theoretical results summarized above to estimate the effects of a weak concentration of magnetic impurities \[$\approx 1\;\mathrm{%
ppm}$, corresponding to $\zeta \approx 10^{-4}$\] in Aluminum-based Qubits and Josephson-junction arrays \[see Fig.(\[fig:fig1\])\]. Our estimates indicate that the presence of a small concentration of magnetic impurities is mostly relevant for large qubits, designed around the $10\;{\mu m}$ scale. Under the conditions above, Cooper pair boxes of the size $10\times 10\times 0.1\;%
\mathrm{\mu m^{3}}$ are expected to have an average of $\approx 10^{2}-10^{3}
$ two dimensional subgap localized states per electrode. At a base temperature of $10\;\mathrm{mK}$, the lifetime of a non-equilibrium quasiparticle localized in one of such states is expected to be as long as $%
100s$. We estimate that this would limit that quality factor of the qubit operating at the frequency $f=10Ghz$ by $Q\approx 10^{4}-10^{5}$.
Acknowledgments.
================
We would like to thank also N. Andrei, M. Feigel’man, Y. Gefen, A. Lamacraft, A. Schiller, and especially B. Altshuler, M. Gershenson, E. Lebanon and M. Müller for discussions. This work is supported by NSF grant DMR 0210575.
Analogies to the problem of band tails in disordered conductors.
================================================================
In the following, we illustrate how the variational reasoning employed in the main text works in the context of the problem of Lifshits tails, and how the exact solution in one dimension for the leading exponential dependence of the DOS tails relative to this problem is recovered [Halperin,Lax,Zittarz]{}.
Consider the Schrödinger equation $$\begin{aligned}
\label{SchLif}
\left[-\frac{\nabla^2}{2}+V(r)\right]\;y(r)=-E\;y(r),\end{aligned}$$ where $V(r)$ is a white noise potential $$\begin{aligned}
\langle V(r)V(r^{\prime}) \rangle=\gamma\;\delta(r-r^{\prime}).\end{aligned}$$ Well below the band edge ($E=0$), the leading contribution to the DOS associated to Eq.(\[SchLif\]) comes from rare local fluctuations of $V(r)$ able to generate bound states. In order to find the leading exponential dependence of the average DOS as a function of energy, we have to find the most probable potential $V(r \mid E)$ admitting a bound state at energy $-E$, i.e the optimal fluctuation. The average DOS is then $$\begin{aligned}
\langle \rho \rangle &\propto& e^{-S}, \nonumber \\
S &=& \frac{1}{2\gamma}\int \;dr\;(V(r \mid E))^2.\end{aligned}$$
Let us now look for a ground state at energy $-E$, i.e. assume that the wave function $y(r)$ does not have nodes \[and can therefore be chosen to be positive\]. For every given $y(r)$, we can use the Schrödinger equation to obtain $$\begin{aligned}
\label{Sch4}
V(r)=\frac{\nabla^2 y}{2y}-E.\end{aligned}$$ This equation specifies for every positive, non-vanishing, and smooth $y(r)$, the potential well $V(r)$ that admits it as a ground state of energy $-E$. Notice that $y(r)$ must tend to $0$ as $\mid r \mid \rightarrow +\infty$. The same is true for $V(r)$, otherwise its action would be infinite. Thus $$\begin{aligned}
\frac{\nabla^2 y}{2y}\rightarrow E\;\;\;\mathrm{as}\;\;\;\mid r
\mid\rightarrow +\infty.\end{aligned}$$
Consider now the functional $$\begin{aligned}
I[V(r)]=\int \;dr\;(V(r))^2.\end{aligned}$$ Let use Eq.(\[Sch4\]) to rewrite this functional as a functional of $y(r)$. One obtains $$\begin{aligned}
\label{cipolla}
I[y(r)]=\int \;dr\;\left(\frac{\nabla^2 y}{2y}-E\right)^2.\end{aligned}$$ The problem of finding the most probable $V(r)$ admitting a ground state at energy $-E$ is now reduce to the problem of finding the stationary points of $I[y(r)]$. Since the argument of the integral is squared, one might be tempted to find the most obvious stationary point by setting $$\begin{aligned}
\label{ippo2}
\frac{\nabla^2 y}{2y}-E=0.\end{aligned}$$ This procedure is obviously incorrect, since it corresponds to the search a bound state in a flat potential landscape $V(r)=0$. In mathematical terms, Eq.(\[ippo2\]) does not admit nontrivial solutions satisfying the appropriate boundary conditions introduced above.
Therefore, let us go back to Eq.(\[cipolla\]) and find the nontrivial saddle points. This program can be definitely completed in one dimension. First of all set $$\begin{aligned}
f(r)= \frac{\dot{y}}{y}\end{aligned}$$ Obviously, $$\begin{aligned}
\frac{\ddot{y}}{y}=\dot{f}+f^2\end{aligned}$$ Let us now rewrite the functional in Eq.(\[cipolla\]) in terms of $f(r)$. We obtain $$\begin{aligned}
I[f(r)]=\frac{1}{2}\int\;dr\;\left(\dot{f}+f^2-2E\right)^2\end{aligned}$$ Using the Euler-Lagrange equation, one may show that the nontrivial saddle point is given by the solution of $$\begin{aligned}
\ddot{f}-2f^3+4\;E\;f=0\end{aligned}$$ This equation describes the motion in an inverted double well potential with maxima at $f_{\pm}=\pm\sqrt{2E}$. The corresponding instanton is a trajectory going from $f_{+}$ to $f_{-}$. Integrating this equation one obtains $$\begin{aligned}
f(r)=-\sqrt{2E}\tanh(\sqrt{2E}r)\end{aligned}$$ Correspondingly we find $$\begin{aligned}
y_0(r) \propto \mathrm{sech}(\sqrt{2E}x)\end{aligned}$$ Finally let us evaluate the action corresponding to the potential relative to $y_0(r)$. We have to calculate $S=1/(2\gamma)I[y_0(r)]$. Inserting the expression for $y_0$ and evaluating the integral one obtains $$\begin{aligned}
S=\frac{2}{3}\frac{1}{\gamma}(2E)^{3/2}\end{aligned}$$ This is corresponds to the action obtained by Halperin in its exact solution of the 1d problem, Ref.\[\].
Gaussian fluctuations.
======================
In this appendix we outline the calculation of the full expression of the subgap tails of the DOS, including gaussian fluctuations.
For every function $f(r)$, we select the center of the $r^{\prime}$ of the optimal fluctuation in such a way as to minimize the functional $$\begin{aligned}
\label{functional}
D(f\mid r^{\prime})=\int dr \left[f(r)-f_0(r-r^{\prime}) \right].\end{aligned}$$ Of all the solutions, we select the $r^{\prime}$ closest to $r$, the point at which the DOS is evaluated. Subsequently, the function $f(r)$ is expanded in an complete set $$\begin{aligned}
\label{expansion}
f(r)=\sum_{n}^{} \xi_n \phi_n(r),\end{aligned}$$ where the first $d+1$ functions of the expansion are selected according to $$\begin{aligned}
\phi_0(r)&=&a\;f_0(r-r^{\prime}), \\
\mathbf{\phi}_1(r)&=&b \nabla f_0(r-r^{\prime}).\end{aligned}$$ Here the constant $a$ and $b$ are given by $$\begin{aligned}
\label{a}
a&=& \left[ \int dr (f_0(r))^2 \right]^{1/2}, \\
b&=& \left[\frac{1}{d} \int dr (\nabla f_0(r))^2 \right]^{-1/2}. \label{b}\end{aligned}$$
Having in mind this expansion, we write $$\begin{aligned}
&&\langle y(r) \rangle= \langle \int_{\sigma} dr^{\prime\prime} y(f(r))
\delta(r^{\prime\prime}-r^{\prime})\rangle \nonumber \\
&&= \langle \int_{\sigma} dr^{\prime\prime} y(f(r)) \delta(\nabla D(f\mid
r^{\prime\prime})) \mid \mathrm{det}\nabla \nabla D \mid \rangle,\end{aligned}$$ where $\sigma$ is a region of space containing at most a single optimal fluctuation, $y(f(r))$ the appropriate solution of Eq.(\[saddlepoint1\])-(\[saddlepoint2\]) relative to $f(r)$, and in the last line we operated a change of variables from $r^{\prime\prime}$ to $D$.
Using Eq.(\[functional\])-(\[expansion\]) we can estimate $$\begin{aligned}
\nabla D&=& \frac{2 \mathbf{\xi}_1}{b}, \\
\mid {\det \nabla \nabla D} \mid &\simeq& \frac{2^d}{b^{2d}}.\end{aligned}$$ At the same time, while to lowest order $y(f(r)) \propto y_0(r-r^{\prime})$, the coefficient of proportionality is set by fluctuations of $f$ around $f_0$. Using Eq.(\[saddlepoint1\]), Eq.(\[saddlepoint2\]), together with Eq.(\[expansion\]), one obtains $$\begin{aligned}
y(f(r)) &\approx& A\;y_0(r), \\
A&=& \sqrt{\frac{\xi_0-a^{-1}}{\int dr \phi_0(r)[y_0(r)]^2}}.\end{aligned}$$
Now using the results above, it is easy to evaluate $$\begin{aligned}
\label{finalalmost}
&&\langle y \rangle \simeq \frac{1}{b^d} \langle \int dr^{\prime\prime}
y(f(r)) \delta(\mathbf{\xi}_1) \rangle \nonumber \\
&&= \frac{1}{(2\pi \eta)^{3/2}\;b^d}\int dr^{\prime\prime} \frac{d\xi_0}{%
(2\pi \eta)^{1/2}} \;A\;y_0 e^{-\frac{(\xi_0)^2}{2\eta}} \nonumber \\
&&= \frac{I_0}{(2\pi \eta)^{2}}\int_{a^{-1}}^{+\infty} d\xi_0 \sqrt{\xi_0
-a^{-1}}\;e^{-\frac{(\xi_0)^2}{2\eta}},\end{aligned}$$ where $$\begin{aligned}
\label{I0}
I_0=\frac{\int dr y_0(r)}{\sqrt{\int dr \phi_0(r) [y_0(r)]^2}}\;\frac{1}{b^d}%
.\end{aligned}$$ The last object we have to evaluate is the integral describing the contribution to the DOS due to gaussian fluctuations of the height of the optimal fluctuation, i.e. $$\begin{aligned}
\label{I}
I&=&\int_{a^{-1}}^{+\infty} d\xi_0 \sqrt{\xi_0 -a^{-1}}\;e^{-\frac{(\xi_0)^2%
}{2\eta}} \nonumber \\
&=&\eta^{3/4}\Gamma(3/2) \exp[-S/2]\;D_{-3/4}(\sqrt{2\;S}) \nonumber \\
&\simeq& \eta^{3/4}\Gamma(3/2)\left(\frac{1}{2S}\right)^{3/4} e^{-S}\end{aligned}$$ where $S$ is the action at the saddle point calculated before, $D_{-3/4}$ is a parabolic cylinder function, and the last line is the lowest order asymptotic expansion for $S \gg 1$.
Using now Eq.(\[finalalmost\])-(\[I0\])-(\[I\]), and Eq.(\[DOS1\]) one obtains the final result Eq.(\[finalDOS\]).
Estimate of the Qubit relaxation rate
=====================================
\[appc\]
In this Appendix, we estimate roughly the Qubit relaxation rate due to the interaction between a single localized quasiparticle and the phase across the Josephson junction. In the following, we will have in mind large Josephson junctions of area $A\approx 10\times 10\;\mathrm{\mu m}$ \[see Fig. (\[fig:fig1\])\] with electrodes of thickness $l_{z}\approx 0.1\;{\mu m}$, operating at a linear frequency of $f=10\;\mathrm{GHz}$, corresponding to a circular frequency of $\omega _{0}\approx 6\;10^{10}\;\mathrm{s^{-1}}$.
Let us consider the phase $\varphi$ across the junction evolving in time according to $\varphi(t)=\cos(\omega_0 t)$. Correspondingly, the voltage $V$ across the junction will evolve as $V(t)=\hbar/2e\;\dot{\varphi}$. For typical junctions, having oxide layers of thickness $~\approx 10\;\mathring{A%
}$, the potential drop is located mostly in a thin layer of width $d\approx
1\;\mathring{A}$ within the superconducting electrodes. The rate at which a quasiparticle localized on a region of linear size $\lambda \approx 1\;{\mu m%
}$ is excited due its interaction with this time dependent potential, can be estimated as $$\begin{aligned}
\Gamma_{qp}\simeq \frac{2\pi }{\hbar} \sum_{f} \mid \int dr \psi_{i}(r)\;
eV(r)\; \psi^{*}_{f}(r) \mid^2 \nonumber \\
\delta(\epsilon_{f}-\epsilon_{i}-\hbar \omega_0),\end{aligned}$$ where $\psi_{i,f}$ and $\epsilon_{i,f}$ are the wave functions and energies of the initial and final state respectively. Using the fact that the interaction is mostly concentrated close to the surface $S$ of the electrode, we can write $$\begin{aligned}
~ \Gamma_{qp} \simeq \frac{\pi}{2\hbar} (\hbar\omega_0)^2 \left(\frac{d}{l_z}%
\right)^2\;\gamma(\omega_0), \label{gammaqp}\end{aligned}$$ where $$\begin{aligned}
\gamma \simeq l_z^2 \sum_{f} \mid \int_{S} dr \psi_{i}(r) \psi^{*}_{f}(r)
\mid^2 \delta(\epsilon_{f}-\epsilon_{i} - \hbar \omega_0).\end{aligned}$$ Since the support of the wave function of the initial state is concentrated on an area $A \simeq \lambda\times\lambda \approx 1 \times 1\;{\mu m^2}$, $%
\gamma$ can be roughly estimated having in mind a three dimensional diffusive quantum dot of dimensions $\lambda \times \lambda \times l_z$ as $$\begin{aligned}
\label{gamma}
\gamma \approx \frac{ l_z }{ 4\pi^2 \nu_0\;\lambda^2}\;\int_S\;dr dr^{\prime}%
\left[G(r,r^{\prime},\epsilon)\right]_{-} \left[G(r^{\prime},r,\epsilon+%
\hbar\omega_0)\right]_{-},\end{aligned}$$ where $\nu_0$ is the DOS per unit volume, and $$\begin{aligned}
[G]_{-}=-i(G^a-G^r),\end{aligned}$$ $G^{r,a}$ being the QD’s retarded/advanced Green’s functions.
Performing a disorder average [@Aleiner2] on Eq.( \[gamma\]), one arrives at $$\begin{aligned}
\Gamma_{qp} \simeq \hbar \omega_0^2 \left(\frac{d}{l_z}\right)^2\left[
(\nu_0 \lambda^2 l_z)\; \left(\frac{\lambda_F}{\lambda}\right)^2 + \frac{1}{6%
} \frac{\;l_z^2}{\hbar D} \right],\end{aligned}$$ where $D$ is the diffusion constant. Finally, inserting this expression into Eq.(\[gammaqp\]), and using the typical parameters reported above \[$%
\lambda_F \approx 1\;\mathring{A},\; D \approx 50 \mathrm{cm^2/s},\; \nu_0
\approx 1.5\;10^{22}\;\mathrm{cm^{-3}\;eV^{-1}}$\], one arrives at $%
\Gamma_{qp} \approx 10^3 \;\mathrm{Hz}$.
[99]{} P. Mohanty, E. M. Q. Jariwala and R. A. Webb, Phys. Rev. Lett. **78**, 3366 (1997).****
F. Pierre and N. O. Birge, Phys. Rev. Lett.** 89**, 206804 (2002).
F. Pierre, A. B. Gougam, A. Anothore, H. Pothier, D. Esteve, and N. O. Birge, Phys. Rev. B **68** , 085413 (2003).
A. Kaminsky and L. I. Glazman, Phys. Rev. Lett. 86 , 2400 (2001).
H. Pothier, S. Gueron, N. O. Birge, D. Esteve, and M. H. Devoret, Phys. Rev. Lett. **79** , 3490 (1997).
F. Pierre, H. Pothier, D. Esteve, M. H. Devoret, A. Gougam, and N. O. Birge, in *Kondo Effect and Dephasing in Low-Dimensional Metallic Systems,*
edited by V. Chandrasekar, C. Van Haesendonck and A. Zawadowsky, (Kluwer, Dordrecht, 2001) ; cond-mat/0012038.
A. A. Abrikosov, and L. P. Gorkov, JETP **12** , 1243 (1961).
A. V. Balatsky, and S. A. Trugman, Phys. Rev. Lett **79**, 3767 (1997).****
I. S. Beloborodov, B. N. Narozhny, and I. L. Aleiner, Phys. Rev. Lett. **85**, 816 (2000).
A. V. Shytov, I. Vekhter, I. A. Gruzberg, and A. V. Balatsky, Phys. Rev. Lett. **90** , 147002 (2003); I. Vekhter, A. V. Shytov, I. A. Gruzberg, A. V. Balatsky, Physica B** 329-333** , 1446 (2003).
M. G. Vavilov, P. W. Brouwer, V. Ambegaokar, and C. W. J. Beenakker, Phys. Rev. Lett. **86** , 874 (2001).
A. Lamacraft, and B. D. Simons, Phys. Rev. Lett. **85** , 4783 (2000); A. Lamacraft, and B.D. Simons, Phys. Rev. B **64**, 014514 (2001).
K. M. Lang, S. Nam, J. Aumentado, C. Urbina, and J. M. Martinis, IEEE Transactions on Applied Superconductivity **13**, 989 (2003).
P. W. Anderson, J. Phys. Chem. Solids **11**, 26 (1959).
Lu Yu, Phys. Sin. **21** , 75 (1965).
A. I. Rusinov, JETP Lett. 9 , 85 (1969).
M. I. Salkola, A. V. Balatsky, and J. R. Schrieffer, Phys. Rev. B 55 , 12648 (1997).
H. Shiba, Progr. Theor. Phys. 40 , 435 (1968).
A. I. Rusinov, Sov. Phys. JETP, 29 , 1101 (1969).
I. M. Lifshits, Sov. Phys. Usp. 7 , 549 (1965).
B. Halperin, Phys. Rev. 139 , A104 (1965).
B. Halperin and M. Lax, Phys. Rev. 148 , 722 (1966).
J. Zittarz and J. S. Langer, Phys. Rev. 148 , 741 (1966).
L. D. Larkin and Yu. N. Ovchinikov, JETP 34 , 1144 (1972).
C. A. Tracy, and H. Widom, Comm. Math. Phys. 159 , 151 (1994); 177 , 727 (1996).
A rather extreme limit of this phenomenon was considered in Ref.\[\], where the authors studied the effect of fluctuations of the concentration of magnetic impurities exceeding the AG critical concentration $n_{c}$, and therefore generating gapless regions optimal fluctuations surrounded by a gap-full matrix.
K. Usadel, Phys. Rev. Lett. 25 , 507 (1970).
N. Kopnin, Theory of Nonequilibrium Superconductivity, (Clarendon Press, Oxford, 2001).
J. S. Meyer, and B. D. Simons, Phys. Rev. B 64 , 134516 (2001).
P. M. Ostrovsky, M. A. Skvortsov, M. V. Feigel’man, Phys. Rev. Lett 87 , 027002 (2001).
D.E. Khmelnitskii, in New Directions in Mesoscopic Physics (Towards Nanoscience), (NATO ASI, Edt. R. Fazio, V. F. Gantmakher, and Y. Imry, Kluwer, Dordrecht, 2003).
The Euler-Lagrange equation Eq.(\[NonTrivialSaddle\]) is of fourth order in the gradients of $x$, while the equation we solve, Eq.([EL1]{}), is of second order. By reducing the order of the equations, some solutions could be lost. In order to prove that the analytical solution that we found is indeed the optimal fluctuation, we minimized numerically the action Eq.(\[actionmin\]) in three dimensions, using a variational ansatz with two free parameters. The numerical result for the constant $\tilde{a}%
_{d}$ \[see Eq.(\[3d\]), and Eq.(\[finalDOS\])\] is $~64$, within 2 percent of the result obtained analytically.
Y. Nakamura, Yu. A. Pashkin, and J. S. Tsai, Nature 398 , 786 (1999).
C. H. van der Wal, A. C. J. ter Haar, F. K. Wilhelm, R. N. Schouten, C. J. P. M. Harmans, T. P. Orlando, S. Lloyd, and J. E. Mooij Science 290 , 773 (2000).
D. Voin, A. Aassime, A. Cottet, P. Joyez, H. Pothier, C. Urbina, D. Esteve, and M. H. Devoret, Science 296, 886 (2002); see also A. Cottet, Ph. D. Thesis (Paris, 2002, unpublished).
J. H. Martinis, S. Nam, J. Aumentado, and C. Urbina, Phys. Rev. Lett. 89, 117901 (2002).
A. Burin, J. Low. Temp. Phys. **100 , 309 (1995).**
M. Gershenson, private communication.
Notice that, in contrast with the case of Lifshits tails, in the present situation each optimal fluctuation contains typically a number $%
N\gg 1$ of quasiparticle states.
Indeed, we expect $\tau _{tr}$ to be of the same order of magnitude as the phonon limited quasiparticle lifetime $\tau _{ph}$ at the Fermi level in Al. In the diffusive limit [@Lawrence], the latter can be estimated to be approximately to be $1/\tau _{ph}\simeq \lambda \pi
^{4}/2\;(k_{B}T/\omega _{D})^{3}(k_{F}l)(k_{D}/k_{F})^{3}\;k_{B}T/\hbar $, where $k_{D}$ and $\omega _{D}$ are the Debye momentum and frequency, $k_{F}$ is the Fermi momentum, $l$ is the mean free path, and $\lambda $ the electron phonon coupling constant. At $T=10\;mK$, using $\lambda \approx 1$,$%
\omega _{D}\approx 1000\;K$, $k_{D}/k_{F}\approx 1$, and $k_{F}l\approx
10^{2}$ one obtains $\tau _{ph}\approx 10^{2}\;s$.
A. Sergeev and V. Mitin, Phys. Rev. B 61 , 6041 (2000).
I.L. Aleiner, P.W. Brouwer, L.I. Glazman, Phys. Rep. 358 , 309 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive the explicit form of the $(g-2)(g-3)/2$ linearly independent relations among the products of pairs in a basis of holomorphic abelian differentials in the case of canonical curves of genus $g\geq4$. It turns out that Petri’s relations remarkably match in determinantal conditions. We explicitly express the volume form on the moduli space ${\hat{\mathcal{M}}_g}$ of canonical curves induced by the Siegel metric, in terms of the period Riemann matrix only. By the Kodaira-Spencer map, the relations lead to an expression of the induced Siegel metric on ${\hat{\mathcal{M}}_g}$, that corresponds to the square of the Bergman reproducing kernel. A key role is played by distinguished bases for holomorphic differentials whose properties also lead to an immediate derivation of Fay’s trisecant identity.'
address: |
Dipartimento di Fisica “G. Galilei” and Istituto Nazionale di Fisica Nucleare\
Università di Padova, Via Marzolo, 8 –\
35131 Padova, Italy
author:
- Marco Matone
- Roberto Volpato
date: September 2006
title: |
Linear relations among holomorphic quadratic\
differentials and induced Siegel’s metric on $\mathcal{M}_g$
---
[^1]
Introduction {#intro}
============
In spite of the remarkable progresses in understanding the Schottky problem, the characterization of the Schottky locus as the zero set of modular forms on the Siegel’s upper half-space remains a fundamental open question. Such a question is strictly related to the problem of characterizing the Schottky locus by means of $(g-2)(g-3)/2$ linearly independent conditions. As suggested also by Mumford (see pg. 241 of [@mumfordd]), a possible solution of that problem should follow by a deeper understanding of Petri’s construction [@petriuno; @ottimo]. Actually, since Petri’s theorem determines the ideal of canonical curves of genus $g\geq4$ by means of linear relations among holomorphic abelian differentials, it seems the natural framework for such an investigation.
Let $\{\eta_1,\ldots,\eta_g\}$ be the Petri’s basis for $H^0(K_C)$, with $C$ a canonical curve of genus $g$. In Petri’s work the coefficients $C_{ij}^k$ in the relationships among quadratic differentials $\sum_{i,j}^g C_{ij}^k\eta_i\eta_j=0$, $k=1,\ldots,
(g-2)(g-3)/2$, are not determined. Finding such coefficients is a necessary condition for an explicit characterization of the ideal of canonical curves. Here, we express Petri’s relation in determinantal form, so that, besides the explicit determination of the coefficients, it is shown that the locus of canonical curves corresponds to a determinantal variety.
We introduce modular invariant bases for holomorphic differentials, leading to a refinement of Petri’s basis and to an immediate derivation of Fay’s trisecant identity [@Fay]. A key point is the introduction of a indexing, which includes the combinatorics of the Petri construction, mapping the components of matrices in the Siegel upper half-space to vector components. This provides the volume form on the moduli space ${\hat{\mathcal{M}}_g}$ of canonical curves induced by the Siegel metric which, remarkably, is expressed in terms of the period Riemann matrix only. By the Kodaira-Spencer map, the above relations lead to an expression of the metric on ${\hat{\mathcal{M}}_g}$, induced by the Siegel metric, that corresponds to the square of the Bergman reproducing kernel.
In the case of branched covering of the torus, corresponding to Jacobians with a distinguished complex multiplication [@MatoneUY], the derived relations should lead to identities of number theoretical interest. Our results, of interest also in superstring theory [@MatoneVM], provide the key for the $(g-2)(g-3)/2$ combinatorial $\theta$-identities in [@MatoneVO].
Determinantal characterization of canonical curves {#sec:1}
==================================================
Let $C$ be a canonical curve of genus $g\ge 4$ and $\{\omega_i\}_{i\in I_g}$, $I_n:=\{1,\ldots,n\}$, a basis of $H^0(K_C)$, with $K_C$ the canonical line bundle of $C$. Denote by $\hat{\mathcal{M}}_g$ the corresponding locus in the moduli space ${\mathcal{M}}_g$ of compact Riemann surfaces. Each element of $H^0(K_C^2)$ can be written as a linear combination of the $M:=g(g+1)/2$ elements in $$\mathcal{S}:=\{\omega_i\omega_j|i\leq j\in I_g\}\ .$$ Since $N:= h^0(K_C^2)=3g-3$, there are $M-N=(g-2)(g-3)/2$ linearly independent relations among the quadratic differentials $\omega_i\omega_j$.
Let $p_1,\ldots,p_g$ and $q_1,\ldots,q_{2g-2}$ be two sets of arbitrary points on $C$. Choose a local trivialization of the canonical line bundle and set $$a_{ij,r}:=\det\,\omega(p_1,\ldots,p_{i-1},q_r,p_{i+1},\ldots,p_g)\det\,\omega(p_1,\ldots,p_{j-1},q_r,p_{j+1},\ldots,p_g)\ ,$$ where $\det\,\omega(x_1,\ldots,x_g):= \det\, \omega_i(x_j)$. Set $$A(k,l):=
\begin{pmatrix} a_{12,1} &\ldots &a_{1g,1} &a_{23,1} &\ldots &a_{2g,1} &a_{kl,1} \\ a_{12,2} &\ldots &a_{1g,2} &a_{23,2}
&\ldots &a_{2g,2} &a_{kl,2} \\ \vdots &\ddots &\vdots
&\vdots &\ddots &\vdots &\vdots
\\ a_{12,2g-2} &\ldots
&a_{1g,2g-2} &a_{23,2g-2} &\ldots &a_{2g,2g-2}
&a_{kl,2g-2}\end{pmatrix}\ ,$$ $3\le k<l\le g$, $g\ge 4$.
\[teo\]$$\det\,A(k,l)=0\ ,\qquad 3\le k<l\le g\ .$$
Set $$\Delta_{m\, n}:=(-)^{m+n}\det_{^{i\neq m}_{j\neq n}}
A(k,l)_{ij}\ , \qquad D_{p\, q}:=(-)^{p+q}\det_{^{i\neq p}_{j\neq
q}}\omega_j(p_i)\ ,$$ and denote by $A_{ij,r}(k,l)$, $i,j\in I_g$, $r\in I_{2g-2}$, the matrix obtained from $A(k,l)$ by replacing the row $(a_{12,r},\ldots,a_{kl,r})$ with $(D_{1i}D_{2j},\ldots,D_{k\, i} D_{l\, j})$.
\[corolla\] For each $r\in I_{2g-2}$, the following relations $$\sum_{i,j=1}^g\frac{\det\,A_{ij,r}(k,l)}{\Delta_{r\, 2g-2}}\omega_i\omega_j=0\ ,$$ $3\le k<l\le g$, provide $(g-2)(g-3)/2$ linearly independent conditions on $\mathcal{S}$ which are independent of the points $q_i$, $i\in I_{2g-2}$.
Distinguished bases of $H^0(K_C^n)$ {#sec:2}
===================================
Set $N_n:=(2n-1)(g-1)+\delta_{1n}$, $n\ge 1$, with $\delta_{ij}$ the Kronecker delta. Note that $N_1=g$ and $N_2\equiv N$. Fix a system of local coordinates on $C$.
\[prop:base\] Fix $n\ge 1$ and let $p_1,\ldots,p_{N_n}$ be a set of points in $C$ such that $$\det\,\phi(p_1,\ldots,p_{N_n})\ne 0\ ,$$ for an arbitrary basis $\{\phi_i\}_{i\in I_{N_n}}$ of $H^0(K_C^n)$. Then $$\label{basendiff}\gamma^n_i(z):=
\frac{\det\,\phi(p_1,\ldots,p_{i-1},z,p_{i+1},\ldots,p_{N_n})}{
\det\,\phi(p_1,\ldots,p_{N_n})}\ ,$$ $i\in I_{N_n}$, determines a basis of $H^0(K_C^n)$ which is independent of the choice of the basis $\{\phi_i\}_{i\in
I_{N_n}}$ and, up to normalization, of the local coordinates on $C$.
The matrix $[\phi]_{ij}:=\phi_i(p_j)$ is non-singular. Then $\gamma^n_i(z)=\sum_j[\phi]^{-1}_{ij}\phi_j(z)$, $i\in I_{N_n}$, is a basis of $H^0(K_C^n)$.
Note that $\gamma_i^n(p_j)=\delta_{ij}$, furthermore $$\label{preludetofay}
\det\, \gamma^n(p_1,\ldots,p_{j-1},z,p_{j+1},\ldots,p_{N_n})=\gamma^n_j(z) \ .$$
As we will see, the Fay trisecant identity [@Fay] directly follows by expressing Eq. in terms of theta functions.
For $n=1$, the choice of $g$ points $p_1,\ldots,p_g\in C$, with $\det\,\omega_i(p_j)\ne 0$, determines the basis $\{\sigma_i\}_{i\in
I_g}$ of $H^0(K_C)$, where $$\label{newbasis}\sigma_i(z):=\gamma^1_i(z)\ ,\qquad i\in I_g \ .$$
We now introduce a refinement of Petri’s basis for $H^0(K_C^2)$ [@petriuno; @ottimo] which provides a modular invariant construction. Let us assume that the points $p_1,\ldots,p_g$ are in “general position" and that the effective divisor $(\sigma_1)+(\sigma_2)-\sum_{i=3}^gp_i$ consists of $3g-3$ distinct points. Consider the following $M$ elements of $H^0(K_C^2)$ $$v_i:=\begin{cases}\sigma_i^2\ , &i\in I_g\
,\\ &\\ \sigma_{j+k}\sigma_j\ , &i=k+j(2g-j+1)/2\ \ ,\quad
j\in I_{g-1}\ ,\quad k\in I_{g-j}\ . \end{cases}$$
$\{v_i\}_{i\in I_N}$ is a basis of $H^0(K_C^2)$.
Set $D:=\sum_{i=3}^{g}p_i$ and let us consider the effective divisors $D_i:=(\sigma_i)-D$, $i=1,2$. Let us first prove that $\sigma_i$ is the unique $1$-differential, up to normalization, vanishing at $D_i$, $i=1,2$. Any $1$-differential $\sigma'_i\in H^0(K_C)$ vanishing at $D_i$, corresponds to an element $\sigma'_i/\sigma_i$ of $H^0({\mathcal{O}}(D))$, the space of meromorphic functions $f$ on $C$ such that $(f)+D$ is an effective divisor. Suppose that there exists $\sigma'_i$ such that $\sigma'_i/\sigma_i$ is not a constant, so that $h^0({\mathcal{O}}(D))\ge 2$. By Riemann-Roch theorem $$h^0(K_C\otimes {\mathcal{O}}(-D))=h^0({\mathcal{O}}(D))-\deg
D-1+g\ge 3\ ,$$ so that there exist at least $3$ linearly independent $1$-differentials vanishing at $D$ and, in particular, there exists a linear combination of such differentials vanishing at $p_1,\ldots,p_g$. This implies that $\det\eta(p_1,\ldots,p_g)=0$ for an arbitrary basis $\{\eta_i\}_{i\in I_g}$ of $H^0(K_C)$, contradicting the hypotheses. Fix $\zeta_i,\zeta_{1i},\zeta_{2i}\in
{\mathbb{C}}$ such that $$\sum_{i=3}^g\zeta_i\sigma_i^2+\sum_{i=1}^g\zeta_{2i}
\sigma_1\sigma_i+\sum_{i=2}^g\zeta_{1i}\sigma_2\sigma_i=0\ .$$ Evaluating this relation at the point $p_j$, $3\le j\le g$, yields $\zeta_j=0$. Set $$\label{leti}t_1:=-\sum_{j=2}^g\zeta_{1j}\sigma_j\ ,\quad
t_2:=\sum_{j=1}^g\zeta_{2j}\sigma_j\ ,$$ so that $\sigma_1t_2=\sigma_2 t_1$. Since $D$, $D_1$ and $D_2$ consist of pairwise distinct points, $t_i$ vanishes at $D_i$, $i=1,2$ and then $t_1/\sigma_1=t_2/\sigma_2=\zeta\in{\mathbb{C}}$. By $$\zeta\sigma_1+\sum_{j=2}^g\zeta_{1j}\sigma_j=0\ ,\quad
\zeta\sigma_2-\sum_{k=1}^g\zeta_{2k}\sigma_k=0\ ,$$ and, by linear independence of $\sigma_1,\ldots,\sigma_g$, we have $\zeta=\zeta_{1j}=\zeta_{2k}=0$, $2\le j\le g$, $k\in I_g$.
Proofs of Theorem \[teo\] and Corollary \[corolla\] {#sec:3}
===================================================
Let $W(P)$ be the Wronskian $W(v_1,\ldots,v_N)(P)$ of the basis $\{v_i\}_{i\in I_N}$ at a generic point $P\in C$, and $\hat{W}_{ij}(P):=W(v_1,\ldots,v_{i-1},v_j,v_{i+1},\ldots,v_N)(P)$.
\[illemma\] The $(g-2)(g-3)/2$ linearly independent relations $$\label{lemma}v_i(z)
W(P) =\sum_{j=1}^Nv_j(z) \hat{W}_{ji}(P)\ ,$$ $i=N+1,\ldots,M$, hold $\forall z\in C$.
Immediate consequence of the Cramer rule.
The ratio $\hat{W}_{ij}(P)/W(P)$ does not depend on $P$.
Since for $i\in I_g$ $$\label{devv}\begin{cases}v_j(p_i)=\delta_{ji}\ , &j\in I_g\ ,\\
&\\ v_j(p_i)=0\ ,&j=g+1,\ldots,M\ ,\end{cases}$$ it follows that for $z=p_i$ Eq. gives $\hat{W}_{ij}(P)=0$ for $i\in I_g$ and $j=N+1,\ldots,M$.
\[lastrem\] The Wronskians in the expansion can be replaced by the corresponding determinant $\det v_j(x_i)$, where $x_1,\ldots,x_{3g-3}$ are arbitrary points on $C$.
[of Theorem \[teo\]]{} Assume that $\det\,\omega(p_1,\ldots,p_g)\ne 0$. Define $$x_i:=\begin{cases}
p_i\ ,&i\in I_g\ ,\\ &\\ q_{i-g}\ ,&i=g+1,\ldots,N+1\
.\end{cases}$$ Fix $k,l$ with $3\le k<l\le g$ and consider the matrix
$$\begin{pmatrix}
v_1(x_1) &\ldots &v_{N}(x_1) &\sigma_k(x_1)\sigma_l(x_1)\\ \vdots
&\ddots &\vdots &\vdots\\ v_1(x_{N+1}) &\ldots &v_{N}(x_{N+1})
&\sigma_k(x_{N+1})\sigma_l(x_{N+1})\end{pmatrix}\ .$$
By , this matrix has ${\rm
diag}\,(1,\ldots,1)$ in the $g\times g$ upper left corner, $0$ in the $g \times (2g-2)$ upper right corner and $(\det\,\omega(p_1,\ldots,p_g))^{-2}A(k,l)$ in the $(2g-2)\times
(2g-2)$ lower right corner. On the other hand, by Lemma \[illemma\] and by Remark \[lastrem\] the determinant of this matrix vanishes. Since such relations hold for $(p_1,\ldots,p_g)$ in a dense subset of $C^g$, they hold $\forall(p_1,\ldots,p_g)\in C^g$ and the theorem follows.
[of Corollary \[corolla\]]{} Divide the relations in Theorem \[teo\] by $\Delta_{i\, 2g-2}$ and note that $$a_{mn,r}=\sum_{i,j=1}^g D_{mi}D_{nj}\omega_i(q_r)\omega_j(q_r)\ .$$ Independence of the points $q_i$, $i\in I_{2g-2}$, follows by noting that the coefficients of $\omega_i\omega_j$ in the relations are functions of $q_i$ with no zeroes or poles.
Define $$({\mathfrak{1}}_i,{\mathfrak{2}}_i):=\begin{cases}(i,i)\ , &1\le i\le g\ ,\\
(1,i-g+1)\ , &g+1\le i\le 2g-1\ ,\\
(2,i-2g+3)\ , &2g\le i\le 3g-3\
,\\ \hfill\vdots \hfill& \hfill\vdots\hfill\\
(g-1,g)\ , &i=g(g+1)/2\ ,\end{cases}$$ so that ${\mathfrak{1}}_i{\mathfrak{2}}_i$ is the $i$-th element in the $M$-tuple $(11,22,\ldots,gg,12,\ldots,1g,23,\ldots)$. $\forall u\in{\mathbb{C}}^g$ and for all the $g\times g$ matrices $A$, set $$\label{doubleindex}uu_i:=u_{{\mathfrak{1}}_i}u_{{\mathfrak{2}}_i}\ ,\quad\;
(AA)_{ij}:=\frac{A_{{\mathfrak{1}}_i{\mathfrak{1}}_j}A_{{\mathfrak{2}}_i{\mathfrak{2}}_j}+A_{{\mathfrak{1}}_i{\mathfrak{2}}_j}A_{{\mathfrak{2}}_i{\mathfrak{1}}_j}}{1+\delta_{{\mathfrak{1}}_j{\mathfrak{2}}_j}}\ ,\quad\;
A_i:=A_{{\mathfrak{1}}_i{\mathfrak{2}}_i}\ ,$$ $i,j\in I_M$. In the following we will repeatedly make use of the identities $$\label{AA}\sum_{i,j=1}^gf(i,j)=\sum_{i\le j}^g\frac{f(i,j)+f(j,i)}{1+\delta_{ij}}
=\sum_{k=1}^M\frac{f({\mathfrak{1}}_k,{\mathfrak{2}}_k)+f({\mathfrak{2}}_k,{\mathfrak{1}}_k)}{1+\delta_{{\mathfrak{1}}_k{\mathfrak{2}}_k}}\
.$$ In particular, if $f(i,j)=f(j,i)$, then $$\label{BB}\sum_{i,j=1}^gf(i,j)=\sum_{k=1}^M(2-\delta_{{\mathfrak{1}}_k{\mathfrak{2}}_k})f({\mathfrak{1}}_k,{\mathfrak{2}}_k)\ ,$$ where we used the identity $$2-\delta_{ij}=\frac{2}{1+\delta_{ij}}\ .$$ With this notation, and observing that $\sigma_i=\sum_{j=1}^g[\omega]^{-1}_{ij}\omega_j$, we have $$\label{vomega}v_i=\sigma\sigma_i=
\sum_{j=1}^M([\omega]^{-1}[\omega]^{-1})_{ij}\omega\omega_j\ ,
\qquad i\in I_M\ .$$ Set $w_{ij}:=W_{ij}/W$, where ${W}_{ij}(P):=W(v_1,\ldots,v_{i-1},\omega\omega_j,v_{i+1},\ldots,v_N)(P)$, and note that $$\label{teor}\omega\omega_i=\sum_{j=1}^N
w_{ji}v_j\ ,\qquad i\in I_M\ .$$
Siegel’s induced measure on $\hat{\mathcal M}_g$ and Bergman reproducing kernel {#sec:4}
===============================================================================
Let $${\mathfrak{H}}_g:=\{Z\in M_g({\mathbb{C}})\,|\, {}^tZ=Z,\ Y>0\}\ , \qquad Y:={\Im}Z\ ,$$ be the Siegel upper half-space and $\{\alpha_1,\ldots,\alpha_g,\beta_1,\ldots,\beta_g\}$ a symplectic basis of $H_1(C,{\mathbb{Z}})$. Denote by $\{\omega_i\}_{i\in I_g}$ the basis of $H^0(K_C)$, dual of $H_1(C,{\mathbb{Z}})$, so that $\oint_{\alpha_i}\omega_j=\delta_{ij}$, $i,j\in I_g$. Let $\tau_{ij}:=\oint_{\beta_i}\omega_j\in {\mathfrak{H}}_g$ be the Riemann period matrix of $C$. Under the symplectic transformation $$\begin{pmatrix}\tilde\alpha \\ \tilde\beta\end{pmatrix}=\begin{pmatrix}D & C \\ B & A\end{pmatrix}
\begin{pmatrix}\alpha\cr\beta\end{pmatrix}\ ,\qquad\qquad \begin{pmatrix}A & B\\ C & D\end{pmatrix}\in Sp(2g,{\mathbb{Z}})\ ,$$ we have $\tilde\omega={}^t(C\tau+D)\cdot\omega$, with $\tilde\tau_{ij}$ and $\tau_{ij}$ related by the modular transformation $$\tilde\tau=(A\tau+B)\cdot(C\tau+D)^{-1}\ .$$ Note that the basis $\{\sigma_i\}_{i\in I_g}$, defined in Eq., is independent of the choice of the basis of $H_0(K_C)$ and therefore is modular invariant.
The Siegel metric $$\label{siegelon}
ds^2:=\operatorname{Tr}\, (Y^{-1}dZ Y^{-1}d\bar Z)\ ,$$ defines the volume form $$d\nu=\frac{i^M}{2^g}\frac{ \bigwedge_{i\le j}^g ({\rm d} Z_{ij}\wedge{\rm d}\bar Z_{ij}) }{
\det Y^{g+1} } \ .$$
We use the indexing introduced in Eq. to express the Siegel metric on ${\mathfrak{H}}_g$ where now the matrix elements $Z_{ij}$, $i,j\in I_M$, are seen as the components of the $M$-dimensional vectors $Z:=(Z_1,\ldots,Z_M)$.
$$\label{sieggmetr}ds^2=\sum_{i,j=1}^Mg_{ij}dZ_id\bar Z_j\ ,$$
where $$\label{metric}
g_{ij}:=(2-\delta_{{\mathfrak{1}}_i{\mathfrak{2}}_i})(Y^{-1}Y^{-1})_{ij}\ .$$
By and $$\begin{aligned}
ds^2&=\sum_{i,j,k,l=1}^gY^{-1}_{ij}d Z_{jk}Y^{-1}_{kl}d\bar Z_{li}\\
&=\sum_{i,l=1}^gd\bar Z_{li}\sum_{m=1}^M\frac{Y^{-1}_{i{\mathfrak{1}}_m}Y^{-1}_{l{\mathfrak{2}}_m}+
Y^{-1}_{i{\mathfrak{2}}_m}Y^{-1}_{l{\mathfrak{1}}_m}}{1+\delta_{{\mathfrak{1}}_m{\mathfrak{2}}_m}}d Z_{{\mathfrak{1}}_m{\mathfrak{2}}_m}\\
&=\sum_{m,n=1}^M(2-\delta_{{\mathfrak{1}}_n{\mathfrak{2}}_n})d\bar Z_{{\mathfrak{1}}_n{\mathfrak{2}}_n}\frac{Y^{-1}_{{\mathfrak{1}}_n{\mathfrak{1}}_m}
Y^{-1}_{{\mathfrak{2}}_n{\mathfrak{2}}_m}+
Y^{-1}_{{\mathfrak{1}}_n{\mathfrak{2}}_m}Y^{-1}_{{\mathfrak{2}}_n{\mathfrak{1}}_m}}{1+\delta_{{\mathfrak{1}}_m{\mathfrak{2}}_m}}d Z_{{\mathfrak{1}}_m{\mathfrak{2}}_m}\\
&=\sum_{m,n=1}^M(2-\delta_{{\mathfrak{1}}_n{\mathfrak{2}}_n})(Y^{-1}Y^{-1})_{nm}d Z_md\bar Z_n\ .\end{aligned}$$
Let $k$ be the Kodaira-Spencer map identifying the quadratic differentials on $C$ with the fiber of the cotangent of the Teichmüller space at $C$. We have $$k(\omega_i\omega_j)={(2\pi i)}^{-1}d\tau_{ij} \ .$$ By Corollary \[corolla\] it follows that $$\sum_{i,j=1}^g\frac{\det\,A_{ij,r}(k,l)}{ \Delta_{r\, 2g-2}}d\tau_{ij}=0\ ,
\label{hnice}$$ $3\le k<l\le g$. Set $d\tau_i:=d\tau_{{\mathfrak{1}}_i{\mathfrak{2}}_i}$, $i\in I_M$. Eq. yields an explicit expression for the volume form on $\hat{\mathcal{M}}_g\hookrightarrow{\mathfrak{H}}_g/Sp(2g,{\mathbb{Z}})$ induced by the modular invariant Siegel metric on ${\mathfrak{H}}_g$. Set $\tau_2:=\Im\tau$, and let $\left|\tau_2^{-1}\tau_2^{-1}\right|^{i_1\ldots
i_N}_{j_1\ldots j_N}$, with $i_k,j_k$, $k\in I_N$, distinct elements of $I_M$, be the determinant of the $N\times N$ submatrix of $(\tau_2^{-1}\tau_2^{-1})_{ij}$, built by taking the rows $i_1,\ldots,i_N$ and the columns $j_1,\ldots,j_N$.
The volume form on $\hat{\mathcal{M}}_g$ induced by the Siegel metric is $$\label{basicc}d\nu_{|\hat{\mathcal{M}}_g}=\Bigl(\frac{i}{2}\Bigr)^N\sum_{\substack{i_N>\ldots>i_1=1\\
j_N>\ldots>j_1=1}}^M\left|\tau_2^{-1}\tau_2^{-1}\right|^{i_1\ldots
i_N}_{j_1\ldots j_N}\prod_{k=1}^N(2-\delta_{{\mathfrak{1}}_{i_k}{\mathfrak{2}}_{i_k}})\bigwedge_1^N(d\tau_{i_k}\wedge
d\bar\tau_{j_k})\ ,$$ so that $$\label{basicccc}
{\rm Vol}(\hat{\mathcal{M}}_g)=\int_{\hat{\mathcal{M}}_g}d\nu_{|\hat{\mathcal{M}}_g}\ .$$
Let $$\omega:=\frac{i}{2}\sum_{i,j=1}^Mg_{ij}dZ_i\wedge d\bar Z_j\ ,$$ be the $(1,1)$-form associated to the Siegel metric on ${\mathfrak{H}}_g$. By Wirtinger’s theorem [@GH], the volume form on a $d$-dimensional complex submanifold $S$ is $$\frac{1}{d!}\omega^d\ ,$$ so that the volume of $S$ is expressed as the integral over $S$ of a globally defined differential form on ${\mathfrak{H}}_g$. Set $g^\tau_{ij}:=(2-\delta_{{\mathfrak{1}}_i{\mathfrak{2}}_i})(\tau_2^{-1}\tau_2^{-1})_{ij}$, $i,j\in I_M$, and note that $$\begin{aligned}
d\nu_{|\hat{\mathcal{M}}_g}=&\frac{i^N}{2^NN!}\sum_{\substack{i_1,\ldots,i_N=1\\j_1,\ldots,j_N=1}}^M
\prod_{k=1}^Ng^\tau_{i_kj_k}\bigwedge_{k=1}^N(d\tau_{i_k}\wedge
d\bar \tau_{j_k})\\
=&\frac{i^N}{2^NN!}\sum_{\substack{i_N<\ldots<i_1=1\\j_N<\ldots<j_1=1}}^M\sum_{r,s\in{\mathcal{P}}_N}
\epsilon(r)\epsilon(s)\prod_{k=1}^Ng^\tau_{i_{r(k)}j_{s(k)}}\bigwedge_{k=1}^N(d\tau_{i_k}\wedge
d\bar \tau_{j_k})\ ,\end{aligned}$$ where ${\mathcal{P}}_N$ is the group of permutations of $N$ elements and $\epsilon(s)$ is the sign of the permutation $s$. The theorem then follows by the identity $$\sum_{r,s\in{\mathcal{P}}_N}
\epsilon(r)\epsilon(s)\prod_{k=1}^Ng^\tau_{i_{r(k)}j_{s(k)}}=N!\left|\tau_2^{-1}\tau_2^{-1}\right|^{i_1\ldots
i_N}_{j_1\ldots j_N}\prod_{k=1}^N(2-\delta_{{\mathfrak{1}}_{i_k}{\mathfrak{2}}_{i_k}})\ .$$
Petri’s basis of $H^2(K_C^2)$ corresponds, through the Kodaira-Spencer map, to a basis for the cotangent space of the Teichmüller space. Setting $d\,\Xi_i:=2\pi i\,k(v_i)$, $i\in I_N$, it follows by Eq. that $$\label{domega}d\tau_i=\sum_{j=1}^Nw_{ji}
d\,\Xi_j \ ,\qquad i\in I_M\ .$$
Fix the points $p_1,\ldots,p_g\in C$ in general position, so that $\{v_i\}_{i\in I_N}$, is a basis of $H^0(K_C^2)$. The metric on $\hat{\mathcal{M}}_g$ induced by the Siegel metric is $$\label{ds} ds_{|\hat{\mathcal{M}}_g}^2=
\sum_{i,j=1}^Ng^{\Xi}_{ij}d\,\Xi_id\,\bar\Xi_j \ ,$$ where $g^{\Xi}_{ij}:=\sum_{k,l=1}^M(2-\delta_{{\mathfrak{1}}_k{\mathfrak{2}}_k})w_{ik}(\tau_2^{-1}\tau_2^{-1})_{kl}
\bar{{w}}_{jl}$.
Immediate.
By using a suitable basis of $H^0(K_C^2)$ and its image under the Kodaira-Spencer map, it turns out that the metric $g$ is related to the Bergman reproducing kernel. Fix the points $z_1,\ldots,z_N\in C$ satisfying the conditions of Proposition \[prop:base\]. The basis $\{\gamma_i\}_{i\in I_N}$ of $H^0(K_C^2)$, with $\gamma_i\equiv\gamma^2_i$, $i\in I_N$, defined by Eq. in the case $n=2$, satisfies the relations $$\omega\omega_i=\sum_{j=1}^N\omega\omega_i(z_j)\gamma_j\ ,\qquad
v_i=\sum_{j=1}^Nv_i(z_j)\gamma_j\ ,\quad i\in I_M\ .$$ Set $\Gamma_i:=(2\pi i)^{-1}k(\gamma_i)$ and $[v]_{ij}:=v_i(z_j)$, $i,j\in I_N$.
$$\label{siegelbergman}
ds^2_{|\hat{\mathcal{M}}_g}= \sum_{i,j=1}^NB^2(z_i,\bar
z_j)d\,\Gamma_id\,\bar\Gamma_j\ ,$$
in particular, the Siegel induced modular invariant volume form on $\hat{\mathcal{M}}_g$ is $$\label{volumee}d\nu_{|\hat{\mathcal{M}}_g}=\Bigl(\frac{i}{2}\Bigr)^N\det\,
B^2(z_i,\bar z_j)\,\textstyle{\bigwedge_1^N}(d\,\Gamma_i\wedge d\,\bar \Gamma_i) \
,$$ where $$B(z,\bar w):=\sum_{i,j=1}^g\omega_i(z)(\tau_2^{-1})_{ij}\bar\omega_j(w)\ ,$$ $z,w\in C$, is the Bergman reproducing kernel.
Use $d\tau_i=\sum_{j=1}^N\omega\omega_i(z_j)d\Gamma_j$, $i\in I_g$, and the identity $$\sum_{k,l=1}^M(2-\delta_{{\mathfrak{1}}_k{\mathfrak{2}}_k})\omega\omega_k(z_i)(\tau_2^{-1}\tau_2^{-1})_{kl}
\bar\omega\bar\omega_{l}(z_j)=B^2(z_i,\bar
z_j)\ ,\qquad i,j\in I_N\ .$$ Note that by $\sum_{k,l=1}^N[v]_{ki}g^{\Xi}_{kl}[\bar v]_{lj}=B^2(z_i,\bar z_j)$, which also follows by .
Fay’s trisecant identity from the distinguished basis of $H^0(K_C^n)$ {#sec:5}
=====================================================================
Let $I_i(p):=\int_{p_0}^p\omega_i$, $p_0,p\in C$, $i\in I_g$, be the Abel-Jacobi map, which extends to a map from divisors of $C$ to the Jacobian $J(C):={\mathbb{C}}^g/({\mathbb{Z}}_g+\tau{\mathbb{Z}}^g)$. We consider Riemann $\theta$-functions $\theta(D+e):=\theta (I(D)+e,\tau)$, $e\in J(C)$, evaluated at some $0$-degree divisor $D$ of $C$. By the Riemann vanishing theorem, there is a divisor class $\Delta$ of degree $g-1$ with $2\Delta = K$, such that $-I(\Delta)$ is the vector of Riemann constants. Let $E(z,w)$ be the prime form, and set $$\sigma(z):=\exp\bigg(-\sum_{i=1}^g\oint_{\alpha_i}\omega_i(w)\ln E(z,w)\bigg)\ .$$
\[thdettheta\] Fix $n\in{\mathbb{N}}$ and let $\{\phi_i^n\}_{i\in I_{N_n}}$ be an arbitrary basis of $H^0(K_C^n)$, $n\geq 1$. Let $y,x_1,\ldots,x_{N_n}$ be arbitrary points of $C$. Then, for $n=1$ $$\label{dettheta}\det\phi^1_i(x_j)=\kappa_1[\phi^1]\frac{\theta\bigl(\sum_{1}^gx_i-y-\Delta\bigr)
\prod_{i<j}^gE(x_i,x_j)\prod_1^g\sigma(x_k)}{
\sigma(y)\prod_1^gE(y,x_i)}\ ,$$ whereas for $n>1$ $$\label{detthetaii}\det\phi^{n}_i(x_j)=\kappa_{n}[\phi^n]
\theta\Bigl(\sum_{1}^{N_n} x_i-(2n-1)\Delta\Bigr)\prod_{i<j}^{N_n}
E(x_i,x_j)\prod_{i=1}^{N_n}\sigma(x_i)^{2n-1}\ ,$$ where $\kappa_1[\phi^1]$ and $\kappa_n[\phi^n]$ are constants depending only on the choice of the bases.
$\kappa_1[\phi^1]$ is a nowhere vanishing section in $x_j$, $j\in I_g$, and $\theta(\sum_{1}^gx_i-y-\Delta)=0$ for $y=x_1,\ldots,x_g$, so that it is also a nowhere vanishing section in $y$ and since it has trivial monodromy it must be a constant. Eq. follows by a similar proof.
Set $w:=\sum_1^{N_n}p_i-(2n-1)\Delta$, $n>1$, and assume that $p_1,\ldots,p_{N_n}\in C$ satisfy the hypothesis of Proposition \[prop:base\]. By Proposition \[thdettheta\] we have $$\label{glindiff}\gamma^n_i(z)=\frac{\theta(w+z-p_i)\sigma(z)^{2n-1}\prod_{^{k=1}_{k\neq
i}}^{N_n}E(z,p_k)}{ \theta(w)\sigma(p_i)^{2n-1}\prod_{^{k=1}_{k\neq
i}}^{N_n}E(p_i,p_k)}\ , \qquad i\in I_{N_n}\ .$$
\[Faytris\] Propositions \[prop:base\] and \[thdettheta\] imply the Fay trisecant identity [@Fay] $$\frac{\theta(w+\sum_{i=1}^m(x_i-y_i))\prod_{i<j}E(x_i,x_j)E(y_i,y_j)}{
\theta(w)\prod_{i,j}E(x_i,y_j)}=(-)^{\frac{m(m-1)}{2}}\det\nolimits_{ij}\frac{\theta(w+x_i-y_j)}{
\theta(w)E(x_i,y_j)}\ ,$$ $m\ge 2$, $\forall
x_1,\ldots,x_m,y_1,\ldots,y_m\in C$, $w\in J(C)$.
Fix $m\ge 2$, $x_1,\ldots,x_m,y_1,\ldots,y_m\in C$ and $w\in J(C)$, with $\theta(w)\neq 0$. Choose $y_1,\ldots,y_m$ distinct, otherwise the identity is trivial. Set $p_i:= y_i$, $i\in I_m$, and fix $n\in {\mathbb{Z}}$, $N_n\ge m$, and $p_{m+1},\ldots,p_{N_n}\in C$, so that $w=\sum_1^{N_n}p_i-(2n-1)\Delta$. By Jacobi inversion theorem such a choice is always possible, provided that $N_n-m\ge g$. It is also clear that, for $n$ large enough, $p_{m+1},\ldots,p_{N_n}$ can be chosen pairwise distinct and distinct from $y_1,\ldots,y_m$. Eq. implies that, by construction, $\det\phi_i^n(p_j)\neq 0$, for any basis $\{\phi_i^n\}_{i\in I_{N_n}}$ of $H^0(K_C^n)$, since the points $p_1,\ldots,p_{N_n}$ are pairwise distinct and $\theta(w)\neq 0$. Therefore, one can define a basis $\{\gamma^n_i\}_{i\in I_{N_n}}$ of $H^0(K_C^n)$ by and consider $\det\gamma^n(x_1,\ldots,x_m,p_{m+1},\ldots,p_{N_n})$, which can be expressed by or by , with $\kappa_n[\gamma^n]$ determined by applying to $\det\gamma^n_i(p_j)=1$. Comparing these formulas the theorem follows.
[99]{}
D. Mumford, [*The red book of varieties and schemes*]{}, Springer Lecture Notes 1358 (1999).
K. Petri, [*$\ddot{U}$ber die invariante darstellung algebraischer funktionen einer ver$\rm\ddot{a}$nderlichen*]{}, Math. Ann. [**88**]{} (1922) 242 – 289.
E. Arbarello, M. Cornalba, P.A. Griffiths and J. Harris, [*Geometry of algebraic curves, I.*]{} Grundlehren Math. Wiss., vol. 267, Springer-Verlag, 1985.
J. Fay, [*Theta Functions on Riemann surfaces*]{}, Springer Lecture Notes 352.
M. Matone, [*Eigenfunctions of the Laplacian acting on degree zero bundles over special Riemann surfaces*]{}, Trans. Am. Math. Soc. [**356**]{} (2004) 2989 – 3004. M. Matone and R. Volpato, [*Higher genus superstring amplitudes from the geometry of moduli spaces*]{}, Nucl. Phys. B [**732**]{} (2006) 321 – 340. M. Matone and R. Volpato, [*Determinantal characterization of canonical curves and combinatorial theta identities*]{}, math.ag/0605734. P.A. Griffiths and J. Harris, [*Principles of Algebraic Geometry*]{}, Wiley, 1978.
[^1]: Work partially supported by the European Community’s Human Potential Programme under contract MRTN-CT-2004-005104 “Constituents, Fundamental Forces and Symmetries of the Universe".
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the rate of convergence to equilibrium of the solution of a Fokker–Planck type equation introduced in [@Tos06] to describe opinion formation in a multi-agent system. The main feature of this Fokker–Planck equation is the presence of a variable diffusion coefficient and boundaries, which introduce new challenging mathematical problems in the study of its long-time behavior.'
address:
- 'DIGIP, University of Bergamo, viale Marconi 5, 24044 Dalmine, Italy'
- 'Department of Mathematics, University of Pavia, via Ferrata 1, Pavia, 27100 Italy'
- 'Department of Mathematics, University of Milan, via Saldini 50, 20133 Milano, Italy '
- 'Department of Mathematics, University of Pavia, via Ferrata 1, Pavia, 27100 Italy'
author:
- GIULIA FURIOLI
- ADA PULVIRENTI
- ELIDE TERRANEO
- GIUSEPPE TOSCANI
bibliography:
- 'biblio-wright-fisher.bib'
title: 'WRIGHT–FISHER–TYPE EQUATIONS FOR OPINION FORMATION, LARGE TIME BEHAVIOR AND WEIGHTED logarithmic-SobOLEV INEQUALITIES'
---
Introduction {#intro}
============
Kinetic models for (continuous) opinion formation have been first introduced and discussed in [@Tos06], starting from the study of a multi-agent system in which agents undergo binary interactions so that the personal opinion could be changed by means of compromise and self-thinking [@BKR03; @BKVR03; @FPTT17]. In most of the problems related to socio-economic studies of multi-agent systems [@NPT; @PT13], the variable is assumed to vary in an unbounded domain (mainly the positive half-line). On the contrary, the opinion variable is assumed to take values in the bounded interval $\CI= (-1, 1)$, the values $\pm 1$ denoting the extremal opinions. Among the various models introduced in [@Tos06] (cf. also [@Bou; @DMPW]), one Fokker–Planck type equation has to be distinguished in view of its equilibrium configurations, which are represented by Beta-type probability densities supported in the interval $(-1, 1)$. This Fokker–Planck equation for the opinion density $v(t,y)$, with $|y| < 1$, is given by \[op-FP\] = 2((1-y\^2) v(t,y)) + ((y -m)v(t,y)). In [(\[op-FP\])]{}, $\lambda$ and $m$ are given constants, with $\lambda >0$ and $-1 <m<1$. Suitable boundary conditions at the boundary points $y = \pm 1$ then guarantee conservation of mass and momentum of the solution [@FPTT17]. Equation [(\[op-FP\])]{} possesses steady states which solve $$ \frac
\lambda 2\frac{d}{d y}\left((1-y^2)v(y)\right) + (y -m) v(y)= 0.$$ In case a mass density equal to unity is chosen, the steady state equals a probability density of Beta type, given by \[beta\] v\_[m,]{}(y)= C\_[m,]{} (1-y)\^[-1 + ]{} (1+y)\^[-1 + ]{}. In [(\[beta\])]{} the constant $C_{m,\lambda}$ is such that the mass of $v_{m,\lambda}$ is equal to one. Since $-1 <m<1$, $v_{m,\lambda}$ is integrable on $\CI$. Note that $v_{m,\lambda}$ is continuous on $\CI$, and as soon as $\lambda > 1+|m|$ tends to infinity as $y \to \pm 1$.
A better understanding of the social meaning of the parameters $\lambda$ and $m$ appearing in [(\[op-FP\])]{} comes from the microscopic description of the opinion change in a multi-agent system through binary interactions among agents, leading to the Boltzmann type kinetic equation considered in [@Tos06]. Given a pair of agents with opinions $x$ and $x_\ast$, it was assumed in [@Tos06] that any elementary interaction between them modifies the entering opinions according to $$\begin{split}
x'&=x+\gamma (x_\ast-x)+D(x)\eta, \\
x_\ast'&=x_\ast+\gamma (x-x_\ast)+D(x_\ast)\eta_\ast.
\label{eq:binary}
\end{split}$$ The right-hand side of [(\[eq:binary\])]{} describes the modification of the opinion in terms of the quantity $\gamma(x_\ast-x)$ (respectively $\gamma(x_\ast-x)$), that measures the *compromise* between opinions with intensity $\gamma$, $0<\gamma<1$, and a random contribution, given by the random variable $\eta$ (respectively $\eta_\ast$), modelling stochastic fluctuations induced by the *self-thinking* of the agents. $D(\cdot)\geq 0$ is an opinion-dependent diffusion coefficient modulating the amplitude of the stochastic fluctuations, that is the variance of $\eta$ and $\eta_\ast$. In [@Tos06] the two random variables were assumed to be independent and identically distributed with zero mean and variance $\sigma^2$. Let us further set \[lam\] = . Then, interactions of type [(\[eq:binary\])]{} with small values of $\lambda$ characterize compromise dominated societies, while interactions with large values of $\lambda$ characterize self-thinking dominated societies.
Introducing the distribution function $f=f(t,\,x):{\mathbb R}_+\times [-1,\,1]\to{\mathbb R}_+$, such that $f(t,\,x)dx$ is the fraction of agents with opinion in $[x,\,x+dx]$ at time $t$, the binary rules give rise to a Boltzmann-type kinetic equation, that in weak form reads $$\begin{gathered}
\frac{d}{dt}\int_{-1}^1\varphi(x)f(t,\,x)\,dx \\
=\frac{1}{2}\int_{-1}^1\int_{-1}^1{\langle\varphi(x')+\varphi(x_\ast^\prime)-\varphi(x)-\varphi(x_\ast)\rangle}f(t,\,x)f(t,\,x_\ast)\,dx\,dx_\ast,
\label{eq:boltz}\end{gathered}$$ where $\varphi:[-1,\,1]\to{\mathbb R}$ is an arbitrary test function, i.e. any observable quantity depending on the microscopic state of the agents, and where we denoted by $\langle \cdot \rangle$ the mathematical expectation. Choosing $\varphi(x)=1$, one shows that the integral of $f$ with respect to $x$ is constant in time, i.e. that the total number of agents is conserved. This also implies that $f$ can be thought as a probability density for every $t>0$. Choosing instead $\varphi(x)=x$, and considering that [(\[eq:binary\])]{} implies $$\langle x^\prime+x^\prime_\ast \rangle = x + x_\ast,$$ one concludes that $$\frac{d}{dt}\int_{-1}^1 xf(t,\,x)\,dx= 0.
\label{eq:mean}$$ Therefore the mean opinion $m:=\int_{-1}^1 xf(t,\,x)\,dx$ is conserved in time. As shown in [@Tos06], one can recover an explicit expression of the asymptotic distribution function at least in the so-called *quasi-invariant regime*, i.e. the one in which the variation of the opinion in each binary interaction is small. To describe such a regime, one scales the parameters $\gamma$, $\sigma^2$ in as $$\gamma\to\epsilon\gamma, \qquad \sigma^2\to\epsilon\sigma^2,
\label{eq:scaling}$$ where $\epsilon>0$ is an arbitrarily small scaling coefficient. Moreover, to study the large time behavior of the system, one introduces the new time scale $t \to\epsilon t$ and scales the distribution function as $v(t,\,x):=f(\frac{t}{\epsilon},\,x)$. In this way, at every fixed $t>0$ and in the limit $\epsilon\to 0^+$, $v$ describes the large time trend of $f$. Moreover, as shown in [@Tos06], if $D(x) = \sqrt{1-x^2}$, $v(t,x)$ satisfies the Fokker–Planck equation [(\[op-FP\])]{}.
Since the value of $\lambda$ is left unchanged by the scaling [(\[eq:scaling\])]{} leading from the Boltzmann-type equation [(\[eq:boltz\])]{} to the Fokker–Planck type equation [(\[op-FP\])]{}, the parameter $\lambda$ maintains its meaning also in the target equation. The roles of the constants $\lambda$ and $m$ are evident also by looking at the shape of the steady Beta distribution [(\[beta\])]{}. We can observe that, by fixing for example $m>0$, increasing the values of $\lambda$, and consequently moving from a compromise dominated to a self-thinking dominated society, such a distribution may depict a transition from a strong consensus around the mean to a milder consensus, and further to a radicalisation in the extreme opinion $x=1$ up to the appearance of a double radicalisation in the two opposite extreme opinions $x=\pm 1$.
In view of the described social meaning, a relevant problem related to the solution to the Fokker–Planck equation [(\[op-FP\])]{} is to understand at which speed the solution $v(t)$ converges to its equilibrium configuration, and to reckon how this rate depends on the parameters $\lambda$ and $m$. Indeed, as outlined before, it is easily recognized that different values of these parameters give raise to situations in which the extremal opinions are not attracting, and this happens for $\lambda < 1-|m|$, or situations in which opinions are polarized around the extreme ones ($\lambda > 1+|m|$). Also, it is not clear if these (different) steady states are reached very quickly in time, independently of the values of the parameters.
As discussed in [@FPTT17], in analogy with the methods developed for the classical Fokker–Planck equation [@T99], the large-time behavior of the solution to [(\[op-FP\])]{} can be fruitfully studied by resorting to entropy methods [@MTV]. This corresponds to the study of the evolution in time of various Lyapunov functionals, the most known being the Shannon entropy of the solution relative to the steady state. We recall here that the relative Shannon entropy of two probability densities $f$ and $g$ supported on the bounded interval $\CI$ is [defined]{} by the formula \[relH\] H(f,g)= \_ f(x) dx. Note that $H(f,g)$ can be alternatively written as $$\int_\CI \left( \frac{f(x)}{g(x)} \log \frac{f(x)}{g(x)} - \frac{f(x)}{g(x)} +1\right ) g(x)dx,$$ which is the integral of a nonnegative function.
As shown in [@FPTT17], the relative entropy $H(v(t), v_{m,\lambda})$ decreases in time, and its time variation can be expressed by the *entropy production* term \[ep\] I(v(t), v\_[m,]{}) = \_ 2 (1-y\^2) (\_y )\^2 v(t,y) dy. While for the classical Fokker–Planck equation [@T99], exponential in time convergence at explicit rate follows in consequence of the logarithmic Sobolev inequality, the results in presence of the weight in [(\[ep\])]{} are less satisfactory. Various convergence results have been obtained in [@FPTT17] by resorting to a generalization of the so-called Chernoff inequality with weight, first proven by Klaassen [@Kla]. The main consequence of this inequality [@FPTT17], was to show that exponential convergence to equilibrium with an explicit rate holds at least for initial values $v_0$ for [(\[op-FP\])]{} close to the steady state [(\[beta\])]{} in the weighted $L^2$-norm \[L2\] v\_0-v\_[m,]{}\_\*\^2 := \_|v\_0(y)-v\_[m,]{}(y)|\^2 v\_[m,]{}(y)\^[-1]{} dy. Also, a weaker convergence result was proven for general initial data, by showing that the standard $L^1$-distance decays to zero at a polynomial rate (without any explicit rate of convergence).
Related results have been obtained by Epstein and Mazzeo in [@EM10] for the adjoint equation \[ad-FP\] = 2 (1-x\^2) - (x-m) , t>0,x . Indeed, the Fokker–Planck equation [(\[op-FP\])]{} is naturally coupled to [(\[ad-FP\])]{} since, at least formally, if $v$ is a solution of [(\[op-FP\])]{}, then \[rel1\] u(t,x) = is a solution of [(\[ad-FP\])]{} (remark that the notation we have chosen for the solutions $v(t,y)$ of [(\[op-FP\])]{} and $u(t,x)$ of is the same as in the paper [@EM10] to which we will often refer in the sequel of the paper). Among other results, in [@EM10] exponential convergence in $L^1(\CI)$ of $v(t)$ towards $v_{m,\lambda}$ has been proven (without rate) by resorting to classical analysis of semigroups.
In this paper we aim at proving that entropy methods can also produce exponential convergence in $L^1(\CI)$ towards equilibrium with an explicit rate, at least in some range of the parameters $\lambda$ and $m$. The result follows from a new weighted logarithmic-Sobolev inequality satisfied by the Beta functions [(\[beta\])]{} when they belong to $L^2(\CI)$. In this case, we will prove that there exists an explicitly computable constant $K_{m,\lambda} >0$ such that, for any probability density $\varphi \in L^1(\CI)$ absolutely continuous with respect to $v_{m,\lambda}$ \[ok\] H(, v\_[m,]{}) K\_[m,]{} I(, v\_[m,]{}). Inequality [(\[ok\])]{} requires that $\lambda >0$, $m\in \CI$ be such that $$1-\frac \lambda 2 >0, \quad {\rm if}\,\, m=0, \quad
1-\frac \lambda 2 \geq |m|, \quad {\rm if}\,\, m\neq 0.$$ and allows us to obtain exponential convergence in relative entropy with an explicitly computable rate.
In more details, this is the plan of the paper: we will start by recalling in Section \[sol\] an existence result for the initial-boundary value problem for the Fokker–Planck equation [(\[op-FP\])]{}, as follows from the analysis of Wright–Fisher type equations presented in [@EM10] for the adjoint equation [(\[ad-FP\])]{}. Then, the proof of the new logarithmic-Sobolev inequality for Beta functions and its consequences on the large-time behavior of the solution to equation [(\[op-FP\])]{} will be studied in Section \[LS\]. Last, in Sections \[dist\] and \[concl\] we will discuss the case $m=0$, $\lambda =1$ which leads to a uniform density at equilibrium, and we will address some concluding remarks.
Existence and properties of solutions {#sol}
=====================================
For given constants $\lambda >0$ and $m\in \CI$, let us consider the initial-boundary value problem \[main\] {
&\_t v(t,y)= 2 \_y\^2 ((1-y\^2) v(t,y)) +\_y ((y-m)v(t,y)),t>0,y\
&v(0,y)=v\_0(y) 0 L\^1(),
. with boundary conditions \[bc-mom\] \_[y-1\^+]{} (1-y\^2)v(t,y)= \_[y1\^-]{} (1-y\^2)v(t,y)= 0, t>0 and \[bc-mass\] {
&\_[y-1\^+]{} (y-m)v(t,y) + ((1-y\^2) v(t,y)) = 0, t>0\
&\_[y1\^-]{} (y-m)v(t,y) + ((1-y\^2) v(t,y)) = 0,t>0 .
. Conditions and are suggested by the nature of the problem, since they imply momentum and mass conservation of the (possible) solution to the Fokker–Planck equation. While condition [(\[bc-mom\])]{} is automatically satisfied for a sufficiently regular density $v$, condition [(\[bc-mass\])]{} requires an exact balance between the so-called advective and diffusive fluxes on the boundaries $y= \pm 1$. This condition is usually referred to as the *no-flux* boundary condition [@FPTT17].
The linear Fokker–Planck equation in [(\[main\])]{} has a variable diffusion coefficient and the variable $y$ belongs to the bounded interval $\CI$, and this requires to consider boundary conditions. An alternative formulation would be to consider the pure initial value problem on the whole real line, by introducing the diffusion coefficient $(1-y^2) \chi(\CI)$, where $\chi(X)$ denotes the characteristic function of the set $X\subseteq {\mathbb R}$. The initial value problem for Fokker–Planck equations with general non smooth coefficients has been recently considered by Le Bris and Lions [@LL08]. However, diffusion coefficients as $(1-y^2) \chi(\CI)$ are not included in their analysis, and the results in [@LL08] do not apply. For such a problem a general theory about existence, uniqueness and continuous dependence on initial data still does not exist.
On the other hand, a quite general theory has been recently developed by Epstein and Mazzeo in [@EM10] for the equation [(\[ad-FP\])]{}. Their results give some insight also on our Fokker–Planck equation [(\[op-FP\])]{}, subject to no-flux boundary conditions as given in [(\[bc-mass\])]{}.
Equation [(\[ad-FP\])]{} is a Wright–Fisher type equation, of the form $$\partial_t u(t,x)= a(x) \partial_x^2 u(t,x) +b(x) \partial_x u(t,x),\quad t>0,\quad x \in (A,B)$$ where $A$, $B \in {\mathbb R}$, $a\in C^\infty([A,B])$, $b \in C^\infty([A,B])$ with $$a(x)=(x-A)(B-x)\tilde a(x),\quad \tilde a \in C^\infty([A,B]), \quad \tilde a(x) >0 \text{\ for all \ } x\in [A,B],$$ and $$b(A)\geq 0,\quad b(B)\leq 0.$$ Since our results heavily depend on the precise analysis by Epstein and Mazzeo on the solutions of the Wright–Fisher–type equations, we collect in the next Theorem the results we need about these solutions. All the details can be extracted from [@EM10]. In the rest, we will use as usual the notation $\bar\CI = [-1,1]$.
\[EM\] For all constants $\lambda >0$ and $m\in \CI$ let us consider the initial-boundary value problem [(\[main\])]{} with no-flux boundary conditions, as given by [(\[bc-mass\])]{}. Then, there exists a kernel $q_t(x,y):\{t>0, x\in \bar\CI, y\in \CI\} \rightarrow \mathbb R$ such that \[Qt\] Q\_tv\_0(y):= \_[-1]{}\^1 q\_t(x,y) v\_0(x) dx is a classical solution of the Cauchy problem. The kernel $q_t(x,y)$ satisfies the properties
1) $q_t(x,y) \in C^\infty\left((0,\infty)\times \bar\CI\times \CI\right)$;
2) $q_t(x,y) >0$ on $(0,\infty)\times \bar\CI\times \CI$;
3) for $y\to -1^+$ we have $q_t (x,y) \sim (1+y)^{-1+\frac {1+m}\lambda } \varphi(t,x)$ for all $t>0$, $x\in \bar\CI$ with $\varphi \in C^\infty$;
4) for $y\to 1^-$ we have $q_t (x,y) \sim (1-y)^{-1+\frac {1-m}\lambda } \tilde \varphi(t,x)$ for all $t>0$, $x\in \bar\CI$ with $\tilde \varphi \in C^\infty$;
5) for all $t>0$ and all $x\in \bar\CI$ we have $$\begin{aligned}
&\lim_{y\to -1^+} \left(\frac \lambda 2 \partial_y \left((1-y^2) q_t(x,y)\right ) +(y-m)q_t(x,y)\right )=0\\
&\lim_{y\to 1^-} \left(\frac \lambda 2 \partial_y \left((1-y^2) q_t(x,y)\right ) +(y-m)q_t(x,y)\right )=0.
\end{aligned}$$
As a consequence, the solution $v(t,y)= Q_tv_0(y)$ satisfies
1. $v(t,y)\in C^\infty\left((0,\infty)\times \CI\right)$;
2. $v(t,y) >0$ on $(0,\infty)\times \CI$;
3. for $y\to -1^+$ we have $v(t,y) \sim (1+y)^{-1+\frac {1+m}\lambda } \psi(t)$ for all $t>0$ with $\psi\in C^\infty$;
4. for $y\to 1^-$ we have $v(t,y) \sim (1-y)^{-1+\frac {1-m}\lambda } \tilde \psi(t)$ for all $t>0$ with $\tilde \psi \in C^\infty$;
5. for all $t>0$ we have (no flux boundary conditions) $$\begin{aligned}
&\lim_{y\to -1^+} \left(\frac \lambda 2 \partial_y \left((1-y^2) v(t,y)\right ) +(y-m)v(t,y)\right )=0\\
&\lim_{y\to 1^-} \left(\frac \lambda 2 \partial_y \left((1-y^2) v(t,y)\right ) +(y-m)v(t,y)\right )=0.
\end{aligned}$$
Moreover, $v\in C((0,\infty), L^1(\CI))$ and $$\lim_{t\to 0^+} \|v(t) -v_0\|_{L^1} =0.$$
In consequence of the validity of no-flux boundary conditions (property *5’)*) conservation of mass follows. Hence, since $v_0$ is a probability density, the solution $v(t)= Q_tv_0$ remains a probability density for all $t>0$. Indeed $$\begin{aligned}
\frac d{dt} \int_{-1}^1 v(t,y) dy & = \int_{-1}^1 \partial_t v(t,y) dy = \int_{-1}^1 \partial_y \left(\frac \lambda 2 \partial_y \left((1-y^2) v(t,y)\right ) +(y-m)v(t,y)\right )dy\\
& = \left[\frac \lambda 2 \partial_y \left((1-y^2) v(t,y)\right ) +(y-m)v(t,y)\right ]_{-1}^1 =0.
\end{aligned}$$ The steady states for equation are given by the Beta densities [(\[beta\])]{}.
Some remarks are in order. First of all, by means of $\emph{3')}$ and $\emph{4')}$ of Theorem \[EM\] we conclude that, for any given initial datum $v_0$ that is a probability density, the solution $v(t)=Q_tv_0$ has the same behavior at the boundary of $\CI$ of the corresponding steady state $v_{m,\lambda}$.
Consequently, in reason of the regularity of both functions, the probability density $v(t)$, solution of the initial value problem, is absolutely continuous with respect to the steady state $v_{m,\lambda}$ for all times $t >0$, \[ac\] C\_b\^() and it can be continuously extended to $\bar\CI$. In addition, if the condition \[cond1\] 1->|m| is satisfied, both the steady state and the solution $v(t)$ vanish on the boundary of the domain.
Weighted logarithmic-Sobolev inequalities and large time behavior. {#LS}
==================================================================
As briefly discussed in the Introduction, our main goal is concerned with the study of the large-time behavior of the solution to the Fokker–Planck equation [(\[op-FP\])]{}. This problem has been considered by Epstein and Mazzeo [@EM10], who studied the large-time behavior of equation [(\[ad-FP\])]{}, and used this to prove exponential convergence in $L^1$ for large times of the solution $v(t)=Q_t v_0$ of the Cauchy problem to the corresponding steady state $v_{m, \lambda}$ for the whole range of the allowed parameters $m\in \CI$ and $\lambda >0$. While their result, obtained by classical semigroup arguments is very general, the rate of the exponential convergence was not explicitly computed. A stronger result was recently obtained in [@FPTT17]. This result has been shown to hold for a large class of Fokker–Planck equations with non constant diffusion coefficients and bounded domains, by resorting to classical entropy type inequalities. Different Lyapunov functionals can be actually evaluated along the solution of the Fokker–Planck equation [(\[op-FP\])]{} and, in presence of some regularity of the solution itself, can be proven to be monotone decreasing in time. Among them, the relative Shannon entropy defined in [(\[relH\])]{}, the Hellinger distance, the reverse relative Shannon entropy, and the weighted $L^2$-distance.
Thanks to Theorem \[EM\], we know that the solution of the opinion formation equation fulfills the conditions which allow the application of the formal results contained in [@FPTT17]. In particular, the following result about exponential convergence to equilibrium follows.
\[l2\] Let $\lambda >0$ and $m\in \CI$ . Let $v_0$ a probability density satisfying \[vic\] v\_0-v\_[m,]{} \_\*\^2 = \_ dy < where $v_{m,\lambda}$ is the stationary solution of the Fokker–Planck equation . Then, the solution $v(t,y)= Q_tv_0(y)$ of defined in converges exponentially in time towards the steady state, and the following holds true \[L22\] v(t)-v\_[m,]{} \_\*\^2 e\^[-2t]{} v\_0-v\_[m,]{} \_\*\^2 , t >0.
Inequality [(\[L22\])]{} implies exponential convergence in $L^1$. Indeed by Cauchy–Schwartz inequality, for any pair $f$, $g$ of probability densities on $\CI$ it holds $$\begin{aligned}
\int_\CI |f(y)-g(y)|\, dy &= \int_\CI \frac{|f(y)-g(y)| }{\sqrt{v_{m,\lambda}(y)} }\sqrt{v_{m,\lambda}(y)}\, dy\\
&\leq
\left( \int_\CI \frac {\left(f(y)-g(y)\right )^2 }{v_{m,\lambda}(y)} \, dy \right )^{\frac 12}\left( \int_\CI {v_{m,\lambda}}(y) \, dy\right )^{\frac 12}\\
&\leq
\left( \int_\CI \frac {\left(f(y)-g(y)\right )^2 }{v_{m,\lambda}(y)} \, dy \right )^{\frac 12}.
\end{aligned}$$ Hence, implies \[conv-expL1-L2\] v(t)-v\_[m,]{}\_[L\^1]{} e\^[-t]{} ( \_ dy)\^[12]{} for the whole set of allowed parameters $m\in \CI$ and $\lambda >0$.
It is important to outline that condition [(\[vic\])]{}, at least when $v_{m,\lambda}$ is equal to zero at the boundaries, is quite restrictive, and requires the initial data $v_0$ to be very close to the steady state. On the contrary, if $\left(v_{m,\lambda}\right )^{-1}$ is bounded (and this happens when $\lim_{y \to -1^+} v_{m,\lambda}(y)= \lim_{y\to 1^-}v_{m,\lambda}(y)=+\infty$), condition [(\[vic\])]{} is satisfied any time $v_0$ is close to $v_{m,\lambda}$ in the $L^2$ distance.
In what follows, we will prove that exponential convergence in $L^1$ can be obtained also for initial values more general than the ones satisfying Theorem \[l2\]. To this extent, we will show that the Beta functions [(\[beta\])]{}, in a certain well defined range of the parameters $\lambda$ and $m$, satisfy a weighted logarithmic-Sobolev inequality. The result allows us to apply to our Fokker–Planck equation for opinion formation the same strategy one can apply to the classical Fokker–Planck equation [@AMTU].
Let us briefly recall the main steps of the (entropy) method for the classical one-dimensional Fokker–Planck equation. Given the initial value problem \[FP\] {
&\_t f(t,x)= \_x\^2 f(t,x) + \_x(xf(t,x)),x[R]{}, t>0\
&f(0,x)=f\_0(x) 0 L\^1([R]{})
. where the initial value is a probability density function, one studies the evolution of the relative entropy functional $H(f(t), M)$, given by \[entr-class\] H(f(t), M) = \_[R]{}f(t,x) dx where $M$ is the Maxwellian (Gaussian) \[Maxw\] M(x)= 1 e\^[-]{}, which can be easily recognized as the unique steady state of equation [(\[FP\])]{}. It is well known (cf. for example [@McK66]) that, if $f(t)$ is a solution of the Cauchy problem , the relative entropy is monotone nonincreasing, and its time derivative is given by \[derivata\] d[dt]{} H(f(t), M) = - I(f(t), M), t>0 where $I(f(t), M)$ is the relative Fisher information (the entropy production) defined as \[fisher-class\] I(f(t), M) = \_[[R]{}]{} ( \_x )\^2 f(t,x) dx. Relation coupled with the logarithmic-Sobolev inequality (cf. for example [@T99]) $$H(f(t), M) \leq \frac 12 I (f(t),M),\quad t>0$$ leads to the exponential decay to zero of the relative entropy [@T99; @T13] with explicit rate. Last, resorting to the well-known Csiszár–Kullback–Pinsker inequality [@C] \[CK\] f-g\_[L\^1]{}\^2 2 H(f,g), f,gL\^1 one obtains exponential convergence in $L^1$ to the Maxwellian density (always with sub-optimal explicit rate).
Going back to our problem, let us assume that the entropy of the initial value relative to the Beta steady state is bounded \[Hfinita\] H(v\_0, v\_[m,]{}) <. Evaluating the time derivative of the relative entropy (cf. the computations in [@FPTT17]), one obtains for the solution to the Fokker–Planck equation a relation analogous to , which now reads \[derivata-peso\] d[dt]{} H(v(t), v\_[m,]{}) = - I(v(t), v\_[m,]{}), t>0. In [(\[derivata-peso\])]{} $\tilde I$ defines the weighted Fisher information \[w-fisher\] I(v(t), v\_[m,]{}) = \_2 (1-y\^2) (\_y )\^2 v(t,y) dy. As one can easily verify, the weight $\frac \lambda 2 (1-y^2)$ is due to the variable diffusion coefficient in equation . It is clear that, if one can prove that, for some universal constant $C >0$ the relative entropy is bounded by $$H(v, v_{m,\lambda}) \leq C \tilde I(v, v_{m,\lambda}),$$ one obtains, as in the classical case, the exponential convergence to equilibrium of the relative entropy of the solution at the explicit rate $C$.
We prove indeed that the following holds.
\[LS-theo\] Let $\lambda >0$, $m\in \CI$ be such that \[cond\]
& 1-2 >0, m=0\
& 1-2 |m|, m0.
and let $v_{m,\lambda}$ be the Beta function on $\CI$ defined by [(\[beta\])]{}. Then, there exists an explicit constant $K_{m,\lambda} >0$ such that, for any probability density $\varphi \in L^1(\CI)$ absolutely continuous with respect to $v_{m,\lambda}$ it holds \[LS-peso\] H(, v\_[m,]{}) K\_[m,]{} I(, v\_[m,]{}). The constant $K_{m,\lambda} >0$ is explicitly computable and equals \[costante\] K\_[m,]{} = (1-2 + )\^[-1]{}.
\[remark\] It is worth underlying that conditions are equivalent to the condition that the corresponding Beta-type function $v_{m,\lambda}$ belongs to $L^2(\CI)$.
A direct consequence of Theorem \[LS-theo\] is the following
Let the parameters $\lambda >0$, $m\in \CI$ satisfy the conditions [(\[cond\])]{} of Theorem \[LS-theo\], and let $v(t)=Q_tv_0$ be the solution to the initial-boundary value problem [(\[main\])]{} with no-flux boundary conditions, and initial data $v_0 \in L^1(\CI)$ a probability density such that the relative entropy $H(v_0, v_{m,\lambda})$ is finite. Then, the relative entropy decays exponentially to zero at an explicit rate, and \[conv\] v(t) -v\_[m,]{}\_[L\^1]{} 2 e\^[- 1 [2K\_[m,]{}]{} t]{} . In [(\[conv\])]{} $K_{m,\lambda}>0$ is given by [(\[costante\])]{}.
We already stressed in that, starting from the initial condition $v_0$, the result by Epstein and Mazzeo implies that the solution $v(t)=Q_t v_0$ defined in is absolutely continuous with respect to $v_{m,\lambda}$ for all $t>0$. Therefore, we can apply and then the weighted logarithmic-Sobolev inequality with $\varphi(y)= v(t,y)$ for all $t>0$ to get $$\frac d{dt} H(v(t), v_{m,\lambda}) \leq - \frac 1{K_{m,\lambda} }H(v(t), v_{m,\lambda}), \quad t>0$$ and this gives $$H(v(t), v_{m,\lambda}) \leq e^{-\frac 1{K_{m,\lambda} } t} H(v_0, v_{m,\lambda}).$$ Then by the Csiszár–Kullback–Pinsker inequality we obtain \[conv-expL1-logsob\] v(t) -v\_[m,]{}\_[L\^1]{} 2 e\^[- 1[2 K\_[m,]{} ]{} t]{} , t>0.
Let us come back to the proof of Theorem \[LS-theo\]. The starting point is the well known Bakry–Emery result about logarithmic-Sobolev inequality.
Let $M$ be a smooth, complete manifold and let $d\nu =e^{-\Psi}dx$ be a probability measure on $M$, such that $\Psi \in C^2(M)$ and $D^2\Psi + Ric \geq \rho I_n$, $\rho>0$. Then, for every probability measure $\mu$ absolutely continuous with respect to $\nu$, we have \[BE\] H(,) 1[2]{} I(,) where $$H(\mu,\nu)=\int_M \log \frac {d \mu}{d \nu} d\mu$$ and $$I(\mu, \nu) = \int_M \left|\nabla \log \frac{d \mu}{d \nu} \right |^2 d \mu .$$
If $M=[a,b]$ is an interval of the real line, $d \nu = g dx$ and $ d\mu= f dx$, with $f$ and $g$ probability densities, the assumptions in Bakry–Emery criterion read as follows
&g(x)=e\^[-(x)]{},\
&C\^2(\[a,b\])\
&\_[\[a,b\]]{}”(x)>0.
Then, for any $f$ probability density on $[a,b]$ absolutely continuous with respect to $g$, inequality becomes \[BE-int\] \_a\^b f(x) dx 1[2]{} \_a\^b (d[dx]{} )\^2f(x) dx. Of course this is a non–weighted logarithmic-Sobolev result. We are going to identify who will play the role of $\mu$ and $\nu$. If we take $M=\CI$ and $\nu= v_{m,\lambda} dx$ then two problems appear. The first one is that the open interval $\CI$ is not a complete manifold and the other one is that even if we prove that $v_{m,\lambda}(y) = e^{- \Psi (y)}$ with $\Psi$ satisfying Bakry–Emery Theorem, then for any $\varphi$ probability density absolutely continuous with respect to $v_{m,\lambda}$ we would get the logarithmic-Sobolev inequality $$\int_{-1}^1 \varphi(x) \log \frac {\varphi(x)}{v_{m,\lambda}(x)} dx \leq \frac 1{2\rho} \int_{-1}^1 \left( \frac d{dx} \log \frac {\varphi(x)}{v_{m,\lambda}(x)}\right )^2 \varphi(x) dx.$$ This is not enough to obtain since $\frac \lambda 2 (1-y^2) \leq 1$ for $\lambda <2$ (which is implied by conditions ). It turns out that actually $v_{m,\lambda}$ satisfies $v_{m,\lambda}(y) = e^{- \Psi (y)}$ with $\Psi$ fulfilling Bakry–Emery conditions. Since we are going to prove a stronger inequality in a different way, we leave the details to the interested reader. [**Proof of Theorem \[LS-theo\].**]{} The main idea is to resort to a change of variable which transforms the weighted logarithmic-Sobolev inequality we are looking for into a usual logarithmic-Sobolev inequality for a different probability density which satisfies the assumptions of the Bakry–Emery criterion. Given the partial differential equation $$\partial_t v(t,y)= \frac \lambda 2 \partial_y^2 \left((1-y^2) v(t,y)\right ) +\partial_y \left((y-m)v(t,y)\right ),\quad t>0,\quad y \in \CI$$ with steady state $v_{m,\lambda}$, its adjoint equation reads \[main-adj\] \_t u(t,x)= 2 (1-x\^2) \_x\^2 u(t,x) - (x-m)\_x u(t,x), t>0,x . If we now set in [(\[main-adj\])]{} $$f(t,s)=u(t,x)$$ where $$\frac {ds}{dx} = \frac 1{\sqrt {1-x^2}}, \quad x\in \CI,$$ equation is transformed into a Fokker–Planck equation with constant diffusion, given by \[FP-adj-free\] \_t f(t,s)= 2 \_s\^2 f(t,s) - \_s f(t,s), t>0, s(-2, 2). The adjoint equation of is in turn \[FP-free\] \_t g(t,z)= 2 \_z\^2 g(t,z) + \_z ( g(t,z)), t>0, z(-2, 2). We denote \[w’\] W\_[m,]{}’(z):= and $$W_{m,\lambda}(z)= \int _{0}^z W_{m,\lambda}'(\sigma) d \sigma.$$ The steady states of Equation are \[exp\] g\_[m,]{}(z)= C\_[m,]{} e\^[-2 W\_[m,]{}(z)]{} = e\^[-(2 W\_[m,]{}(z) -P\_[m,]{} )]{} for $C_{m,\lambda} >0$ as in and explicitly \[beta-2\] g\_[m,]{}(z)= C\_[m,]{} 1[(z)\^[ 1-2]{}]{} . One can check that $$\begin{aligned}
& g_{m,\lambda}(z) \sim R_{m,\lambda}\left( \frac \pi 2-z\right )^{\frac 2\lambda -1-\frac {2m}\lambda},\quad z\to \frac \pi 2^-\\
& g_{m,\lambda}(z) \sim \tilde R_{m,\lambda}\left( \frac \pi 2+z\right )^{\frac 2\lambda -1+\frac {2m}\lambda},\quad z\to -\frac \pi 2^+
\end{aligned}$$ with $R_{m,\lambda}$, $\tilde R_{m,\lambda}$ positive constants. Moreover, we have \[stati-staz\] = v\_[m,]{}(y),ywith $v_{m,\lambda}$ as in or, equivalently, $$g_{m,\lambda} (z)= v_{m,\lambda}(\sin z) \cos z, \quad z\in \left( -\frac \pi 2, \frac \pi 2\right ).$$ It is immediate to show that $g_{m,\lambda}$ satisfies the assumptions of Bakry–Emery criterion on $ \left(-\frac \pi 2, \frac \pi 2\right )$. Since the latter is an open interval (and so it is not a complete manifold), we will overcome this difficulty by a suitable approximation argument. Resorting to , we need to evaluate $\frac 2\lambda W_{m,\lambda}''(z)$. We obtain $$\frac 2\lambda W_{m,\lambda}''(z)= \frac 2\lambda \frac d{dz} W_{m,\lambda}'(z) =\frac 2\lambda \frac{\left(1-\frac \lambda 2\right )+ m\sin z}{\cos^2 z}, \quad z\in \left( -\frac \pi 2, \frac \pi 2\right ).$$ Therefore, provided $1-\frac \lambda 2 \geq |m|$, $$\inf_{ \left( -\frac \pi 2, \frac \pi 2\right )} W_{m,\lambda}''(z) \ge 0$$ and the function $W_{m,\lambda}(z)$ is convex on $ \left( -\frac \pi 2, \frac \pi 2\right )$. If $m=0$, \[rho\_0\] \_[( -2, 2)]{} W\_[0,]{}”(z)= W\_[0,]{}”(0)= 1-2:= \_[0,]{}. Consequently, in order to apply Bakry–Emery criterion, we have to assume $1-\frac \lambda 2 >0$.
Let us now set $m \not=0$. Since $$\frac d {dz} W_{m,\lambda}''(z)= \left(-\frac 1{\cos^3 z}\right )\left(m\sin^2 z +(\lambda-2)\sin z+m\right ),$$ for any given $m\in \CI$, $m\neq 0$ and $\lambda$ such that $1-\frac \lambda 2 \geq |m|$, there exists $\bar z_{m,\lambda} \in \left(-\frac \pi 2,\frac \pi 2\right )$ such that \[rho\_mlambda\] \_[( -2, 2)]{} W\_[m,]{}”(z)= W\_[m,]{}”(|z\_[m,]{})= 12 (1-2 + ) := \_[m,]{}>0. If we could apply Bakry–Emery criterion directly on $ \left(-\frac \pi 2,\frac \pi 2\right )$ we would obtain, for all $f$ probability densities on $ \left(-\frac \pi 2,\frac \pi 2\right )$ absolutely continuous with respect to $g_{m,\lambda}$, the logarithmic Sobolev inequality $$\int_{-\frac \pi 2}^{\frac \pi 2} f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} dz \leq \frac \lambda{4 \rho_{m,\lambda}} \int_{-\frac \pi2}^{\frac \pi 2} \left(\frac d{dz} \log \frac{f(z)}{g_{m,\lambda}(z)} \right )^2f(z) dz,$$ where the explicit constants $\rho_{m,\lambda}$ are defined in and . Since $\left(-\frac \pi 2,\frac \pi 2\right )$ is not a complete manifold we perform an approximation argument. Let us fix $m\in \CI$ and $\lambda >0$ satisfying and let $f$ be a probability density on $\left(-\frac \pi 2,\frac \pi 2\right )$ absolutely continuous with respect to $g_{m,\lambda}$. For $\eps >0$ let us define $$\begin{aligned}
&
f_\eps=\frac 1{A_\eps} f\chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}, \text{ with } A_\eps = \int_{-\frac \pi 2 +\eps} ^{\frac \pi 2 -\eps} f(z) dz\\
&
g_\eps = \frac 1{B_\eps}g_{m,\lambda}\chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}, \text{ with } B_\eps = \int_{-\frac \pi 2 +\eps} ^{\frac \pi 2 -\eps} g_{m,\lambda}(z) dz.\end{aligned}$$ Of course $f_\eps$ and $g_\eps$ are probability densities and $A_\eps \to 1$, $B_\eps \to 1$ for $\eps \to 0$. Moreover by $$g_\eps (z)= e^{-\left(\frac 2 \lambda W_{m,\lambda}(z) -\log P_{m,\lambda} + \log B_\eps \right )} \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z)$$ and $f_\eps$ is absolutely continuous with respect to $g_\eps$ on $\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]$. For all $\eps >0$ we have $$\frac {d^2}{dz^2} \left(\frac 2 \lambda W_{m,\lambda}(z) -\log P_{m,\lambda} + \log B_\eps\right ) = \frac 2 \lambda W''_{m,\lambda}(z) \geq \frac 2\lambda \rho_{m,\lambda}.$$ Since $g_\eps$ satisfies the assumptions of Bakry–Emery criterion on $\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]$, we get for all $\eps >0$ \[BE-approx\] \_[-2+]{}\^[2-]{} f\_(z) dz \_[-2+]{}\^[2-]{} (d[dz]{} )\^2f\_(z) dz. Now assume that \[Fisher-bounded\] \_[-2]{}\^[2]{} (d[dz]{} )\^2f(z) dz <. As far as the right hand side of is concerned, by Lebesgue’s dominated convergence theorem we get for $\eps \to 0$ $$\begin{aligned}
&\int_{-\frac \pi2+\eps}^{\frac \pi 2-\eps} \left(\frac d{dz} \log \frac{f_\eps(z)}{g_\eps(z)} \right )^2f_\eps(z) dz\\
& = \frac 1 {A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2} \left(\frac d{dz} \log \left(\frac {f(z)}{A_\eps} \frac {B_\eps}{g_{m,\lambda}(z)} \right ) \right )^2 f(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}dz\\
& = \frac 1 {A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2} \left(\frac d{dz} \left(\log \frac{f(z)}{g_{m,\lambda}(z)} + \log \frac {B_\eps}{A_\eps}\right ) \right )^2 f(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz\\
& = \frac 1 {A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2} \left(\frac d{dz} \log \frac{f(z)}{g_{m,\lambda}(z)} \right )^2 f(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz \to
\int_{-\frac \pi2}^{\frac \pi 2} \left(\frac d{dz} \log \frac{f(z)}{g_{m,\lambda}(z)}\right )^2f(z) dz.
\end{aligned}$$ Letting $\eps\to 0$, for the left hand side we obtain $$\begin{aligned}
&\int_{-\frac \pi 2+\eps}^{\frac \pi 2-\eps} f_\eps(z) \log \frac {f_\eps(z)}{g_\eps(z)} dz = \int_{-\frac \pi 2}^{\frac \pi 2} \frac {f(z)}{A_\eps} \log \left(\frac {f(z)}{A_\eps} \frac {B_\eps}{g_{m,\lambda}(z)} \right ) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z)dz\\
& = \frac 1{A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2}f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz +
\frac 1{A_\eps} \log \frac {B_\eps} {A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2} f(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz \\
&\to \int_{-\frac \pi 2}^{\frac \pi 2} f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} dz.
\end{aligned}$$ Indeed, by Lebesgue’s dominated convergence theorem $$\frac 1{A_\eps} \log \frac {B_\eps} {A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2} f(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz \to 0, \quad \eps \to 0,$$ and thanks to the identity $$\begin{aligned}
& \int_{-\frac \pi 2}^{\frac \pi 2}f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz\\
&= \int_{-\frac \pi 2}^{\frac \pi 2}\left( \frac{f(z)}{ g_{m,\lambda}(z)} \log \frac {f(z)}{g_{m,\lambda}(z)} -\frac{f(z)}{ g_{m,\lambda}(z)} +1\right ) g_{m,\lambda}(z) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz\\
&\quad\quad + \int _{-\frac \pi 2}^{\frac \pi 2}\left(f(z)- g_{m,\lambda}(z)\right ) \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz,
\end{aligned}$$ by the Lebesgue’s dominated and monotone convergence theorems we conclude $$\frac 1{A_\eps} \int_{-\frac \pi 2}^{\frac \pi 2}f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} \chi_{\left[-\frac \pi 2+\eps,\frac \pi 2-\eps \right]}(z) dz \to \int_{-\frac \pi 2}^{\frac \pi 2}f(z) \log \frac {f(z)}{g_{m,\lambda}(z)} dz, \quad \eps \to 0.$$ Finally, for all $f$ probability densities on $ \left(-\frac \pi 2,\frac \pi 2\right )$ absolutely continuous with respect to $g_{m,\lambda}$ it holds \[logg\] \_[-2]{}\^[2]{} f(z) dz \_[-2]{}\^[2]{} (d[dz]{} )\^2f(z) dz, where $\rho_{m,\lambda}$ are defined as in and . Going back to the original functions, by means of the change of variables $$z=\arcsin y$$ the logarithmic Sobolev inequality [(\[logg\])]{} transforms into a weighted logarithmic-Sobolev inequality. In fact, for any $f$ probability density on $\left(-\frac \pi 2,\frac \pi 2\right )$ absolutely continuous with respect to $g_{m,\lambda}$ $$\begin{gathered}
\int_{-1}^{1} f(\arcsin y) \log \frac {f(\arcsin y)}{g_{m,\lambda}(\arcsin y)} \frac 1{\sqrt{1-y^2}} dy \leq \\
\frac 1{2\rho_{m,\lambda}} \int_{-1}^{1} \frac \lambda 2 \left(\frac d{dy} \log \left(\frac{f(\arcsin y)}{g_{m,\lambda}(\arcsin y)}\right ) \sqrt{1-y^2} \right )^2f(\arcsin y) \frac 1{\sqrt{1-y^2}} dy.\end{gathered}$$ Now, by we get $$\begin{gathered}
\int_{-1}^{1} \frac {f(\arcsin y) } {\sqrt{1-y^2}} \log \frac { \frac {f(\arcsin y) } {\sqrt{1-y^2}} }{v_{m,\lambda}(y)} dy \leq \\
\frac 1{2\rho_{m,\lambda}} \int_{-1}^{1} \frac \lambda 2(1-y^2) \left(\frac d{dy} \log \left(\frac{ \frac {f(\arcsin y) } {\sqrt{1-y^2}} }{v_{m,\lambda}(y)}\right ) \right )^2 \frac {f(\arcsin y) }{\sqrt{1-y^2}} dy.\end{gathered}$$ In order to complete the proof of inequality it enough to observe that $\varphi \in L^1(\CI)$ is a probability density absolutely continuous with respect to $v_{m,\lambda}$ if and only if $\varphi(y)=\frac {f(\arcsin y) } {\sqrt{1-y^2}}$ with $f \in L^1 \left(\left(-\frac \pi 2,\frac \pi 2\right )\right )$ is a probability density absolutely continuous with respect to $g_{m,\lambda}$. Inequality is then proven with \[Kmlambda\] K\_[m,]{} = 1[2\_[m,]{}]{}.
$\square$
It is worth comparing the results of exponential convergence in $L^1$ contained in and . By , for $m$ and $\lambda$ satisfying conditions we get : $$\|v(t) -v_{m,\lambda}\|_{L^1} \leq \sqrt 2 e^{- \rho_{m,\lambda} t} \sqrt {H(v_0, v_{m,\lambda})}, \quad t>0,$$ with $\rho_{m,\lambda}$ as in . On the other hand, for all $m\in \CI$ and $\lambda >0$ we get : $$\left \| v(t)-v_{m,\lambda}\right \|_{L^1} \leq e^{-t} \left( \int_{-1}^1 \frac {(v_0(y)-v_{m,\lambda}(y))^2}{v_{m,\lambda}(y)} \, dy\right )^{\frac 12},\quad t>0.$$ Since $\rho_{m,\lambda}\leq 1$ for all $m$, $\lambda$ satisfying conditions , the rate of exponential convergence in the second estimate is sharper than the first one. Let us compare now the assumptions $$H(v_0, v_{m,\lambda}) = \int_{-1}^1 v_0(y)\log \frac {v_0(y)}{v_{m,\lambda}(y)}\, d y< \infty$$ and $$\int_{-1}^1 \frac {(v_0(y)-v_{m,\lambda}(y))^2}{v_{m,\lambda}(y)} dy <\infty$$ for the values of the parameters which fulfill conditions . Since $$x\log x \geq x-1 + \frac 12 (x-1)^2\chi_{\left\{x\leq 1\right\} }(x), \quad x>0$$ we get $$\begin{aligned}
&\int_{-1}^1 v_{m,\lambda}(y) \frac {v_0(y)}{v_{m,\lambda}(y)}\log \frac {v_0(y)}{v_{m,\lambda}(y)}\, d y\\
& \geq \int_{-1}^1 v_{m,\lambda}(y)\left( \frac {v_0(y)}{v_{m,\lambda}(y)} -1\right )\, d y + \frac 12 \int_{-1}^1 v_{m,\lambda}(y)\left( \frac {v_0(y)}{v_{m,\lambda}(y)} -1\right )^2 \chi_{\left\{v_0(y) \leq v_{m,\lambda}(y) \right \}}(y) \, d y\\
&= \frac 12 \int_{-1}^1 \frac {(v_0(y)-v_{m,\lambda}(y))^2}{v_{m,\lambda}(y)} \chi_{\left\{v_0(y) \leq v_{m,\lambda}(y) \right \}}(y) \, d y.
\end{aligned}$$ So for $v_0 \leq v_{m,\lambda}$ the rate of convergence contained in is stronger than that in . Moreover, $$x\log x \leq x-1 + \frac 12 (x-1)^2, \quad x\geq 1$$ and so for $v_0 \geq v_{m,\lambda}$ we get $$\frac 12 \int_{-1}^1 \frac {(v_0(y)-v_{m,\lambda}(y))^2}{v_{m,\lambda}(y)} \, d y \geq \int_{-1}^1 v_{m,\lambda}(y) \frac {v_0(y)}{v_{m,\lambda}(y)}\log \frac {v_0(y)}{v_{m,\lambda}(y)}\, d y.$$ In this case, the convergence obtained by the new weighted logarithmic-Sobolev inequality could be the only one available. Of course, in all the other cases the two conditions seem not to be comparable.
A distinguished case {#dist}
====================
From Theorem \[LS-theo\] one can extract some interesting consequences. The case $m =0$, $\lambda = 1$ corresponds to the uniform density $$v_{0,1}(x) = \frac 12, \quad x \in \CI.$$ Hence, considering that $K_{0,1} = 1$, for a given probability density $h$ on $\CI$, inequality [(\[LS-peso\])]{} takes the form \[poi\] \_h(x) h(x) dx + 2 12 \_(1-x\^2) dx. A more suitable form is obtained by setting $h(x) = f^2(x)$ into [(\[poi\])]{}. One obtains the inequality \[LS-d\] \_f\^2(x) f\^2(x) dx + 2 2 \_(1-x\^2) (f’(x))\^2 dx, satisfied by all functions $f$ in $L^2(\CI)$ of $L^2$-norm equal to one. Inequality [(\[LS-d\])]{} is the analogous of the standard *Euclidean logarithmic-Sobolev inequality* established in Gross [@Gro], which in one-dimension reads \[LSI\] \_[R]{}f\^2(x) f\^2(x) dx + 12 (2e\^2) 2 \_[R]{}(f’(x))\^2 dx, and it is valid for all functions $f$ such that $$\int_{\mathbb R}f(x)^2\, dx = \int_{\mathbb R}x^2f^2(x) \, dx = 1.$$ Note that the main difference between the logarithmic-Sobolev inequality [(\[LSI\])]{} and the new inequality [(\[LS-d\])]{}, apart from the different interval of integration, is the presence of the weight on the right-hand side.
Clearly, the constraint $\|f\|_2 = 1$ can be easily cut to give the (general) inequality \[poi1\] \_w\^2(x) w\^2(x) dx - w\_2\^2 2 2 \_(1-x\^2) (w’(x))\^2 dx, which is valid for any function $w \in L^2(\CI)$.
Numerical experiments {#nume}
=====================
In this short Section, we will focus on some numerical experiments that illustrate the time-evolution of the weighted logarithmic Sobolev inequality [(\[LS-peso\])]{} for various values of the parameter $\lambda$, and $m =0$. To this extent, we make use of numerical schemes for the Fokker–Planck equation [(\[op-FP\])]{}, recently considered in [@PZ], that preserve the structural properties, like non negativity of the solution, entropy dissipation and large time behavior. These properties are essential for a correct description of the underlying physical problem.
The experiments have been done by choosing as initial density a bimodal normal distribution centered in $\pm 1/2$, normalized in the interval $(-1,1)$. It is clearly shown in Figure [(\[fig:test1\])]{} that inequality [(\[LS-peso\])]{} gives a better approximation to the entropy decay towards equilibrium for small values of the parameter. In all cases, however, exponential in time decay follows.
\
In Figure [(\[fig:test2\])]{} it is shown that the numerical method correctly reproduce the equilibrium Beta density [(\[beta\])]{} of the Fokker–Planck equation [(\[op-FP\])]{} for any value of the parameter $\lambda$.
\
Conclusions {#concl}
===========
In this paper, we investigated the large-time behavior of the solution of a Fokker–Planck type equation arising in the study of opinion formation. The same equation, in adjoint form, is well-known under the name of Wright–Fisher equation, and has been exhaustively studied, among others, in a recent paper by Epstein and Mazzeo [@EM10] from the point of view of semigroup theory. Our approach to the analysis of the large-time behavior of the solution is different, and relies on the classical study of the evolution of the relative Shannon entropy, which is of common use in the field of kinetic theory. The study of lower bounds for the relative entropy production leads to a new type of logarithmic-Sobolev inequality with weight, satisfied by the Beta-type densities, which allow us in various cases to conclude with exponential convergence to the equilibrium with an explicit rate.
The case in which the Beta-type density reduces to a uniform variable separates in a natural way from the others, and gives rise to the corresponding of the Euclidean logarithmic-Sobolev inequality.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work has been written within the activities of GNFM and GNAMPA groups of INdAM (National Institute of High Mathematics).
The support of the Italian Ministry of Education, University and Research (MIUR) through the “Dipartimenti di Eccellenza Program (2018–2022)” - Dept. of Mathematics “F. Casorati”, University of Pavia, is kindly acknowledged.
The authors also kindly acknowledge R. Mazzeo for fruitful explanations on the paper [@EM10], and M. Zanella, who performed the numerical experiments of Section \[nume\] by means of the entropic numerical scheme introduced in [@PZ].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We solve the advection-diffusion equation for a stochastically stationary passive scalar $\theta$, in conjunction with forced 3D Navier-Stokes equations, using direct numerical simulations in periodic domains of various sizes, the largest being $8192^3$. The Taylor micro-scale Reynolds number varies in the range $140-650$ and the Schmidt number $Sc \equiv \nu/D$ in the range $1-512$, where $\nu$ is the kinematic viscosity of the fluid and $D$ is the molecular diffusivity of $\theta$. Our results show that turbulence becomes less effective as a mixer when $Sc$ is large. First, the mean scalar dissipation rate $\langle \chi \rangle = 2D \langle |\nabla \theta|^2\rangle$, when suitably non-dimensionalized, decreases as the inverse of $\log Sc$. Second, 1D cuts through the scalar field indicate increasing density of sharp fronts on larger scales, oscillating with large excursions and reduced mixing. The scaling exponents of the scalar structure functions in the inertial-convective range saturate with respect to the moment order and the saturation exponent approaches unity as $Sc$ increases, qualitatively consistent with 1D cuts of the scalar.'
author:
- Dhawal Buaria
- 'Matthew P. Clay'
- 'Katepalli R. Sreenivasan'
- 'P. K. Yeung'
title: Turbulence is an ineffective mixer when Schmidt numbers are large
---
#### Introduction:
An important property of fluid turbulence is that it mixes substances extremely well [@tl72]. Thus, any circumstances in which turbulence loses that property is important to study and understand, from both theoretical and practical perspectives. This Letter examines one instance in which turbulence loses its ability to perform mixing. By analyzing a massive database generated through state-of-the-art direct numerical simulations (DNS) of the governing equations, we show that turbulence becomes a weak mixer in high Reynolds number turbulence when the Schmidt number, $Sc \equiv \nu/D$, is rendered very large. Here $\nu$ is the kinematic viscosity of the fluid and $D$ is the molecular diffusivity of the mixing substance.
The rate of mixing of the scalar $\theta$ is related to the average ‘dissipation’ rate $\langle \chi \rangle$ of its variance, defined as $\langle \chi \rangle= 2D \langle |\nabla \theta|^2\rangle$. There is a general claim that $\chi$ remains finite even when molecular diffusivity $D \to 0$. This claim derives from the analogy with the mean dissipation rate of turbulent kinetic energy, whose essential character is that it is independent of viscosity when the latter is sufficiently small [@taylor:1935; @K41]. There is concrete empirical evidence that this result is essentially correct [@sreeni84; @sreeni98; @pearson02; @kaneda03]. The question is whether the analogous property holds for passive scalars advected by turbulence for weak diffusivity [@batch1959a; @MY.II; @ShrSig00; @Donzis05]. We show that it does not hold.
There is some evidence scalar dissipation indeed becomes independent of $D$ for small $D$ when $Sc = \mathcal{O}(1)$ [@Donzis05]—i.e., $\nu$ and $D$ are both equally small. Figure \[fig:sc1diss\] shows that $\langle \chi\rangle$, non-dimensionalized by the large-scale quantity $\langle\theta^2\rangle u^\prime/L$, becomes independent of $D$ for small $D$, or for large Taylor-scale Péclet number $Pe_\lambda = u^\prime \lambda/D$; here $u^\prime$ is the root-mean-square (rms) velocity fluctuations and $\lambda = u'/\sqrt{\langle(\partial u/\partial x)^2\rangle}$ is the Taylor microscale, and $L/u^\prime$ is the time characteristic of the large scale $L$. There was already some suggestion in [@Donzis05] that the asymptotically constant values of $(\langle \chi \rangle L)/(\langle\theta^2\rangle u^\prime)$ become smaller with increasing $Sc$, but the data available at that time were limited and inconclusive.
![ The plot shows the behavior of scalar dissipation rate as a function of $Pe \propto 1/D$. The data in (blue) triangles are from [@Donzis05]; the new data are described below. In Ref. [@Donzis05], the fit to the data, shown by the dashed line, had the functional form shown in the legend, with $c_1=0.36$ and $c_2=31$. []{data-label="fig:sc1diss"}](chi_sc1.eps){width="7.5cm"}
#### Direct numerical simulations:
The data examined in this work were generated using the canonical DNS setup of isotropic turbulence in a periodic domain [@Ishihara09; @BPBY2019], forced at large scales to maintain statistical stationarity. The passive scalar is obtained by simultaneously solving the advection-diffusion equation in the presence of mean uniform gradient $\nabla \Theta = (G,0,0)$ along one of the Cartesian directions, $x$ [@yeung02]. For $Sc=1$, we utilize the conventional Fourier pseudo-spectral methods for both the velocity and scalar fields. For $Sc=4$ and higher, we utilize a hybrid approach [@gotoh12a; @clay.cpc1; @clay.cpc2], where the velocity field is obtained pseudo-spectrally, focused on resolving the Kolmogorov length scale $\eta$, and the scalar field by using compact finite differences on a finer grid to adequately resolve the smaller Batchelor scale $\eta_B = \eta/\sqrt{Sc}$. The database is summarized in Table \[tab:data\].
$\re$ $Sc$ $N_v^3$ $k_{max}\eta$ $N_\theta^3$ $k_{max}\eta_B$
------- ------ ---------- --------------- -------------- -----------------
140 1 $512^3$ 3 $512^3$ 3
140 4 $512^3$ 3 $1024^3$ 3
140 4 $512^3$ 3 $2048^3$ 6
140 8 $256^3$ 1.5 $1024^3$ 2
140 8 $512^3$ 3 $1024^3$ 2
140 8 $512^3$ 3 $2048^3$ 4
140 16 $256^3$ 1.5 $1024^3$ 1.5
140 16 $256^3$ 1.5 $2048^3$ 3
140 16 $512^3$ 3 $1024^3$ 1.5
140 16 $512^3$ 3 $2048^3$ 3
140 32 $512^3$ 3 $2048^3$ 2
140 32 $512^3$ 3 $2048^3$ 2
140 32 $1024^3$ 6 $4096^3$ 4
140 64 $512^3$ 3 $2048^3$ 1.5
140 64 $1024^3$ 6 $4096^3$ 3
140 128 $512^3$ 3 $4096^3$ 2
140 256 $1024^3$ 6 $8192^3$ 3
140 512 $1024^3$ 6 $8192^3$ 2
240 1 $1024^3$ 3 $1024^3$ 3
390 1 $2048^3$ 3 $2048^3$ 3
390 8 $2048^3$ 3 $8192^3$ 4
650 1 $4096^3$ 3 $4096^3$ 3
: Simulation parameters for the DNS runs used in the current work: the Taylor-scale Reynolds number $\re$, the Schmidt number $Sc$, the number of grid points for the velocity and scalar fields, $N_v^3$ and $N_\theta^3$, and the spatial resolution for the velocity and scalar fields, $k_{max}\eta$ and $k_{max}\eta_B$, in that order. Most of the simulations at $\re=140$ were previously reported in [@clay_thesis]. All simulations were run for at least 10 large-scale turnover times in the statistically stationary state, so all results presented here show excellent statistical convergence. The data for $Sc=256$ is from an independent, unpublished simulation by K. Ravikumar at Georgia Tech, and corresponds to 5 turnover times. []{data-label="tab:data"}
#### The mean scalar dissipation and the reduction of mixing around the diffusive scale:
Here we explore the influence of $Sc$ on mean scalar dissipation rate, $\langle \chi \rangle$. We see in Fig. \[fig:scdiss\] that the asymptotic value of scalar dissipation continually decreases with $Sc$. In fact, using arguments based on functional form of the scalar spectrum, the authors of refs. [@BSXDY; @Donzis05] showed that the inverse scalar dissipation rate $(\langle\theta^2\rangle u^\prime)/(\langle \chi \rangle L)$ varies as $\log Sc$. In order to see this behavior clearly, we plot the inverse dissipation versus $\log Sc$ in the inset of Fig. \[fig:scdiss\]. The data are in excellent agreement with expectations.
![ Test for scalar dissipation anomaly at $\re=140$ with increasing $Sc$. The mean scalar dissipation rate is normalized as in Fig. \[fig:sc1diss\]. The dashed line corresponds to $1/\log Sc$ dependence. The inset shows the inverse of these data versus $Sc$ on log-linear axes, affirming the $\log Sc$ dependence. []{data-label="fig:scdiss"}](chi_re140.eps){width="7.5cm"}
The observation that the normalized scalar dissipation tends to zero in the limit $Sc \to \infty$, albeit logarithmically, suggests that the diffusivity is ultimately incapable of smoothing the scalar fluctuations and that there is no mixing at small scales. This picture can be intuitively understood from a Lagrangian perspective by considering trajectories of individual scalar particles [@falkovich01; @sj2010; @SP.2013]. Physically, mixing occurs when some local concentration of scalar particles eventually disperses through the fluid under the combined action of turbulence and molecular diffusion. If we consider two coincident scalar particles, the diffusivity is necessary to create some finite separation, thereafter allowing turbulence to take over; however, in the limit of $D\to0$, they cannot separate and the action of turbulence does not manifest [@BSY.2015; @BYS.2016]. In fact, based on Lagrangian data of diffusing particles, an indication of decreasing scalar dissipation with increasing $Sc$ was already present in [@BYS.2016]. Another useful inference is that the time to mix a passive scalar increases like log$Sc$; in the limit $Sc \to \infty$, the mixing becomes infinitely slow.
#### Reduced mixing at larger scales:
Figure \[fig:rc\] shows typical 1D cuts of the scalar field in the direction of the mean gradient. The upper panel corresponds to fixed $Sc=1$ and increasing $\re$. The well known ramp-cliff structures (see [@KRS91; @HS94; @CLMV01; @sreeni19]) are clearly evident in all traces, with disorganized small-scale fluctuations superimposed on them. With increasing $\re$, small-scale fluctuations become more conspicuous as one expects, but the steep cliffs remain. In the lower panel, the cuts are for fixed $\re=140$ with $Sc$ varying from $1$ to $512$. For low to moderate $Sc$, the ramp-cliff structures stand out as before, but superimposed on them are oscillations in the scalar field that become stronger with increasing $Sc$. The ramp-cliff structures continue to be present even at the highest $Sc$ ($=512$), but are overwhelmed by sharp oscillations essentially between the smallest and largest concentrations, generating intermediate levels of the scalar $\theta$ far less frequently (i.e., producing reduced mixing).
The standard interpretation of energy dissipation is that it is the ultimate manifestation of energy flux from the large scales through intermediate ranges to the smallest. A similar interpretation holds for scalar dissipation. The vanishing of the scalar dissipation not only suggests that mixing is reduced at the smallest scales, but also that there is reduced flux of the scalar variance at larger scales where the advective action of turbulence overwhelms the role of diffusivity.
![ Typical one-dimensional cuts of the scalar field, normalized by the rms, in the direction of the imposed mean gradient ($x$), highlighting the changing character of the signals. $L_0 = 2\pi$ is the domain length. The curves in the upper panel correspond to fixed $Sc=1$ and $\re=140$, $390$ and $650$ from top to bottom; those in the lower panel are for fixed $\re=140$ and $Sc=1$, $8$, $64$ and $512$ from top to bottom. The curves are shifted for clarity, as indicated by dotted horizontal lines. []{data-label="fig:rc"}](rc_re.eps "fig:"){width="7.2cm"}\
![ Typical one-dimensional cuts of the scalar field, normalized by the rms, in the direction of the imposed mean gradient ($x$), highlighting the changing character of the signals. $L_0 = 2\pi$ is the domain length. The curves in the upper panel correspond to fixed $Sc=1$ and $\re=140$, $390$ and $650$ from top to bottom; those in the lower panel are for fixed $\re=140$ and $Sc=1$, $8$, $64$ and $512$ from top to bottom. The curves are shifted for clarity, as indicated by dotted horizontal lines. []{data-label="fig:rc"}](rc_sc.eps "fig:"){width="7.2cm"}
#### Structure functions:
To elaborate this point further, we consider the scalar increment $\Delta_r \theta$ between two points separated by distance $r$, whose moments are the scalar structure functions. In the so-called inertial-convective range, the $p$-th order structure function is expected to follow a power law of the form $\langle (\Delta_r \theta)^p \rangle \sim r^{\zeta_p}$, where $\zeta_p$ is anomalous with respect to the Kolmogorov phenomenology (i.e., $\zeta_p = p/3$) [@MY.II; @ZW00; @GY.2013]. In order to extract $\zeta_p$, we have followed an analysis similar to the recent work [@KI18] where $\zeta_2$ was obtained by a power law fit in the inertial-convective range, and higher order moments were extracted through extended self-similarity. Although not shown here, we have performed extensive tests (as in [@KI18]) to establish statistical convergence of structure functions of high orders.
![ The scalar increment exponent, $\zeta_p$, as a function of the moment order $p$ for various combinations of $\re$ and $Sc$ shown in the legend. The error bars indicate $95\%$ confidence interval. The dotted lines at 1.2 and 1.1 correspond to saturation values for the present data at $Sc = 1$ and $8$, respectively, whereas the dotted line at 1 is the likely saturation value in the limit of $Sc\to\infty$, as will be determined momentarily. The dashed line, $\zeta_p=p/3$, is the Kolmogorov phenomenology. []{data-label="fig:exp"}](scstrfn_exp.eps){width="45.00000%"}
The scaling exponents $\zeta_p$ are plotted against the moment order in Fig. \[fig:exp\], for $\re \ge 390$. The results for $\re=650$ and $Sc=1$ are virtually identical to those of [@KI18], and reaffirm that the scalar exponents saturate to $\lim_{p\to\infty} \zeta_p = \zeta_\infty \approx1.2$. In comparison, the exponents for $\re=390$ and $Sc=1$ are mostly identical to those at $\re=650$, but differ somewhat for $p\ge12$ (possibly due to a slightly smaller scaling range from which the exponents were extracted). The more important result is that for $\re=390$ and $Sc=8$ the exponents are consistently smaller than those for $Sc=1$ and appear to saturate at a smaller value of $\zeta_\infty \approx1.1$. Evidently, the smaller saturation value for larger $Sc$ invites the question as to whether it is bounded as $Sc \to \infty$.
![ The local slope of $p$-th order scalar structure functions at $\re=140$ and $Sc=32,128$ and $512$. The curves are shown for $p=4,8,12$ and $16$. They are shifted vertically for clarity and the corresponding dashed lines represent a local slope of unity. []{data-label="fig:ls"}](ls_140.eps){width="45.00000%"}
To obtain a definitive answer, one needs to obtain data for higher $Sc$ for at least $\re = 390$ for which convincing scaling exists (as demonstrated in [@KI18] for $Sc = 1$), but they are currently unattainable for large $Sc$ and unlikely to be obtained anytime soon. We have therefore analyzed the data at the lower $\re$ of 140, for which inertial range characteristics just begin to manifest [@yeung97; @Ishihara09]. In Fig. \[fig:ls\], we show the local slope of the $p$-th order structure function for $p=$4, 8, 12 and 16 and $Sc$=32, 128 and 512; the curves for different values of $p$ are shifted for clarity and the dashed lines represent local slopes of unity. With increasing $p$, the curves for all $Sc$ progressively get closer to their respective dashed lines. If we focus on the region $r/\eta \gtrsim 30$, which nominally corresponds to onset of the inertial-convective range, it also appears that the local slope for all $Sc$ are approximately equal for highest $p$ values, and close to unity—hinting that the high-order exponents saturate at about $1$ as $Sc \to \infty$.
![ Compensated plot of $N(r)$, the number of cubes of side $r$ containing the scalar fronts satisfying the threshold condition $|\partial \theta/\partial x| \ge 0.2 \theta_{rms}/\eta_B$ [@KI18]. Curves are shown for $Sc$=8, 32, 128 and 512 for $\re=140$. $N^3$ is the total number of grid points. $D_F=3-\zeta_\infty$ is the fractal co-dimension. We set $D_F=2$ corresponding to $\zeta_\infty=1$. []{data-label="fig:boxc"}](boxc_140.eps){width="45.00000%"}
#### Co-dimension result:
Finally, we turn to quantifying the fractal dimension of the sharp scalar fronts and understanding how it relates to the saturation exponent. In the recent work [@KI18], the authors found that the saturation exponent $\zeta_\infty$ and the box counting dimension $D_F$ of the sharp scalar fronts, satisfying the threshold $|\partial \theta/\partial x| \ge 0.2 \theta_{rms}/\eta_B$, add up to the Euclidean dimension of the flow, i.e., $\zeta_\infty+D_F = 3$. In that same spirit, we perform box counting of the strong scalar gradients corresponding to sharp fronts, given by $N(r)$ for various cubes of edge size $r$. For the saturation exponent $\zeta_\infty=1$, the co-dimension $D_F=2$. In Fig. \[fig:boxc\], we plot the $N(r)/N^3$ compensated by $r^{D_F}$ with $D_F=2$, for the same cases shown in Fig. \[fig:ls\]. Remarkably, the curves at the highest $Sc$ exhibit an extended plateau for small scales, consistent with a fractal dimension of 2. For large $r$, all curves are consistent with $D_F=3$, as expected by the space filling nature at large scales. This consolidates the result that the fractal dimension of sharp fronts is the co-dimension of the saturation exponent of scalar structure functions.
#### Conclusions:
We have demonstrated by several means that turbulence becomes an ineffective mixer in the limit of large Schmidt numbers. The scalar dissipation becomes smaller, and the scalar field oscillates effectively between the largest and smallest concentrations without producing as many intermediate levels. This is a qualitatively different behavior for the case of unity Schmidt numbers, and shows that turbulence mixes the passive scalar poorly when $Sc$ is large. We find that the exponents of the scalar structure functions saturate for high-order moments, and that the saturation value appears to be bounded by unity. We confirm this result by showing that large excursions in $\partial \theta/\partial x$ have a co-dimension of 2. We believe that these results form an important ingredient in a fuller understanding of turbulent mixing, and note that models like 1D Burgers equation [@bec07] and the Kraichnan’s passive scalar [@balk98] have the same behavior of saturated exponents for large moment orders, leveling off at unity.
Acknowledgments: We thank Kartik Iyer and Jörg Schumacher for useful discussions and Kiran Ravikumar for providing the $Sc = 256$ datapoint used in Fig. \[fig:scdiss\]. This research used resources of the Oak Ridge Leadership Computing Facility (OLCF), which is a Department of Energy (DOE) Office of Science user facility supported under Contract DE-AC05-00OR22725. We acknowledge the use of advanced computing resources at the OLCF under 2017 and 2018 INCITE Awards. Parts of the data analyzed in this work were obtained through National Science Foundation (NSF) Grant ACI-1036170, using resources of the Blue Waters sustained petascale computing project, which was supported by the NSF (awards OCI- 725070 and ACI-1238993) and the State of Illinois.
[34]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{}, Vol. (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} **, @noop [Ph.D. thesis]{}, () @noop [ ****, ()]{} @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The first-order correction of the perturbative solution of the coupled equations of the quadratic gravity and nonlinear electrodynamics is constructed, with the zeroth-order solution coinciding with the ones given by Ayón-Beato and García and by Bronnikov. It is shown that a simple generalization of the Bronnikov’s electromagnetic Lagrangian leads to the solution expressible in terms of the polylogarithm functions. The solution is parametrized by two integration constants and depends on two free parameters. By the boundary conditions the integration constants are related to the charge and total mass of the system as seen by a distant observer, whereas the free parameters are adjusted to make the resultant line element regular at the center. It is argued that various curvature invariants are also regular there that strongly suggests the regularity of the spacetime. Despite the complexity of the problem the obtained solution can be studied analytically. The location of the event horizon of the black hole, its asymptotics and temperature are calculated. Special emphasis is put on the extremal configuration.'
author:
- Waldemar Berej
- Jerzy Matyjasek
- Dariusz Tryniecki
- Mariusz Woronowicz
title: Regular black holes in quadratic gravity
---
Introduction
============
One of the most important and intriguing questions of the black hole physics is the problem of singularities that reside in their internal region, hidden to an external observer by the event horizon. In the vast majority of papers singularities are treated as symptoms of illness of the theory rather than its health (see however Ref. [@Horowitz] for a different point of view), and, consequently, a great deal of efforts were directed to constructing singularity-free models. However, a subtle point is that the Einstein field equations loose their predicative power and cannot be trusted when the curvature of the manifold approaches the Planck regime. Indeed, according to our present understanding a proper description of the gravitational phenomena should be given by the quantum gravity, being perhaps a part of a more fundamental theory. And although at the present stage we have no clear idea how this theory looks like, we expect that the action functional describing its low-energy approximation should consist of higher order terms constructed from the curvature and its covariant derivatives to some required order. It means that in the full theory the analogs of the nonsingular solution of the Einstein gravity may loose their nonsingular character as well as the singular ones their singularities.
Among various modifications of the general relativity the prominent role is played by the quadratic gravity [@Utiyama; @Stelle1; @Stelle2; @Weyl1; @Weyl2; @Pauli; @sirArthur]. Motivations for introducing into the action functional terms which are quadratic in curvature are numerous. For example, when invented, the equations of quadratic gravity have been treated as an exact formulation of the theory of gravitation. For historic informations and important references the reader is referred to Ref. [@HJS]. It may be considered, quite naturally, as truncation of series expansion of the action of the more general theory. Such terms appear generically in one-loop calculations of the quantum field theory in curved background [@Birrell]. Moreover, from the point of view of the semi-classical gravity it might be treated (in certain circumstances) as some sort of a poor man’s stress- energy tensor, allowing in a relatively simple way to mimic, especially when the application of the full stress- energy tensor would produce extremely complicated or even intractable results, the fairly more complex source term of the field equations.
Analyses of the spherically-symmetric and static solutions to the higher derivative theory has been carried out in [@Stelle1; @Lousto1; @Lousto2; @Lousto3; @Holdom1; @Tryn]. For example, in Ref [@Stelle1] it has been shown that the weak-field limit of the quadratic gravity involves, beside the ordinary Newtonian term, also the terms with the Yukawa-like potential. Series solutions near the $r=0$ have been investigated in [@Stelle1; @Holdom1]. Such solutions are limited to the closest neighborhood of the center and should be matched to the appropriate solutions valid for large $r$ that require numerical integration. However, an important lesson that follows from this calculations is that the regular solutions of the equations of the higher order theories are quite common [@Stelle1].
As compared with the General Relativity, the Lagrangian of the quadratic gravity in four dimensions requires two additional terms $\alpha R_{ab}R^{ab}+\beta R^{2},$ where $\alpha $ and $\beta $ are the coupling constants. Since the coupling constants are expected to be very small, one can easily device a perturbative approach to the problem treating the classical solution of the Einstein field equations as the zeroth-order of the approximation. Successive perturbations are therefore solutions of the differential equations of ascending complexity. It should be noted that although the method is clear the calculations beyond the first- order may be intractable.
According to the well-known theorem, if the Lagrangian $\mathcal{L}
\left( F\right) $ has Maxwell asymptotics for weak fields, i.e., $\mathcal{L}\left( F\right) \sim F$ and $\mathcal{L}_{F}=d\mathcal{L}/dF\rightarrow $ $1$ as $ F\rightarrow 0$, then any static and spherically-symmetric solutions to the system of coupled equations of nonlinear electrodynamics and general relativity characterized by electric charge, $Q_{e}$, cannot have a regular center [@Bronnikov3; @Bronnikov1; @Bronnikov2]. This no-go theorem does not forbid existence of the solutions with magnetic charge, $Q$, as well as some hybrid solutions in which the electric field does not extend to the central region. On the other hand, however, it has been argued in Ref. [@Burinskii1] that the Maxwell asymptotics at great distances are essentially different from these at the center, and, consequently, the condition $ \mathcal{L}_{F}\rightarrow 1$ as $F\rightarrow 0$ is too restrictive.
The issue of the regular black holes in general relativity has a long and interesting history. For example, one of the method that can be used in construction of such configurations is simply replacing the singular black hole interior by a regular core. This idea appeared almost forty years ago, in mid sixties [@Sakharov; @Gliner; @Bardeen:68] and is actively investigated today. In models considered in Refs. [@frolov1; @frolov2] part of the region inside the event horizon is joined through a thin boundary layer to de Sitter geometry. Such a geometric surgery certainly does not exhaust all interesting possibilities: of equal importance are the regular geometries with suitably chosen profile functions, or, better, exact solutions constructed for specific, physically reasonable sources [@Irina1; @Irina2; @Borde; @Mars; @Ayon-Beato:2000zs; @Ayon-Beato:2004ih]. One of the most intriguing regular solutions to the coupled equations of the nonlinear electrodynamics and gravity have been constructed by Ayón-Beato and García [@ABG] and by Bronnikov [@Bronnikov1]. We shall refer to solutions of this type as ABGB geometry. The former describes a regular static and spherically symmetric configuration with an electric charge whereas the latter describes a similar geometry characterized by the total mass $M$ and $
Q$. For certain values of the parameters both solutions describe a black hole. It should be noted that the electric solution does not contradict the non existence theorem as the formulation of nonlinear electrodynamics employed by Ayón-Beato and García ($\mathcal{P}$ framework in the nomenclature of Refs. [@Bronnikov1; @Bronnikov2]) is not the one to which one refers in the proviso of the no-go theorem. Indeed, the solution of Ayón-Beato and García has been constructed in a formulation of the nonlinear electrodynamics obtained from the original one ($\mathcal{F}$ framework) by means of a Legendre transformation (see Ref. [@Bronnikov2] for details). An attractive feature of this solutions that certainly simplifies calculations is possibility to express the location of the horizons in terms of the known special functions [@Kocio1; @Kocio2].
The natural question that arises in connection with the foregoing discussion is whether or not it is possible to construct an analog of the solution of the ABGB-type in the quadratic gravity which shares with its classical counterpart regularity at $r=0.$ And although the full, detailed answer is beyond our capabilities, it is possible to provide an affirmative answer to the restricted problem. Indeed, since the complexity of the coupled equations of the quadratic gravity and nonlinear electrodynamics, even in the simplest case of spherically-symmetric and static geometries, hinders construction of the exact solution, one has to refer to the analytical approximations or numerical methods.
In this paper we shall employ perturbative methods to construct the first-order solution to the equations of the quadratic gravity with the source term being generalization of the stress-energy tensor of the Bronnikov type. The zeroth-order solution coincides, as expected, with the ABGB line element whereas the first-order correction can be elegantly expressed in terms of the polylogarithm functions. An interesting feature of this very solution is its regularity for $r = 0.$ The Kretschmann scalar and other curvature invariants are also regular at the center that suggests regularity of the underlying geometry. It should be emphasized that the demonstration of the regularity of the full solution would require either profound understanding of the perturbation series to any required order or construction of the full, physically acceptable nonperturbative solution. On the other hand, the perturbative approach is expected to yield reasonable results especially for higher derivative dynamical equations. In fact it may be the only method to deal with them. Indeed, since the quadratic gravity involves fourth-order derivatives of the metric their nonperturbative solutions may appear to be spurious and one has to invent a method for systematic selecting physical ones. It seems that the acceptable solutions, when expanded in powers of the small parameter, should reduce to those obtained within the framework of perturbative approach [@Simon1; @Simon:1991jn; @Parker1].
The paper is organized as follows. In Sec. \[sec2\] we introduce basic equations and briefly sketch employed method. We choose the line element in the form propounded by Visser [@Matt1], which has proved to be a very useful representation considerably simplifying calculations. In Sec. \[sec3\] we introduce the Lagrangian of the nonlinear electrodynamics, construct solutions to the first-order equations and establish regularity of the thus obtained line element. Subsequently we study the limit of the vanishing $Q$ and analyse the regularity of various curvature invariants. Corrections to the location of the inner and outer horizons and to temperature are given in Sec. \[sec4\]. The position of the horizons of the ABGB spacetime is given in terms of the real branches of the Lambert functions. The extremal configuration is studied in Sec. \[sec5\]. Specifically, we calculate modifications of the location of the degenerate horizon caused by quadratic gravity and analyse relations between $Q$ and the total mass as seen by a distant observer. Sec. \[sec6\] contains final remarks. In Appendix we have collected useful formulas and presented a brief description of the method of integration of the field equations in terms of the polylogarithm functions. Throughout the paper the geometric system of units has been adopted and the the sign conventions are taken to be that of MTW [@MTW].
The equations {#sec2}
=============
In absence of the cosmological term, the coupled system of the nonlinear electrodynamics and the quadratic gravity is described by the action $$S=\frac{1}{16\pi G}S_{g}+S_{m},$$ where $$S_{g}=\int \left( R+\alpha R^{2}+\beta R_{ab}R^{ab}\right) \sqrt{-g}\,d^{4}x,$$ and $$S_{m}=-\dfrac{1}{16\pi }\int \mathcal{L}\left( F\right) \sqrt{-g}\,d^{4}x.$$ Here $\mathcal{L}\left( F\right) $ is some functional of $F=F_{ab}F^{ab}$ (its exact form will be given later) and all symbols have their usual meaning. The third possible term constructed form the Kretschmann scalar, $\gamma R_{abcd}R^{abcd}$, may by removed from the Lagrangian with the help of the Gauss-Bonnet invariant. Of numerical parameters $\alpha $ and $\beta $ we assume, as usual, that they are small and of comparable order, otherwise they would lead to the observational consequences within our solar system. Their ultimate values should be determined form observations of light deflection, binary pulsars and cosmological data [@Odylio; @accioly:2001cc; @Mijic:1986iv]. Moreover, following Ref. [@Lousto3], we shall restrict ourselves to spacetimes of small curvatures, for which the conditions $$|\alpha R| <<1, \hspace{5mm} |\beta R_{ab}|<<1$$ hold. Although demanding that the mass scales associated with the linearized equations are real may place additional constraints [@Steve; @Whitt:1984pd; @Audretsch:1993kp] on $\alpha$ and $\beta,$ we shall treat them as small but arbitrary.
The tensor $F_{ab}$ satisfies equations
$$\nabla _{a}\left( \dfrac{d\mathcal{L}\left( F\right) }{dF}F^{ab}\right) =0,$$
$$\nabla _{a}\,^{\ast }F^{ab}=0,$$
and the asterix denotes the Hodge duality. The stress-energy tensor defined as $$T^{ab}=\frac{2}{\sqrt{-g}}\frac{\delta }{\delta g_{ab}}S_{m} \label{tensep}$$ is given therefore by $$T_{a}^{b}=\dfrac{1}{4\pi }\left( \dfrac{d\mathcal{L}\left(
F\right) }{dF}F_{ca}F^{cb}-\dfrac{1}{4}\delta _{a}^{b}\mathcal{L}\left(
F\right) \right) .$$ Differentiating functionally the action $S$ with respect to the metric tensor one has $$L^{ab}=G^{ab}-\alpha I^{ab}-\beta J^{ab}=8\pi T^{ab}, \label{2nd_order}$$ where $$I^{ab}=2R^{;\,ab}-2RR^{ab}+\frac{1}{2}g^{ab}\left( R^{2}-4\Box R\right)$$ and $$J^{ab}=R^{;\,ab}-\Box R^{ab}-2R_{cd}R^{cbda}+\frac{1}{2}g^{ab}\left(
R_{cd}R^{cd}-\Box R\right) .$$
Let us consider the spherically symmetric and static configuration described by the line element of the form $$ds^{2}=-e^{2\psi \left( r\right)}f(r) dt^{2}+\frac{dr^{2}}{f(r) } + r^{2}d\Omega
^{2} \label{el_gen},$$ where $$f(r) = 1 - \frac{2 M(r)}{r}.$$ The spherical symmetry places restrictions on the components of $F_{ab}$ tensor and its only nonvanishing components compatible with the assumed symmetry are $F_{01}$ and $F_{23}$. Simple calculations yield $$F_{23}=Q\sin \theta$$ and $$r^{2}e^{-2\psi }\dfrac{d\mathcal{L}\left( F\right) }{dF}F_{10}=Q_{e},$$ where $Q$ and $Q_{e}$ are the integration constants interpreted as the magnetic and electric charge, respectively. In the latter we shall assume that the electric charge vanishes, and, consequently, $F$ is given by $$F=\dfrac{2Q^{2}}{r^{4}}. \label{postacF}$$ The stress-energy tensor (\[tensep\]) calculated for this configuration is $$T_{t}^{t}=T_{r}^{r}=-\dfrac{1}{16\pi }\mathcal{L}\left( F\right) \label{t1}$$ and $$T_{\theta }^{\theta }=T_{\phi }^{\phi }=\dfrac{1}{4\pi }\dfrac{d\mathcal{L}
\left( F\right) }{dF}\dfrac{Q^{2}}{r^{4}}-\dfrac{1}{16\pi }\mathcal{L}\left(
F\right) .$$
To simplify calculations and to keep control of the order of terms in complicated series expansions we shall introduce a dimensionless parameter $
\varepsilon $ substituting $\alpha \rightarrow \varepsilon \alpha $ and $
\beta \rightarrow \varepsilon \beta $. We shall put $\varepsilon =1$ at the final stage of calculations. Of functions $M\left( r\right) $ and $\psi
\left( r\right) $ we assume that they can be expanded as $$M\left( r\right) =M_{0}\left( r\right) +\varepsilon M_{1}\left( r\right) +
\mathcal{O}\left( \varepsilon ^{2}\right) \label{Mser}$$ and $$\psi \left( r\right) =\varepsilon \psi _{1}\left( r\right) +\mathcal{O}
\left( \varepsilon ^{2}\right) . \label{psiser}$$
Consider the left hand side of Eq. (\[2nd\_order\]) calculated for the line element (\[el\_gen\]) first. Making use of the above expansions and collecting the terms with the like power one obtains $$L_{t}^{t}=-\frac{2}{r^{2}}(M_{0}^{\prime }+\varepsilon M_{1}^{\prime
}-\varepsilon S_{t}^{t}), \label{1st}$$ where $$\begin{aligned}
S_{t}^{t}& =\beta \left( \frac{2\,M_{0}^{\prime }}{r^{2}}-\frac{
8\,M_{0}\,M_{0}^{\prime }}{r^{3}}+\frac{2\,{M_{0}^{\prime }}^{2}}{r^{2}}-
\frac{2\,M_{0}^{\prime \prime }}{r}+\frac{5\,M_{0}\,M_{0}^{\prime \prime }}{
r^{2}}-\frac{M_{0}^{\prime }\,M_{0}^{\prime \prime }}{r}\right. \notag \\
& \left. +\frac{{M_{0}^{\prime \prime }}^{2}}{2}+M_{0}^{(3)}-\frac{
M_{0}\,M_{0}^{(3)}}{r}-M_{0}^{\prime
}\,M_{0}^{(3)}+r\,M_{0}^{(4)}-2\,M_{0}\,M_{0}^{(4)}\right) \notag \\
& -\alpha \left( \frac{24\,M_{0}\,M_{0}^{\prime }}{r^{3}}-\frac{
8\,M_{0}^{\prime }}{r^{2}}-\frac{4\,{M_{0}^{\prime }}^{2}}{r^{2}}+\frac{
8\,M_{0}^{\prime \prime }}{r}-\frac{18\,M_{0}\,M_{0}^{\prime \prime }}{r^{2}}
-{M_{0}^{\prime \prime }}^{2}\right. \notag \\
& \left. +\frac{2\,M_{0}^{\prime }\,M_{0}^{\prime \prime }}{r}
-4\,M_{0}^{(3)}+\frac{6\,M_{0}\,M_{0}^{(3)}}{r}+2\,M_{0}^{\prime
}\,M_{0}^{(3)}-2\,r\,M_{0}^{(4)}+4\,M_{0}\,M_{0}^{(4)}\right) \notag \\
& \label{1sta}\end{aligned}$$ and $M_{0}^{\prime},$ $M_{0}^{\prime \prime}$ and $M^{(i)}_{0}$ for $i \geq 3 $ denote first, second and $i-$th derivatives with respect to the radial coordinate. On the other hand, a simple combination of the components of $L_{a}^{b}$ tensor $$L_{r}^{r}-L_{t}^{t}=0 \label{2nd}$$ can be easily integrated to yield $$\psi _{1}(r)\,=\,(2\alpha +\beta )M_{0}^{(3)}-{\frac{4}{r^{2}}}(3\alpha
+\beta )M_{0}^{\prime }+C_{1}, \label{pseq}$$ where $C_{1}$ is the integration constant. It should be noted that contrary to the case of coupled system of the Maxwell equations and quadratic gravity considered in Refs. [@Lousto1; @Lousto2; @Lousto3; @Tryn] now we have explicit dependence on the parameter $\alpha.$ This together with the nonlinear character of the source term (that will be specified below) results in substantial complications of the first-order equations.
To develop the model further we shall express solutions of the system of differential equations consisting of the time component of Eqs. (\[2nd\_order\]) and Eq. (\[2nd\]) in terms of the total mass $\mathcal{M}$ as seen by a distant observer $$\lim_{r\rightarrow \infty }M\left( r\right) =\mathcal{M,} \label{bound1}$$ whereas for the function $\psi \left( r\right) $ we shall adopt the natural condition $$\lim_{r\rightarrow \infty }\psi \left( r\right) =0.
\label{bound2}$$
Solutions {#sec3}
=========
Further considerations require specification of the Lagrangian $\mathcal{L}\left( F\right) .$ Following Ayón-Beato, García and Bronnikov let us chose it in the form $$\mathcal{L}\left( F\right) \,=F\left[ 1-\tanh ^{2}\left( s\,\sqrt[4]{\frac{
Q^{2}F}{2}}\right) \right] , \label{labg}$$ where $$s=\frac{\left| Q\right| }{2b}, \label{sabg}$$ and the free parameter $b$ will be adjusted to guarantee regularity at the center.
Before proceeding further we shall briefly discuss the question of regularity, postponing presentation of the technical details for a while. First, observe that the zeroth-order solution coincides with a general ABGB line element that depends on a free parameter $b$ and an integration constant $C_{2}$. On the other hand, the first-order solution written in the suitable representation approaches, as we shall see, a constant value, say $\mu ,$ as $r \to \infty,$ and, consequently, by the boundary conditions (\[bound1\]) one has $C_{2} = \mathcal{M} - \varepsilon \mu.$ The thus obtained line element is generally singular at the center, and the only way to make it regular consists in appropriate choice of the free parameter. The regularity is understood here as the regularity of the line element rather than regularity of the spacetime itself as the latter requires the curvature invariants to be regular. We shall return to this issue later.
Now we present the calculations in a more systematic form. In order to guarantee sufficient generality of our considerations we shall take the parameter $b$ in the form $$b=b_{1}+\varepsilon b_{2}, \label{bgen}$$ with $b_{2}\neq 0.$ * * Since we have assumed the expansions of $M\left( r\right) $ and $\psi \left(
r\right) $ in the form given by Eqs. (\[Mser\]) and (\[psiser\]), respectively, we shall rewrite the boundary conditions as $$\lim_{r\rightarrow \infty }M_{0}\left( r\right) =\mathcal{M,} \label{M0}$$ with vanishing $M_{1}(\infty)$ and $\psi_{1}(\infty).$ Inserting Eq. (\[sabg\]) into (\[labg\]) and makig use of Eq. (\[postacF\]) one has $$\mathcal{L}\left( F\right) =\frac{2Q^{2}}{r^{4}}\left( 1-\tanh ^{2}\frac{
Q^{2}}{2br}\right) .$$ Subsequently expanding the right hand side of Eqs. (\[2nd\_order\]) with respect to $\varepsilon $ and retaining the linear terms only yields $$8\pi T_{t}^{t}=8\pi T_{r}^{r}=-\frac{Q^{2}}{r^{4}}\left( 1-\tanh ^{2}\frac{
Q^{2}}{2b_{1}r}\right) -\varepsilon \frac{b_{2}Q^{4}}{b_{1}^{2}r^{5}}\left(
\cosh ^{-2}\frac{Q^{2}}{2b_{1}r}\,\tanh \frac{Q^{2}}{2b_{1}r}\right) .
\label{tep}$$ The zeroth-order equation $$M_{0}^{\prime }\left( r\right) =\frac{Q^{2}}{2r^{2}}\left( 1-\tanh ^{2}\frac{
Q^{2}}{2b_{1}r}\right)$$ can be easily integrated $$M_{0}\left( r\right) =C_{2}-b_{1}\tanh \frac{Q^{2}}{2b_{1}r}, \label{mm0}$$ where $C_{2}$ is the integration constant. Making use of the condition (\[M0\]) gives $C_{2}=\mathcal{M}$. On the other hand, demanding of the regularity of the line element as $r\rightarrow 0$ yields $b_{1}=\mathcal{M,}
$ and, consequently, $M_{0}\left( r\right) $ reads $$M_{0}\left( r\right) =\mathcal{M}\left( 1-\tanh \frac{Q^{2}}{2\mathcal{M}r}
\right) .
\label{eMzero}$$ The zeroth-order solution has an interesting property, which, as we shall see, is crucial in our subsequent analysis: $M_{0}(r)$ as well as its derivatives vanish in the limit $r\to 0.$ Moreover, it should be emphasized that although the integration constant and the free parameter in the thus constructed solution are equal to the total mass of the system as seen by a distant observer their status is different: the interpretation of the former is an inevitable consequence of the boundary conditions whereas the latter should be postulated.
For small values of $|Q|/\mathcal{M}$ as well as at great distances the ABGB line element resembles that of Reissner-Nordström. It can be easily seen by expanding the function $M_{0}(r)$ $$g_{tt}\,=-1+\frac{2\mathcal{M}}{r} - \frac{Q^2}{r^{2}}\,
+\,\frac{Q^{6}}{12\mathcal{M}^{2}r^{4}}\,+\,...
,$$ whereas for $r \to 0$ one has $$f \sim 1 -\frac{4\mathcal{M}}{r}\exp\left(\frac{-Q^{2}}{\mathcal{M}r}\right).$$ Noticeable differences appear for the configurations near the extremality limit and in the internal region in the vicinity of the center. Indeed, for the ABGB solution $g_{tt}$ tends to $-1$ as $ r \to 0$ whereas the $(00)$ component of the metric tensor of the Reissner-Nordström solution diverges in that limit as $-r^{-2}.$
From (\[pseq\]) one concludes that the zeroth-order solution, $M_{0}\left(r\right) $, is sufficient to determine the function $\psi _{1}\left(r\right) .$ Indeed, substituting (\[eMzero\]) into Eq. (\[pseq\]) and making use of the condition (\[bound2\]) one obtains $$\begin{aligned}
\psi _{1}\left( r\right) &=&\left[ \frac{\beta Q^{2}}{r^{4}}+\frac{\left(
2\alpha +\beta \right) Q^{6}}{2\mathcal{M}^{2} r^{6}}\tanh ^{2}\frac{Q^{2}}{2
\mathcal{M}r}-\frac{3\left( 2\alpha +\beta \right) Q^{4}}{\mathcal{M}r^{5}}
\tanh \frac{Q^{2}}{2\mathcal{M}r}\right] \cosh ^{-2}\frac{Q^{2}}{2\mathcal{M}
r} \notag \\
&&-\frac{\left( 2\alpha +\beta \right) Q^{6}}{4\mathcal{M}^{2}r^{6}}\cosh
^{-4}\frac{Q^{2}}{2\mathcal{M}r}.
\label{psi_zeroo}\end{aligned}$$ It could be easily demonstrated that that $\psi_{1}(r)$ vanishes at $r=0$ and inspection of Eq. (\[pseq\]) reveals similar behaviour of its derivatives.
The solution of the first-order equation $$\frac{2}{r^{2}}(M_{1}^{\prime }-S_{t}^{t})=\frac{b_{2}Q^{4}}{b_{1}^{2}r^{5}}
\left( \cosh ^{-2}\frac{Q^{2}}{2b_{1}r}\,\tanh \frac{Q^{2}}{2b_{1}r}\right)
\label{f_o}$$ is more complicated and when combined with the appropriate boundary conditions it could be written as $$M_{1}\left( r\right) =\int_{\infty }^{r}S_{t}^{t}\left( r\right) dr+\frac{
b_{2}Q^{4}}{2\mathcal{M}^{2}}\int_{\infty }^{r}\frac{1}{r^{3}}\left( \cosh
^{-2}\frac{Q^{2}}{2\mathcal{M}r}\,\tanh \frac{Q^{2}}{2\mathcal{M}r}\right)
dr. \label{eqM1}$$ Let us consider the first integral in the right hand side of (\[eqM1\]) first. The result can be expressed in terms of the hyperbolic functions and polylogarithms $\mathrm{Li}_{i}\left( s\right) $. Indeed, utilizing formulas collected in Appendix one can construct the solution which has the general structure $$\begin{aligned}
M_{1}^{\left( 0\right) }\left( r\right) &=&-\int_{\infty
}^{r}S_{t}^{t}\left( r\right) dr=\frac{1}{\mathcal{M}}\sum_{i=0}^{1}
\sum_{j=0}^{2}\left( \alpha \widetilde{f}_{ij}^{\left( \alpha \right)
}+\beta \widetilde{f}_{ij}^{\left( \beta \right) }\right) \tanh ^{i}\dfrac{
Q^{2}}{2\mathcal{M}r}\sec \mathrm{h}^{2j}\dfrac{Q^{2}}{2\mathcal{M}r} \notag
\\
&&+\frac{1}{\mathcal{M}}\sum_{i=1}^{6}\left( \alpha \widetilde{h}
_{i}^{\left( \alpha \right) }+\beta \widetilde{h}_{i}^{\left( \beta \right)
}\right) \mathrm{Li}_{i}\left( -\exp \left( -\frac{Q^{2}}{\mathcal{M}r}
\right) \right) -\mu . \label{sol1}\end{aligned}$$ Here $\mu $ is given by $$\mu =\frac{\pi ^{2}\mathcal{M}^{5}}{Q^{6}}\left[ \alpha \left( 8-\frac{31}{
315}\pi ^{4}\right) +\beta \left( \frac{8}{3}+\frac{7}{45}\pi ^{2}-\frac{31}{
630}\pi ^{4}\right) \right]\,=\,\frac{\pi^{2} \mathcal{M}^{5}}{Q^{6}}\sigma \label{mu},$$ and since the terms in the square brackets will appear frequently in our subsequent analyses we have singled them out and denoted by $\sigma .$ The functions $\tilde{f}$ and $\tilde{h}$ are simple polynomials of $r^{-1};$ their actual form will not be displayed here as we shall readily rewrite them in a slightly modified form. It should be noted, however, that a careful term-by-term analysis of Eq. (\[sol1\]) reveals that $M_{1}^{(0)}(r)$ approaches $-\mu$ as $r \to 0.$ Since for $b=b_{1}$ the functions $M_{1}(r)$ and $M_{1}^{(0)}$ coincide it is impossible to obtain a regular solution at the center, and the remedy is to retain $b_{2}$ in the calculations.
Converting all hyperbolic functions appearing in the solution to the exponents, introducing a new function $\xi $ defined as $$\xi \left( r\right) =\exp \left( -\frac{Q^{2}}{\mathcal{M}r}\right)
\label{xi1}$$ and finally making use of the elementary properties of the function $\rm{Li}_{0}(-\xi)$ (see Eq. (\[AA\]) of the Appendix) one has $$M_{1}^{\left( 0\right) }\left( r\right) =\sum_{i=0}^{6}\left( \alpha
f_{i}^{\left( \alpha \right) }+\beta f_{i}^{\left( \beta \right) }\right)
\mathrm{Li}_{i}\left( -\xi \right) -\mu . \label{tylda}$$ The term $\mu $ has been singled out for convenience and $f_{i}^{\left(
\alpha \right) }$ and $f_{i}^{\left( \beta \right) }$ are given respectively by $$\begin{aligned}
&&f_{0}^{\left( \alpha \right) }=-\frac{48\mathcal{M}^{3}}{Q^{2}r^{2}}+\frac{
4\mathcal{M}Q^{2}\left( \xi -1\right) }{r^{4}\left( 1+\xi \right) ^{2}}-
\frac{16\mathcal{M}^{2}}{r^{3}\left( 1+\xi \right) }-\frac{4Q^{6}\left(
1-4\xi +\xi ^{2}\right) }{\mathcal{M}^{2}r^{5}\left( 1+\xi \right) ^{3}}
\notag \\
&&-\frac{4Q^{4}\left[ \mathcal{M}\left( 1+86\xi -89\xi ^{2}\right)
+25r\left( \xi ^{2}-1\right) \right] }{5\mathcal{M}r^{5}\left( 1+\xi \right)
^{3}}+\frac{2Q^{6}\xi \left( 75-425\xi +125\xi ^{2}+\xi ^{3}\right) }{15
\mathcal{M}r^{6}\left( 1+\xi \right) ^{4}},\end{aligned}$$ $$f_{1}^{\left( \alpha \right) }=\frac{4Q^{4}}{5r^{5}}-\frac{96\mathcal{M}^{4}
}{Q^{4}r},\qquad f_{2}^{\left( \alpha \right) }=\frac{4\mathcal{M}Q^{2}}{
r^{4}}-\frac{96\mathcal{M}^{5}}{Q^{6}},$$ $$\frac{6r^{3}\mathcal{M}^{3}}{Q^{6}}f_{3}^{\left( \alpha \right) }=\frac{
2r^{2}\mathcal{M}^{2}}{Q^{4}}f_{4}^{\left( \alpha \right) }=\frac{r\mathcal{M
}}{Q^{2}}f_{5}^{\left( \alpha \right) }=f_{6}^{\left( \alpha \right) }=\frac{
96\mathcal{M}^{5}}{Q^{6}}$$ and $$\begin{aligned}
f_{0}^{\left( \beta \right) } &=&-\frac{16\mathcal{M}^{3}}{Q^{2}r^{2}}-\frac{
16\mathcal{M}^{2}}{3r^{3}\left( 1+\xi \right) }-\frac{2Q^{2}\left[ 12r\left(
1+\xi \right) +\mathcal{M}\left( 3-36\xi +\xi ^{2}\right) \right] }{
3r^{4}\left( 1+\xi \right) ^{2}} \notag \\
&&-\frac{2Q^{6}\left( 1-4\xi +\xi ^{2}\right) }{\mathcal{M}^{2}r^{5}\left(
1+\xi \right) ^{3}}+\frac{Q^{6}\xi \left( 75-425\xi +125\xi ^{2}+\xi
^{3}\right) }{15\mathcal{M}r^{6}\left( 1+\xi \right) ^{4}} \notag \\
&&-\frac{2Q^{4}\left[ \mathcal{M}\left( 1+96\xi -109\xi ^{2}\right)
+30r\left( \xi ^{2}-1\right) \right] }{5\mathcal{M}r^{5}\left( 1+\xi \right)
^{3}},\end{aligned}$$ $$f_{1}^{\left( \beta \right) }=\frac{2Q^{4}}{5r^{5}}-\frac{8\mathcal{M}^{2}}{
3r^{3}}-\frac{32\mathcal{M}^{4}}{Q^{4}r},\qquad f_{2}^{\left( \beta \right)
}=-\frac{32\mathcal{M}^{5}}{Q^{6}}+\frac{2\mathcal{M}Q^{2}}{r^{4}}-\frac{8
\mathcal{M}^{3}}{Q^{2}r^{2}},$$ $$f_{3}^{\left( \beta \right) }=\frac{8\mathcal{M}^{2}}{r^{3}}-\frac{16
\mathcal{M}^{4}}{Q^{4}r},\qquad f_{4}^{\left( \beta \right) }=-\frac{16
\mathcal{M}^{5}}{Q^{6}}+\frac{24\mathcal{M}^{3}}{Q^{2}r^{2}},$$ $$\frac{r\mathcal{M}}{Q^{2}}f_{5}^{\left( \beta \right) }=f_{6}^{\left( \beta
\right) }=\frac{48\mathcal{M}^{5}}{Q^{6}}.$$ The second quadrature in the right hand side of Eq. (\[eqM1\]) is elementary and gives $$M_{1}^{\left( 1\right) }=\frac{b_{2}Q^{4}}{2\mathcal{M}^{2}}\int_{\infty
}^{r}\frac{1}{r^{3}}\left( \cosh ^{-2}\frac{Q^{2}}{2\mathcal{M}r}\,\tanh
\frac{Q^{2}}{2\mathcal{M}r}\right) dr=\frac{b_{2}Q^{2}}{2\mathcal{M}r}\cosh
^{-2}\frac{Q^{2}}{2\mathcal{M}r}-b_{2}\tanh \frac{Q^{2}}{2\mathcal{M}r}
\label{M_1_1}$$ or equivalently $$M_{1}^{(1)}\,=\,b_{2}\left[\frac{1-\xi}{\xi}
- \frac{2Q^{2}}{\mathcal{M}r (1+\xi)} \right] {\rm Li}_{0}(-\xi) .$$ Now, in order to obtain a regular solution at the center one has to adjust suitably the parameter $b_{2}.$ To accomplish this let us observe that $M_{1}^{\left(1\right) }\left( 0\right) =-b_{2}.$ Further, taking the results of the discussion presented below Eq. (\[mu\]) into account one finds $$b_{2}=-\mathcal{\mu} .$$ It could be easily seen that with such a choice of $b_{2}$ the function $M_{1}\left(r\right) $ vanishes at the center and as $r\rightarrow \infty $ as required.
It should be noted that the representation of the thus obtained line element is by no means unique. For example one can express the result in terms of $\rm{Li}_{n}(-1/\xi)$ rather than $\rm{Li}_{n}(- \xi)$ employing identity [@Levin1] $${\rm Li}_{n}(-\xi) + (-1)^{n} {\rm Li}_{n}(-1/ \xi)\,=\,-\frac{1}{n!} \ln^{n} \xi +
2 \sum_{r=1}^{[n/2]} \frac{\ln^{n-2r}\xi}{(n-2r)!}\, {\rm Li}_{2 r}(-1) ,$$ where $[n/2]$ is the greatest integer contained in $n/2$ and the constants ${\rm Li}_{2 r}(-1)$ are related to the Bernoulli numbers, $B_{2 r},$ according to the formula $${\rm Li}_{2 r}(-1)\,=\frac{2^{2r-1}}{(2r)!} B_{2 r} \pi^{2 r}.$$ Regardless of the chosen representation, after imposing boundary conditions, both solutions are, of course, equivalent.
As the ABGB geometry reduces to the Schwarzschild solution in the limit $Q=0,$ it is the charge, no matter how small, that secures regularity. The vanishing of the charge leads therefore to dramatic changes in the geometry of the black hole interior. On the other hand, the terms calculated to ${\mathcal O}(Q^{2})$ for the ABGB spacetime coincide with the Reissner-Nordström solution. Let us consider a series expansion of the function $M_{1}^{\left( 1\right)
}\left( r\right) $ for small $q =|Q|/\mathcal{M}.$ After some algebra one has $$\begin{aligned}
M_{1}&=&\left(\frac{2}{x^{3}} - \frac{3}{x^{4}} \right)\frac{\beta}{\mathcal{M}}q^{2} +
\frac{6\beta}{5\mathcal{M}}q^{4} - \left[\left( \frac{3}{x^{5}}
-\frac{11}{2 x^{6}} \right)\alpha
+ \left(\frac{9}{4 x^{5}} - \frac{4}{x^{6}} \right)\beta\right]
\frac{q^{6}}{\mathcal{M}}\nonumber \\
&-&\left(\frac{5}{2}\alpha + \frac{13}{7}\beta \right)\frac{q^{8}}{\mathcal{M}x^{7}}\,+\,...
-b_{2} \left(\frac{q^{6}}{12 x^{3}} - \frac{q^{10}}{60 x^{5}}
+ \frac{17q^{14}}{6720 x^{7}} \right) + ...,
\label{exp_M1}\end{aligned}$$ where $x=r/\mathcal{M}.$ It could be easily checked that for the regular line element the leading term of the expansion is proportional to $b_{2}Q^{6}.$ Since the coefficient $b_{2}$ is proportional to $Q^{-6}$ the constructed line element does not approach the Schwarzschild solution in the limit $Q \to 0$. Indeed, simple calculations yield $$ds^{2}=-\left( 1-\frac{2\mathcal{M}}{r}-\frac{2\mathcal{M}^{2}}{r^{4}}
k\right) dt^{2}
+\left[\left( 1-\frac{2\mathcal{M}}{r}\right)^{-1}\,+\,
\left( 1-\frac{2\mathcal{M}}{r}\right)^{-2}
\frac{2\mathcal{M}^{2}}{r^{4}
}k \right]dr^{2}
+r^{2}d\Omega ^{2}, \label{limitt}$$ where $k$ is given by $$k=\frac{\pi ^{2}}{12}\sigma \approx 5.781\alpha -0.486\beta .$$ Consequently, one has either Schwarzschild asymptotic of a singular line element with $b_{2} =0$ ($k =0$) or a regular line element with non-Schwarzschild $Q = 0$ limit for $b_{2} = - \mu.$ It is possible, of course, to accept other values of the parameter $b_{2}$ but it seems that they are of lesser importance.
The approximate location of the event horizon of the line element (\[limitt\]) lies near its Schwarzschild value and is approximately given by $$r_{+}\approx 2\mathcal{M}\left( 1+\frac{k}{8\mathcal{M}^{2}}\right) .$$ As is well-known, the Hawking temperature, $T_{H},$ can be easily calculated from the metric tensor without referring to the field equations. The standard by now method of obtaining $T_{H}$ relies on the Wick rotation. The Euclidean line element has no conical singularity provided the time coordinate is periodic with a period $P$ given by $$P=4\pi \lim_{r\rightarrow r_{+}}\left( g_{tt}g_{rr}\right) ^{1/2}\left(
\frac{d}{dr}g_{tt}\right) ^{-1} .
\label{period}$$ Its reciprocal is identified with the black hole temperature, which, in the case in hand, reads $$T_{H}\,=\,\frac{1}{8\pi\mathcal{M}}\left(1 + \frac{k}{4\mathcal{M}^{2}} \right) .
\label{th}$$ Note that for $\alpha$ and $\beta$ satisfying $$\alpha = \frac{1680 + 98\pi^{2} - 31\pi^{4}}{2 (31 \pi^{4} - 2520)}\beta,$$ both (\[limitt\]) and (\[th\]) reduce to their Schwarzschild counterparts.
Finally, we shall investigate the important issue of regularity. However, before we proceed further a few words of comments are in order. It should be emphasized that we have constructed a linearized solution to the coupled system of equations of quadratic gravity and nonlinear electrodynamics only. Consequently, when calculating the curvature invariants we are to restrict ourselves to $O(\varepsilon)$ terms. As it has been stressed earlier, the regularity of the line element does not necessarily entail the regularity of the underlying geometry: the latter requires various curvature invariants to be regular in the interesting region. First, let us concentrate on the Kretschmann scalar, $K.$ It could be easily shown that for the constructed line element it consists of terms involving products of the functions $M_{0}(r),$ $M_{1}(r),$ $\psi_{1}(r)$ and their derivatives. Indeed, after some algebra one obtains $K = K_{0} + \varepsilon \Delta K,$ where $K_{0}$ is the Kretschmann scalar of the ABGB spacetime $$K_{0} = \frac{48 M_{0}^{2}}{r^{6}} +\frac{16 M_{0}^{\prime\prime}M_{0}}{r^{4}}
+\frac{32 {M_{0}^{\prime}}^{2}}{r^{4}} - \frac{64 M_{0}^{\prime} M_{0}}{r^{5}}
+\frac{4 {M_{0}^{\prime\prime}}^{2}}{r^{2}} -
\frac{16 M_{0}^{\prime\prime}M_{0}^{\prime}}{r^{3}}$$ and $\Delta K$ is given by $$\begin{aligned}
\Delta K &=&\left( \frac{128 M_{0} M_{0}^{\prime}}{r^{4}}+\frac{16 M_{0}}{r^{4}}
-\frac{80 M_{0}^{2}}{r^{5}}
-\frac{16 M_{0}^{\prime}}{r^{3}} - \frac{24 M_{0} M_{0}^{\prime\prime}}{r^{3}} +
\frac{24 M_{0}^{\prime} M_{0}^{\prime\prime}}{r^{2}} -\frac{48 {M_{0}^{\prime}}^{2}}{r^{3}}
\right)\psi_{1}^{\prime}\nonumber \\
&+&
\left(-\frac{8 M_{0}^{\prime\prime}}{r} +
\frac{16 M_{0} M_{0}^{\prime\prime}}{r^{2}}+\frac{32 M_{0}^{2}}{r^{4}}
+ \frac{16 M_{0}^{\prime}}{r^{2}} - \frac{32 M_{0}M_{0}^{\prime}}{r^{3}}
- \frac{16 M_{0}}{r^{3}}\right)
\psi_{1}^{\prime\prime}\nonumber\\
&+&\left(-\frac{16 M_{0}^{\prime\prime}}{r^{4}}
-
\frac{96 M_{0}}{r^{6}} + \frac{64 M_{0}^{\prime}}{r^{5}} \right)M_{1}
+
\left(\frac{16 M_{0}^{\prime}}{r^{3}} -\frac{16 M_{0}}{r^{4}}
-\frac{8 M_{0}^{\prime\prime}}{r^{2}} \right)M_{1}^{\prime\prime}\nonumber \\
&+&
\left(\frac{16 M_{0}^{\prime\prime}}{r^{3}} -
\frac{64 M_{0}^{\prime}}{r^{4}} +\frac{64 M_{0}}{r^{5}} \right) M_{1}^{\prime} .\end{aligned}$$ The regularity of $K_{0}$ has been demonstrated in Ref [@ABG]; in order to establish regularity of $\Delta K$ one has to establish regularity of its building blocks. To analyse behaviour of the $k-$th derivative of $M_{1}$ we shall employ the first order equation. From Eq. (\[1sta\]) one clearly sees that all derivatives of $S_{t}^{t}$ vanish as $r\to 0.$ It is simply because $k-$th derivative of $M_{0}$ has the asymptotic form $$\frac{d^{k}M_{0}}{dr^{k}} \sim
\sum_{i=k+1}^{2k}\frac{c_{i}}{r^{i}} \exp\left(-\frac{Q^{2}}{\mathcal{M}r}\right),$$ as $r \to 0,$ where the coefficients $c_{i}$ depend on $\mathcal{M}$ and $Q.$ It follows then that $k$-th derivative behaves as $r^{-2k} \exp(-Q^{2}/\mathcal{M}r)$ near the center. One encounters a similar behaviour in the second term in the right hand side of Eq. (\[eqM1\]). Pulling this all together one concludes that the Kretschmann scalar is indeed regular at the center. Moreover, one can easily establish the regularity of $R^{2}$ and $R_{ab} R^{ab}.$ Similar arguments can be used in demonstration that higher invariants constructed from the curvature are also regular.
Geometry {#sec4a}
========
Generally speaking a black hole belongs to one of the two distinct classes: it may be either extremal (when horizon has at least twofold degeneracy) or nonextremal. In the former case the family of ultraextremal black holes, i. e. configurations with triple (or even higher) degeneracy can be singled out [@dirty_bh; @semi]. Since we have restricted ourselves to the case of vanishing cosmological constant, we expect that the ABGB black hole possesses at most two horizons. (Generalization of the ABGB solution to $\Lambda \neq 0$ case is straightforward.) Indeed, depending on $q=|Q|/\mathcal{M}$ the ABGB spacetime has two, one, or has no horizons at all. Simple calculations indicate that for $q<q_{c},$ where $q_{c}$ is a critical value of $q,$ it has both the inner and outer horizon, whereas for $q>q_{c\text{ }
}$ the metric potential $g_{tt}(r)$ has no real roots for $r \geq 0.$ For $q=q_{c\text{ }}$ the event and inner horizons merge and the configuration becomes extremal. It should be noted that the extremal ABGB geometry is distinct from, say, the geometry of the extremal Reissner-Nordström solution. Indeed, because of the regularity of the former, the configurations with $q>q_{c}$ are not forbidden by a cosmic censor.
Nonextreme black holes {#sec4}
----------------------
In order to determine the location of the inner and outer horizon we shall analyse equation $
g_{tt}\left( r\right) =0$ and construct the iterative solution restricting ourselves to the terms linear in $\varepsilon .$ As the solution can be written as $$r_{\pm }=r_{0}^{\left( \pm \right) }+\varepsilon r_{1}^{\left( \pm \right) },$$ one has to solve the simple system of two equations. First of them $$1-\frac{2M_{0}\left( r_{0}^{\left( \pm \right) }\right) }{r_{0}^{\left( \pm
\right) }}=0 \label{hor+1}$$ admits solutions expressible in terms of the Lambert special functions. Indeed, it has been demonstrated in Ref. [@Kocio1] that $x_{0}^{\left( \pm\right) }=r_{0}^{\left( \pm \right) }\slash\mathcal{M}$ are given by $$x_{0}^{\left( \pm \right) }=-\frac{4q^{2}}{4W_{\pm }\left( -\frac{q^{2}}{4}\exp \left( \frac{q^{2}}{4}\right) \right) -q^{2}}, \label{x1}$$ where $W_{+}\left( s\right) \equiv W(0,s)$ and $W_{-}\left( s\right) \equiv
W(-1,s)$ are the only real branches of the Lambert functions [@Knuth]. On the other hand, $r_{1}$ satisfies algebraic equation which can be easily solved. After some manipulations one has $$r_{1}^{\left( \pm \right) }=\frac{2M_{1}\left( r_{0}^{\left( \pm \right)
}\right) }{1-2M_{0}^{\prime }\left( r_{0}^{\left( \pm \right) }\right) }.
\label{hor+2}$$ Now, inserting the zeroth-order solutions (\[x1\]) into Eq. (\[hor+2\]) one can easily determine $r_{1}^{\left( \pm \right) },$ and the result can be further simplified with the substitution $$x_{0}=\frac{4\xi }{\xi +1},$$ where $\xi $ (defined as in Eq. (\[xi1\])) is calculated at $x=x_{0}.$ Specifically, denominator in the right hand side of Eq. (\[hor+2\]) reduces to the simple expression: $$1-2M_{0}^{\prime }\left( r_{0}^{\left( \pm \right) }\right) =\frac{4\xi
-q^{2}}{4\xi }.$$ To avoid unnecessary proliferation of long formulas we shall not present the results for $r_{1}^{\left( \pm \right) }$ here as they could be readily obtained by a direct substitution of the equation (\[x1\]) and the expression describing $M_{1}$ at $ r^{(\pm)}_{0} $ into Eq. (\[hor+2\]).
Although the complexity of the expression describing $r_{+}$ makes its direct analytical application rather intricate, one can obtain interesting and important information analyzing limiting cases. Below we derive expansion valid for small $q.$ The extremality limit will be analyzed in the next subsection. Expanding $r_{+}$ and collecting the terms with like powers of $q,$ after massive simplifications, one has $$\begin{aligned}
r_{+} &=& 2\mathcal{M} - \frac{q^{2}}{2}\mathcal{M} -
\frac{q^{4}}{8}\mathcal{M} - \frac{5q^{6}}{96}\mathcal{M} +\, ...\,
+ \frac{1}{4\mathcal{M}}k +\left(\frac{1}{4\mathcal{M}}k
+\frac{\beta}{8\mathcal{M}} \right)q^{2}
\nonumber \\
&+& \left(\frac{71}{320 \mathcal{M}}k+\frac{17}{160\mathcal{M}}\beta \right)q^{4}
+ \left(\frac{123}{640\mathcal{M}}k +
\frac{47}{640\mathcal{M}}\beta -\frac{1}{64\mathcal{M}}\alpha \right)q^{6} + ...
\label{rozw1}\end{aligned}$$ Analogous expression for the event horizon of the Reissner-Nordström solution in quadratic gravity reads $$\begin{aligned}
r_{+} &=& 2\mathcal{M} - \frac{1}{2}\mathcal{M}q^{2} - \frac{1}{8}\mathcal{M}q^{4}
- \frac{1}{16}\mathcal{M}q^{6} + ... \nonumber \\
&+& \left(\frac{1}{8\mathcal{M}}q^{2} + \frac{17}{160\mathcal{M}}q^{4}
+ \frac{57}{640\mathcal{M}}q^{6} + ... \right)\beta .
\label{rozw2}\end{aligned}$$ Inspection of (\[rozw1\]) and (\[rozw2\]) shows that for $b_{2} = 0$ (or, equivalently, $k =0$) both expansion coincide up to $q^{4}$ terms.
One of the most important characteristics of the black hole is its temperature. To investigate how it is modified by the quadratic terms, we employ a general expression (\[period\]), which, in the present context, can be rewritten in the form $$T_{H} = \frac{1}{4\pi} \lim_{r \to {r_{+}}} e^{-\varepsilon \psi_{1}}
\left( \frac{1}{r}- \frac{2\varepsilon}{r}S^{t}_{t} + 8\pi r T^{t}_{t} \right).$$ Further, expanding $T_{H},$ in the auxiliary parameter, collecting the terms with like powers of $\varepsilon$ and linearizing the thus obtained result one has $$T_{H} = T_{0} + \varepsilon \Delta T ,$$ where $T_{0}$ coincides with the temperature of the nonextremal ABGB black hole $$T_{0} = \frac{1}{4\pi}\left[ \frac{1}{r_{0}} -
\frac{Q^{2}}{\mathcal{M} r_{0}^{2}}\left( 1 - \frac{r_{0}}{4 \mathcal{M}} \right)\right]$$ and $\Delta T$ is given by $$\Delta T = r_{1}\left(2 r_{0}
\frac{d T^{t}_{t}}{dr}_{| r_{0}} -\frac{1}{2 \pi r_{0}^{2}} +
\frac{T_{0}}{r_{0}} \right) - \frac{1}{2 \pi r_{0}} S_{t}^{t} +
T_{0} \psi_{1} (r_{0}) +2 r_{0} T^{t}_{t}.$$ Here $r_{0} = r_{0}^{(+)},$ whereas $T^{t}_{(0)t}$ and $T_{(1)t}^{t} $ are given by the first and the second term in the right hand side of Eq. (\[tep\]), respectively.
Of course, the general expression for temperature is also too complicated to be displayed here. We shall, therefore, calculate its limiting behaviour for small $q$ precisely in the same manner as it has been done with the location of the event horizon $r_{+}.$ As the temperature of the extremal configuration is zero, this condition can be used in establishing relation between $Q$ and $\mathcal{M}.$ This will be done in the next subsection. The expansion of the temperature of the nonextremal black hole reads $$\begin{aligned}
T_{H} &=& \frac{1}{8\pi \mathcal{M}} - \frac{q^{4}}{128\pi\mathcal{M}} -
\frac{5 q^{6}}{768\pi \mathcal{M}} + ...\nonumber\\
&+&\frac{k}{32\pi \mathcal{M}^{3}} + \frac{q^{2}}{64\pi \mathcal{M}^{3}}(3 k -\beta)
+\frac{q^{4}}{80\pi\mathcal{M}^{3}}\left(\frac{33}{8}k - \beta \right) + ...\end{aligned}$$ whereas for the Reissne-Norström geometry in quadratic gravity one has $$T_{H} = \frac{1}{8\pi\mathcal{M}} - \frac{q^{4}}{128\pi\mathcal{M}} -\frac{q^{6}}{128\pi\mathcal{M}}
+\, ... \,+\left( \frac{q^{4}}{320 \mathcal{M}} + \frac{3 q^{6}}{512 \mathcal{M}}\,+ ...\right)\beta .$$
Extreme black holes {#sec5}
-------------------
In this section we shall investigate the important issue of extremal black holes. When $r_{+}$ and $r_{-}$ of the classical ABGB black hole merge into the one degenerate horizon one has the extremal configuration characterized by vanishing surface gravity (Hawking temperature). The geometry of the vicinity of the degenerate horizon belongs to the ${\rm AdS}_{2}\times {\rm S}^{2}$ class with different modules of curvatures of subspaces [@Kocio3]. One of the peculiarities of the general extremal solution is infinite proper distance between two points, one of which resides on the event horizon. The degeneracy means that in the vicinity of the horizon the leading behaviour of the metric potential $g_{tt}$ is of the form $g_{tt}\sim \left( r-r_{+}\right) ^{2},$ or equivalently $$g_{tt}\left( r_{+}\right) =g_{tt}^{\prime }\left( r_{+}\right) =0.
\label{eqqss}$$ Combining these equations one obtains simple algebraic equation that could be easily solved $$x_{+} = \frac{4 q_{c}^{2}}{4-q_{c}^{2}}.$$ It can be demonstrated that the parameters $q$ and $x_{x}$ of the classical extremal ABGB black hole can be expressed in terms of the principal branch of the Lambert function evaluated at $e^{-1}.$ Employing definition of $W_{+}$ one has $$q_{c}=2\sqrt{W_{+}\left( e^{-1}\right) }\approx 1.0554,$$ whereas the location of the degenerate horizon is given by $$x_{c}=\frac{4W_{+}\left( e^{-1}\right) }{1+W_{+}\left( e^{-1}\right) }\approx 0.8712 .$$ Here and below $x$ (with or without sub or superscripts) denotes appropriate radial coordinate divided by $\mathcal{M}.$
Now, we shall examine modifications caused by quadratic gravity. Lengthy but routine calculations show that the extremal canfiguration is possible for $$q^{2}=q_{c}^{2} + \delta ,$$ where the correction $\delta$ is given by $$\delta =\frac{1}{\mathcal{M}^{2}}\sum_{i=0}^{6}\left( \alpha h_{i}^{\left(
\alpha \right) }+\beta h_{i}^{\left( \beta \right) }\right) \mathrm{Li}_{i}\left( w\right) \,+\,\frac{2(3w-1)}{\mathcal{M}}b_{2}
\label{delta}$$ with $$h_{0}^{\left( \alpha \right) }=-\frac{(1+w)^{3}}{240 w^{3}}
\left( 516-69w+11w^{2}-5w^{3}-w^{4}\right)
- \frac{(1+w)^{2}\pi^{2}}{10080w^{4}}\left(31\pi^{2}-2520 \right) ,$$ $$h_{1}^{\left( \alpha \right) }=-\frac{\left( 1+w\right) ^{2}}{40w^{3}}\left(
119-4 w-6 w^{2}-4 w^{3}-w^{4}\right) ,$$ $$h_{2}^{\left( \alpha \right) }=-\frac{1}{8w^{3}}\left(1+w \right)\left(
23-4 w-6 w^{2}-4 w^{3}-w^{4}\right) ,$$ $$\frac{2}{\left( 1+w\right) ^{3}}h_{3}^{\left( \alpha \right) }=\frac{2}{\left( 1+w\right) ^{2}}h_{4}^{\left( \alpha \right) }=\frac{1}{\left(
1+w\right) }h_{5}^{\left( \alpha \right) }=h_{6}^{\left( \alpha \right) }=\frac{3\left( 1+w\right) }{w^{3}},$$ $$h_{0}^{\left( \beta \right) }=-\frac{(1+w)^{3}}{480w^{3}}
\left(356+21w+21w^{2}-5w^{3}-w^{4} \right)
+ \frac{(1+w)^{2}\pi^{2}}{20160w^{4}}\left(1680 + 98\pi^{2} - 31 \pi^{4} \right) ,$$ $$h_{1}^{\left( \beta \right) }=-\frac{\left( 1+w\right) ^{2}}{240w^{3}}\left(
257+28w+2w^{2}-12w^{3}-3w^{4}\right) ,$$ $$h_{2}^{\left( \beta \right) }=-\frac{1}{16w^{3}}\left(1+w\right)\left(
19+4w+2w^{2}-4w^{3}-w^{4}\right) ,$$ $$h_{3}^{\left( \beta \right) }=-\frac{1}{4w^{3}}\left( 1+w \right)^{2}\left(
1-2w-w^{2}\right) ,\qquad h_{4}^{\left( \beta \right) }=\frac{1}{
4w^{3}}\left(1+w \right)\left( 1+6w+3w^{2}\right) ,$$ $$\frac{1}{1+w}h_{5}^{\left( \beta \right) }=h_{6}^{\left( \beta \right) }=
\frac{3\left( 1+w\right) }{2w^{3}},$$ and $$w=W_{+}\left( e^{-1}\right) .$$ The term containing $b_{2} (=- \mu),$ i. e., the last term in right hand side of Eq. (\[delta\]) has been singled out for convenience. For the extreme black hole $\mu$ reads $$\mathcal{\mu }=\frac{\pi ^{2}}{64\mathcal{M}w^{3}}\sigma,$$ where $\sigma$ is defined through the relation (\[mu\]). The location of the degenerate horizon is given by $$x=x_{c}+ \Delta,$$ where $ \Delta$ can be compactly expressed in the form $$\Delta =\frac{\delta w}{\left( 1+w\right) ^{2}}+
\frac{2\alpha+\beta}{16w\mathcal{M}^{2}}(w+3)(w-1)
-\frac{4w(w-1)}{\mathcal{M}(1+w)^{2}}b_{2},$$ with $\delta$ given by Eq. (\[delta\]). As both $\delta$ and $\Delta$ depend on the particular value of the Lambert function at $e^{-1}$ one can easily calculate their numerical values. Indeed, simple calculations yield $$\delta = -\frac{0.81451}{\mathcal{M}^{2}}\beta-\frac{3.67845}{\mathcal{M}^{2}}\alpha$$ and $$\Delta =
\frac{1.40646}{\mathcal{M}^{2}}\beta + \frac{3.88206}{\mathcal{M}^{2}}\alpha ,$$ for $b_{2} = - \mu,$ whereas with $b_{2} =0$ one has $$\delta = -\frac{0.57553}{\mathcal{M}^{2}}\beta-\frac{0.05121}{\mathcal{M}^{2}}\alpha$$ and $$\Delta =
-\frac{0.43288}{\mathcal{M}^{2}}\beta - \frac{1.05314}{\mathcal{M}^{2}}\alpha .$$ Note that for particular choices of the coupling parameters it is possible to reduce either $x_{+}$ or $q_{c}$ to its general relativistic values. Obtained corrections can be contrasted with the analogous results calculated for the extreme Reissner-Nordström black hole in quadratic gravity: $$q_{c}^{2} = 1 + \frac{2}{5\mathcal{M}^{2}}\beta$$ and $$x_{+} = 1 + \frac{1}{5\mathcal{M}^{2}}\beta.$$
Conclusion and summary {#sec6}
======================
In this paper we have constructed perturbative solutions describing spherically symmetric and static black holes to the coupled equations of fourth-order gravity with the source term given by the stress- energy tensor of the nonlinear electrodynamics. The Lagrangian of the nonlinear field is a natural generalization of the Bronnikov’s Lagrangian. Because of technical complexities we have restricted ourselves to the first-order corrections, i. e. the terms proportional to $\varepsilon.$ The obtained line element is parametrized by two integration constants and free parameters. Integration constants are related to the (magnetic) charge and a total mass of the system as seen by a distant observer whereas the free parameters are adjusted to make the resulting solution regular at the center. The regularity heavily relies on the form of the zeroth-order solution which coincides with the solution of the ABGB-type and the special properties of the polylogarithms ${\rm Li}_{i}.$ The metric potentials thus computed enabled construction of the basic characteristics of the geometry and its asymptotics: location of the inner and outer horizon, Hawking temperature and the relation between $Q$ and $\mathcal{M}$ for the extremal configurations.
If one is interested in the solution possessing Schwarzschild asymptotics as $Q\to 0$ or intend to study the external region without any relations to the issue of regularity it suffices to put $b_{2} =0$ throughout the paper. We intentionally presented basic formulas in a form allowing for a different choices of the parameter $b_{2}.$ It should be stressed that the only choice leading to regularity of the solution at the center is $b_{2}\,=\,-\mu.$
The calculations and results presented in this paper suggest some generalizations and obvious extensions. First, one may attempt to go beyond first order calculations in order to establish full regularity of the solution. Of course it would be interesting and desirable, with all conceptual limitations of the method, to demonstrate it in a nonperturbative way. However, the technical complexities may be a real obstacle in this regard. Further, it is possible to analyse behaviour of the test quantum fields in the ABGB background with the special emphasis put on the back reaction on the metric. Indeed, since the general expression describing the stress-energy tensor of quantized massive scalar, spinor and vector fields in the large mass limit is known [@Matyjasek:1999an; @Kocio1] it is possible to investigate the resultant quantum-corrected geometry even in the vicinity of the center of the black hole. Recently such a programme has been carried out in the case of the quantum-corrected Reissner-Nordström black holes (see Refs. [@jirinek01b; @jirinek03b] and the references cited therein). Finally, it should be noted that another important choice of the boundary conditions $$\begin{aligned}
M_{i}({r_{+}})\,=\,\begin{cases}
\frac{{r_{+}}}{2} & \text{if $i=0$},\\
0 & \text{if $i\geq 1$}
\end{cases}\end{aligned}$$ which is, in turn, related to the horizon defined mass of the black hole is not considered in this paper. Here we remark only that parametrizing solution by the exact location of the event horizon and the charge is quite natural and certainly deserves further study. A lesson that follows from investigations of the corrected Reissner-Nordström solution is that some relations, as for example these for the extremal configurations, are simpler and physically more transparent. We intend to return to these problems elsewhere.
In this appendix we sketch the method of evaluating the indefinite integrals appearing in the solution of the equation (\[f\_o\]). After substitution of auxiliary variable $u=x^{-1}$ we have to find an expression for the integral of the form: $$\int u^{p} (\tanh u)^{s} du,$$ with natural $p, s$. We show that when $s>1$ the above integral can be reduced to the integral with $s=1$ and then we will express the latter by a sum containing polylogarithm functions. At the beginning we use the formulas (1.4.22.3) and (1.4.22.4) from the very extensive tables of integrals [@Prud]:
$$\int u^{p} (\tanh u)^{2n} du = \frac{1}{p+1}u^{p+1}
+\sum_{k=0}^{n-1}(-1)^{n+k}\binom{n}{k}\int\frac{u^{p}}{(\cosh u)^{2n-2k}}
du,$$
$$\begin{aligned}
\int u^{p} (\tanh u)^{2n+1} du &=& \int u^{p} \tanh u\,du
+\sum_{k=0}^{n-1}
\frac{(-1)^{n+k}}{2n-2k}\binom{n}{k} \left[-\frac{u^{p}}{(\cosh u)^{2n-2k}}\right.
\nonumber \\
&&\left. +\, p\int\frac{u^{p-1}}{(\cosh u)^{2n-2k}} du\right].\end{aligned}$$
Evaluation of integrals containing even powers $2n-2k>2$ of sech$\,$ function can be further reduced to the case with the square of ${\rm sech} $ in the integrand by repeated use of the formula (1.4.24.1) of [@Prud]: $$\begin{aligned}
\int \frac{u^{p}}{(\cosh u)^{q}} du &= & \frac{pu^{p-1}}{(q-1)(q-2)(\cosh
u)^{q-2}} + \frac{u^{p}}{(q-1)\sinh u(\cosh u)^{q-1}} \nonumber\\
& & - \frac{p(p-1)}{(q-1)(q-2)}\int\frac{u^{p-2}}{(\cosh u)^{q-2}} du +
\frac{q-2}{q-1}\int\frac{u^{p}}{(\cosh u)^{q-2}} du.\end{aligned}$$ As a simple integration by parts yields: $$\int \frac{u^{p}}{(\cosh u)^{2}} du = u^{p}\tanh u - p\int u^{p-1}\tanh
u\,du,$$ we clearly see that to accomplish our calculation we have to find the integral: $$\int u^{p}\tanh u\,du$$ which requires special treatment. It will be proven below that it is expressible in terms of the polylogarithm functions $\mathrm{Li}_{n}(z),
n=1, 2,..., p+1,$ with a properly chosen argument $z$. There is considerable literature on analytical and numerical properties of these functions [@Levin1; @Levin2] being a generalization of Euler’s dilogarithm. In the unit circle polylogarithm of integral order $m>1$ can be defined by $$\mathrm{Li}_{m}(z) = \sum_{k=1}^{\infty}\frac{z^{k}}{k^{m}}.$$ From this series we easily derive the important recurrence relation: $$\frac{d\:}{dz}\mathrm{Li}_{n}(z) = \frac{1}{z}\mathrm{Li}_{n-1}(z)
\label{recc}$$ which will also be used in the integral form: $$\mathrm{Li}_{n}(z) = \int\frac{\mathrm{Li}_{n-1}(z)}{z} dz +C.$$ Using this property and the integral representation of dilogarithm $$\mathrm{Li}_{2}(z)=-\int_{0}^{z}\frac{\ln(1-t)}{t}\,dt$$ one can extend the definition of polylogarithms to low orders: $$\mathrm{Li}_{1}(z)=-\ln(1-z),\:\:\mathrm{Li}_{0}(z)=\frac{z}{1-z}.
\label{AA}$$ Starting from the recurrence relation in the integral form we repeatedly integrate by parts using the expression for the derivative of a polylogarithm (\[recc\]) at every step: $$\begin{aligned}
\mathrm{Li}_{n}(z)& =& C+\int (\ln z)^{\prime}\,\mathrm{Li}_{n-1}(z) dz =...=\nonumber
\\
&=& C+\ln z\,\mathrm{Li}_{n-1}(z) -\frac{1}{2}(\ln z)^{2}\,\mathrm{Li}
_{n-2}(z) +\frac{1}{2\cdot 3}(\ln z)^{3}\,\mathrm{Li}_{n-3}(z)+... \nonumber \\
& & +\frac{(-1)^{n-1}}{(n-2)!}(\ln z)^{n-2}\,\mathrm{Li}_{2}(z) +\frac{
(-1)^{n-1}}{(n-1)!}(\ln z)^{n-1}\,\ln(1-z) +\frac{(-1)^{n-1}}{(n-1)!}\int
\frac{(\ln z)^{n-1}}{1-z}dz,\nonumber\\
\label{AAA}\end{aligned}$$ where in the final step we have taken into account the form of $\mathrm{Li}
_{1}(z)$. If we now set $z=-e^{-2u}$ we recover the required integral in the last term: $$(-1)^{n-1}\int\frac{(\ln z)^{n-1}}{1-z}dz =
2^{n}\int u^{n-1}\frac{e^{-u}}{e^{u}+e^{-u}}du = 2^{n-1}
\left[\frac{u^{n}}{n} -\int u^{n-1}\tanh u\,du \right].$$ After rearrangement of the sum in the identity (\[AAA\]) and change $
n\!-\!1\rightarrow p$ we get the expression: $$\begin{aligned}
\int u^{p}\tanh u\,du &=& C + \frac{u^{p+1}}{p+1} + u^{p}\ln(1+e^{-2u}) -
\sum_{k=1}^{p}\frac{p!}{2^{k}(p-k)!}\,u^{p-k}\,\mathrm{Li}_{k+1}(-e^{-2u}).\nonumber\\\end{aligned}$$ The hyperbolic functions and monomials in $u$ variable arising in our calculation can also be cast in the unified way when the special forms of low-order polylogarithms (\[AA\]) are used.The case studied here falls into a wide range of expressions which can be integrated with the help of polylogarithms. Some occurrences of these functions in physical problems are mentioned in the Levin’s book [@Levin1]. It is remarkable that they arise in Feynman diagrams integrals, in particular in computation of quantum electrodynamics corrections to the electron gyromagnetic ratio [@Laporta].\
[55]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, (), .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, ****, (), .
, **, Cambridge Monographs on Mathematical Physics (, , ), ISBN .
, ****, (), .
, , , ****, (), .
, , , ****, (), .
, ****, (), .
, ****, (), .
, , , , ****, ().
, ****, ().
, ****, (), .
, ****, (), .
, ****, ().
, ****, ().
, in ** (, ).
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, (), .
, ****, (), .
, , , ****, ().
, ****, (), .
, ****, (), .
, ****, (), .
, ****, (), .
, ****, (), .
, ****, ().
, ****, ().
, ****, (), .
, ****, (), .
, , , ** (, , ).
, , , ****, (), .
, ****, (), .
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, (), .
, ** (, , ).
, ****, (), .
, ****, (), .
, , , , , ****, ().
, ****, (), .
, ****, (), .
, ****, (), .
, ****, ().
, , , ** (, , ).
, ** (, , ).
, ****, (), .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $p \geqslant 3$ be a prime number and let $n \geqslant 0$ be an integer such that $p-1$ divides $n.$ In this short note we construct a family of $(p,n)$-gonal Riemann surfaces of maximal genus $2np+(p-1)^2$ with more than one $(p,n)$-gonal group.'
address: 'Departamento de Matemática y Estadística, Universidad de La Frontera, Avenida Francisco Salazar 01145, Temuco, Chile.'
author:
- 'Sebastián Reyes-Carocca'
title: |
A family of $(p,n)$-gonal Riemann surfaces with\
several $(p,n)$-gonal groups
---
[^1]
Introduction and statement of the result
========================================
Let $S$ be a compact Riemann surface of genus $g \geqslant 2$ and let $\mbox{Aut}(S)$ denote its automorphism group. If $p \geqslant 2$ is a prime number and $n \geqslant 0$ is an integer then $S$ is called [*$(p,n)$-gonal*]{} if there exists a group of automorphisms $$C_p \cong H \leqslant \mbox{Aut}(S)$$such that the corresponding orbit space $S/H$ has genus $n.$ The group $H$ is called a [*$(p,n)$-gonal group*]{} of $S.$
Each compact Riemann surface with non-trivial automorphisms is $(p,n)$-gonal for suitable of $p$ and $n.$ This simple fact shows that to study $(p,n)$-gonal Riemann surfaces and their automorphisms is equivalent to study the singular locus of the moduli space of compact Riemann surfaces.
$(p,n)$-gonal Riemann surfaces and their automorphisms have been extensively considered over the last century as they generalize important and well-studied classes of Riemann surfaces, such as $(2,0)$-gonal or [*hyperelliptic*]{}, $(p,0)$-gonal or [*$p$-gonal*]{} and $(2,n)$-gonal or [*$n$-hyperelliptic*]{} Riemann surfaces, among others.
Let $S$ be a $p$-gonal Riemann surface of genus $g \geqslant 2.$ By the classical Castelnuovo-Severi inequality (see Accola’s book [@accola]), if $$\label{ec}g > (p-1)^2$$then the $p$-gonal group is unique in the automorphism group of $S$. A family of $p$-gonal Riemann surfaces of maximal genus $g=(p-1)^2$ endowed with two $p$-gonal groups was constructed in [@ciy], showing that the bound is sharp. Furthermore, in the general case, following [@gabino], if $S$ has two $p$-gonal groups then they are conjugate in the automorphism group of $S$. An upper bound for the number of such groups was obtained in [@grego].
For $(p,n)$-gonality with $n \geqslant 1,$ the Castelnuovo-Severi inequality ensures that if $S$ is a $(p,n)$-gonal Riemann surface of genus $g\geqslant 2$ and $$\label{ec2}g > 2pn + (p-1)^2$$then the $(p,n)$-gonal group is unique in the automorphism group of $S$. In the general case, it was proved in [@greww] that if $S$ is a $(p,n)$-gonal Riemann surface of genus $g$ and $p > 2n + 1$ then all its $(p, n)$-gonal groups are conjugate in the automorphism group of $S;$ an upper bound for the size of the corresponding conjugacy class was also determined in the same paper. Later, in [@andreas], the uniqueness of the $(p,n)$-gonal group was proved to be true under the assumptions that the $(p,n)$-gonal group acts with fixed points and $p > 6n-6.$
This short note is devoted to provide a family of $(p,n)$-gonal Riemann surfaces of maximal genus $g=2pn+(p-1)^2$ with two $(p,n)$-gonal groups. The existence of this family shows that the bound is sharp, for infinitely many pairs $(p,n).$
\[t1\] Let $p \geqslant 3$ be a prime number and let $n \geqslant 0$ be an integer such that $p-1$ divides $n.$ Set $$d=n/(p-1) + 1.$$Then there exists a complex $d$-dimensional family of $(p,n)$-gonal Riemann surfaces $S$ of genus $$g=2np+(p-1)^2$$with automorphism group of order $4p^2$ acting on $S$ with signature $$(0; 2,2,2,p, \stackrel{d}{\ldots},p)$$in such a way that each $S$ has more than one $(p,n)$-gonal group.
It is worth mentioning here the following observations which will follow from the proof of the theorem.
1. The result remains true if $p=2$ and $n$ is odd.
2. If $n=0$ our family agrees with the family constructed in [@ciy].
Proof of the Theorem
====================
Let $\Delta$ be a Fuchsian group of signature $(0; 2,2,2,p, \stackrel{d}{\ldots},p)$ canonically presented as $$\Delta=\langle \gamma_1, \ldots, \gamma_{d+3} : \gamma_1^2=\gamma_2^2=\gamma_3^2=\gamma_4^p=\cdots =\gamma_{d+3}^p=\gamma_1\cdots\gamma_{d+3}=1\rangle$$and consider the group $G=\mathbf{D}_p \times \mathbf{D}_p$ (where $\mathbf{D}_p$ denotes the dihedral group of order $2p$) presented in terms of generators $s_1, s_2, r_1, r_2$ and relations$$s_1^2=s_2^2=r_1^p=r_2^p=(s_1r_1)^2=(s_2r_2)^2=[s_1, r_2]=[s_1, s_2]=[r_1, r_2]=[r_1, s_2]=1.$$
[*Existence of the family.*]{} By virtue of the classical Riemann’s existence theorem, the existence of the desired family follows after verifying that the Riemann-Hurwitz formula holds and after providing a surface-kernel epimorphism $\theta$ from $\Delta$ onto $G.$
Note that the equality $$2(g-1)=4p^2(-2+3(1-\tfrac{1}{2})+d(1-\tfrac{1}{p}))$$shows that the Riemann-Hurwitz formula is satisfied for a branched $4p^2$-fold covering map from a compact Riemann surface of genus $g=2np+(p-1)^2$ onto the projective line, ramified over three values marked with 2 and $d$ values marked with $p$.
In addition, if $d$ is odd we can choose the surface-kernel epimorphism $\theta$ as $$\theta(\gamma_1)= s_1, \,\, \theta(\gamma_2)=s_2, \,\, \theta(\gamma_3)= s_1s_2r_1r_2 \, \mbox{ and }\, \theta(\gamma_i)= \left\{ \begin{array}{ll}
(r_1r_2)^{-1} & \textrm{if $i$ is even}\\
\,\,\, r_1r_2 & \textrm{if $i$ is odd,}
\end{array} \right.$$and if $d$ is even we can choose $\theta$ as $$\theta(\gamma_1)= s_1, \,\, \theta(\gamma_2)= s_2, \,\, \theta(\gamma_3)=s_1s_2(r_1r_2)^{-d/2} \, \mbox{ and }\, \theta(\gamma_i)= \left\{ \begin{array}{ll}
r_1 & \textrm{if $i$ is even}\\
r_2 & \textrm{if $i$ is odd}
\end{array} \right.$$where $i \in \{4, \ldots, d+3\}.$
The complex dimension of the family agrees with the complex dimension of the Teichmüller space associated to $\Delta;$ namely, its dimension is $d.$
[*$(p,n)$-gonal groups.*]{} We denote the branched regular covering map given by the action of $G$ on $S$ by $\pi: S \to S/G$ and its branch values by $y_1,y_2, y_3, z_1 \ldots, z_{d}$, where each $y_k$ is marked with 2 and each $z_k$ is marked with $p.$
Assume $d$ odd. Consider the cyclic subgroups of order $p$ $$H_1 = \langle r_1r_2 \rangle \,\, \mbox{ and }\,\, H_2 = \langle r_1^{-1}r_2 \rangle$$of $G.$ We denote by $\pi_1$ and $\pi_2$ the branched regular covering maps given by the action of $H_1$ and $H_2$ on $S$ respectively. We observe that the fiber of $\pi$ over each $y_k$ does not contain any branch value of $\pi_1$ and $\pi_2.$ In addition, for each $k$, the fiber of $\pi$ over $z_k$ has $4p$ elements; the isotropy group of $2p$ of them is isomorphic to $H_1$ and the remaining ones have isotropy group isomorphic $H_2.$ It follows that $\pi_1$ and $\pi_2$ ramify over $2pd$ values, each of them marked with $p.$ Equivalently, the signature of the action of $H_j$ on $S$ is $(n_j; p, \stackrel{2dp}{\ldots}, p)$ where $n_j$ is the genus of $S/H_j$. We now consider the Riemann-Hurwitz formula to see that $$2(g-1)=p[2n_j-2+2pd(1-\tfrac{1}{p})]$$and, after straightforward computations, one obtains that $n_j=n$ for $j=1,2$.
Assume $d$ even. Consider the cyclic subgroups of order $p$ $$H_1 = \langle r_1 \rangle \,\, \mbox{ and }\,\, H_2 = \langle r_2 \rangle$$of $G$ and let $\pi_1$ and $\pi_2$ be as before. As in the previous case, the fiber of $\pi$ over each $y_k$ does not contain any branch value of $\pi_1$ and $\pi_2.$ For each $k$ the fiber of $\pi$ over $y_k$ has $4p$ elements; the isotropy group of them is isomorphic to $H_1$ if $k$ is odd and is isomorphic to $H_2$ if $k$ is even. It follows that $\pi_1$ and $\pi_2$ ramify over $2pd$ values, each of them marked with $p.$ Equivalently, the signature of the action of $H_j$ on $S$ is $(n_j; p, \stackrel{2dp}{\ldots}, p)$ where $n_j$ is the genus of $S/{H_j}$. Similarly as previously done, the Riemann-Hurwitz formula ensures that $n_j=n$ for $j=1,2$.
In both cases, $H_1$ and $H_2$ are two $(p,n)$-gonal groups of $S,$ as desired.
Note that if $d$ is odd then the $(p,n)$-gonal groups are conjugate, but if $d$ even then they are not.
[99]{}
Lecture Notes in Math. [**1595**]{}, Springer (1994).
Geom. Dedicata [**147**]{} (2010), 139–147.
Ann. Mat. Pura Appl. (4) [**168**]{} (1995), 1–15.
Rocky Mountain J. Math. [**40**]{} (2010), no. 4, 1221–1226.
Geom. Dedicata [**149**]{} (2010), 1–14.
Arch. Math. (Basel) [**98**]{} (2012), no. 6, 591–598.
[^1]: Partially supported by Fondecyt Grants 11180024, 1190991 and Redes Grant 2017-170071
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a recent experiment Campbell *et al.* \[, 020406 (2006)\] observed degenerate four-wave mixing of matter-waves in a one-dimensional optical lattice, a process with potential for generating entanglement among atoms. We analyse the essential quantum features of the experiment to show that entanglement is created between the quadratures of the two scattered atomic clouds and is a true many-body (rather than two-body) effect. We demonstrate a significant violation of entanglement inequalities that is robust to a moderate level of coherent seeding. The system is thus a promising candididate for generating macroscopically entangled atomic samples.'
author:
- 'M. K. Olsen and M. J. Davis'
title: 'Entanglement properties of degenerate four-wave mixing of matter-waves in a periodic potential'
---
*Introduction:* A recent development in the field of quantum atom optics has been the proposal of degenerate four-wave mixing of a Bose-Einstein condensate (BEC) in a periodic potential [@KarenMH]. The usual quadratic matter wave dispersion relation in free space ordinarily prevents collisional processes in a single condensate generating new momentum components. However, the dispersion relation of a periodic potential can allow phase-matched two-body collision processes within a single condensate that conserve quasimomentum and energy. Hence condensates with new momenta can be spontaneously generated. The proposal of Hilligsøe and Mølmer [@KarenMH] was recently implemented experimentally by Campbell *et al.* [@Wolfgang], with a BEC loaded into a one-dimensional optical lattice. When the phase-matching conditions for energy and quasi-momentum were satisifeld both spontaneous and stimulated scattering were observed, with an initial state with one quasimomentum being scattered into two different quasimomentum states. Recent experimental work by Gemelke *et al.* [@gemelke] has also investigated phase-matched scattering processes of matter waves in a driven optical lattice.
Hilligsøe and Mølmer [@KarenMH] used the mean-field Gross-Pitaevskii equation (GPE) to analyse a one-dimensional BEC moving in an optical lattice. Although multi-mode, the GPE cannot describe spontaneous scattering processes and their numerical analysis required the initial state to be seeded by hand. In addition, a mean-field approach cannot be used to determine the quantum correlations that are necessary to show entanglement between the scattered momentum states.
In this paper we complement the approach of Ref. [@KarenMH] by performing fully quantum analyses of the dynamics resulting from a simple three-mode description of the degenerate four-wave scattering process. While idealised, our analysis allows us to demonstrate continuous variable entanglement without the complications arising from a full multi-mode, multi-dimensional analysis. Our results suggest that degenerate four-wave mixing could be an efficient way to generate highly entangled atomic samples.
Campbell *et al.* suggest that “parametric amplification could also be an efficient means of producing pairs of momentum entangled atoms for quantum information applications.” In fact, although individual pairs are correlated in the scattering process, they are bosonic and scatter into occupied modes, so that we cannot know which pair a given atom belongs to. It is also difficult to know how individual pairs, if detected, could be used for quantum information processes as this generally requires some other degree of freedom, such as spin [@Pu; @Duansqueeze], which is not present in this experiment. However, the coherent nature of the scattering allows for the build up of many-body entanglement between the field quadratures, which, while presently difficult to measure experimentally, is robust to losses and seeding. Although other methods have been proposed and demonstrated for the generation of pair-correlated and entangled atoms from BEC [@Pu; @Duansqueeze; @Vogels; @Kherunts1; @Haine; @Kherunts2], the method of Refs. [@KarenMH; @Wolfgang] appeals because of its relative simplicity.
*Formalism:* We consider a condensate loaded into a one dimensional periodic potential in a single Bloch state with quasi-momentum $\hbar k_0$. For particular combinations of lattice depth and $k_0$, there exists a phase-matched process that conserves both energy and quasi-momentum such that $$2k_0 = k_1 + k_2,
\qquad
2\epsilon(k_0) = \epsilon(k_1) + \epsilon(k_2),$$ where the generated Bloch modes are $k_1$ and $k_2$, and $\epsilon(k_i)$ is the energy of mode $k_i$. Expanding the full Hamiltonian in terms of Bloch states and using a rotating wave approximation, the interaction picture Hamiltonian is $${\cal H}_{int}=i\hbar\chi\left[\hat{a}_0^{2}\hat{a}_1^{\dag}\hat{a}_2^{\dag}
-\hat{a}_0^{\dag\;2}\hat{a}_1\hat{a}_2\right],
\label{eq:hamiltoniano}$$ where $\hat{a}_i$ is the annihilation operator for quasi-momentum mode $k_{i}$, and we have made the transformation $\hat{a}_0 \rightarrow \hat{a}_0 e^{i\pi/4}$. The nonlinear interaction is represented by $\chi$, and is given by $$\chi = \frac{U_0}{\hbar A_\perp} \int dx \;u_{0}\!(x)^2 u_{1}\!(x) u_{2}\!(x),$$ where $U_0 = 4\pi \hbar^2 a/m$, with $a$ the $s$-wave scattering length, $A_\perp$ is the cross-sectional area of the system, and $u_{i}(x)$ is the Bloch state of mode $k_i$ (assumed to be real).
We have made several approximations here. The first is that the Bloch states are a good approximation to the eigenstates of the system. This will be true as long as the effective interaction strength $\chi$ is sufficiently small. Using mean-field Bloch states appropriate to the effective potential of the lattice plus initial density will extend the regime of validity for short times while the overall density is unchanged, but will alter the phase matching conditions due to the energy shifts of the Bloch modes. The three-mode reduction is also only appropriate for short times before scattering into other modes becomes significant. The dimensional reduction is also only appropriate for times short enough that there are no appreciable dynamics in the perpendicular dimensions.
To demonstrate entanglement we calculate both the Duan criteria [@Duan] (see also Simon [@Simon]) and a set of Einstein-Podolsky-Rosen (EPR) criteria developed by Reid [@eprMDR], both of which establish the presence of continuous variable bipartite entanglement. These many-body continuous-variable criteria are more appropriate to the present case than any consideration of entanglement between the individual atoms of each scattered pair. This type of many-body entanglement is also more robust to atomic losses [@Kherunts2]. For quadrature entanglement, criterion have been outlined by Dechoum *et al.* [@ndturco], which follow from inequalities developed by Duan *et al.* [@Duan], based on the inseparability of the system density matrix, and a method to demonstrate the EPR paradox [@EPR] using quadratures was developed by Reid [@eprMDR]. We will briefly outline these criteria here, using the field quadrature operators $$\begin{aligned}
\hat{X}_{i} = \hat{a}_i + \hat{a}_i^\dag,
\qquad
\hat{Y}_{i} = -i(\hat{a}_i - \hat{a}_i^\dag).\end{aligned}$$ To demonstrate entanglement between the modes, we define the combined quadratures $\hat{X}_{\pm}=\hat{X}_{1}\pm\hat{X}_{2}$ and $\hat{Y}_{\pm}=\hat{Y}_{1}\pm\hat{Y}_{2}$. Following the treatment of Ref. [@ndturco], entanglement is guaranteed provided that $$V(\hat{X}_{\pm})+V(\hat{Y}_{\mp})<4.
\label{eq:inequalityduan}$$ To examine the utility of the system for the production of states which exhibit the EPR paradox, we assume that a measurement of the $\hat{X}_{1}$ quadrature, for example, will allow us to infer, with some error, the value of the $\hat{X}_{2}$ quadrature, and similarly for the $\hat{Y}_{i}$ quadratures. By minimising the rms error in these estimates, we find the inferred variances, $$\begin{aligned}
V^{inf}(\hat{X}_{1}) &=& V(\hat{X}_{1})-\frac{[V(\hat{X}_{1},\hat{X}_{2})]^{2}}{V(\hat{X}_{2})},\nonumber\\
V^{inf}(\hat{Y}_{1}) &=& V(\hat{Y}_{1})-\frac{[V(\hat{Y}_{1},\hat{Y}_{2})]^{2}}{V(\hat{Y}_{2})},
\label{eq:EPROPA}\end{aligned}$$ with those for the $k_2$ quadratures being found by swapping the indices $1$ and $2$. As the $\hat{X}_{i}$ and $\hat{Y}_{i}$ operators do not commute, the products of the actual variances obey a Heisenberg uncertainty relation, with $V(\hat{X}_{i})V(\hat{Y}_{i})\geq 1$. Hence we find a demonstration of the EPR paradox whenever $$V^{inf}(\hat{X}_{i})V^{inf}(\hat{Y}_{i}) < 1.
\label{eq:demonstration}$$
*Analytic results:* By analogy to the parametric or undepleted-pump approximation of quantum optics (also sometimes used with coupled atomic and molecular BEC [@Yurovsky]) we can set $\kappa=\chi\langle\hat{a}_{0}^{2}\rangle$ in the Hamiltonian (\[eq:hamiltoniano\]) with $\kappa$ real, and find the following Heisenberg equations of motion $$\begin{aligned}
\frac{d\hat{a}_1}{dt} = \kappa\hat{a}_2^{\dag},
\qquad \frac{d\hat{a}_2}{dt} = \kappa\hat{a}_1^{\dag},
\label{eq:heom}\end{aligned}$$ along with their Hermitian conjugates. The solutions to these equations are well known from quantum optics [@Roy], and give us all the operator moments needed to calculate the entanglement criteria. With the two modes $k_1$ and $k_2$ initially unpopulated, we find $$\begin{aligned}
V(\hat{X}_{-})+V(\hat{Y}_{+}) &=& 4\left(\cosh 2\kappa t-2\cosh\kappa t\sinh\kappa t\right),\nonumber\\
V^{inf}(\hat{X}_{i})V^{inf}(\hat{Y}_{i}) &=& \frac{1}{\cosh^{2}2\kappa t}\; <\; 1\;\mbox{ for } t > 0,
\label{eq:ancorrelations}\end{aligned}$$ which obviously violate the appropriate inequalities and can be compared to the numerical results obtained below.
![The mode occupations as a function of time. The solid line and the lower dashed line are the averages of $4.34\times 10^{6}$ trajectories of the positive-P representation equations for $N_{0}$, $N_{1}$ and $N_{2}$ in the spontaneous case. Note that $N_{1}=N_{2}$ and that the Wigner results are indistinguishable. The dotted line and the upper dashed lines are the Wigner results ($6.6\times 10^{5}$ trajectories), for $N_{0}$, $N_{1}$ and $N_{2}$ with an initial seed, $N_{1}(0)=100$. All quantities plotted in this and subsequent figures are dimensionless.[]{data-label="fig:numbers"}](number.eps){width="0.9\columnwidth"}
![The Duan correlation, $V(\hat{X}_{-})+V(\hat{Y}_{+})$, with and without an injected seed. The lower three lines are for the spontaneous case, with the dotted line being analytical, the dash-dotted line being the positive-P prediction, and the full line being the Wigner prediction. The dashed line is the Wigner prediction with $N_{1}(0)=100$.[]{data-label="fig:Duan"}](Duan.eps){width="0.9\columnwidth"}
*Numerical results:* The Hamiltonian equations of motion may be mapped exactly onto stochastic differential equations in the positive-P representation [@P+], following the usual methods [@QNCrispin]. Making the correspondences $\hat{a}_{0}\rightarrow\alpha,\:\:\hat{a}_{1}\rightarrow\beta,\:\:\hat{a}_{2}\rightarrow\gamma$, the stochastic equations are found as $$\begin{aligned}
\frac{d\alpha}{dt} &=& -2\chi\alpha^{+}\beta\gamma+\sqrt{-\chi\beta\gamma}\;\eta_{1},\nonumber\\
\frac{d\alpha^{+}}{dt} &=& -2\chi\alpha\beta^{+}\gamma^{+}+\sqrt{-\chi\beta^{+}\gamma^{+}}\;\eta_{2},\nonumber\\
\frac{d\beta}{dt} &=& \chi\alpha^{2}\gamma^{+}+\sqrt{\chi\alpha^{2}/2}\left(\eta_{3}+i\eta_{5}\right),\nonumber\\
\frac{d\beta^{+}}{dt} &=& \chi\alpha^{+\;2}\gamma+\sqrt{\chi\alpha^{+\;2}/2}\left(\eta_{4}+i\eta_{6}\right),\nonumber\\
\frac{d\gamma}{dt} &=& \chi\alpha^{2}\beta^{+}+\sqrt{\chi\alpha^{2}/2}\left(\eta_{3}-i\eta_{5}\right),\nonumber\\
\frac{d\gamma^{+}}{dt} &=& \chi\alpha^{+\;2}\beta+\sqrt{\chi\alpha^{+\;2}/2}\left(\eta_{4}-i\eta_{6}\right),
\label{eq:Pplus} \end{aligned}$$ where the $\eta_{j}$ are real Gaussian noise terms with the correlations $$\overline{\eta_{j}}=0,
\qquad
\overline{\eta_{j}(t)\eta_{k}(t')}=\delta_{jk}\delta(t-t').
\label{eq:noiseP}$$ These equations are integrated over many trajectories, with stochastic averages of the variables becoming equal to normally ordered operator expectation values.
![The EPR correlation, $V^{inf}(\hat{X}_{1})V^{inf}(\hat{Y}_{1})$. The dotted line is the spontaneous analytical solution, the dash-dotted line is the spontaneous positive-P result, the solid line is the spontaneous Wigner result, and the dashed line is the stimulated Wigner result with $N_{1}(0)=100$.[]{data-label="fig:EPR"}](EPR.eps){width="0.9\columnwidth"}
Initially we chose $\overline{|\alpha(0)|^{2}}=10^{4}$, with these momentum $k_{0}$ atoms in a coherent state, and $\beta(0)=\gamma(0)=0$. The numerical solutions are parametrised by $\xi=\chi|\alpha(0)|^{2}t$. In practice, we found that integration became unstable for times greater than $\xi=6$, and was probably not trustworthy after $\xi\approx 5$. However, this covers the region of maximum violation of the inequalities (\[eq:inequalityduan\],\[eq:demonstration\]).
It is also of interest to compare the positive-P solutions to those of the approximate, but stable, truncated Wigner representation. This has been used with some success in investigations of BEC [@mjs; @Sinatra2000a; @Sinatra2001a; @Sinatra2002a; @qstate1; @qstate2; @Fock], and allows for the calculation of symmetrically-ordered operator moments. Again following standard procedures [@QNCrispin], we can map the system Hamiltonian onto a generalised Fokker-Planck equation for the Wigner pseudoprobability distribution, which has third order derivatives and hence no equivalent stochastic differential equations. Although methods exist for a mapping onto stochastic difference equations [@nossoEPL], these seldom result in equations which are simple to integrate numerically. We will therefore truncate the third-order terms and map the resulting Fokker-Planck equation onto differential equations for the Wigner variables. This is justified here since the number of particles is much larger than the number of modes. This results in the set of equations $$\begin{aligned}
\frac{d\alpha}{dt} = -2\chi\alpha^{\ast}\beta\gamma,\quad
\frac{d\beta}{dt} = \chi\alpha^{2}\gamma^{\ast},\quad
\frac{d\gamma}{dt} = \chi\alpha^{2}\beta^{\ast}.
\label{eq:Wminus} \end{aligned}$$ Note that, although these equations seem deterministic, the initial variables are chosen from the appropriate Wigner distribution, so that quantum noise is included in the initial conditions. We begin our integrations with coherent states, with $\overline{|\alpha|^{2}}=10^{4}$, $\overline{|\gamma|^{2}}=0$ and $\beta$ either being vacuum or a small coherent seed of $\overline{|\beta(0)|^{2}}=100$ atoms.
The results are presented in the three figures, which allow us to compare the predictions of the approximate analytic solutions, the formally exact positive-P representation solutions, and those of the truncated Wigner representation. We show the atom numbers in the three modes in Fig. \[fig:numbers\] for both the spontaneous and seeded situations. For the spontaneous case we find that the positive-P and Wigner methods give almost identical results over most of the range shown, thus we can be confident of the Wigner solutions in this regime. We see that a seed in the $k_{1}$ mode with only $1\%$ of the number of atoms in the $k_{0}$ mode gives appreciably faster scattering into modes $k_{1}$ and $k_{2}$, with almost full occupation at a time when the spontaneous case has seen approximately $10\%$ of the atoms scattered. A similar effect was observed in the experiment of Campbell *et al.* [@Wolfgang]. As we are interested in the quantum correlations between these two modes, we must also consider the effect of the seeding on these. In the case of the nondegenerate optical parametric amplifier (OPA) with an injected signal, for example, an injection level of $1\%$ had a much less noticeable effect on the mean fields, but was sufficient to almost destroy some quantum correlations [@injectOPA]. In the present case, however, as we can see in Fig. \[fig:Duan\] and Fig. \[fig:EPR\], the seeding still allows for considerable violation of the inequalities, especially in the EPR case. We note here that the descriptions are only similar once the parametric approximation is made, which results in a quadratic Hamiltonian for both cases. It may be of considerable practical interest that the slightly lesser violations occur for large numbers of atoms in the two scattered modes once the system is seeded, as can be seen by comparison with Fig. \[fig:numbers\]. Also of interest is that the EPR product demonstrates entanglement with seeding in a region where the form of the Duan inequality that we have used does not. This is not a contradiction as the most general form of the Duan inequality [@Duan] does not define the quadratures in the symmetric manner we have used here, so that a more appropriate form would be violated.
*Conclusions:* We have calculated the entanglement properties of a Hamiltonian which gives a simplified description of the lattice four-wave mixing experiment of Campbell [*et al.*]{}. We have shown that the system exhibits entanglement between the scattered modes and is a candididate for a demonstration of the EPR paradox with massive particles. Seeding of one of the scattered modes allows for substantially quicker conversion than in the spontaneous case. The entanglement is between the entire modes rather than between individual pairs of scattered atoms and thus is robust to reasonable levels of seeding and is not destroyed by small rates of atomic loss. A fully quantum spatial analysis that could more closely describe experiments is currently being investigated.
This research was supported by the Australian Research Council and the Queensland State government.
[99]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a family $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$, define the $\delta$-shadow of $A$ to be the set obtained from $A$ by removing from any of its vectors one coordinate that equals zero. Given the size of $A$, how should we choose $A$ to minimise its $\delta$-shadow? Our aim in this paper is to show that, for any $r$, the family of all sequences with at most $r$ zeros has minimal $\delta$-shadow. We actually give the exact best $A$ for every size.'
author:
- 'Eero R[ä]{}ty[^1]'
title: Coordinate deletion of zeroes
---
Introduction
============
The classical Kruskal-Katona theorem is concerned with the lower shadow of set systems. For $A\subseteq\left\{ 0,1\right\} ^{n}$, define the *lower shadow* of $A$ to be the set of sequences obtained from any of its vectors by flipping one of its 1-entries to 0. The *rank* of a sequence $x\in\left\{ 0,\dots,n\right\} ^{k}$ is defined to be $\left|x\right|=\sum_{i=1}^{k}x_{i}$. Note that the lower shadow operator decreases the rank of a sequence by 1. For given $r$, it is natural to ask how to choose a family $A\subseteq\left\{ 0,1\right\} ^{n}$ of given size containing only vectors with rank $r$, which minimises the lower shadow. This question was answered by Kruskal [@key-5] and Katona [@key-4].
Define the *colexicographic order* on $\left\{ x\in\left\{ 0,1\right\} ^{n}\,:\,\left|x\right|=r\right\} $ by $x\leq_{colex}y$ if $\max\left(X\Delta Y\right)\in Y$. Here $X=\left\{ i\,:\,x_{i}=1\right\} $ and $Y=\left\{ i\,:\,y_{i}=1\right\} $ . The Kruskal-Katona theorem states that for a set $A\subseteq\left\{ 0,1\right\} ^{n}$ containing only sequences of rank $r$, the lower shadow is minimised when $A$ is chosen to be an initial segment of colexicographic order.
Instead of changing the coordinates, it is also natural to define an operator which acts by deleting coordinates. For $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ define the *coordinate deletion shadow* $\Delta A$ to be the set of those sequences obtained from any of its vectors by deleting one coordinate. For example $\Delta\left(\left\{ 000,001,002,121\right\} \right)=\left\{ 00,01,02,12,11,21\right\} $.
Again it is natural to ask that which sets minimises the coordinate deletion shadow. Define the *simplicial order* $\leq_{sim}$ on $\left\{ 0,1\right\} ^{n}$ by $$x\leq_{sim}y\text{ if }\left|x\right|<\left|y\right|\text{ or }\left|x\right|=\left|y\right|\text{ and }\min(X\Delta Y)\in X.$$
It was proved by Danh and Daykin that for subsets of $\left\{ 0,1\right\} ^{n}$, $\Delta A$ is minimised by an initial segment of the simplicial order [@key-3]. They also conjectured a certain order as best in $\left\{ 0,1,\dots\right\} ^{n}$, but Leck [@key-6] showed that this turned out to be false and in fact there is no order in general whose all initial segments have minimal coordinate deletion shadow.
Bollob[á]{}s and Leader [@key-1] pointed out that for $k\geq2$ the sets $A_{t}=\left\{ 0,\dots,t-1\right\} ^{n}\subseteq\left\{ 0,\dots,k\right\} ^{n}$ are extremal for $\Delta$. Indeed, suppose that $B\subseteq\left\{ 0,\dots,k\right\} ^{n}$ is extremal with $\left|B\right|=t^{n}$. Define $B_{\left[n\right]\setminus\left\{ i\right\} }$ to be the projection of $B$ onto the hyperplane excluding the $i^{th}$ direction. Suppose that $\left|\Delta B\right|<t^{n-1}$. Since $B_{\left[n\right]\setminus\left\{ i\right\} }\subseteq\Delta B$ for all $i$, it follows that $\left|B_{\left[n\right]\setminus\left\{ i\right\} }\right|<t^{n-1}$ for all $i$. Thus the Loomis-Whitney inequality [@key-7] implies that $\left|B\right|^{n-1}\leq\left(\prod_{i=1}^{n}B_{\left[n\right]\setminus\left\{ i\right\} }\right)<t^{n\left(n-1\right)}$, which contradicts $\left|B\right|=t^{n}$. Hence $\left|\Delta B\right|\geq t^{n-1}$ and since $\Delta A_{t}=t^{n-1}$, it follows that each $A_{t}$ is extremal.
Bollob[á]{}s and Leader also made the following conjecture that certain other type of sets are also extremal.\
**Conjecture 1** [@key-1]**.** For each $t\leq k$ and $r\leq k$, let $B_{r,t}\subseteq\left\{ 0,\dots,k\right\} ^{n}$ be the subset containing all sequences with at most $r$ zeros, and with all coordinates in $\left\{ 0,\dots,t\right\} $. Then the sets $B_{r,t}$ have extremal $\Delta$-shadow in $\left\{ 0,\dots,k\right\} ^{n}$.\
\
Even the case $t=k$ in the conjecture in unknown.
There is, however, a notion that comes ’between’ the lower shadow and the coordinate deletion shadow. The usual lower shadow operator decreases the rank by 1 and preserves the dimension $n$, while the coordinate deletion shadow decreases the dimension by 1 but there is no control on how it changes the rank. So it is natural to consider the following operator which preserves the rank, but reduces the dimension by one.
Define the $\delta$*-shadow* of $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ to be the set of sequences in $\left\{ 0,\dots,k\right\} ^{n-1}$ obtained by removing one coordinate that equals 0 from any of the vectors in $A$. Denote this set by $\delta A$. Thus for example $\delta\left(\left\{ 00011,00101\right\} \right)=\left\{ 0011,0101\right\} $ and $\delta\left(\left\{ 112,113,123\right\} \right)=\emptyset$.
How can we find sets $A$ with minimal $\delta$-shadow? If $\left|A\right|\leq k^{n}$ then the question is trivial, as one can take any subset of $\left\{ 1,\dots,k\right\} ^{n}$ of given size. In general, it is natural to choose $A$ to contain sequences with as few zeros as possible. Furthermore, it is natural to guess that for each $0\leq i\leq n$, the sets containing all sequences with at most $i$ zeros have minimal $\delta$-shadow.
Our main result in this paper is to find an order on $\left\{ 0,\dots,k\right\} ^{n}$ whose initial segments have minimal $\delta$-shadow. In particular, it follows that the sets containing all sequences with at most $i$ zeros have minimal $\delta$-shadow.
In order to state the main result, we need a few definitions. For a sequence $x\in\left\{ 0,\dots,k\right\} ^{n}$, let $R\left(x\right)=\left\{ i\,:\,x_{i}=0\right\} $ and let $w\left(x\right)=\left|R\left(x\right)\right|$. Let $L_{r}\left(n\right)=\left\{ x\in\left\{ 0,\dots,k\right\} ^{n}\,:\,w\left(x\right)=r\right\} $. Note that $\delta$ maps sequences in $L_{r}\left(n\right)$ to sequences in $L_{r-1}\left(n-1\right)$.
For $x\in\left\{ 0,\dots,k\right\} ^{n}$, define its *reduced sequence* to be the sequence obtained by removing all coordinates from $x$ that equal 0. Denote the reduced sequence of $x$ by $re(x)$. Note that for any sequence $s$ and for any $t\in\delta s$ we have $re(s)=re(t)$, as removing a coordinate which equals 0 does not change the reduced sequence. Hence each $L_{r}\left(n\right)$ splits into disjoint components based on the reduced sequences.
We will start by proving that inside a component one should choose sequences $x$ for which the sets $R\left(x\right)$ form an initial segment of colex. This is a straightforward consequence of the work of Danh and Daykin in [@key-3].
Since we know that $\left\{ 0,\dots,k\right\} ^{n}$ splits into components based on the reduced sequences, and we know that initial segments of the colexicographic order minimises the $\delta$-shadow inside each component, we are left with the question on how to split the sequences into different components in order to minimise the shadow.
We go on to prove that in order to minimise the shadow of a subset in $\left\{ 0,\dots,k\right\} ^{n}$, one should first choose sequences in components in $L_{r}\left(n\right)$ rather than in $L_{s}\left(n\right)$ for all $r<s$, and inside $L_{r}\left(n\right)$ one should choose all sequences from a component before taking any sequences from another component. As a consequence we obtain an order whose initial segments minimises the $\delta$-shadow.
For $r\in\left\{ 0,1,\dots,k\right\} $ define $R_{r}\left(x\right)=\left\{ i\,:\,x_{i}=r\right\} $ and $w_{r}\left(x\right)=\left|R_{r}\left(x\right)\right|$. Note that $R=R_{0}$ and $w=w_{0}$. For all $r$ we define an order $\leq_{c}$ on $\left\{ 1,\dots,k\right\} ^{r}$ as follows. For distinct $x$, $y\in\left\{ 1,\dots,k\right\} ^{r}$ let $i$ be minimal such that $R_{i}\left(x\right)\neq R_{i}\left(y\right)$. We say that $x\leq_{c}y$ if and only if $\max\left(R_{i}\left(x\right)\Delta R_{i}\left(y\right)\right)\in R_{i}\left(y\right)$.
Define an order $\leq$ on $\left\{ 0,\dots,k\right\} ^{n}$ as follows. For distinct $x$, $y\in\left\{ 0,\dots,k\right\} ^{n}$ we set $x\leq y$ if
1. $w_{0}\left(x\right)<w_{0}\left(y\right)$
2. $w_{0}\left(x\right)=w_{0}\left(y\right)$, $re\left(x\right)\neq re\left(y\right)$ and $re\left(x\right)\leq_{c}re\left(y\right)$
3. $w_{0}\left(x\right)=w_{0}\left(y\right)$, $re\left(x\right)=re\left(y\right)$ and $R_{0}\left(x\right)\leq_{colex}R_{0}\left(y\right)$
Now we are ready to state our main theorem.\
**Theorem 1.** Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and let $B$ be an initial segment of $\leq$ with $\left|B\right|=\left|A\right|$. Then $\left|\delta A\right|\geq\left|\delta B\right|$.\
In particular, it follows that the sets of the form $L_{\leq r}\left(n\right)=\bigcup_{i=0}^{r}L_{i}\left(n\right)$ are extremal. Note that for fixed $r$, every component of $L_{r}\left(n\right)$ behaves in the same way. Hence for any fixed $r$, one could replace the $\leq_{c}$ order by any other order on $\left\{ 1,\dots,k\right\} ^{r}$ in the definition of the $\leq$-order.
The plan of the paper is as follows. In Section 2 we prove that inside a component the sets of sequences whose associated sets $R\left(x\right)$ form an initial segment of colex have minimal $\delta$-shadow. In Section 3 we prove Theorem 1. In Section 4 we generalise the $\delta$-shadow to allow deleting coordinates in some set $\left\{ 0,\dots,r\right\} $ instead of just deleting only coordinate which equals 0. In this case we will show that sets $\left\{ x\,:\,\sum_{i=0}^{r}w_{i}\left(x\right)\leq s\right\} $, which are analogous to the sets $L_{\leq s}\left(n\right)$, are extremal for all $0\leq s\leq n$. In this general case we do not know what happens for sets of other sizes.
We write use the standard notation $\left[n\right]=\left\{ 1,\dots,n\right\} $ and $\left[n\right]^{(r)}=\left\{ A\subseteq\left[n\right]\,:\,\left|A\right|=r\right\} $. We write $L_{r}$ instead of $L_{r}\left(n\right)$ if the dependence on $n$ is clear. When $k=1$ we may also write $\left\{ 0,1\right\} _{r}^{n}$ instead of $L_{r}\left(n\right)$. This notation will be used to highlight that we are working with $\left\{ 0,1\right\} $-sequences.
Deletion on $\left\{ 0,1\right\} $-sequences
============================================
In [@key-3] Danh and Daykin proved the following result for the coordinate deletion shadow $\Delta$ on $\{0,1\}^{n}$.\
**Theorem 2 (Danh, Daykin).** Let $A\subseteq\left\{ 0,1\right\} ^{n}$ and let $B$ be an initial segment of simplicial order with $\left|B\right|=\left|A\right|$. Then $\left|\Delta A\right|\geq\left|\Delta B\right|$. $\square$\
There is a natural correspondence between the sequences $\left\{ 0,1\right\} ^{n}$ and the power-set ${\cal P}\left(\left\{ 1,\dots,n\right\} \right)$. For our purposes it will be convenient to choose this correspondence to be given by mapping a sequence $\left(x_{i}\right)$ to the set $R_{0}\left(x\right)=\left\{ i\,:\,x_{i}=0\right\} $.
In this way we can identify set $A\subseteq L_{r}\left(n\right)$ with a set system ${\cal A}\subseteq\left[n\right]^{\left(r\right)}$ by taking ${\cal A}$ to be the images of the elements of $A$ under this bijection. This enables us to translate questions on $\delta$ to questions related to properties of set systems ${\cal A}\subseteq\left[n\right]^{\left(r\right)}$ instead. We start by proving that the subsets $A$ of $L_{r}$ with minimal shadow are the ones whose corresponding set ${\cal A}$ is an initial segment of colex.\
**Lemma 3.** Let $A\subseteq\left\{ 0,1\right\} _{r}^{n}$, and let $B\subseteq\left\{ 0,1\right\} _{r}^{n}$ be a set of same size for which ${\cal B}$ is an initial segment of colex. Then $\left|\delta A\right|\geq\left|\delta B\right|$.\
**Proof.** Define $C_{1}=A\cup L_{>r}\left(n\right)$ and $C_{2}=B\cup L_{>r}\left(n\right)$, where $L_{>r}\left(n\right)=\bigcup_{i=r+1}^{n}\left\{ 0,1\right\} _{i}^{n}$ . Now $C_{2}$ is isomorphic to initial segment of simplicial order, and the isomorphism is the map which reverses the sequences. Since this map preserves the size of $\Delta$-shadow, Theorem 2 implies that $\Delta C_{2}$ is minimal and hence $$\left|\Delta C_{2}\right|\leq\left|\Delta C_{1}\right|.$$ Note that $\Delta C_{1}=L_{>r}(n-1)\cup\delta A$ and $\Delta C_{2}=L_{>r}(n-1)\cup\delta B$. Indeed, $L_{>r}(n-1)$ is certainly subset of both of these sets, and the only contribution to elements not in $L_{>r}(n-1)$ comes from removing 0 from a sequence which contains exactly $n-r$ 1’s. Hence $$\left|\Delta C_{1}\right|=\left|L_{>r}\left(n-1\right)\right|+\left|\delta A\right|$$ and $$\left|\Delta C_{2}\right|=\left|L_{>r}\left(n-1\right)\right|+\left|\delta B\right|.$$ Combining (1), (2) and (3) yields that $\left|\delta A\right|\ge\left|\delta B\right|$. $\square$\
Lemma 3 implies that inside $\left\{ 0,1\right\} _{r}^{n}$ the colexicographic order minimises the size of the shadow. Before moving on to general $k$ from $k=1$, we find a way to relate the size of $\delta A$ to the associated family ${\cal A}$. For convenience, from now on we say that $A\subseteq\left\{ 0,1\right\} _{r}^{n}$ is an initial segment of colex if the associated set system ${\cal A}$ is an initial segment of colex. For ${\cal A}\subseteq{\cal P}\left(\left\{ 1,\dots,n\right\} \right)$ define ${\cal A}_{1}=\left\{ B\in{\cal A}\,:\,1\in B\right\} $.\
**Lemma 4.** Let $A\subseteq\left\{ 0,1\right\} _{r}^{n}$ be an initial segment of colex associated to ${\cal A}$. Then $\left|\delta A\right|=\left|{\cal A}_{1}\right|$.\
**Proof.** The proof is by induction on $\left|A\right|$, the case $\left|A\right|=1$ is clear. Let $B$ be an initial segment with $\left|B\right|=\left|A\right|+1$, say $B=A\cup\left\{ x\right\} $ with $x=x_{1}\dots x_{n}$. First we will prove that $x_{2}\dots x_{n}$ is the only element which could be in $\delta B\setminus\delta A$.
Indeed, suppose $t\in\delta B\setminus\delta A$ and that it is obtained by removing the $k^{th}$ coordinate of $x$. Hence $t=x_{1}\dots x_{k-1}x_{k+1}\dots x_{n}$ and $x_{k}=0$. Let $i=\min\left\{ j\,:\,x_{j}=1\right\} $ and set $y=0t=0x_{1}\dots x_{k-1}x_{k+1}\dots x_{n}$. If $i\leq k$, then $y_{j}=x_{j}$ for all $j\leq i-1$ but $y_{k}=x_{k-1}=0\neq1=x_{k}$ so $y<_{colex}x$. But then $t\in\delta y\subseteq\delta A$ which contradicts $t\in\delta B\setminus\delta A$.
Hence we must have $i>k$. But in this case $x_{1}=\dots x_{k}=0$ and therefore $t=0\dots0x_{k+1}\dots x_{n}=x_{2}\dots x_{n}$. Hence $\delta B\setminus\delta A$ is either empty or contains only $x_{2}\dots x_{n}$.
Note that $0x_{2}\dots x_{n}$ is the least element in colex which has $x_{2}\dots x_{n}$ contained in its $\delta$-shadow. Hence $0x_{2}\dots x_{n}\in\delta B\setminus\delta A$ if and only if $x=0x_{2}\dots x_{n}$. Thus $$\left|\delta B\right|=\begin{cases}
\begin{array}{c}
\left|\delta A\right|+1\\
\left|\delta A\right|
\end{array} & \begin{array}{c}
\text{if }x_{1}=0\\
\text{if }x_{1}=1
\end{array}\end{cases}.$$ Also ${\cal B}={\cal A}\cup R_{0}(x)$, and the set $R_{0}(x)$ contains $1$ if and only if $x_{1}=0$. Thus $$\left|{\cal B}_{1}\right|=\begin{cases}
\begin{array}{c}
\left|{\cal A}_{1}\right|+1\\
\left|{\cal A}_{1}\right|
\end{array} & \begin{array}{c}
\text{if }x_{1}=0\\
\text{if }x_{1}=1
\end{array}\end{cases}$$ and hence $\left|\delta B\right|=\left|{\cal {\cal B}}_{1}\right|$ by induction. $\square$
The main theorem
================
Let $H$ be the bipartite graph with vertex set $\left\{ 0,\dots,k\right\} ^{n}\cup\left\{ 0,\dots,k\right\} ^{n-1}$ and whose edges are precisely those pairs $s,\,t$ with $s\in\left\{ 0,\dots,k\right\} ^{n}$ and $t\in\delta s$. Then for $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$, $\delta A$ is just the neighbourhood of $A$ in the graph $H$. Note that both classes can be partitioned as $\left\{ 0,\dots,k\right\} ^{n}=\bigcup_{i=0}^{n}L_{i}\left(n\right)$ and $\left\{ 0,\dots,k\right\} ^{n-1}=\bigcup_{i=0}^{n}L_{i}\left(n-1\right)$, and by definition of $\delta$ it is clear that there are edges only between $L_{i}\left(n\right)$ and $L_{i-1}\left(n-1\right)$, with the convention $L_{-1}=\emptyset$.
Let $C$ be a connected component in $H$. Suppose $C$ is non-trivial, i.e. $\left\{ 0,\dots,k\right\} ^{n}\cap C\subseteq L_{i}(n)$ for some $i>0$. Recall that for all $x$ and for any $y\in\delta x$, $x$ and $y$ have the same reduced sequences. But since $C$ is a connected component, this means that every $x\in C$ has the same reduced sequence. Conversely it is easy to check that for $i>0$ all sequences $x\in L_{i}(n)\cup L_{i-1}(n-1)$ with the same reduced sequence are in the same connected component. Thus we can deduce that the connected components in $H$ are given as follows.\
**Lemma 5.** For $s\in\bigcup_{i=0}^{r}\left\{ 1,\dots,k\right\} ^{i}$ define $C_{s}=\left\{ x\in\left\{ 0,\dots,k\right\} ^{n}\,:\,re(x)=s\right\} $ and $D_{s}=\left\{ x\in\left\{ 0,\dots,k\right\} ^{n-1}\,:\,re(x)=s\right\} $. Then $C_{s}\cup D_{s}$ are the connected components of $H$. $\square$\
Broadly speaking, we need to only understand how to minimise $\delta$ inside a connected component and to determine how to distribute the sequences into different connected components in order to minimise $\delta$. It turns out that inside connected component one should choose sequences $x$ whose sets $R_{0}\left(x\right)$ forms initial segment of colex.\
**Lemma 6.** Let $C\subseteq L_{i}(n)\cup L_{i-1}(n-1)$ be a connected component corresponding to a reduced word $x=x_{1}\dots x_{n-i}$. Let $B\subseteq L_{i}\cap C$ and let $A\subseteq L_{i}\cap C$ be a set of sequences chosen such that $\left|A\right|=\left|B\right|$ and $\left\{ R_{0}\left(x\right)\,:\,x\in A\right\} $ is an initial segment of colex. Then $\left|\delta B\right|\geq\left|\delta A\right|$.\
Proof Note that the behaviour of the connected component depends only on $n-i$ and in particular not on the sequence $x_{1}\dots x_{n-i}$, as the reduced sequence and the order of coordinates in the reduced sequence is preserved under taking $\delta$-shadow. In particular, all such connected components has the same size and they all behave in the same way under taking $\delta$-shadow. Hence it suffices to consider only the component with $x_{1}=\dots=x_{n-i}=1$. But this component is just $\left\{ 0,1\right\} _{i}^{n}$ and hence the result follows from Lemma 3. $\square$\
Hence it remains to understand how to fill different connected components. Our aim is to show that it is optimal to first choose all sequences in a component before taking sequences from another component, and also to prefer a component in $L_{i}(n)$ over a component in $L_{i+1}(n)$.
From now on we call the sets $C_{s}$ connected components, i.e. by a connected component we refer to the intersection of a connected component with $\left\{ 0,\dots,k\right\} ^{n}$.
For $s,\,t\in\bigcup_{i=0}^{r}\left\{ 1,\dots,k\right\} ^{i}$ define the *$s,t$-compression* operator as follows. For $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ its compression $B=C_{s,t}\left(A\right)$ is given by setting
1. $B\cap C_{s}$ to be an initial segment of colex of length $\min\left(\left|A\cap\left(C_{s}\cup C_{t}\right)\right|,\left|C_{s}\right|\right)$
2. $B\cap C_{t}$ to be an initial segment of colex of length $\max\left(0,\left|A\cap\left(C_{s}\cup C_{t}\right)\right|-\left|C_{s}\right|\right)$
3. $B\setminus\left(C_{s}\cup C_{t}\right)=A\setminus\left(C_{s}\cup C_{t}\right)$
It is clear that $\left|C_{s.t}\left(A\right)\right|=\left|A\right|$ for all $s$ and $t$. As usual we say that $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ is *$s,t$-compressed* if $C_{s,t}\left(A\right)=A$.
In order to prove Theorem 1, we will need the following two Lemmas.\
**Lemma 7.** Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ be a set and let $s,\,t\in\left\{ 1,\dots,k\right\} ^{n-i}$ for some $i$. Then $\left|\delta A\right|\geq\left|\delta C_{s,t}\left(A\right)\right|$.\
**Lemma 8.** Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ be a set and let $s\in\left\{ 1,\dots,k\right\} ^{n-i}$, $t\in\left\{ 1,\dots,k\right\} ^{n-i-1}$ for some $i$. Then $\left|\delta A\right|\geq\left|\delta C_{s,t}\left(A\right)\right|$.\
In order to prove these Lemmas, we will relate them to the appropriate questions on the subsets of $\left[n\right]^{(i)}$. We will now state these results, but the proof is presented after the proofs of Lemma 7 and Lemma 8.
Define ${\cal B}\subseteq\left[n\right]^{(r)}$ to be a *segment* if there exists initial segments ${\cal I}$ and ${\cal J}$ of colex such that ${\cal A}={\cal I}\setminus{\cal J}$.\
**Lemma 9.** The following claims are true.\
**Claim 1.** Let ${\cal A}\subseteq\left[n\right]^{(i)}$ be a segment and ${\cal I}\subseteq\left[n\right]^{(i)}$ be an initial segment of colex with $\left|{\cal A}\right|=\left|{\cal I}\right|$. Then $\left|{\cal A}_{1}\right|\leq\left|{\cal I}_{1}\right|$\
**Claim 2.** Let ${\cal I}\subseteq\left[n\right]^{(i)}$ and ${\cal J}\subseteq\left[n\right]^{(i+1)}$ be initial segments of colex with $\left|{\cal I}\right|=\left|{\cal J}\right|$. Then $\left|{\cal I}_{1}\right|\leq\left|{\cal J}_{1}\right|$\
**Claim 3.** Let ${\cal A}\subseteq\left[n\right]^{(r)}$ be a segment and let ${\cal I}=\left[n\right]^{(r)}\setminus{\cal J}$, where ${\cal J}$ is an initial segment of colex chosen such that $\left|{\cal I}\right|=\left|{\cal A}\right|$. Then $\left|{\cal A}_{1}\right|\geq\left|{\cal I}_{1}\right|$.\
**Claim 4.** Let ${\cal I}_{*}$ and ${\cal J}_{*}$ be initial segments of colex chosen such that ${\cal I}=\left[n\right]^{(i)}\setminus{\cal I}_{*}$ and ${\cal J}=\left[n\right]^{(i+1)}\setminus{\cal J}_{*}$ satisfies $\left|{\cal I}\right|=\left|{\cal J}\right|$. Then $\left|{\cal I}_{1}\right|\leq\left|{\cal J}_{1}\right|$.\
**Proof of Lemma 7.**
Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and $B=C_{s,t}\left(A\right)$. Note that $B$ depends only on $\left|A\cap C_{s}\right|$ and $\left|A\cap C_{t}\right|$. Lemma 6 implies that an initial segment of colex minimises the $\delta$-shadow inside a connected component, so we may assume that $Q=A\cap C_{s}$ and $R=A\cap C_{t}$ are initial segments of colex.
Let $S=B\cap C_{s}$ and $T=B\cap C_{t}$. Let ${\cal Q}$, ${\cal R}$, ${\cal S}$ and ${\cal T}$ be the associated families in $\left[n\right]^{(i)}$. Since $B\setminus\left(C_{s}\cup C_{t}\right)=A\setminus\left(C_{s}\cup C_{t}\right)$, it follows that $\left|\delta B\right|\geq\left|\delta A\right|$ is equivalent to $\left|\delta Q\right|+\left|\delta R\right|\geq\left|\delta S\right|+\left|\delta T\right|$. By applying Lemma 4, this can be rewritten as $\left|{\cal Q}_{1}\right|+\left|{\cal R}_{1}\right|\geq\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|$.\
**Case 1. $\left|Q\right|+\left|R\right|\leq\left|C_{s}\right|$**\
By definition of $B$, it follows that $T=\emptyset$ and $\left|S\right|=\left|Q\right|+\left|R\right|$. Let ${\cal I}={\cal S}\setminus{\cal Q}$. Since ${\cal S}$ and ${\cal Q}$ are initial segments of colex, it follows that ${\cal I}$ is a segment of length $\left|{\cal R}\right|$. Thus $\left|{\cal I}_{1}\right|\leq\left|{\cal R}_{1}\right|$ by Claim 1 and hence $$\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|=\left|{\cal Q}_{1}\right|+\left|{\cal I}_{1}\right|\leq\left|{\cal Q}_{1}\right|+\left|{\cal R}_{1}\right|$$ as required.\
**Case 2. $\left|Q\right|+\left|R\right|>\left|C_{s}\right|$**\
In this case $S=C_{s}$ and hence $|T|<|R|$. Thus we can write ${\cal I}={\cal R}\setminus{\cal T}$, which is a segment as ${\cal R}$ and ${\cal T}$ are initial segments of colex. Also set ${\cal J}={\cal S}\setminus{\cal Q}=\left[n\right]^{(i)}\setminus{\cal Q}$, which is a segment as well. Since $\left|{\cal S}\right|+\left|{\cal T}\right|=\left|{\cal R}\right|+\left|{\cal Q}\right|$ it follows that $\left|{\cal I}\right|=\left|{\cal J}\right|$. Thus Claim 3 implies that $\left|{\cal J}_{1}\right|\leq\left|{\cal I}_{1}\right|$.
Combining this together with the definitions of ${\cal I}$ and ${\cal J}$ implies that $$\left|{\cal Q}_{1}\right|+\left|{\cal R}_{1}\right|=\left|{\cal Q}_{1}\right|+\left|{\cal I}_{1}\right|+\left|{\cal T}_{1}\right|\leq\left|{\cal Q}_{1}\right|+\left|{\cal J}_{1}\right|+\left|{\cal T}_{1}\right|=\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|$$ as required, which completes the proof of Lemma 7. $\square$\
**Proof of Lemma 8.**
Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and $B=C_{s,t}\left(A\right)$. By Lemma 5 we may assume that both $A\cap C_{s}$ and $A\cap C_{t}$ are initial segments of colex. As in the proof of Lemma 7, set $Q=A\cap C_{s}$ , $R=A\cap C_{t}$, $S=B\cap C_{s}$ and $T=B\cap C_{t}$. Let ${\cal Q}$ and ${\cal S}$ be the associated set systems in $\left[n\right]^{(i)}$, and ${\cal R}$ and ${\cal T}$ be the associated set systems in $\left[n\right]^{(i+1)}$. By Lemma 4 it suffices to prove that $\left|{\cal Q}_{1}\right|+\left|{\cal R}_{1}\right|\geq\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|$.\
**Case 1. $\left|{\cal Q}\right|+\left|{\cal R}\right|\leq\left|C_{s}\right|$**
By definition of $B$, it follows that ${\cal S}$ is an initial segment of colex of length $\left|{\cal Q}\right|+\left|{\cal R}\right|$ in $\left[n\right]^{(i)}$, and ${\cal T}=\emptyset$. Let ${\cal I}$ be an initial segment of colex of length $\left|{\cal R}\right|$ in $\left[n\right]^{(i)}$, and set ${\cal J}={\cal S}\setminus{\cal Q}$. Then $\left|{\cal J}\right|=\left|{\cal R}\right|=\left|{\cal I}\right|$ and ${\cal J}$ is a segment, as ${\cal S}$ and ${\cal Q}$ are initial segments of colex. Thus Claim 1 implies that $\left|{\cal J}_{1}\right|\leq\left|{\cal I}_{1}\right|$. On the other hand, Claim 2 implies that $\left|{\cal R}_{1}\right|\geq\left|{\cal I}_{1}\right|$. Combining these two yields $\left|{\cal R}_{1}\right|\geq\left|{\cal J}_{1}\right|$. Hence $$\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|=\left|{\cal S}_{1}\right|=\left|{\cal J}_{1}\right|+\left|{\cal Q}_{1}\right|\leq\left|{\cal R}_{1}\right|+\left|Q_{1}\right|$$ as required.\
**Case 2. $\left|{\cal Q}\right|+\left|{\cal R}\right|>\left|C_{s}\right|$**
By definition of $B$ it follows that ${\cal S}=\left[n\right]^{(i)}$. Note that since $\left|{\cal S}\right|\geq\left|{\cal Q}\right|$, it follows that $\left|{\cal R}\right|\geq\left|{\cal T}\right|$. Hence ${\cal I}={\cal R}\setminus{\cal T}\subseteq\left[n\right]^{(i+1)}$ is a segment and it satisfies ${\cal R}={\cal I}\cup{\cal T}$. Let ${\cal I}_{*}\subseteq\left[n\right]^{(i+1)}$ be an initial segment of colex chosen such that ${\cal K}=\left[n\right]^{(i+1)}\setminus{\cal I}_{*}$ is a segment of size $\left|{\cal I}\right|$. Define ${\cal J}=\left[n\right]^{(i)}\setminus{\cal Q}={\cal S}\setminus{\cal Q}$. Hence ${\cal J}$ is a segment of size $\left|{\cal S}\right|-\left|{\cal Q}\right|=\left|{\cal R}_{1}\right|-\left|{\cal T}_{1}\right|=\left|{\cal I}\right|$.
Claim 3 implies that $\left|{\cal I}_{1}\right|\geq\left|{\cal K}_{1}\right|$ and Claim 4 implies that $\left|{\cal K}_{1}\right|\geq\left|{\cal J}_{1}\right|$. Thus combining these results yields that $\left|{\cal I}_{1}\right|\geq\left|{\cal J}_{1}\right|$. Using the definitions of ${\cal I}$ and ${\cal J}$ it follows that $$\left|{\cal S}_{1}\right|+\left|{\cal T}_{1}\right|=\left|{\cal J}_{1}\right|+\left|{\cal Q}_{1}\right|+\left|{\cal T}_{1}\right|\leq\left|{\cal I}_{1}\right|+\left|{\cal Q}_{1}\right|+\left|{\cal T}_{1}\right|=\left|{\cal R}_{1}\right|+\left|{\cal Q}_{1}\right|$$ as required. This completes the proof of Lemma 8. $\square$\
**Proof of Lemma 9.**
We will first start by proving Claim 1, and then we will prove that other claims can be deduced from Claim 1.\
**Proof of Claim 1.** Since ${\cal A}$ is a segment, there exists initial segments ${\cal I}_{A}$ and ${\cal J}_{A}$ of colex with ${\cal A}={\cal I}_{A}\setminus{\cal J}_{A}$, and denote their associated sets of sequences by $I_{A}$ and $J_{A}$. Let $C$ be obtained from $J_{A}$ by reversing all the sequences and by adding $2n$ $1$’s at the start of each reversed sequence. Let $D$ be obtained from $I$ by adding $2n$ 1’s at the end of each sequence in $I$, where $I$ is the set of sequences associated to ${\cal I}$ . Set $B=C\cup D$.
Due to the additional 1’s at the start of the elements of $C$ and at the end of the elements in $D$, it follows that $\delta C$ and $\delta D$ are disjoint sets. Also note that reversing all the sequences and adding $1$’s to every sequence do not change the size of the shadow. Hence $\left|\delta B\right|=\left|\delta C\right|+\left|\delta D\right|=\left|\delta I\right|+\left|\delta J_{A}\right|$. On the other hand, since ${\cal I}$ and ${\cal J}_{A}$ are initial segments of colex, Lemma 4 implies that $\left|\delta I\right|=\left|{\cal I}_{1}\right|$ and $\left|\delta J_{A}\right|=\left|\left({\cal J}_{A}\right)_{1}\right|$. Thus $$\left|\delta B\right|=\left|{\cal I}_{1}\right|+\left|\left({\cal J}_{A}\right)_{1}\right|.$$
Since ${\cal I}_{A}$ is an initial segment of colex, Lemma 4 implies that $\left|\delta I_{A}\right|=\left|\left({\cal I}_{A}\right)_{1}\right|$. But ${\cal I}_{A}$ is a disjoint union of ${\cal J}_{A}$ and ${\cal A}$ so $$\left|\delta I_{A}\right|=\left|\left({\cal I}_{A}\right)_{1}\right|=\left|\left({\cal J}_{A}\right)_{1}\right|+\left|{\cal A}_{1}\right|.$$ Since ${\cal I}_{A}$ is an initial segment of colex, the corresponding set of sequences $I_{A}$ has minimal shadow inside a connected component. Since $\left|{\cal B}\right|=\left|{\cal I}_{A}\right|$, it follows that $$\left|\delta B\right|\geq\left|\delta I_{A}\right|.$$ Thus combining (4), (5) and (6) yields $$\left|{\cal I}_{1}\right|\geq\left|{\cal A}_{1}\right|$$ as required. $\square$\
**Claim 1 $\Rightarrow$ Claim 3.** Let ${\cal A}$ and ${\cal I}$ be as in Claim 3. Define $\overline{{\cal A}}=\left\{ A^{c}\,:\,A\in{\cal A}\right\} $ and define $\overline{{\cal I}}$ similarly. Note that $\left|\overline{{\cal A}}\right|=\left|{\cal A}\right|$ and $\overline{{\cal A}}\subseteq\left[n\right]^{(n-r)}$. It is easy to check that if ${\cal B}\subseteq\left[n\right]^{(r)}$ is an initial segment of colex, then so is $\overline{\left(\left[n\right]^{(r)}\setminus{\cal B}\right)}$ . Thus $\overline{{\cal I}}$ is an initial segment of colex.
Since ${\cal A}$ is a segment, there exists initial segments ${\cal K}$ and ${\cal L}$ such that ${\cal A}={\cal K}\setminus{\cal L}$. This can be rewritten as ${\cal A}=\left(\left[n\right]^{\left(r\right)}\setminus{\cal L}\right)\setminus\left(\left[n\right]^{\left(r\right)}\setminus{\cal K}\right)$ and hence $\overline{{\cal A}}=\overline{\left(\left[n\right]^{\left(r\right)}\setminus{\cal L}\right)\setminus\left(\left[n\right]^{\left(r\right)}\setminus{\cal K}\right)}=\overline{\left(\left[n\right]^{\left(r\right)}\setminus{\cal L}\right)}\setminus\overline{\left(\left[n\right]^{\left(r\right)}\setminus{\cal K}\right)}$. As $\overline{\left(\left[n\right]^{\left(r\right)}\setminus{\cal L}\right)}$ and $\overline{\left(\left[n\right]^{\left(r\right)}\setminus{\cal K}\right)}$ are initial segments of colex, it follows that $\overline{{\cal A}}$ is a segment as well.
Hence $\overline{{\cal A}}$ and $\overline{{\cal I}}$ satisfies the conditions of Claim 1, and therefore $$\left|\left(\overline{{\cal I}}\right)_{1}\right|\geq\left|\left(\overline{{\cal A}}\right)_{1}\right|.$$
Note that for any set system ${\cal B}$, we have $\left|{\cal B}\right|=\left|{\cal B}_{1}\right|+\left|\left(\overline{{\cal B}}\right)_{1}\right|$ as for every $A\in{\cal B}$ exactly one of $1\in A$ and $1\in A^{c}$ is satisfied. Thus $$\left|{\cal I}\right|=\left|{\cal I}_{1}\right|+\left|\left(\overline{{\cal I}}\right)_{1}\right|$$ and $$\left|{\cal A}\right|=\left|{\cal A}_{1}\right|+\left|\left(\overline{{\cal A}}\right)_{1}\right|.$$ Combining (8), (9) and (10) with $\left|{\cal I}\right|=\left|{\cal A}\right|$ yields that $$\left|{\cal A}_{1}\right|\geq\left|{\cal I}_{1}\right|$$ which completes the proof of Claim 3.$\square$\
**Claim 1 $\Rightarrow$ Claim 2.** Let ${\cal I}$ and ${\cal J}$ be as in Claim 2. For $i+1\leq j\leq n$ let ${\cal S}_{j}=\left\{ A\setminus\{j\}\,:\,A\in{\cal J},\,\max A=j\right\} $. Thus ${\cal S}_{j}\subseteq\left[j-1\right]^{\left(i\right)}\subseteq\left[n\right]^{\left(i\right)}$ for all $i$. Since ${\cal J}$ is an initial segment of colex, it follows that ${\cal S}_{j}$ is an initial segment of colex in $\left[j-1\right]^{\left(i\right)}$ for all $j$. ${\cal S}_{j}$ is an initial segment of colex also in $\left[n\right]^{(i)}$ as initial segments of colex are not affected by adding new larger elements to the ground set. Note that we can express ${\cal J}$ as a disjoint union ${\cal J}=\bigcup_{j=i+1}^{n}\left({\cal S}_{j}+\left\{ j\right\} \right)$. Hence $$\left|{\cal J}_{1}\right|=\sum_{j=i+1}^{n}\left|\left({\cal S}_{j}+\left\{ j\right\} \right)_{1}\right|=\sum_{j=i+1}^{n}\left|\left({\cal S}_{j}\right)_{1}\right|.$$
Since each ${\cal S}_{j}$ is an initial segment of colex in $\left[n\right]^{(i)}$ and we have$\sum_{j=i+1}^{n}\left|{\cal S}_{j}\right|=\left|{\cal J}\right|=\left|{\cal I}\right|$, a repeated application of Claim 1 implies that $\left|{\cal J}_{1}\right|\geq\left|{\cal I}_{1}\right|$. $\square$\
**Claim 2 $\Rightarrow$ Claim 4.**
Let ${\cal I}$, ${\cal J}$, ${\cal I}_{*}$ and ${\cal J}_{*}$ be as in Claim 4. Since ${\cal I}_{*}$ and ${\cal J}_{*}$ are initial segments of colex, the observation pointed out in the proof of Claim 1 $\Rightarrow$ Claim 3 implies that $\overline{{\cal I}}\subseteq\left[n\right]^{(n-i)}$ and $\overline{{\cal J}}\subseteq\left[n\right]^{(n-i-1)}$ are initial segments of colex as well. Thus Claim 2 implies that $$\left|\left(\overline{{\cal I}}\right)_{1}\right|\geq\left|\left(\overline{{\cal J}}\right)_{1}\right|.$$ Combining this with $$\left|{\cal I}\right|=\left|{\cal I}_{1}\right|+\left|\left(\overline{{\cal I}}\right)_{1}\right|$$ and $$\left|{\cal J}\right|=\left|{\cal J}_{1}\right|+\left|\left(\overline{{\cal J}}\right)_{1}\right|$$ yields $\left|{\cal J}_{1}\right|\geq\left|{\cal I}_{1}\right|$ as required. $\square$
This completes the proof of Lemma 9. $\square$\
We are now ready to deduce Theorem 1. For convenience, we will recall the definition of the order $\leq$ and restate Theorem 1. For distinct $x$, $y\in\left\{ 0,\dots,k\right\} ^{n}$ we set $x\leq y$ if
1. $w_{0}\left(x\right)<w_{0}\left(y\right)$
2. $w_{0}\left(x\right)=w_{0}\left(y\right)$, $re\left(x\right)\neq re\left(y\right)$ and $re\left(x\right)\leq_{c}re\left(y\right)$
3. $w_{0}\left(x\right)=w_{0}\left(y\right)$, $re\left(x\right)=re\left(y\right)$ and $R_{0}\left(x\right)\leq_{colex}R_{0}\left(y\right)$
**Theorem 1.** Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and let $B$ be an initial segment of $\leq$ with $\left|B\right|=\left|A\right|$. Then $\left|\delta A\right|\geq\left|\delta B\right|$.\
**Proof.** Let $A$ be a subset of $\left\{ 0,\dots,k\right\} ^{n}$ of given size with minimal $\delta A$. Define $$v(A)=\sum_{j=0}^{n}j\left|A\cap L_{j}\left(n\right)\right|.$$
If possible, choose $l\in\left[n\right]$, $s\in\left\{ 1,\dots,k\right\} ^{n-l}$ and $t\in\left\{ 1,\dots,k\right\} ^{n-l-1}$ for which $C_{s,t}\left(A\right)\neq A$. Then by Lemma 8, $B=C_{s,t}\left(A\right)$ satisfies $\left|\delta A\right|\geq\left|\delta B\right|$ and by minimality of $\delta A$ it follows that $\delta B$ is also minimal. We also have $v\left(A\right)>v\left(B\right)$, which follows from the definition of $C_{s,t}\left(A\right)$ and from the fact that $C_{s,t}\left(A\right)\neq A$.
Repeating this process we obtain a set $A_{1}$ of size $\left|A\right|$ with minimal $\delta A_{1}$ for which $C_{s,t}\left(A_{1}\right)=A_{1}$ for all $i$, $s\in\left\{ 1,\dots,k\right\} ^{n-l}$ and $t\in\left\{ 1,\dots,k\right\} ^{n-l-1}$. This follows from the fact that $v\left(B\right)$ is always a non-negative integer which strictly decreases on each step. Since $C_{s,t}\left(A_{1}\right)=A_{1}$ for all $l\in\left[n\right]$, $s\in\left\{ 1,\dots,k\right\} ^{n-l}$ and $t\in\left\{ 1,\dots,k\right\} ^{n-l-1}$, it is easy to check that there exists $i$ such that $L_{j}\left(n\right)\subseteq A_{1}$ for all $j<i$ and $L_{j}\left(n\right)\cap A_{1}=\emptyset$ for all $j>i$.
Let $C_{s_{1}},\,\dots,\,C_{s_{t}}$ be the connected components in $L_{i}\left(n\right)$ with $s_{j}\leq_{c}s_{k}$ for $j\leq k$. Define $$w\left(B\right)=\sum_{j=1}^{t}j\left|C_{s_{j}}\cap B\right|.$$ If possible, choose $j<k$ for which $C_{s_{j},s_{k}}\left(A_{1}\right)\neq A_{1}$, and set $B=C_{s_{j},s_{k}}\left(A_{1}\right)$. Now $\left|\delta A_{1}\right|\geq\left|\delta B\right|$ by Lemma 7 and hence $\delta B$ is also minimal. Also $w\left(A_{1}\right)>w\left(B\right)$ follows directly from the definition of the compression operator and from the definition of $B$. Repeating this process we obtain a set $A_{2}$ for which
1. $\delta A_{2}$ is minimal
2. There exists $i$ such that $L_{j}\left(n\right)\subseteq A_{2}$ for all $j<i$ and $L_{j}\left(n\right)\cap A_{2}=\emptyset$ for all $j>i$
3. $C_{s_{j},s_{k}}\left(A_{2}\right)=A_{2}$ for all $j<k$
Note that the process must terminate as $w\left(B\right)$ is always a non-negative integer which strictly decreases on each step. Since $C_{s_{j},s_{k}}\left(A_{2}\right)=A_{2}$ for all $j<k$ it follows that there exists $p$ for which $C_{s_{k}}\subseteq A$ for all $k<p$ and $C_{s_{k}}\cap A=\emptyset$ for all $k>p$.
Let $D=A_{2}\cap C_{s_{p}}$ and let $A_{3}$ be set obtained from $A_{2}$ by taking $A_{3}\cap C_{s_{p}}$ to be the set corresponding to an initial segment of colex of length $\left|D\right|$, and taking $A_{3}\setminus C_{s_{p}}=A_{2}\setminus C_{s_{p}}$. Then Lemma 3 implies that $\left|\delta A_{2}\right|\geq\left|\delta A_{3}\right|$ so $\delta A_{3}$ is minimal. On the other hand, by the construction of $A_{3}$ it is clear that it is an initial segment of $\leq$. Hence an initial segment of $\leq$ minimises $\delta$.$\square$
An extremal result for the generalised shadow
=============================================
So far we have considered operator which allows us to delete a coordinate which equals 0. It is natural to ask what happens if we generalise this set-up and allow the deletion of any coordinate that is in some chosen set.
Define $\delta_{r}$-shadow of $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ to be the subset of sequences in $\left\{ 0,\dots,k\right\} ^{n-1}$ obtained from any of its vectors by removing exactly one coordinate that is one of $\left\{ 0,\dots,r\right\} $. Thus $\delta=\delta_{0}$ and $\Delta=\delta_{k}$. Define $v_{r}\left(x\right)=\sum_{i=0}^{r}w_{i}(x)$. That is, $v_{r}\left(x\right)$ is the number of coordinates of $x$ in the set $\left\{ 0,\dots,r\right\} $. Define $L_{s}\left(n\right)=\left\{ x\in\left\{ 0,\dots,k\right\} ^{n}\,:\,v_{r}\left(x\right)=s\right\} $ and $L_{\leq s}\left(n\right)=\bigcup_{i=0}^{s}L_{i}\left(n\right)$. The aim of this section is to prove that the sets $L_{\leq s}\left(n\right)$ are extremal for $\delta_{r}$. This follows directly from the following Proposition.\
**Proposition 10.** Let $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and let $A_{s}=A\cap L_{s}\left(n\right)$. Then $$\left|\delta A\right|\geq\frac{1}{n\left(r+1\right)}\sum_{s=0}^{n}s\left|A_{s}\right|.$$\
**Proof.** Let $X=\left\{ 0,\dots,k\right\} ^{n}$, $Y=\left\{ 0,\dots,k\right\} ^{n-1}$, let $H$ be defined as in Section 3 and let ${\cal H}$ be a bipartite multigraph on $X\cup Y$ with edges given as follows. For each $x\in X\cap L_{s}\left(n\right)$ there are exactly $s$ coordinates $x_{i_{1}},\dots,x_{i_{s}}$ which are elements of $\left\{ 0,\dots,r\right\} $. Define $y_{j}$ to be the sequence obtained by deleting the coordinate $x_{i_{j}}$. Then certainly $y_{j}\in\delta x$ and some of the $y_{j}$ may be equal. Define the edges of ${\cal H}$ to be the edges $xy_{j}$ for all $1\leq j\leq s$ counting with multiplicities. For example, when $r=1$ the sequence $x=00121$ is connected by two edges to $0121$, and by one edge to both $0012$ and $0021$.
It is easy to verify that for all $y\in Y$, $y$ has degree $n\left(r+1\right)$ as this corresponds to adding any element of $\left\{ 0,\dots,r\right\} $ to any of the $n$ possible places in the sequence $y$. Note that for all $x\in X$ we have $\Gamma_{{\cal H}}\left(x\right)=\delta x$, and hence for any $A\subseteq X$ we have $\delta A=\Gamma_{{\cal H}}\left(A\right)$. By the definition of ${\cal H}$ we have $d\left(x\right)=s$ for all $x\in L_{s}\left(n\right)$, and as observed earlier we have $d\left(y\right)=n\left(r+1\right)$ for all $y\in Y$. Since the connected components of ${\cal H}$ are contained in the sets $L_{s}\left(n\right)\cup L_{s-1}\left(n-1\right)$, we have $\Gamma_{{\cal H}}\left(A\right)\cap L_{s-1}\left(n-1\right)=\Gamma_{{\cal H}}\left(A\cap L_{s}\left(n\right)\right)$ and therefore
$$\left|\Gamma_{{\cal H}}A\right|=\sum_{s=0}^{r}\left|\Gamma_{{\cal H}}\left(A_{s}\right)\right|.$$
For a set $B\subseteq L_{s}\left(n\right)$ we have $$s\left|B\right|=e\left(B,\Gamma_{{\cal H}}\left(B\right)\right)\leq e\left(\Gamma_{{\cal H}}\left(B\right),X\right)=\left|\Gamma_{{\cal H}}\left(B\right)\right|n\left(r+1\right)$$ and hence $$\left|\Gamma_{{\cal H}}\left(B\right)\right|\geq\frac{s}{n\left(r+1\right)}\left|B\right|.$$ Applying (17) to each term of the sum in (16) yields $$\left|\delta A\right|=\left|\Gamma_{{\cal H}}A\right|\geq\frac{1}{n\left(r+1\right)}\sum_{s=0}^{r}s\left|A_{s}\right|$$ as required. $\square$\
Now we are ready to conclude that the sets $L_{\leq s}\left(n\right)$ are extremal.\
**Corollary 11.** If $A\subseteq\left\{ 0,\dots,k\right\} ^{n}$ and $\left|A\right|=\left|L_{\leq s}\left(n\right)\right|$, then $\left|\delta A\right|\geq\left|\delta L_{\leq s}\left(n\right)\right|$ with equality if and only if $A=L_{\leq s}\left(n\right)$.\
**Proof.** Let $B=L_{\leq s}\left(n\right)$. We will first check that the equality holds for $B$ in (18). Note that $B_{i}=L_{i}\left(n\right)$ for all $i\leq s$ and $B_{i}=\emptyset$ for all $i>s$. For $i\le s$, $\left|B_{i}\right|=\left|L_{i}\left(n\right)\right|={n \choose i}\left(r+1\right)^{i}(k-r)^{n-i}$ and $\left|\delta B_{i}\right|=\left|L_{i-1}\left(n-1\right)\right|={n-1 \choose i-1}\left(r+1\right)^{i-1}\left(k-r\right)^{n-i}$. Therefore $\left|\delta B_{i}\right|=\frac{i}{n\left(r+1\right)}\left|B_{i}\right|$ holds for all $i\leq s$, and in fact also for $i>s$ as in this case both sides are 0. Hence the equality holds in (17) for all $i$, and thus the equality holds in (18) as well.
Given a set $A$ of fixed size with $\left|A_{i}\right|\leq\left|L_{i}\left(n\right)\right|$ for all $i$, it is easy to see that $\frac{1}{n\left(r+1\right)}\sum_{t=0}^{r}t\left|A_{t}\right|$ is minimised if and only if $A=L_{\leq n}\cup B$ for suitably chosen $n$ and for any $B\subseteq L_{n+1}$ of suitable size. Hence given $A$ with $\left|A\right|=\left|L_{\leq s}\left(n\right)\right|$, the quantity $\frac{1}{n\left(r+1\right)}\sum_{t=0}^{r}t\left|A_{t}\right|$ attains its minimum value uniquely when $A=L_{\le s}\left(n\right)$.
Thus $$\left|\delta A\right|\geq\frac{1}{n\left(r+1\right)}\sum_{t=0}^{r}t\left|A_{t}\right|\ge\frac{1}{n\left(r+1\right)}\sum_{t=0}^{s}t\left|L_{t}\left(n\right)\right|=\left|\delta L_{\leq s}\left(n\right)\right|$$ and the second inequality holds if and only if $A=L_{\leq s}\left(n\right)$, as required. $\square$
[1]{} B. Bollob[á]{}s, I. Leader, personal communication, 2017
T. N. Danh, D. E. Daykin, Ordering integer vectors for coordinate deletion, Journal of the London Mathematical Society **55**(3) (1997), 417-426
G. Katona, A theorem of finite sets, in: Theory of graphs (P. Erd�s and G. Katona eds.), Akademial Kiado and Academic Press, Budapest and New York (1968), 187-207
J. B. Kruskal, The number of simplices in a complex, in: Mathematical Optimization Techniques (R. Bellman ed.), Univ. of Californa Press, Berkeley, Los Angeles (1963), 251-278
U. Leck, Nonexistence of a Kruskal-Katona type theorem for subword orders, Combinatorica, **24** (2004), 305-312.
L.H. Loomis, H. Whitney. An inequality related to the isoperimetric inequality. Bull. Amer. Math. Soc., **55** (1949), 961962.
[^1]: Centre for Mathematical Sciences, Wilberforce Road, Cambridge CB3 0WB, UK, [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss several open problems on spectrally bounded operators, some new, some old, adding in a few new insights.'
address: |
Department of Pure Mathematics, Queen’s University Belfast,\
Belfast BT7 1NN, Northern Ireland\
[email protected]
author:
- MARTIN MATHIEU
title: |
A COLLECTION OF PROBLEMS ON\
SPECTRALLY BOUNDED OPERATORS [^1]
---
*Dedicated to Professor Rainer Nagel on the occasion of his retirement.*
Introduction {#sec:intro}
============
Let $A$ be a unital complex Banach algebra. A linear mapping $T\colon E\to B$ from a subspace $E\subseteq A$ into another unital complex Banach algebra $B$ is called *spectrally bounded* if there is a constant $M\geq0$ such that $r(Tx)\leq M\mkern.6mu r(x)$ for all $x\in E$. Here, and in what follows, $r(x)$ stands for the spectral radius of a Banach algebra element $x$.
This concept evolved in Banach algebra theory, and especially automatic continuity, over time in the 1970’s and 1980’s but the terminology was only introduced in [@MM1994a], together with its companions *spectrally infinitesimal*: $M=0$; *spectrally contractive*: $M=1$; and *spectrally isometric*: $r(Tx)=r(x)$ for all $x$. It follows from [@Aup82], see also [@AupMM], Lemma A, that the separating space of every surjective / $T$ on a closed subspace $E$ is contained in the radical of $B$; thus $T$ is bounded if $B$ is semisimple. This was used by Aupetit in [@Aup82] to give a new proof of Johnson’s uniqueness-of-the-complete-norm-topology theorem; cf. also [@Ran89]. It was further exploited in [@AupMM] to investigate continuity properties of Lie epimorphisms.
A systematic study of /s was begun in [@Schick], with its main results published in [@MMS1] and [@MMS2]. Since then, the interest in the topic has steadily grown and by now there is a sizeable literature on a variety of aspects. The present paper’s aim is to discuss several of the open problems on /s along with some new results that should make it clear why these questions are natural and important.
Section \[sec:auto-cont\] is devoted to a recapitulation of probably the most important and deepest problem on /s, the non-commutative Singer–Wermer conjecture. No substantial progress seems to have been made on this over the last few years. A standard method to reduce a more sophisticated problem in Banach algebra theory to a simpler one is to use quotients. When dealing with operators between Banach algebras, we, of course, need invariant ideals to perform this. In Section \[sec:quotients\] we discuss the interplay between properties of a (/) operator and the operator it induces on suitable quotient Banach algebras in a fairly general setting and recover recent results for operators preserving the essential spectral radius on $B(H)$ in [@BBS08].
The identity element plays a distinguished role for /s. Suppose that $T\colon A\to B$ is a surjective / between unital s $A$ and $B$. Since $T$ restricted to the centre $Z(A)$ of $A$ is a \*-isomorphism onto the centre $Z(B)$ of $B$ ([@MM05], Proposition 2.3), $T1$ is a central unitary in $B$. Therefore, whenever we study a / $T$, we can without loss of generality assume that $T$ is unital (i.e., $T1=1$); see the proof of [@LMM], Corollary 2.6. For a general /, the situation changes and for this reason we shall pay attention in Section \[sec:value-at-one\] to the relevance of the value $T1$. As it turns out, if the domain algebra is sufficiently ‘infinite’, then $T1$ must be a central invertible element in $B$. As a consequence, we are able to extend the main results in [@MMS2] and [@LMM] to the non-unital setting (Theorem \[thm:nonunital-spbdd\] below). The special case $A=B(H)$, $B=B(K)$ for Hilbert spaces $H$ and $K$ was studied previously in [@CuiHou].
In [@MM05], we propose a non-selfadjoint version of Kadison’s classical theorem stating that every unital surjective isometry between unital s is a Jordan \*-isomorphism. In many cases, this conjecture (see Problem \[prob:spiom-conjecture\] below) has been confirmed, however almost always under some strong assumption of ‘infiniteness’ on the domain . In Section \[sec:spec-isoms\], we employ a reduction method developed in [@MMSou] and [@MMRud] to obtain a new result of a similar ilk (Theorem \[thm:no-tracial-states\]). More importantly, perhaps, we propose a new route based on a description of /s in the presence of a trace to solve this open question at least in the setting of 1f/s (Corollary \[cor:type-twoone\]).
Automatic Continuity {#sec:auto-cont}
====================
The *separating space* ${{\mathcal{S}({T})}}$ of an operator $T$ between normed spaces $E$ and $F$ is defined by $${{\mathcal{S}({T})}}=\{y\in F\mid y=\lim_nTx_n\text{ for some }(x_n)_{n\in{{\mathbb N}}}\subseteq E,\ x_n\to0\}$$ and measures the degree of discontinuity of $T$. If $E$ and $F$ are Banach spaces then $T$ is bounded if and only if ${{\mathcal{S}({T})}}=\{0\}$, by the closed graph theorem. For a detailed discussion of the separating space see [@Dal].
It follows from [@Aup82], see also [@Ran89] and [@AupMM], that ${{\mathcal{S}({T})}}\cap TE$ consists of quasinilpotent elements whenever $T\colon E\to B$ is a / on a subspace $E$ of a Banach algebra. In fact, $r(Tx)\leq r(Tx-y)$ for each $x\in E$ and $y\in{{\mathcal{S}({T})}}$. This, together with Zemánek’s characterisation of the Jacobson radical ${{\textup{rad}({A})}}$ of $A$, entails the following result.
\[thm:aupetit-aut-cont\] Let $T\colon E\to B$ be a / on a closed subspace $E$ of a Banach algebra onto a Banach algebra $B$. Then ${{\mathcal{S}({T})}}\subseteq{{\textup{rad}({B})}}$. In particular, if $B$ is semisimple, then $T$ is bounded.
The above estimate on the spectral radii is used in [@AupMM] to describe the continuity of surjective Lie homomorphisms as follows.
\[thm:aut-cont-lie\] The separating space ${{\mathcal{S}({\theta})}}$ of a Lie epimorphism $\theta$ between two Banach algebras $A$ and $B$ is contained in $\mathcal{Z}(B)$, the centre modulo the radical.
The centre modulo the radical, $\mathcal{Z}(B)$, is the pre-image of the centre of $\hat B=B/{{\textup{rad}({B})}}$ under the canonical map $B\to\hat B$ and turns out to be the largest quasinilpotent Lie ideal of $B$. As a consequence, it is invariant under every Lie epimorphism and thus the above result yields the following automatic continuity statement ([@AupMM], Corollary).
Let $\theta\colon A\to B$ be a Lie epimorphism between the Banach algebras $A$ and $B$. The induced Lie epimorphism $\hat\theta\colon\hat A= A/{{\textup{rad}({A})}}\to\hat B$ between the Banach Lie algebras $\hat A$ and $\hat B$ is continuous.
The centre modulo the radical also plays an important role in the automatic continuity of derivations on Banach algebras. It is an open question whether the separating space of each of the iterates $\delta^n$, $n\in{{\mathbb N}}$ of a derivation $\delta\colon A\to A$ on a Banach algebra $A$ is contained in the radical of $A$. If this is the case, the following long-standing problem would have a positive answer.
\[prob:ncsw\] Does $[x,\delta x]\in\mathcal{Z}(A)$ for all $x\in A$ imply that $\delta A\subseteq{{\textup{rad}({A})}}$?
Here, $[x,y]$ of course stands for the commutator $xy-yx$. If $\mathcal{Z}(A)$ is replaced by the proper centre $Z(A)$ of $A$, then Problem \[prob:ncsw\] has an affirmative answer, due to a reduction technique developed in [@MMRun] and Marc Thomas’ famous theorem [@Thom]. The connection to /s is provided by the next result [@BresMM].
A derivation on a Banach algebra $A$ is / if and only if it maps into ${{\textup{rad}({A})}}$.
Hence, Problem \[prob:ncsw\] is equivalent to the following question.
[**Problem 1$^\prime$.**]{} Does $[x,\delta x]\in\mathcal{Z}(A)$ for all $x\in A$ imply that $\delta$ is /?
It appears that no progress on this question has been made in the last decade. For a fuller discussion see [@MM1994a] and [@Dal].
Quotients {#sec:quotients}
=========
In the previous section, we exploited the concept of /s to study linear mappings satisfying some additional algebraic conditions, such as Lie homomorphisms, derivations, etc. Another main direction of research has been on the question which algebraic properties can be derived from the assumption of spectral boundedness. Maybe the most prominent of these problems is the following one from [@Kap]. By a *Jordan epimorphism* we will, of course, understand a surjective linear mapping preserving the Jordan product $x\circ y=\frac12(xy+yx)$.
\[prob:kaplansky\] Let $A$ and $B$ be semisimple unital Banach algebras. Suppose $T\colon A\to B$ is unital, surjective and invertibility-preserving. Is $T$ necessarily a Jordan epimorphism?
Note that a Jordan epimorphism necessarily has all the properties assumed in Problem \[prob:kaplansky\] and hence is bounded, by Theorem \[thm:aupetit-aut-cont\]. Many contributions on Kaplansky’s problem have been made over the past decades but so far it eludes a final answer. From Aupetit’s substantial work on the question we only quote the following beautiful result in [@Aup00]; see also the references therein.
\[thm:aupetit-main\] Let $A$ and $B$ be von Neumann algebras. Then every unital surjective invertibility-preserving linear mapping $T\colon A\to B$ is a Jordan epimorphism.
To the best of our knowledge, Problem \[prob:kaplansky\] is still open even in the case when both domain and codomain are s; cf. also [@Harr01]. The structural investigation of /s is partly motivated by the question to what extent the hypotheses in Kaplansky’s problem can be relaxed while retaining the same goal, to show that the mapping is a Jordan homomorphism. The relevance of the value at $1$ will be discussed in the next section. When considering the possibility of replacing “invertibility-preserving” by “/” we must keep in mind that (a) *every* bounded linear mapping defined on a commutative is / (since spectral radius and norm coincide in spaces of the form $C(X)$); (b) a finite trace on a is a spectral contraction, hence, already on the $n\times n$ matrices there is a unital / onto ${{\mathbb C}}$ which is not a Jordan epimorphism. Notwithstanding this there are satisfactory results in the setting of ‘very infinite’ s; see Sections \[sec:value-at-one\] and \[sec:spec-isoms\].
Many examples are known illustrating the fact that no strong results can be expected for non-surjective /s in general; see, e.g., [@Harr01]. We shall now study the situation when a / is merely surjective ‘up to’ or ‘modulo’ an ideal.
Suppose $I\subseteq A$ and $J\subseteq B$ are proper closed ideals in the unital Banach algebras $A$ and $B$, respectively. We say that a linear mapping $T\colon A\to B$ is *essentially /* (more precisely, $I$-$J$-*essentially /*) if there is a constant $M\geq0$ such that $r(Tx+J)\leq M\mkern.6mu r(x+I)$ for all $x\in A$. We call $T$ *surjective modulo* $J$ if, for every $y\in B$, there is $x\in A$ such such $y-Tx\in J$. If $TI\subseteq J$, we can define the induced linear mapping $\hat T\colon A/I\to B/J$ by $\hat T(x+I)=Tx+J$, $x\in A$. The following proposition relates the properties of $\hat T$ to those of $T$.
\[prop:ess-spec-bdd\] Let $A$ and $B$ be unital Banach algebras. Suppose that $I$ and $J$ are proper closed ideals of $A$ and $B$, respectively, such that $B/J$ is semisimple. For a linear mapping $T\colon A\to B$ the following conditions are equivalent.
1. $T$ is essentially spectrally bounded and surjective modulo $J$;
2. $TI\subseteq J$ and $\hat T$ is / and surjective.
(b)${}\Rightarrow{}$(a) Under the assumption $TI\subseteq J$, $\hat T$ is a well-defined linear mapping which is / if and only if $T$ is essentially /, by definition. Moreover, $\hat T$ surjective precisely means that, for each $y\in B$, $y+J=\hat T(x+I)=Tx+J$ for some $x\in A$; that is, $T$ is surjective modulo $J$.
(a)${}\Rightarrow{}$(b) By the first paragraph in this proof, it suffices to show that $TI\subseteq J$. Take $x\in I$. By hypothesis, for each $y\in B$, there is $x'\in A$ such that $y-Tx'\in J$. Let $\lambda\in{{\mathbb C}}$. We have $$\begin{split}
r(\lambda\,(Tx+J)+y+J) &= r(\lambda\,Tx+y+J)\\
&= r(T(\lambda\,x+x')+J)\\
&\leq M\mkern.6mu r(\lambda\,x+x'+I)\\
&= M\mkern.6mu r(x'+I),
\end{split}$$ for some $M\geq0$. Consequently, the subharmonic function $\lambda\mapsto r(\lambda\,(Tx+J)+y+J)$ is bounded on ${{\mathbb C}}$, hence constant. It follows that $r(Tx+J+y+J)=r(y+J)$ for all $y\in B$ from which we conclude that $Tx+J\in{{\textup{rad}({B/J})}}$, by Zemánek’s characterisation of the radical. As $B/J$ is semisimple, we obtain that $Tx\in J$ as desired.
As an immediate consequence we have the following variant of the main result in [@LMM].
\[cor:pure-infinite-quotient\] Let $T\colon A\to B$ be a unital linear mapping from a unital purely infinite $A$ with real rank zero into a unital Banach algebra $B$. Let $J\subseteq B$ be a proper closed ideal in $B$ such that $B/J$ is semisimple and suppose that, for some proper closed ideal $I\subseteq A$ of $A$, the operator $T$ is $I$-$J$-essentially / and surjective modulo $J$. Then $T$ is a Jordan epimorphism modulo $J$.
By Proposition \[prop:ess-spec-bdd\], $TI\subseteq J$ and the induced mapping $\hat T\colon A/I\to B/J$ is unital, / and surjective. Since $A/I$ is purely infinite and has real rank zero, Corollary 2.5 in [@LMM] implies that $\hat T$ is a Jordan epimorphism; hence, $T$ is a Jordan epimorphism modulo $J$.
In particular, if $T$ is even a spectral isometry modulo $J$ in the above situation, then $\hat T$ provides a Jordan isomorphism between the quotients $A/I$ and $B/J$. This was obtained in the special case $A=B=B(H)$ and $I=J=K(H)$ for an infinite dimensional Hilbert space $H$ in [@BBS08], Theorem 3.1.
In order to make use of quotients, invariant ideals are needed. In the setting of s, so-called Glimm ideals offer themselves as good candidates, see Section \[sec:spec-isoms\] below. However, so far their invariance has only been established under additional hypotheses, for instance for / on von Neumann algebras [@MMSou].
Does every unital / from a unital onto another one leave each Glimm ideal invariant?
The setting of s is favourable also because a / induces a / on every quotient ([@MMSou], Proposition 9). In a more general setting it is easy to see that a / induces a / quotient operator provided the domain algebra is an *SR*-algebra; cf. [@MMRud].
The Relevance of the Value at $1$ {#sec:value-at-one}
=================================
A surjective Jordan homomorphism between unital algebras attains the value $1$ at $1$. In this section, we shall discuss the behaviour of an arbitrary / at $1$ to see how this affects the possible difference from a Jordan homomorphism. Remembering that *every* bounded operator from a space $C(X)$ is /, we have to be careful not to expect too much in a general setting.
The following observation follows directly from our earlier results.
\[prop:basic-fact\] Let $T\colon A\to B$ be a / from a unital $A$ onto a unital semisimple Banach algebra $B$. For every pair $p,\,q$ of mutually orthogonal properly infinite projections in $A$ we have $$\label{eq:basic-fact}
(Ta)\,(Tb)+(Tb)\,(Ta)=0\qquad(a\in pAp,\ b\in qAq).$$
By [@MMS2], Corollary 3.2, $T$ preserves elements of square zero. By [@LMM], Proposition 2.1, every element in the corners $pAp$ and $qAq$ can be written as a finite sum of elements of square zero (in $pAp$ and $qAq$, respectively), since both $p$ and $q$ are properly infinite. The claim thus follows from Lemma 3.3 in [@MMS2].
\[cor:t-of-one\] Let $T\colon A\to B$ be a / from a unital $A$ with real rank zero onto a unital semisimple Banach algebra $B$. Suppose that every non-zero projection in $A$ is properly infinite. Then $T1$ is an invertible element in the centre of $B$.
The basic idea of the argument has been used before in some special cases, see, e.g., [@CuiHou]. Let $p\in A$ be a non-trivial projection. Applying the identity in Proposition \[prop:basic-fact\] to $a=p$ and $b=q=1-p$ we obtain $$(Tp)\,T(1-p)+T(1-p)\,(Tp)=0,$$ that is, $(Tp)\,(T1)+(T1)\,(Tp)=2\,(Tp)^2$. Upon multiplying this identity first on the left, then on the right by $Tp$ and subtracting the resulting two identities we obtain $(Tp)^2\,(T1)=(T1)\,(Tp)^2$ for every (non-trivial) projection $p$.
Let $\{p_1,\ldots,p_n\}$ be an orthogonal family of projections in $A$. By applying (\[eq:basic-fact\]) inductively we find that $\bigl(T\bigl(\sum_i p_i\bigr)\bigr)^2=\sum_i\bigl(Tp_i\bigr)^2$. Since $A$ has real rank zero, it follows that $(T1)\,(Ta)^2=(Ta)^2\,(T1)$ for all $a$ in a dense subset of $A_{sa}$ and since $T$ is bounded, thus for all $a\in A_{sa}$. As $$ab+ba=(a+b)^2-a^2-b^2$$ and $$(a+ib)^2=a^2-b^2+i(ab+ba)$$ for all selfadjoint $a,b\in A$, we conclude that $T1$ commutes with $(Tx)^2$ for every $x\in A$. The surjectivity of $T$ entails that $T1$ commutes with every square of an element in $B$ but since $2y=(1+y)^2-1-y^2$ for each $y\in B$, we finally obtain that $T1$ belongs to the centre of $B$.
Going back to the first paragraph of the proof we therefore have $T(p^2)\,(T1)=(Tp)\,(T1)=(Tp)^2$ for every projection $p\in A$. Using the same argumentation as above and the fact that $A$ has real rank zero another time we conclude that $T(x^2)\,(T1)=(Tx)^2$ for all $x\in A$. As $T$ is surjective, it follows that $T1$ must be invertible.
Whenever $T$ is a / such that $T1$ is an invertible element in the centre of the codomain, the mapping $\tilde T$ defined by $\tilde Tx=(T1)^{-1}\,Tx$, $x\in A$ is a unital /; thus we can apply the results known in this situation.
\[thm:nonunital-spbdd\] Let $T\colon A\to B$ be a / from a unital $A$ onto a unital semisimple Banach algebra $B$. Suppose that
1. $A$ is a properly infinite von Neumann algebra, or
2. $A$ is a purely infinite with real rank zero.
Then there is a unique Jordan epimorphism $J\colon A\to B$ such that $Tx=(T1)\,Jx$ for all $x\in A$. Moreover, $T1$ is a central invertible element in $B$.
Case (ii) is immediate from Corollary \[cor:t-of-one\], since every non-zero projection in a purely infinite is properly infinite. Thus we can define $J$ by $Jx=(T1)^{-1}\,Tx$, $x\in A$, which is a Jordan epimorphism by [@LMM], Corollary 2.5.
The case (i) will not only need the result in the unital case, [@MMS2], Theorem 3.6, in the same manner as just explained but also an elaboration of the projection techniques used in the main technical result to obtain the unital case, which is [@MMS2], Proposition 3.4. Note that, by the proof of Corollary \[cor:t-of-one\], it suffices to show that $T1$ commutes with $(Tp)^2$ for every projection $p\in A$; once this is verified, $T1$ will be a central invertible element. We shall use the same strategy as in Proposition 3.4 of [@MMS2].
Suppose first that both $p$ and $1-p$ are properly infinite. Then the assertion follows immediately from identity (\[eq:basic-fact\]).
Next suppose that $p$ is properly infinite but $1-p$ is not. By the Halving Lemma, there is a subprojection $f$ of $p$ such that $p\sim f\sim p-f$, where $\sim$ denotes Murray–von Neumann equivalence. Hence, all projections $f$, $1-f$, $p-f$ and $1-p+f$ are properly infinite, see the proof of [@MMS2], Proposition 3.4. By our first step we thus have $$(T1)\,(Tf)^2=(Tf)^2\,(T1)\quad\text{and}\quad(T1)\,T(p-f)^2=T(p-f)^2\,(T1).$$ Applying (\[eq:basic-fact\]) to $f$ and $p-f$ we have $(Tp)^2=T(p-f)^2+(Tf)^2$, wherefore $T1$ commutes with $(Tp)^2$.
Suppose now that $p$ is infinite but not properly infinite. Then there is a non-trivial central projection $z$ in $A$ such that $zp$ is properly infinite and $(1-z)p$ is finite. We need the following preliminary observation. Suppose that $q$ is a properly infinite projection in $A$ but $1-q$ is not. Choosing the subprojection $f$ of $q$ as in the last paragraph we have $$\begin{split}
(Tq)\,(T1)+(T1)\,(Tq) &= (Tf)\,(T1)+(T1)\,(Tf)+T(q-f)\,(T1)+(T1)\,T(q-f)\\
&= 2\,(Tf)^2+2\,T(q-f)^2 = 2\,(Tq)^2,
\end{split}$$ where the last two equality signs come from identity (\[eq:basic-fact\]). Multiplying this identity first on the left, then on the right by $T1$ and subtracting the two identities we obtain $$(T1)^2\,(Tq)-(Tq)\,(T1)^2=2\,\bigl((T1)\,(Tq)^2-(Tq)^2\,(T1)\bigr)=0,$$ by the previous paragraph.
To simplify our calculations we will now use the usual commutator notation $[x,y]=xy-yx$. Since $zp$ is a properly infinite projection, the preliminary observation yields $[(T1)^2, T(zp)]=0$. By the proof of [@MMS2], Proposition 3.4, the projection $q=(1-z)(1-p)$ is properly infinite as well; therefore, $[(T1)^2,T((1-z)(1-p))]=0$, too. As a result, $$\begin{split}
[(T1)^2,T((1-z)p)] &= [(T1)^2,Tp]-[(T1)^2,T(zp)] = -\, [(T1)^2,T(1-p)]\\
&=-\, [(T1)^2,T(z(1-p))]-[(T1)^2,T((1-z)(1-p))]\\
&=-\, [(T1)^2,Tz]+[(T1)^2,T(zp)]=0,
\end{split}$$ since every non-zero central projection in $A$, in particular $z$, is properly infinite and thus $[(T1)^2,Tz]=0$ as well.
From identity (\[eq:basic-fact\]) we have $$\label{eq:basic-fact-central}
T(za)\,T((1-z)b)+T((1-z)b)\,T(za)=0\qquad(a,b\in A),$$ since $1-z$ is non-zero, hence properly infinite. In particular, $$(Tz)\,T((1-z)p)+T((1-z)p)\,(Tz)=0.$$ It follows that $$\begin{split}
(Tq)^2 &=T((1-z)(1-p))^2 = \bigl(T(1-z)-T((1-z)p)\bigr)^2\\
&= T(1-z)^2+T((1-z)p)^2-(T(1-z)\,T((1-z)p)+T((1-z)p)\,T(1-z))\\
&=T(1-z)^2+T((1-z)p)^2-((T1)\,T((1-z)p)+T((1-z)p)\,(T1)).
\end{split}$$ Combining this with $$[(T1)^2,Tw]=T1\,[T1,Tw]+[T1,Tw]\,T1=[T1,T1\,Tw+Tw\,T1]\qquad(w\in A)$$ we find that $$\begin{split}
0 &= [T1, (Tq)^2]\\
&= [T1,T(1-z)^2]+[T1,T((1-z)p)^2] - [T1,(T1)\,T((1-z)p)+T((1-z)p)\,(T1)]\\
&= 0 + [T1,T((1-z)p)^2] - [(T1)^2,T((1-z)p)],
\end{split}$$ that is, $[T1,T((1-z)p)^2]=[(T1)^2,T((1-z)p]$. The commutator on the right hand side is zero, as we saw above. Therefore, $$[T1, (Tp)^2]=[T1,T(zp)^2]+[T1,T((1-z)p)^2]=0,$$ where we used (\[eq:basic-fact-central\]) once again.
Finally suppose that $p$ is finite. Then $p'=1-p$ is infinite. Letting $z'$ be the central projection such that $z'p'$ is properly infinite and $(1-z')p'$ is finite, we recollect the necessary information from the arguments above.
1. $[T1,(Tp')^2]=0$;
2. $[(T1)^2, T((1-z')p']=[T1, T((1-z')p')^2]=0$;
3. $[(T1)^2,T(z'p')]=[T1,T(z'p')^2]=0$.
Consequently, $$[(T1)^2,Tp']=[(T1)^2,T(z'p')]+[(T1)^2,T((1-z')p')]=0.$$ As $$\begin{split}
[T1,T(1-p)^2] &=[T1, (T1)^2]+[T1,(Tp)^2]-[T1,(T1)\,(Tp)+(Tp)\,(T1)]\\
&=[T1,(Tp)^2]-[(T1)^2,Tp]=[T1,(Tp)^2]+[(T1)^2,T(1-p)],
\end{split}$$ we conclude that $$[T1,(Tp)^2]=[T1,(Tp')^2]-[(T1)^2,Tp']=0.$$
This completes the argument that $[T1,(Tp)^2]=0$ for every projection $p\in A$ and, as explained above, this is sufficient to prove the result.
Evidently, all we need to assume on $T$ in the proof of the above theorem is that $T$ is surjective, bounded and preserves elements of square zero. Thus, the main result in [@LMM], Theorem 2.4, extends appropriately to the non-unital setting.
As pointed out above, we cannot expect a result like Theorem \[thm:nonunital-spbdd\] for arbitrary s of real rank zero or even von Neumann algebras. However, in a more restricted setting the situation might change.
\[prob:Tone-on-finite-factor\] Let $T$ be a / defined on a finite von Neumann factor (onto a semisimple unital Banach algebra $B$). Is $T1$ a non-zero complex multiple of the identity in $B$?
Spectral Isometries {#sec:spec-isoms}
===================
Generalising the classical Banach–Stone theorem, Kadison showed in [@Kad51] that every unital surjective isometry $T$ between two unital s must be a Jordan \*-isomorphism. In fact, he showed a stronger result in [@Kad52], Theorem 2, namely that it suffices that $T$ maps the selfadjoint part $A_{sa}$ of $A$ isometrically onto the selfadjoint part $B_{sa}$ of $B$. There are many interesting consequences of these results, all relating isometric properties to selfadjointness. As an example, we state and prove a variant of a theorem of Chan ([@Chan], Theorem 3).
Recall that the numerical radius $\nu(x)$ of an element $x$ in a unital $A$ is defined by $\nu(x)=\sup\{|\varphi(x)|\mid\varphi\ \text{a state of }A\}$.
Every unital surjective numerical radius-preserving operator $T$ between two unital s $A$ and $B$ is a Jordan \*-isomorphism.
Since $\nu$ is a norm, $T$ is injective and it is clear that thus $T^{-1}$ is numerical radius-preserving as well. As $\nu(x)=\|x\|$ for all $x\in A_{sa}$, by Kadison’s theorem it suffices to show that $TA_{sa}\subseteq B_{sa}$. Let $a\in A_{sa}$, $\|a\|=1$. Write $Ta=b+ic$ with $b,c\in B_{sa}$. Suppose $c\ne0$; then there is $0\ne\gamma\in\sigma(c)$ and we may assume that $\gamma>0$. For each state ${\varphi}$ on $B$ and each $n\in{{\mathbb N}}$, we have $$|{\varphi}(c+n)|^2 \leq |{\varphi}(b)|^2+|{\varphi}(c+n)|^2 = |{\varphi}(b)+i{\varphi}(c+n)|^2 = |{\varphi}(Ta+in)|^2.$$ Hence, for large $n$, $$\nu(a+in)^2=1+n^2<(\gamma+n)^2\leq\nu(c+n)^2\leq\nu(Ta+in)^2$$ which is impossible as $T$ is numerical radius-preserving. Therefore, $c=0$ and $Ta$ is selfadjoint.
Kadison’s theorem in one direction and an application of the Russo–Dye theorem in the other direction show that a unital surjective spectral isometry is selfadjoint (i.e., maps selfadjoint elements to selfadjoint elements) if and only if it is an isometry, see [@MM05], Proposition 2.4. These results and others, and the analogy to Kaplansky’s question (Problem \[prob:kaplansky\] above), made us ask the following question which we indeed state as a conjecture in [@MMS1].
\[prob:spiom-conjecture\] Is every unital surjective / between unital s a Jordan isomorphism?
By now, there is a fair number of results affirming this conjecture in reasonable, though not full generality. Combining the methods of [@MM04b], [@MMRud] and [@MMSou] we can cover another new case below.
\[thm:no-tracial-states\] Let $T\colon A\to B$ be a unital spectral isometry from a unital $A$ with real rank zero and without tracial states such that ${{\hbox{\it Prim}\kern.05em(A)}}$ is Hausdorff and totally disconnected onto a unital $B$. Then $T$ is a Jordan isomorphism.
The above statement includes in particular the simple case but we will reduce the more general situation to this one. To this end, we need the notion of a Glimm ideal in a .
There are (at least) five structure spaces associated with a unital $A$, which we will briefly discuss. Their relation can be depicted as follows. $${{\hbox{\it Irr}\kern.05em(A)}}\longrightarrow{{\hbox{\it Spec}\kern.05em(A)}}\longrightarrow{{\hbox{\it Prim}\kern.05em(A)}}\,{\buildrel{{\beta}}\over\longrightarrow}\,{{\hbox{\it Glimm}\kern.05em(A)}}\,{\buildrel{{\cong}}\over\longrightarrow}\,{{\hbox{\it Spec}\kern.05em(Z(A))}}$$ The set ${{\hbox{\it Irr}\kern.05em(A)}}$ of all irreducible representations of $A$ maps onto the set ${{\hbox{\it Spec}\kern.05em(A)}}$ of all unitary equivalence classes of such representations. Since equivalent representations have the same kernel, there is a canonical surjection from ${{\hbox{\it Spec}\kern.05em(A)}}$ onto ${{\hbox{\it Prim}\kern.05em(A)}}$, the set of all primitive ideals of $A$. The latter carries a natural topology, the Jacobson topology, which can be pulled back to ${{\hbox{\it Spec}\kern.05em(A)}}$, and then to ${{\hbox{\it Irr}\kern.05em(A)}}$, to turn them into topological spaces in a canonical way. Since $Z(A)$ is a commutative unital , ${{\hbox{\it Spec}\kern.05em(Z(A))}}$ allows several equivalent descriptions of which we choose the one via maximal ideals. For each maximal ideal $M$ of $Z(A)$, the closed ideal $$AM=\Bigl\{\sum_{j=1}^n\,x_jz_j\mid x_j\in A,\,z_j\in M,\,n\in{{\mathbb N}}\Bigr\}$$ is called a *Glimm ideal* of $A$. For each $I\in{{\hbox{\it Glimm}\kern.05em(A)}}$, $M=I\cap Z(A)$ gives the generating maximal ideal in $Z(A)$ back, wherefore there is a bijection between the two sets. Transporting the natural topology of ${{\hbox{\it Spec}\kern.05em(Z(A))}}$ over to ${{\hbox{\it Glimm}\kern.05em(A)}}$ thus turns the latter into a compact Hausdorff space. From the other direction, we can apply the *complete regularisation map* $\beta\colon{{\hbox{\it Prim}\kern.05em(A)}}\to{{\hbox{\it Glimm}\kern.05em(A)}}$ defined by $$P,Q\in{{\hbox{\it Prim}\kern.05em(A)}}\colon\quad P\sim Q \text{ \ if \ } f(P)=f(Q)\qquad\bigl(f\in C_b({{\hbox{\it Prim}\kern.05em(A)}})\bigr)$$ and $\beta(P)$ is the equivalence class with respect to this relation. If $I$ is the Glimm ideal given by $I=A(P\cap Z(A))$ then $\beta(P)$ can be identified with $\bigcap\limits_{Q\supseteq I}Q$. Among the consequences of this is $\bigcap{{\hbox{\it Glimm}\kern.05em(A)}}=\bigcap{{\hbox{\it Prim}\kern.05em(A)}}=\{0\}$, that is, the Glimm ideals separate the points of $A$.
It follows from the definition of the Jacobson’s topology that ${{\hbox{\it Prim}\kern.05em(A)}}$ is T$_1$ if and only if every primitive ideal is maximal. In a similar vein, ${{\hbox{\it Prim}\kern.05em(A)}}$ is T$_2$, i.e., Hausdorff, if and only if every Glimm ideal is maximal ([@MMRud], Lemma 9).
**Proof of Theorem \[thm:no-tracial-states\].** $\square$
In contrast to results like Theorem \[thm:no-tracial-states\], very little is known about the behaviour of / on s carrying a trace. In the remainder of this paper we suggest a possible new route to tackle this situation.
The following example is mentioned with less detail in [@Sem98].
\[exam:spbdd-mnc\] Let $T\colon M_n({{\mathbb C}})\to M_n({{\mathbb C}})$ be a linear mapping. Then $T$ is unital, surjective and / if and only if there are a Jordan automorphism $S$ of $M_n({{\mathbb C}})$ and a non-zero complex number $\gamma$ such that $$\label{equa:canonical-form}
Tx=\gamma\,Sx+(1-\gamma)\,\tau(x)\qquad\bigl(x\in M_n({{\mathbb C}})\bigr),$$ where $\tau$ denotes the normalised centre-valued trace on $M_n({{\mathbb C}})$.
Evidently, the formula (\[equa:canonical-form\]) defines a unital mapping which is /, since the images under $S$ and $\tau$ commute. Moreover, $T$ is surjective: let $y\in{{M_n({{\mathbb C}})}}$ and $z\in{{M_n({{\mathbb C}})}}$ be such that $Sz=y$. Put $x_1=\smash{\frac1\gamma}(z-\tau(z))$ and $x_2=\tau(z)$. Then $x_1\in\ker\tau$. Setting $x=x_1+x_2$ we have $$\begin{split}
Tx
&= T(x_1+x_2)=\gamma Sx_1+\gamma Sx_2+(1-\gamma)\tau(x_1+x_2)\\
&=Sz-\tau(z)+\gamma\tau(z)+(1-\gamma)\tau(z)\\
&=Sz=y.
\end{split}$$
Conversely, if $T$ is spectrally bounded, it leaves $\ker\tau$ invariant, as this space is spanned by the nilpotent matrices and $T$ preserves nilpotency ([@MMS2], Lemma 3.1). Assuming that $T$ is surjective, it is in fact bijective and hence remains injective when restricted to $\ker\tau$. Since the latter is finite dimensional, ${{{T}_{\kern-.5pt|{\ker\tau}}}}$ is bijective from $\ker\tau$ to $\ker\tau$. By [@BPW], there exist a Jordan automorphism $S$ of ${{M_n({{\mathbb C}})}}$ and a non-zero complex number $\gamma$ such that ${{{T}_{\kern-.5pt|{\ker\tau}}}}=\gamma\,{{{S}_{\kern-.5pt|{\ker\tau}}}}$. Hence, for each $x\in{{M_n({{\mathbb C}})}}$, $$T(x-\tau(x))=\gamma\,S(x-\tau(x))$$ which is nothing but identity (\[equa:canonical-form\]), if $T$ is unital.
Specialising the above description of / operators on matrix algebras to / we recover Aupetit’s result from [@Aup93], Proposition 2, which was proved using holomorphic methods.
\[exam:spisom-mnc\] Every unital / $T$ from ${{M_n({{\mathbb C}})}}$ into itself is a Jordan automorphism.
Since $T\colon{{M_n({{\mathbb C}})}}\to{{M_n({{\mathbb C}})}}$ is injective, by [@MMS1], Proposition 4.2, it is surjective as well. By the above description (\[equa:canonical-form\]), ${{{T}_{\kern-.5pt|{\ker\tau}}}}=\gamma\,{{{S}_{\kern-.5pt|{\ker\tau}}}}$ for a Jordan automorphism $S$ of ${{M_n({{\mathbb C}})}}$ and $\gamma\in{{\mathbb C}}\setminus\{0\}$. Suppose first that $n\geq3$ and take $y\in\ker\tau$ with $\sigma(y)=\{1,-\frac12\}$; for instance, $y=e_{11}-\frac12(e_{22}+e_{33})$, where $e_{ij}$, $1\leq i,j\leq n$ denote the standard matrix units. Then $$\sigma(Ty)=\gamma\,\sigma(Sy)=\gamma\,\sigma(y)
=\{\textstyle{\gamma,-\frac{\gamma}{2}}\}.$$ On the other hand, $T$ preserves the peripheral spectrum, by Proposition 4.7 in [@MMS1]. Therefore $\gamma=1$, and the claim follows from (\[equa:canonical-form\]). In the case $n=2$, note that identity (\[equa:canonical-form\]) entails that $\tau(Tx)=\tau(x)$ for all $x\in{{M_n({{\mathbb C}})}}$. Since $T$ always preserves one eigenvalue (in the peripheral spectrum), it follows that $T$ preserves the entire spectrum and hence must be a Jordan automorphism. The case $n=1$ is trivial.
Let us now state a problem motivated by these examples.
\[prob:spsiom-II1f\] Let $A$ be a 1f/ with normalised centre-valued trace $\tau$. Let $T\colon A\to A$ be a unital surjective /. Are there a Jordan automorphism $S$ of $A$ and a non-zero complex number $\gamma$ such that $$\label{equa:canonical-form-factor}
Tx=\gamma\,Sx+(1-\gamma)\,\tau(x)\qquad\bigl(x\in A\bigr)\,?$$
1. Every Jordan epimorphism $S\colon A\to A$ is either multiplicative or anti-multiplicative, by a classical result due to Herstein, as $A$ is simple, hence prime. Therefore its kernel is an ideal of $A$. Since $A$ is simple, $S$ must be injective.
2. By the remark just made together with the final part of the argument in Example \[exam:spbdd-mnc\] it suffices to find a Jordan epimorphism $S$ on $A$ such that ${{{T}_{\kern-.5pt|{\ker\tau}}}}=\gamma\,{{{S}_{\kern-.5pt|{\ker\tau}}}}$ for some non-zero $\gamma\in{{\mathbb C}}$.
3. The “if”-part in Example \[exam:spbdd-mnc\] remains valid in general, so identity (\[equa:canonical-form-factor\]) would in fact be a *characterization* of unital surjective /s on 1f/s.
4. It is easily seen that the representation (\[equa:canonical-form-factor\]) is unique. Suppose that $$T=\gamma\,Sx+(1-\gamma)\,\tau(x)= \gamma'\,S'x+(1-\gamma')\,\tau(x)\qquad\bigl(x\in A\bigr)$$ are two representations of the / $T$, where $\gamma,\gamma'\in{{\mathbb C}}\setminus\{0\}$ and both $S$, $S'$ are Jordan automorphisms of $A$. It suffices to show that $\gamma=\gamma'$. Since $\gamma\,{{{S}_{\kern-.5pt|{\ker\tau}}}}=\gamma'\,{{{S'}_{\kern-.5pt|{\ker\tau}}}}$ and both $S$ and $S'$ are spectral isometries, it follows that $|\gamma|=|\gamma'|$. Let $p\in A$ be a projection with $\tau(p)=\frac13$ and put $x=p-\tau(p)\in\ker\tau$. Then $\sigma(x)=\{0,1\}-\frac13=\{-\frac13,\frac23\}$. As both $S$ and $S'$ preserve the spectrum, it follows that $$\textstyle{\{-\frac{\gamma}{3},\frac{2\gamma}{3}\}}
=\gamma\,\sigma(x)=\gamma'\,\sigma(x)
=\textstyle{\{-\frac{\gamma'}{3},\frac{2\gamma'}{3}\}}$$ and thus $\gamma=\gamma'$.
Adapting the argument in Example \[exam:spisom-mnc\] to the 1f/ situation, we can infer the following result from a positive answer to Problem \[prob:spsiom-II1f\].
\[cor:type-twoone\] Every unital surjective / $T$ from a 1f/ $A$ onto itself is a Jordan automorphism.
Suppose the / $T\colon A\to A$ can be written in the form (\[equa:canonical-form-factor\]). Then ${{{T}_{\kern-.5pt|{\ker\tau}}}}=\gamma\,{{{S}_{\kern-.5pt|{\ker\tau}}}}$ for some non-zero $\gamma\in{{\mathbb C}}$. Since $r(x)=r(Tx)=|\gamma|\,r(Sx)=|\gamma|\,r(x)$ for all $x\in\ker\tau$, it follows that $|\gamma|=1$. Let $p\in A$ be a projection with $\tau(p)=\frac13$. Then $\sigma(p-\tau(p))=\{0,1\}-\frac13=\{-\frac13,\frac23\}$. As in Example \[exam:spisom-mnc\], $\sigma(T(p-\tau(p)))=\{-\frac{\gamma}{3},\frac{2\gamma}{3}\}$. Since $T$ preserves the peripheral spectrum, we conclude that $\gamma=1$.
We know that Problem \[prob:spiom-conjecture\] has a positive answer whenever the domain is a factor of type different from [II]{}$_1$.
[30]{}
, The uniqueness of the complete norm topology in Banach algebras and Banach Jordan algebras, [*J. Funct. Anal.*]{} [**47**]{} (1982), 1–6.
, Spectral characterization of the radical in Banach and Jordan–Banach algebras, [*Math. Proc. Camb. Phil. Soc.*]{} [**114**]{} (1993), 31–35.
, Spectrum-preserving linear mappings between Banach algebras or Jordan–Banach algebras, [*J. London Math. Soc.*]{} [**62**]{} (2000), 917–924.
, The continuity of Lie homomorphisms, [*Studia Math.*]{} [**138**]{} (2000), 193–199.
, Linear maps preserving the essential spectral radius, [*Linear Algebra Appl.*]{} [**428**]{} (2008), 1041–1045.
, Linear transformations that preserve the nilpotent matrices, [*Pacific J. Math.*]{} [**104**]{} (1983), 39–46.
, Derivations mapping into the radical, III, [*J. Funct. Anal.*]{} [**133**]{} (1995), 21–29.
, Numerical radius preserving operators on s, [*Arch. Math.*]{} [**70**]{} (1998), 486–488.
, The spectrally bounded linear maps on operator algebras, [*Studia Math.*]{} [**150**]{} (2002), 261–271.
, Banach algebras and automatic continuity, LMS monographs **24**, Claredon Press, Oxford, 2000.
, Invertibility preserving linear maps of Banach algebras, [*Contemp. Math.*]{} [**364**]{} (2004), 59–66.
, Isometries of operator algebras, [*Annals of Math.*]{} [**54**]{} (1951), 325–338.
, A generalized Schwarz inequality and algebraic invariants for operator algebras, [*Annals of Math.*]{} [**56**]{} (1952), 494–503.
, Algebraic and analytic aspects of operator algebras, CBMS Series Vol. [**1**]{}, Amer. Math. Soc., Providence, RI, 1970.
, Jordan isomorphism of purely infinite s, [*Quart. J. Math.*]{} [**58**]{} (2007), 249–253.
, Where to find the image of a derivation, [*Banach Center Publ.*]{} [**30**]{} (1994), 237–249.
, Spectrally bounded operators on simple s, [*Proc. Amer. Math. Soc.*]{} [**132**]{} (2004), 443–446.
, Spectrally bounded operators on simple s, II, [*Irish Math. Soc. Bull.*]{} [**54**]{} (2004), 33–40.
, Towards a non-selfadjoint version of Kadison’s theorem, in: Proc. Conf. in honour of the 125th birthday of L. Fejér and F. Riesz, Eger, Hungary, June 2005; [*Ann. Math. Inf.*]{} [**32**]{} (2005), 87–94.
, Spectral isometries, II, [*Contemp. Math.*]{} [**435**]{} (2007), 301–309.
, Derivations mapping into the radical, II, [*Bull. London Math. Soc.*]{} [**24**]{} (1992), 485–487.
, First results on spectrally bounded operators, [*Studia Math.*]{} [**152**]{} (2002), 187–199.
, Spectrally bounded operators from von Neumann algebras, [*J. Operator Theory*]{} [**49**]{} (2003), 285–293.
, Hereditary properties of spectral isometries, [*Arch. Math.*]{} [**82**]{} (2004), 222–229.
, A short proof of Johnson’s uniqueness-of-norm theorem, [*Bull. London Math. Soc.*]{} [**21**]{} (1989), 487–488.
, Spectrally bounded operators on Banach algebras, PhD thesis, Queen’s University Belfast, 2001.
, Spectrally bounded linear maps on $B(H)$, [*Quart. J. Math.*]{} [**49**]{} (1998), 87–92.
, The image of a derivation is contained in the radical, [*Annals of Math.*]{} [**128**]{} (1988), 435–460.
[^1]: This paper is an expanded version of a talk given at the conference *Jordan Structures: Nonassociative Analysis and Geometry* on 6 September 2008 at Queen Mary College, London.
| {
"pile_set_name": "ArXiv"
} |
**Hölder Inequalities and QCD Sum-Rule Bounds on the**
**Masses of Light Quarks**
T.G. Steele[^1]
*Department of Physics & Engineering Physics, University of Saskatchewan, 116 Science Place*
*Saskatoon, Saskatchewan, S7N 5E2, Canada*
In this paper we briefly review the development of Hölder inequalities for QCD sum-rules[@sr_holder] and their application to obtain light-quark ($u,d$) mass bounds.[@holder_bounds]
Laplace sum-rules for pseudoscalar currents with quantum numbers of the pion relate a QCD prediction $R_5\left(M^2\right)$ to the integral of the associated hadronic spectral function $\rho_5(t)$ $$R_5\left(M^2\right)=\frac{1}{\pi}\int\limits_{t_0}^\infty \rho(t) \exp{\left(-\frac{t}{M^2}\right)}\,dt \quad ,
\label{basic_sr}$$ where $t_0$ is the physical threshold for the spectral function. Since $\rho_5(t)\ge 0$, the right-hand (phenomenological side) side of (\[basic\_sr\]) must satisfy integral inequalities over a measure $d\mu=\rho_5(t)\,dt$.
Hölder’s inequality over a measure $d\mu$ is $$\biggl|\int_{t_1}^{t_2} f(t)g(t) d\mu \biggr|\! \le \!
\left(\int_{t_1}^{t_2} \big|f(t)\big|^ p d\mu \right)^{\frac{1}{p}}
\!\!\!\left(\int_{t_1}^{t_2} \big|g(t)\big|^q d\mu \right)^{\frac{1}{q}}
~,~
\frac{1}{p}+\frac{1}{q} =1~;~ p,~q\ge 1 \quad ,
\label{holder_ineq}$$ which for $p=q=2$ reduces to the familiar Schwarz inequality, implying that the Hölder inequality is a more general constraint. The Hölder inequality can be applied to Laplace sum-rules by identifying $d\mu=\rho(t)\,dt$, $\tau=1/M^2$ and defining $$S_5\left(\tau\right)=\frac{1}{\pi}\int\limits_{\mu_{th}}^\infty \!\!\rho_5(t) e^{-t\tau}\,dt
\label{s5}$$ where $\mu_{th}$ will later be identified as lying above $m_\pi^2$. Suitable choices of $f(t)$ and $g(t)$ in the Hölder inequality (\[holder\_ineq\]) yield the following inequality for $S_5(t)$:[@sr_holder] $$S_5\left(\tau+(1-\omega)\delta\tau\right)\le S_5^\omega\left(\tau\right)
S_5^{1-\omega}\left(\tau+\delta\tau\right)
\quad ,~ \forall~ 0\le \omega\le 1\quad .
\label{s5_ineq}$$
=7.8truein
The Laplace sum-rule relating QCD and hadronic physics is obtained by applying the Borel transform operator[@SVZ] $\hat B$ to the dispersion relation $$\Pi_{5}\left(Q^2\right)=a+b Q^2
+\frac{Q^4}{\pi}\int\limits_{t_0}^\infty
\frac{ \rho_5(t)}{t^2\left(t+Q^2\right)}\, dt\quad .
\label{dispersion}$$ The quantity $\Pi_5\left(Q^2\right)$ is the QCD prediction for the correlation function of pseudoscalar (pion) currents $$\begin{gathered}
\Pi_{5}\left(Q^2\right)=i\int d^4x\, e^{iq\cdot x}\left\langle O \vert T\left[ J_{5}(x)
J_5(0)\right] \vert O \right\rangle
\label{corr_fn}\\
J_5(x)=\frac{1}{\sqrt{2}}\left(m_u+m_d\right)\left[\bar u(x)i\gamma_5u(x)-
\bar d(x)i\gamma_5d(x)\right]\quad ,
\label{current}\end{gathered}$$ and the theoretically-determined quantity $R_5\left(M^2\right)$ is obtained from the Borel transform of the QCD correlation function. $$R_{5}\left(M^2\right)=M^2\hat B\left[\Pi_{5}\left(Q^2\right)\right]$$
Perturbative contributions to $R_5\left(M^2\right)$ are known up to four-loop order.[@chetyrkin; @BNRY] Infinite correlation-length vacuum effects in $R_5\left(M^2\right)$ are represented by the (non-perturbative) QCD condensate contributions.[@SVZ; @BNRY; @Bagan] In addition to the QCD condensate contributions the pseudoscalar (and scalar) correlation functions are sensitive to finite correlation-length vacuum effects described by direct instantons in the instanton liquid model.[@EVS] The total result for $R_5\left(M^2\right)$ to leading order in the light-quark masses is[@holder_bounds] $$\begin{split}
{ R}_5\left(M^2\right)=&\frac{3m^2M^4}{8\pi^2}\left(
1+4.821098 \frac{\alpha}{\pi}+21.97646\left(\frac{\alpha}{\pi}\right)^2+53.14179\left(\frac{\alpha}{\pi}\right)^3
\right)
\\
& +m^2\left(
-\langle m\bar q q\rangle
+\frac{1}{8\pi}\langle \alpha G^2\rangle
+\frac{\pi\langle{\cal O}_6\rangle}{4M^2}
\right)
\\
& +m^2
{3\rho_c^2 M^6\over{8 \pi^2}} e^{-\rho_c^2M^2/2 }
\left[
K_0\left( {\rho_c^2M^2/2} \right) +
K_1\left( {\rho_c^2M^2/2} \right)
\right]\quad ,
\end{split}
\label{R_5}$$ where $\alpha$ and $m=\left(m_u+m_d\right)/2$ are the $\overline{MS}$ running coupling and quark masses at the scale $M$, and $\rho_c$ represents the instanton size in the instanton liquid model. Note that all the theoretical contributions are proportional to $m^2$, demonstrating that the quark mass sets the scale of the pseudoscalar channel. Higher-loop perturbative contributions in (\[R\_5\]) are significant, and effectively enhance the quark mass with increasing loop order.
To employ the Hölder inequality (\[s5\_ineq\]) we separate out the pion pole by setting $\mu_{th}=9m_\pi^2$ in (\[s5\]). $$S_5\left(M^2\right)=R_5\left(M^2\right)-2f_\pi^2m_\pi^4=\int\limits_{9m_\pi^2}^\infty \rho_5(t)
e^{-t\tau}\,dt
\label{s5_fin}$$ Lower bounds on the quark mass $m$ can then be obtained by finding the minimum value of $m$ for which the Hölder inequality (\[s5\_ineq\]) is satisfied. Introducing further phenomenological contributions ([*e.g.*]{} three-pion continuum) would tend to give a larger mass bound. However, if only the pion pole is separated out, then the analysis is not subject to uncertainties introduced by the phenomenological model.
Standard values of the QCD parameters are employed in the inequality analysis of (\[s5\_ineq\]), and we use $\delta\tau\lesssim 0.1\,{\rm GeV^{-2}}$ for which this analysis becomes local (depending only on the Borel scale $M$).[@sr_holder; @holder_bounds] Validity of QCD predictions at the $\tau$ mass is evidenced by the analysis of the $\tau$ hadronic width, hadronic contributions to $\alpha_{EM}\left(M_Z\right)$ and the muon anomalous magnetic moment,[@braaten_davier] so we impose the inequality (\[s5\_ineq\]) at the $\tau$ mass $M=M_\tau$. The resulting Hölder inequality bound on the $\overline{MS}$ quark masses scaled to $1.0\,{\rm GeV}$ is[@holder_bounds] $$m(1\,{\rm GeV})=\frac{1}{2}\left[ m_u(1\,{\rm GeV})+m_d(1\,{\rm GeV})\right]\ge 3\,{\rm MeV}
\label{final_bound}$$ For comparison with other determinations of the light quark masses, this result has been converted to $$m(2\,{\rm GeV})=\frac{1}{2}\left[ m_u(2\,{\rm GeV})+m_d(2\,{\rm GeV})\right]\ge 2.1\,{\rm MeV}
\label{final_pdg_bound}$$ by the Particle Data Group.[@PDG]
The theoretical uncertainties in the quark mass bound (\[final\_bound\]) from the QCD parameters and (estimated) higher-order perturbative effects are less than 5%, and the result (\[final\_bound\]) is the absolute lowest bound resulting from the uncertainty analysis.[@holder_bounds] for $M\gtrsim M_\tau$ the theoretical uncertainties in the mass bound are $\lesssim 0.1 \, {\rm MeV}$. Thus we have not extracted mass bounds below the energy scale $M\approx M_\tau$ at which theoretical uncertainties first reach a non-negligible level. Finally, compared with the positivity inequality $S_5\left(M^2\right)\ge 0$ (as first used to obtain quark mass bounds from QCD sum-rules[@BNRY]) the Hölder inequality leads to quark mass bounds larger by a factor of 2 for identical theoretical and phenomenological inputs at $M=M_\tau$, demonstrating that the Hölder inequality provides stringent constraints on the quark mass.
[References]{}
[99]{}
M. Benmerrouche, G. Orlandini, T.G. Steele, [Phys. Lett. ]{} [B356]{} (1995) 573.
T.G. Steele, K. Kostuik, J. Kwan, [Phys. Lett. ]{} B451 (1999) 201.
M.A. Shifman, A.I. Vainshtein, V.I. Zakharov, [Nucl. Phys. ]{} [B147]{} (1979) 385.
K.G. Chetyrkin, [Phys. Lett. ]{} [ B390]{} (1997) 309.
C. Becchi, S. Narison, E. de Rafael, F.J. Yndurain, [Z. Phys. ]{}[ C8]{} (1981) [335]{}.
E. Bagan, J.I. Latorre, P. Pascual, [Z. Phys. ]{} [C32]{} (1986) 43.
E.V. Shuryak, [Nucl. Phys. ]{} [B214]{} (1983) [237]{}.
E. Braaten, S. Narison, A. Pich, [Nucl. Phys. ]{} [ B373]{} (1992) 581; M. Davier, A. Höcker, [Phys. Lett. ]{} [B435]{} (1998) 427.
D.E. Groom [*et al.*]{} (Particle Data Group), [Eur. Phys. Jour. ]{} C15 (2000) 1.
[^1]: Research funded by the Natural Science and Engineering Research Council of Canada (NSERC).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the Fermi surface topological transition of the pocket-opening type in a two dimensional Fermi liquid with spin-orbit coupling of Rashba type. We find that the interactions, far from instabilities, drive the transition first order at zero temperature surprisingly in a more pronounced way than in the case of interacting Fermi liquid without spin-orbit coupling. We first gain insight from second order perturbation theory in the self-energy. We then extend the results to stronger interaction, using the self-consistent fluctuation approximation. We discuss existing experimental work on the system BiTeI and suggest further experiments in the light of these results .'
author:
- 'Yury Sherkunov and Joseph J. Betouras'
bibliography:
- 'ybs-2.bib'
title: 'First order Fermi surface topological (Lifshitz) transition in an interacting two-dimensional Rashba Fermi liquid'
---
Introduction
============
The Lifshitz transition [@Lifshitz60] (LT) is a Fermi surface topological transition as a result of the change of the Fermi energy and/or band structure. The possibilities that have been usually considered are either a change of the number of Fermi surfaces in a pocket opening or closing transition, or the connection of two parts of Fermi surface in a neck-opening or closing transition. Recently higher order Fermi surface topological or multicritical transitions have been put forward in order to explain unusual properties of correlated materials [@Efremov18; @Noah19]. The LT can be induced by the variation of an external parameter such as pressure, doping or magnetic field and has been experimentally observed in many systems such as heavy fermions [@Daou06; @Bercx12; @AokiCelrln16], iron-based superconductors [@Liu10; @Xu13; @Khan14], cuprate-based high-temperature superconductors [@Benhabib15; @Wu18; @Bragan18], or other strongly correlated electrons system such as layered material $Na_xCoO_2$ [@Okamoto10; @Slizovskiy15]. It has been also shown to be responsible for the re-entrant superconductivity [@Sherkunov18] in uranium-based ferromagnetic superconductors such as $URhGe$ [@HuxleyLifshitz11]. The interaction-driven LT has also been proposed and observed in ultracold fermionic systems [@Wang12; @vanLoon16; @Quintanilla09].
Of particular interest is the LT in systems with strong spin-orbit coupling (SOC), such as the material BiTeI [@Ye15; @Hamlin14] which has shown quantum magnetotransport signature of a change in the Fermi surface topology, very relevant to the present work. Two other classes of systems that can exhibit the same physics are ultracold gases [@Wang12] and two-dimensional electronic systems that can be formed at the interface between insulating oxides, such as $LaAlO_3$ and $SrTiO_3$ [@Joshua12]. The former represents an important tool in the search for Majorana fermions while the latter exhibits a range of interesting phenomena such as ferromagnetism, superconductivity and unique magnetotransport properties, for which a universal LT is responsible [@Joshua12; @Yin19].
Theoretically, it has been shown that interactions, in the region of paramagnetic fluctuations, can play an important part in the LT in a two-dimensional Fermi liquid changing the order of the transition from the second for non-interacting systems to the first for interacting Fermi liquids in a pocket-opening transition [@Slizovskiy14]. Here we investigate how interactions, outside any phase formation, affect the LT in a two-dimensional Fermi liquid with SOC. In particular, we consider an LT occurring when the bottom of $s=-1$ Rashba subband crosses the level of chemical potential as a result of variations in doping, as shown in Fig. \[Fig1\]. We show that, similarly to [@Slizovskiy14], at zero temperature the LT changes its order from the second to the first due to strong paramagnetic fluctuations manifesting themselves as the non-analyticity of the self-energy in the Fermi momenta, $p_F$, calculated from the bottom of the $s=-1$ Rashba subband (see Fig. \[Fig1\]). As we show, using the second order of perturbation theory, the self-energy correction takes the form $\Sigma_2\propto p_F \log p_F$ in the presence of SOC while in the absence of it the correction is $\Sigma_2\propto p_F^2 \log p_F$ [@Slizovskiy14]. This fact results in an even stronger first order character of the transition in the two-dimensional FL in the presence of Rashba type SOC compared to the absence of it.
In the next sections, we introduce the model and then we gain insight by working in second order perturbation theory. This is followed by summation of diagrams in random phase approximation (RPA) and finally we discuss the wider implications and propose possible experimental directions.
Model
=====
We consider a 2D Fermi liquid with strong Rashba-type SOC described by the Hamiltonian: $$\begin{aligned}
H=H_{0}+H_{int}.\end{aligned}$$ The noninteracting part $H_0$ $$\begin{aligned}
H_0=\Sigma_p b^{\dagger}_p[(\frac{p^2}{2m}-\mu)\sigma_0+\lambda (\sigma_x p_y-\sigma_yp_x)]b_p, \label{H0}\end{aligned}$$ where $b_p=(b_{p\uparrow},b_{p\downarrow})^T$ is the annihilation operators of a particle, $\mu$ is the chemical potential, $\sigma_0$ is the $2\times 2$ unit matrix, $\sigma_x$ and $\sigma_y$ are Pauli matrices, and $\lambda$ is the SOC constant, can be diagonalized in the helicity basis $$\begin{aligned}
|\mathbf p,s\rangle=\frac{1}{\sqrt 2}(1,ise^{-i\phi(\mathbf p)})^T,\label{basis}\end{aligned}$$ where $\phi(\mathbf p)=\arctan(p_y/p_x)$ is the angle between the $x-$ axis and momentum $\mathbf p$, and $s=\pm 1$ is the helicity. The dispersion relations for the two helical branches $$\begin{aligned}
\epsilon_s(p)=\frac{p^2}{2m}+\lambda s|\mathbf p|\label{disp}\end{aligned}$$ is shown in Fig. \[Fig1\]. In this work, we consider the case $\mu<0$, where the Fermi surfaces are in the branch with $s=-1$. The most interesting situation will correspond to $\mu\rightarrow \mu_0=-m\lambda^2/2$, where the density of states is divergent.
![Dispersion relation for two helical branches ($s=\pm 1$) of a non-interacting system with Rashba SOC. []{data-label="Fig1"}](Fig1.pdf){width="40.00000%"}
We take into account short-ranged interactions of Hubbard type in the system, which read: $$\begin{aligned}
H_{int}=U\sum_{\sigma\neq\sigma'}\sum_{\mathbf p,\mathbf k, \mathbf q}b^{\dagger}_{k+q,\sigma}b^{\dagger}_{p-q,\sigma'}b_{p,\sigma'}b_{k,\sigma},\end{aligned}$$ where $\sigma,\sigma'=\{\uparrow,\downarrow\}$ are the spin indices. In the helicity basis Eq. (\[basis\]), $H_{int}$ takes the form: $$\begin{aligned}
H_{int}&=&\sum_{s,s'r,r'}\sum_{\mathbf p,\mathbf k, \mathbf q}U_{r,s}^{r',s'}(\mathbf p,\mathbf k,\mathbf q)a^{\dagger}_{k+q,s'}a^{\dagger}_{p-q,r'}a_{p,r}a_{k,s},\label{Hint}\\
& &U_{r,s}^{r',s'}(\mathbf p,\mathbf k,\mathbf q)=\frac{U}{4}\left(r r' e^{i[\phi(\mathbf p)-\phi(\mathbf p-\mathbf q)]}\right.\nonumber\\
&+&\left. s s' e^{i[\phi(\mathbf k)-\phi(\mathbf k+\mathbf q)]}\right),\label{U}\end{aligned}$$ where $a_{p,s}$ is the annihilation operator of a particle in the state $|\mathbf p,s\rangle$.
Perturbation theory
===================
We start by analysing the lowest orders of perturbation theory valid for small $U$. Using Matsubara Green’s functions, the self-energy correction at first order perturbation theory and at zero temperature is given by the Hartree diagram (the Fock contribution is zero), which leads to $\Sigma_1(p_F,i\omega=0)=Un/2$, where $n$ is the electron density. This contribution is absorbed into $\mu$. The second order contribution to the self-energy is $$\begin{aligned}
& &\Sigma_2(p_F,i\omega=0,s=-1)=\frac{U^2}{2}\int_q\sum_{s'}\{\chi^0_{00}(q)\nonumber\\
&\times& [1+ss'\cos[\phi(\mathbf p)-\phi(\mathbf p-\mathbf q)]]\nonumber\\
&+&\chi^0_{zz}(q)[1-ss'\cos[\phi(\mathbf p)-\phi(\mathbf p-\mathbf q)]]\}g^0_{s'}(p-q),\label{SOPT}\end{aligned}$$ where $q=\{\mathbf q,i\omega\}$ is a four-dimensional momentum, $\int_q\equiv\int \frac{d^2qd\omega}{(2\pi)^3}$, $g^0_{s}(p)=(i\omega-\epsilon_s(p)+\mu)^{-1}$ is the electron Green’s function in the helicity basis. The charge, $\chi^0_{00}$ and spin, $\chi^0_{zz}$ susceptibilities of non-interacting particles are defined as the Fourier transform of $$\begin{aligned}
\chi^0_{ij}(\mathbf r,\mathbf r')=\nonumber\\
-\int_0^\infty d\tau\langle T_\tau\psi^{\dagger}(\mathbf r,\tau)\sigma_i\psi(\mathbf r,\tau)\psi^{\dagger}(\mathbf r',0)\sigma_j\psi(\mathbf r',0)\rangle,\label{suscepdef}
\end{aligned}$$ with $\psi(\mathbf r,\tau)=\sum_{\mathbf p}b_pe^{i\mathbf p\cdot \mathbf r}$, which can be calculated as [@MatiMaslovPRB15] $$\begin{aligned}
\chi^0_{ij}(q)=-Tr\int_p\sigma_iG(p)\sigma_jG(p+q)=\nonumber\\
-\frac{1}{2}\int_pg_r(p)g_s(p+q)F^{ij}_{sr}(\mathbf p+\mathbf q,\mathbf p),\label{Suscept}
\end{aligned}$$ where $G(p)=\sum_s\Omega_s(\mathbf p)g^0_s(p)$ is the Green’s function in the spin basis, with $\Omega_s(\mathbf p)=\frac{1}{2}[\sigma_0+s(\sigma_x\sin\phi(\mathbf p)-\sigma_y\cos\phi(\mathbf p))]$ and $F^{ij}_{sr}(\mathbf p,\mathbf k)=2\langle p,s|\sigma_i|k,r\rangle\langle k,r|\sigma_j|p,s\rangle$ is the overlap factor. In the vicinity of the bottom of $s=-1$ band, where $\mu\rightarrow\mu_0=-m\lambda^2/2$ and at large transferred momenta $|\mathbf q|\gg m\lambda\gg|p_{F1,2}-m\lambda|$, the susceptibilities can be estimated as $$\begin{aligned}
\chi^0_{00}(q)=\chi^0_{zz}(q)=\frac{m}{\pi}\frac{q^2(p_{F2}^2-p_{F1}^2)}{q^4+4m^2\omega^2},\label{chi2order}
\end{aligned}$$ which with logarithmic accuracy in the vicinity $p_{F1}\rightarrow m\lambda$ and $p_{F2}\rightarrow m\lambda$ yields $$\begin{aligned}
\Sigma_2(p_{F1,2},i\omega=0,s=-1)\approx -\frac{u^2\lambda |P_{1,2}|}{2\pi^2}\log\left|\frac{\Lambda}{ P_{1,2}}\right|,\label{Sig2}
\end{aligned}$$ where $u$ is the Hubbard coupling constant in units of $m\lambda^2$, $P_1=m\lambda-p_{F1}$ and $P_2=p_{F2}-m\lambda$ are the Fermi momenta calculated from the bottom of the $s=-1$ band, and $\Lambda$ is the short wave-length cutoff. We checked the validity of Eq. (\[Sig2\]) by fitting the result of numerical calculations (see Fig.\[Fig1p\]).
Eq. (\[Sig2\]) is similar to the self energy obtained for the system without SOC [@Slizovskiy14], however, the main important difference is that $\Sigma$ in Eq. (\[Sig2\]) is proportional to the first power of the Fermi momenta $P_{1,2}$, while for the systems without SOC, it is proportional to the second power of the Fermi momentum.
![Fitting of the self-energy calculated numerically in the second order perturbation theory by $\Sigma_2(P_{1,2}\rightarrow 0)/(|P_{1,2}|u^2\lambda)=a+b\log|\Lambda/P_{1,2}|$, where $\Lambda=10^2m\lambda$, $a=2.1\times 10^{-2}$ and $b=9.1\times 10^{-2}$ in the vicinity of the bottom of the $s=-1$ band []{data-label="Fig1p"}](Fig2.pdf){width="40.00000%"}
For the equation of the chemical potential $\mu\sim\mu_0<0$ we find: $$\begin{aligned}
\mu=\epsilon_{-1}(p_{F1,2})+\Sigma_2(p_{F1,2},i\omega=0,s=-1)\approx \mu_0\nonumber\\
+\frac{P^2_{1,2}}{2m}-\frac{u^2\lambda |P_{1,2}|}{2\pi^2}\log\left|\frac{\Lambda}{ P_{1,2}}\right|,\label{mueq}\end{aligned}$$ thus, for small $|P_{1,2}|$ the third term becomes large compared to the second one giving rise to a first-order LT. Note that because the third term is proportional to $|P_{12}|$ and the second one is quadratic in $|P_{12}|$, the effect could be more pronounced than in the case of no SOC, where both terms are quadratic in the Fermi momentum [@Slizovskiy14].
Random phase approximation
==========================
We now consider larger interaction strengths, but below any Stoner-like instability. Then the effective interaction can be obtained by summing up ring and ladder diagrams [@Slizovskiy14], leading to $$\begin{aligned}
\Sigma(p_F,i\omega&=&0,s=-1)=\Sigma_{ring}+\Sigma_{lad},\label{SigRPA}\\
\Sigma_{ring}&=&\frac{u^2}{2}\int_q\sum_r\left[\frac{\chi^0_{00}(q)}{1+u\chi^0_{00}(q)}F_{sr}^{00}(\mathbf p-\mathbf q,\mathbf p)\right.\nonumber\\
&+&\left.\frac{\chi^0_{zz}(q)}{1-u\chi^0_{zz}(q)}F_{sr}^{zz}(\mathbf p-\mathbf q,\mathbf p)\right]g_r(p-q),\label{ring}\\
\Sigma_{lad}&=&\frac{u^3}{2}\int_q\sum_r\left[\frac{[\chi^0_{0y}(q)]^2}{1-u\chi^0_{yy}(q)}F_{sr}^{00}(\mathbf p-\mathbf q,\mathbf p)\right.\nonumber\\
&-&\left.\frac{[\chi^0_{xz}(q)]^2}{1-u\chi^0_{xx}(q)}F_{sr}^{zz}(\mathbf p-\mathbf q,\mathbf p)\right]g_r(p-q).\label{lad}\end{aligned}$$ Here we put $m=1$ and $\lambda=1$, so all energies and momenta are measured in the unites of $m\lambda^2$ and $m\lambda$. The spin-charge susceptibility matrix is given by Eq. (\[Suscept\]), with the overlap matrix elements given by $$\begin{aligned}
F^{00}&=&1+rs\cos[\phi(\mathbf p)-\phi(\mathbf p+\mathbf q)],\nonumber\\
F^{zz}&=&1-rs\cos[\phi(\mathbf p)-\phi(\mathbf p+\mathbf q)],\nonumber\\
F^{0y}&=&-(r\cos[\phi(\mathbf p)]+s\cos[\phi(\mathbf p+\mathbf q)]),\nonumber\\
F^{yy}&=&1+rs\cos[\phi(\mathbf p)+\phi(\mathbf p+\mathbf q)],\nonumber\\
F^{xx}&=&1-rs\cos[\phi(\mathbf p)+\phi(\mathbf p+\mathbf q)],\nonumber\\
F^{xz}&=&-i(r\cos[\phi(\mathbf p)]-s\cos[\phi(\mathbf p+\mathbf q)]).\end{aligned}$$ The results of numerical calculations of $\Sigma$ and $\mu$ using eq. (\[SigRPA\]) are presented in Fig.\[Fig2\].
![ $\Sigma$ as a function of $p_{F1}$ (left) for $u=0.1$(orange), $u=0.5$ (blue), $u=1$ (red), $u=1.5$ (green), and $u=2$ (black) and $\mu$ as a function of $p_{F1,2}$ (right) for $u=0.5$ (blue), $u=1$ (red), $u=1.5$ (green), and $u=2$ (black). The dashed curve represents the chemical potential of non-interacting Fermi gas. []{data-label="Fig2"}](Fig3.pdf){width="45.00000%"}
A 2D Fermi liquid with Rashba SOC can support four collective modes: one plasmon mode and three chiral spin modes manifested by poles of charge and spin susceptibilities, which coincides with the solutions of the following equations [@MatiMaslovPRB15]: $$\begin{aligned}
1+u\chi^0_{00}=0 \label{plasmons}\end{aligned}$$ for plasmons and $$\begin{aligned}
1-u\chi^0_{jj}=0,\;\;j=x,y,z, \label{snins}\end{aligned}$$ for spin collective modes, giving rise to instabilities in Eq. (\[ring\]) and (\[lad\]). However, the charge plasmon instability manifests itself only for $\mu>0$ [@MatiMaslovPRB15]. Indeed, direct calculation of $\chi^0_{00}$ at $q=0$ and $\omega=0$ for $\mu<0$ leads to $\chi^0_{00}(\mathbf q=0,i\omega=0)=\frac{1}{2\pi}\frac{p_{F1}+p_{F2}}{\sqrt{1+2\mu}}>0$, and, thus, Eq. (\[plasmons\]) has no real solutions. Note, that the divergence of $\chi^0_{00}$ at the bottom of the $s=-1$ band is due to the divergence of the density of states at this point. The spin susceptibilities $\chi^{0}_{jj}$ at $\mathbf q=0$ are given by[@MatiMaslovPRB15] $\chi^0_{zz}(\mathbf q=0,i\omega)=\frac{1}{4\pi}[(p_{F2}-p_{F1})/2+\omega(\arctan\frac{2p_{F1}}{\omega}-\arctan\frac{2p_{F2}}{\omega})]$ and $\chi^0_{xx}(\mathbf q=0,i\omega)=\chi^0_{yy}(\mathbf q=0,i\omega)=\chi^0_{zz}(\mathbf q=0,i\omega)/2$, which take their maximum values $\chi_{zz}=1/\pi$ and $\chi_{xx}=\chi_{yy}=1/2\pi$ at $\omega=0$ and $p_{F1}=0$, leading to chiral-spin instabilities in Eqs. (\[ring\]) and (\[lad\]) at $u_c=\pi$. Thus, in our analysis we consider $u<u_c$.
For $\mu<\mu_0=-m\lambda^2/2$, the $s=-1$ band of non-interacting fermions is empty, however, in the presence of interactions, the effective energy of fermions bents down leading to opening a pocket. In this case, Eq. (\[mueq\]) has four solutions for $p_F$.
In Fig. \[Fig3\] the different values self-energy corrections at the two values of the Fermi momentum $p_F$ is demonstrated, for completeness.
![ The self-energy $\Sigma$ as a function of $p_F$, for $u=1$. []{data-label="Fig3"}](Fig3b.pdf){width="40.00000%"}
In order to estimate which solutions are stable, we estimate the potential $\Omega$ integrating $d\Omega=-nd\mu$ from the point where the phases with the pocket and no pocket merge. The density of states are given by the Luttinger theorem, $n=\frac{1}{2\pi}(p_{F2}^2-p_{F1}^2)$, which is respected. The results suggesting that in a Fermi liquid with SOC one can expect a first order phase transition in the vicinity of the bottom of the $s=-1$ band are shown in Fig. \[Fig4\].
![Top: electron density $n$ as a function of chemical potential for $u=1$. The red line indicates the states with no pocket for $\mu<\mu_0$, the dashed line corresponds to unstable solutions of Eq. (\[mueq\]), with arrow showing the phase transition. Middle: potential $\Delta\Omega$ as a function of chemical potential. The stable solutions correspond to negative $\Delta\Omega$. Bottom: Size of the pocket. The dashed curves correspond to unstable solutions. []{data-label="Fig4"}](Fig4.pdf){width="40.00000%"}
Discussion
==========
We have showed that the LT of a pocket appearing type in 2D fermions with SOC and in the presence of interactions is discontinuous. This first order transition is more pronounced than in the case without SOC due to the different form of the kinetic energy which acquires a linear in momentum term. The reason of the first order behavior is the competition between kinetic and self-energy contribution to the energy.
This knowledge is needed to disentangle the contribution of different processes that may occur simultaneously and, therefore, to explain behavior that may be attributed to quantum criticality. It is also crucial to realise that unconventional behavior such as temperature dependence of specific heat $C \propto T ln(1/T)$ can be also realised in systems with a LT of pocket appearing/disappearing type [@Slizovskiy15].
A prime candidate to show this behavior is the giant Rashba semiconductor BiTeI which exhibits a change in the Fermi surface topology in quantum magnetotransport experiments, upon systematic tuning of the Fermi level E$_F$ [@Ye15]. BiTeI has a conduction band bottom that is split into two sub-bands due to the strong Rashba coupling, resulting in a Dirac point. A marked increase (or decrease) in electrical resistivity is observed when E$_F$ is tuned above (or below) this Dirac node, beyond the quantum limit. The origin of this behavior has been shown convincingly to be the Fermi surface topology and essentially it reflects the electron distribution on low-index Landau levels. The finite bulk $k_z$ dispersion along the $c$ axis and strong Rashba spin-orbit coupling strength of the system enable this measurement. The Dirac node is independently identified by Shubnikov-de Haas oscillations as a vanishing Fermi surface cross section at $k_z=0$. Further measurements of Shubnikov-de Haas oscillations on BiTeI under applied pressures supported the same physics. One high frequency oscillation at all pressures and one low frequency oscillation that emerges between $0.3$ and $0.7$ GPa has been observed [@Hamlin14], indicating the appearance of a second small Fermi surface. The suggested explanation is that the chemical potential starts below the Dirac point in the conduction band at ambient pressure and crosses it as pressure is increased. As a result, the pressure brings the system closer to the predicted topological quantum phase transition.
Regarding experimental observation, there are two promising routes. One is related to recent advances in creating complex oxide heterostructures, with interfaces formed between two different transition metal oxides [@Joshua12; @Sulpizio14], which enables the investigation of new physical phenomena in experimentally controlled systems. There is a universal LT demonstrated already in the prototypical LaAlO$_3$/SrTiO$_3$ interface [@Joshua12], between d-orbitals at the core of the observed transport phenomena in this system. At the LT and the critical electronic density, the transport switches from single to multiple carriers. Although the order of the LT it was beyond the scope of the experiment, there was observed a hysteretic behavior as a function of the applied magnetic field near its value at the metamagnetic transition [@Ilani19]. As a result, more measurements to clarify our predictions are necessary. The second experimental direction is the ultracold atoms where 2D SO coupled atoms in optical lattices can be realised [@Grusdt17].
In a recent work [@Miserev19] a magnetic first order transition has been reported in monolayers of transition metal dichalcogenides with strong SOC, as a result of exchange intervalley scattering. The presence of this interaction would turn our system ferromagnetic. Finally, it is worth mentioning that there has been growing interest in the effects of SOC on the fermionic Hubbard model in a two-dimensional square lattice. In particular in Ref. it was shown that in the strong coupling limit, the inclusion of SOC leads to the rotated antiferromagnetic Heisenberg model, a new class of quantum spin model. Our work explores another aspect of the same Hamiltonian.
[*Acknowledgments.*]{} We acknowledge useful discussions and communications with James Hamlin, Shahal Ilani, Jelena Klinovaja, and Sergey Slizovskiy. The work is supported by EPSRC through the grant EP/P003052/1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of a search for supersymmetry with gauge-mediated breaking and ${\mbox{$\tilde{\chi}_1^0$}}\rightarrow\gamma\Gravitino$ in the $\gamma\gamma$+missing transverse energy final state. In 2.6$\pm$0.2 [$\rm fb^{-1}$]{}of $p{\bar p}$ collisions at $\sqrt{s}$$=$$1.96$ TeV recorded by the CDF II detector we observe no candidate events, consistent with a standard model background expectation of 1.4$\pm$0.4 events. We set limits on the cross section at the 95% C.L. and place the world’s best limit of 149 [GeV/c$^2$]{} on the mass at $\tau_{\tilde{\chi}_1^0}$$\ll$1 ns. We also exclude regions in the mass-lifetime plane for $\tau_{\tilde{\chi}_1^0}$$\lesssim$$2$ ns.'
author:
- 'Eunsin Lee (for CDF Collaboration)'
title: 'Search for Supersymmetry Using Diphoton Events in $p\bar{p}$ Collisions at $\sqrt{s}=1.96$ TeV'
---
Introduction {#introtheory}
============
For theoretical reasons [@gmsb], and because of the ‘$ee\gamma\gamma+$missing transverse energy (${\mbox{${E\!\!\!\!/_T}$}}$)’ candidate event recorded by the CDF detector in RUN I [@run1evnt], there is a compelling rationale to search in high energy collisions for the production of heavy new particles that decay producing the signature of $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$.
An example of a theory that would produce such events is gauge mediated supersymmetry breaking (GMSB) [@gmsb] with ${\mbox{$\tilde{\chi}_1^0$}}\to\gamma\Gravitino$ where the ${\mbox{$\tilde{\chi}_1^0$}}$ is the lightest neutralino and the next-to-lightest supersymmetric particle (NLSP) and the $\Gravitino$ is a gravitino which is the lightest supersymmetric particle (LSP), giving rise to by leaving the detector without depositing any energy. The [$\tilde{G}$]{} also provides a warm dark matter candidate that is both consistent with inflation and astronomical observations [@astro].
In these models, above the current limits from recent experiments [@lep], the is restricted to be well above 100 GeV and is favored to have a lifetime on the order of a nanosecond; the [$\tilde{G}$]{} is restricted to have a mass in the range 0.5$<$${m}_{{\ensuremath{\tilde{G}}}}$$<$1.5 keV$/c^2$ [@cosmology]. At the Tevatron sparticle production is predicted to be primarily into gaugino pairs, and the mass ($m_{\tilde{\chi}_1^0}$) and lifetime ($\tau_{\tilde{\chi}_1^0}$) are the two most important parameters in determining the final states and their kinematics [@gmsb]. Depending on how many of the two ${\mbox{$\tilde{\chi}_1^0$}}$’s decay inside the detector, the event has the signature $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$, $\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ or ${\mbox{${E\!\!\!\!/_T}$}}$ with one or more additional high $E_{T}$ particles from the other gaugino pairs. Different search strategies are required for lifetimes above and below about a nanosecond [@prospects]. Previous searches have been performed for low lifetime models in $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ [@minsuk; @d0search] and nanosecond lifetime models in the delayed $\gamma+jet+{\mbox{${E\!\!\!\!/_T}$}}$ [@delayedPRLD; @lep] final state.
In this analysis we focus on the $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ final state, as recommended in [@prospects], for low lifetime, high-mass models of the ${\mbox{$\tilde{\chi}_1^0$}}$. The new features of our analysis since the last $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ search with 202 pb$^{-1}$ using the CDF detector [@CDFII] are to use the EMTiming system [@nim] and a new [*Met Resolution Model*]{} [@ggXPRD]. We also use 13 times the data (2.6 fb$^{-1}$). These additions significantly enhance our rejection of backgrounds from instrumental and non-collision sources, which allows us to considerably extend the sensitivity of the search for large masses compared to previous Tevatron searches [@d0search]. We also extend the search by considering lifetimes up to 2 ns which are favored for larger $m_{\tilde{\chi}_1^0}$.
Our analysis begins by defining a preselection sample by selecting events with two isolated, central photons with $E_{T} > 13$ GeV. All candidates are required to pass the standard CDF diphoton triggers, global event selection, standard photon ID, and non-collision background rejection requirements [@minsuk; @ggXPRD].
The final signal region for this analysis is defined by the subsample of preselection events that also pass a set of optimized final kinematic requirements. The methods for determining the background in the signal region are based on a combination of data and Monte Carlo (MC) simulation and allow for a large variety of potential final sets of kinematic requirements. We perform an [*a priori*]{} analysis in the sense that we blind the signal region and select the final event requirements based on the signal and background expectations alone. We optimize our predicted sensitivity using a simulation of our GMSB model. We then calculate, for each GMSB parameter point, the lowest, expected 95% C.L. cross section limit in the no-signal scenario [@limitcalc] as a function of the following event variables: MetSig, ${\mbox{$\Delta\phi(\gamma_{1}, \gamma_{2})$}}$, and ${H_{T}}$, each of which will be described in Section \[optimization\].
Data Selection {#dataset}
==============
The analysis is based on 2.59$\pm$0.16 fb$^{-1}$ of data delivered to the CDF detector in Run II. The analysis selection begins with events that pass the CDF diphoton triggers which is effectively 100% efficient for the final diphoton selection requirements [@ggXPRD]. We require both highest-$E_{T}$ photons to be in the fiducial part of the detector with $|\eta|\leq 1.1$, pass the standard photon ID and isolation requirements and have $E_{T}^{\gamma}>$ 13 GeV. In addition to the standard photon ID requirements we have added additional requirements to suppress photomultiplier tube (PMT) high-voltage breakdowns (“spikes”) [@delayedPRLD] and electron rejection requirements [@delayedPRLD] to remove events where an electron fakes a prompt photon (Phoenix tracking rejection). Each event is required to have at least one high quality vertex with $|z_{vx}|\leq$60 cm. The $E_{T}$ of all calorimeter objects (individual towers, photons, electrons, and jets) are calculated with respect to the highest $\sum P_{T}$ vertex. However, an incorrect vertex can be selected when two or more collisions occur in one beam-bunch crossing, making it possible that the highest reconstructed $\sum P_{T}$ vertex does not produce the photons. If assigning the photons to a different vertex lowers the ${\mbox{${E\!\!\!\!/_T}$}}$, we take that and the photon $E_{T}$’s to be from that vertex for all calculations (Vertex Re-assignment).
Additional standard selection requirements are placed to reduce non-collision backgrounds, such as cosmic rays and beam-related (beam halo) effects [@delayedPRLD]. We also apply ${\mbox{${E\!\!\!\!/_T}$}}$ quality requirements (cleanup) to remove events if there is evidence that the second photon ($\gamma_{2}$) or a jet is partially lost in a crack between detector components [@ggXPRD]. Our pre-selection sample consists of 38,053 events left after all the quality, ID and cleanup requirements are applied [@ggXPRD].
Backgrounds {#background}
===========
There are three major sources of background for $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ events: QCD events with fake ${\mbox{${E\!\!\!\!/_T}$}}$, electroweak events with real ${\mbox{${E\!\!\!\!/_T}$}}$, and non-collision events (PMT spikes, cosmic ray or beam-halo events where one or more of the photons and are not related to the collision).
Standard Model QCD sources, $\gamma\gamma$, $\gamma-jet\to\gamma\gamma_{fake}$, and $jet-jet\to\gamma_{fake}\gamma_{fake}$, are the dominant producer of events in the diphoton final state and a major background for $\gamma\gamma$ with fake ${\mbox{${E\!\!\!\!/_T}$}}$. These backgrounds come in two different categories; fake ${\mbox{${E\!\!\!\!/_T}$}}$ due to energy measurement fluctuations in the calorimeter and fake ${\mbox{${E\!\!\!\!/_T}$}}$ due to pathologies such as picking the wrong vertex in events where the true collision did not create a vertex or tri-photon events with a lost photon.
To estimate the background due to energy measurement fluctuations we use the [*Met Resolution Model*]{}. The [*Met Resolution Model*]{} considers the clustered and unclustered energy in the event and calculates a probability, $P({E\!\!\!\!/_{T}~\!\!\!\!^{fluct}}>{\mbox{${E\!\!\!\!/_T}$}})$, for fluctuations in the energy measurement to produce ${E\!\!\!\!/_{T}~\!\!\!\!^{fluct}}$ equivalent to or larger than the measured ${\mbox{${E\!\!\!\!/_T}$}}$. This probability is then used to define MetSig $=-$$\log _{10}
\left(P_{{E\!\!\!\!/_{T}~\!\!\!\!^{fluct}}>{\mbox{${E\!\!\!\!/_T}$}}}\right)$. Events with true and fake ${\mbox{${E\!\!\!\!/_T}$}}$ of the same value should have, on average, different MetSig. For each data event we throw 10 pseudo-experiments to generate a ${\mbox{${E\!\!\!\!/_T}$}}$ and calculate its significance, according to the jets and underlying event configuration. Then we count the number of events in the pseudo-experiments that pass our MetSig and other kinematic requirements. This number, divided by the number of pseudo-experiments, gives us the [*Met Model*]{} prediction for a sample. The systematic uncertainty on the number of events above a MetSig cut is evaluated by comparing the [*Met Model*]{} predictions with the default set of model parameters to predictions obtained with the parameters deviated by $\pm\sigma$. The total uncertainty is estimated by adding the statistical uncertainty on the number of pseudo-experiments passing the cuts and these systematic uncertainties in quadrature.
A source of QCD background that is unaccounted for by the [*Met Model*]{} is diphoton candidate events with event reconstruction pathologies such as a wrong choice of the primary interaction vertex or tri-photon events with a lost photon. To obtain the prediction for the number of events with significant reconstruction pathologies in the QCD background at the same time, we model the kinematics and event reconstruction using a MC simulation of events with in the detector using [pythia]{} [@pythia] and a [geant]{}-based detector simulation [@geant]. We simulate a sample of SM $\gamma\gamma$ events, with large statistics, and normalize to the number of events in the presample to take into account jet backgrounds which should have similar detector response. Then we subtract off the expectations for energy measurement fluctuations in the MC to avoid double counting. The remaining prediction is due to pathologies alone. The systematic uncertainties on this background prediction include the uncertainty on the scale factor and the uncertainty due to MC-data differences in the unclustered energy parameterization and the jet energy scale.
Electroweak processes involving $W$’s and $Z$’s are the most common source of real and significant ${\mbox{${E\!\!\!\!/_T}$}}$ in $p\bar{p}$ collisions. We estimate the background rate from decays into both charged and neutral leptons using a combination of data and MC methods. There are four ways we can get a $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ signature in electroweak events that decay into one or more leptons: 1) from $W\gamma\gamma$ and $Z\gamma\gamma$ events where both photons are real; 2) from $W\gamma$ and $Z\gamma$ events with a fake photon; 3) from $W$ and $Z$ events where both photon candidates are fake photons; and 4) $t\bar{t}$ production and decay. To estimate the contribution from the electroweak backgrounds we use the Baur [@baur] and [pythia]{} MC’s along with a detector simulation, according to their production cross section and k-factors (the ratio of the next-to-leading order (NLO) cross section to the leading order cross section), but normalized to data. To minimize the dependence of our predictions on potential “MC-data” differences, we normalize, using the rate of the number of $e\gamma$ events observed in the data that also pass all signal kinematic cuts, to the number of events observed in MC. This $e\gamma$ sample is derived from diphoton trigger datasets and the events are required to pass the preselection requirements where electrons are required to pass photon-like ID requirements [@ggXPRD]. The uncertainty on the electroweak backgrounds are dominated by the $e\gamma$ normalization factor uncertainty. This includes data and MC statistical uncertainties as well as differences in MC modeling. The total uncertainties also include the MC statistical uncertainties and uncertainties on the normalization factors added in quadrature.
Non-collision backgrounds coming from cosmic rays and beam-related effects can produce $\gamma\gamma$$+$${\mbox{${E\!\!\!\!/_T}$}}$ candidates [@delayedPRLD]. These are estimated using the data. Using the inclusive $\gamma\gamma$ sample selection requirements, but requiring one of the photons to have $t_{\gamma}$$>$$25$ ns we identify a cosmic-enhanced sample. Similarly, we utilize a beam-related background enhanced sample. We estimate the number of these events in the signal region using the ratio of events outside the timing requirements to events inside the signal region and the measured efficiencies of the non-collision rejection requirements [@ggXPRD]. The uncertainties on both non-collision background estimates are dominated by the statistical uncertainty on the number of identified events.
After estimating the MetSig distributions for all the backgrounds, where the QCD is normalized to the data in the low MetSig region where the EWK backgrounds are expected to be negligible, the expected MetSig distribution for the presample is shown in Figure \[fig:bckdist\]. With these tools in hand we are set to estimate the backgrounds for a large variety of kinematic requirements and move to an estimation of the acceptance for GMSB models in the signal region for use in optimization.
[ ![The background predictions of MetSig for the presample. The highest MetSig bin includes all overflow events.[]{data-label="fig:bckdist"}](./metsig_2.6fb_pub.eps "fig:"){width="1.\linewidth"}]{}
GMSB Signal Monte Carlo and Systematic Uncertainties {#acceptance}
====================================================
To estimate the acceptance for GMSB we use the [pythia]{} event generator as well as a full detector simulation. For the purpose of this analysis we consider a GMSB model with parameters fixed on the minimal-GMSB Snowmass slope constraint (SPS 8) that is commonly used [@minsuk; @lep] and take the messenger mass scale $M_{\mathrm{m}}$$=$$2$$\Lambda$, tan($\beta$)$=$15, $\mu$$>$$0$ and the number of messenger fields $N_{\mathrm{m}}$$=$1. The [$\tilde{G}$]{} mass factor and the supersymmetry breaking scale $\Lambda$ are allowed to vary independently. All SUSY production processes are simulated to maximize our sensitivity to the model [@simeon].
Since we estimate the sensitivity of the search to be equal to the expected 95% C.L. cross section limits with the no signal hypothesis, we need the uncertainties for the luminosity, background and acceptance. The systematic uncertainty on the luminosity is taken to be 6% with major contributions from the uncertainties on the CLC acceptance from the precision of the detector simulation [@geant] and the event generator [@pythia]. The background uncertainty is evaluated for every set of cuts in the optimization procedure. The systematic uncertainty on the signal acceptance for an example GMSB point of $m({\mbox{$\tilde{\chi}_1^0$}})$ = 140 GeV and $\tau({\mbox{$\tilde{\chi}_1^0$}})$ $\ll$ 1 ns is estimated to be 6.9% with major contributions from diphoton ID and isolation efficiency (5.4%) and ISR/FSR (3.9%). The uncertainty on the NLO production cross section is dominated by the uncertainty from parton distribution functions (7.6%) and the renormalization scale (2.6%) for a total of 8.0%. All uncertainties are included in the final cross section limit calculation, and we take the acceptance and production cross section uncertainties in quadrature for a total uncertainty of 10.6%.
Optimization and Results {#optimization}
========================
Now that the background is estimated and the signal acceptance is available for a variety of selection requirements, an optimization procedure can be readily employed to find the optimal selection requirements before unblinding the signal region. We optimize for the following kinematic requirements: MetSig, ${H_{T}}$, and ${\mbox{$\Delta\phi(\gamma_{1}, \gamma_{2})$}}$.
As described in earlier section, the MetSig cut gets rid of most of the QCD background with fake ${\mbox{${E\!\!\!\!/_T}$}}$. The ${H_{T}}$ cut separates between the high $E_{T}$, light final state particles produced by GMSB events via cascade decays and SM backgrounds, dominated by QCD and electroweak backgrounds, which do not have lots of high $E_{T}$ objects. The ${\mbox{$\Delta\phi(\gamma_{1}, \gamma_{2})$}}$ cut gets rid of events where two photons are back to back since electroweak backgrounds with large ${H_{T}}$ are typically a high $E_{T}$ photon recoiling against $W\to e\nu$, which means the gauge boson decay is highly boosted. Also the high $E_{T}$ diphoton with large ${H_{T}}$ from QCD background are mostly back-to-back with fake ${\mbox{${E\!\!\!\!/_T}$}}$ or wrong vertex.
By estimating our sensitivity using the 95% C.L. expected cross section limits on GMSB models in the no-signal assumption, we find the optimal set of cuts before unblinding the signal region. We use the standard CDF cross section limit calculator [@junk] to calculate the limits, taking into account the predicted number of background events, the acceptance, the luminosity and their systematic uncertainties.
For each GMSB point the minimum expected cross section limit defines our set of optimal requirements for the mass and lifetime combination. The exclusion region is defined by the region where the production cross section is above the 95% C.L. cross section limit. The mass/lifetime limit is where the two cross. Figure \[fig:xsection\]-(a) shows the expected cross section limit as a function of a kinematic selection requirement after keeping all other requirements fixed at the already optimized values, showing it is at the minimum for a mass-lifetime combination of $m({\mbox{$\tilde{\chi}_1^0$}})=140$ GeV and $\tau({\mbox{$\tilde{\chi}_1^0$}})\ll1$ ns, which is near the exclusion region limit.
We decided to use a single set of optimal requirements before we open the box based on the observation that they will yield the largest expected exclusion region. We chose: MetSig${\mbox{$>$}}3$, ${H_{T}}{\mbox{$>$}}200$ GeV, ${\mbox{$\Delta\phi(\gamma_{1}, \gamma_{2})$}}{\mbox{$<$}}\pi-0.35$ rad. With these requirements we predict a total of 1.38$\pm$0.44 background events. The dominant electroweak contributions are $Z\gamma\to\nu\nu\gamma$ and $Z\gamma\to\mu\mu\gamma$ which produce a total of 0.26$\pm$0.08 and 0.19$\pm$0.10 events respectively. The QCD background is dominated by energy measurement fluctuations in the ${\mbox{${E\!\!\!\!/_T}$}}$, estimated using the [*Met Model*]{}, to have a rate of 0.46$\pm$0.24 events. The non-collision backgrounds are dominated by cosmic ray which have a rate of 0.001$^{+0.008}_{-0.001}$ events.
After all optimal cuts we open the box and observe no events, consistent with the expectation of 1.2$\pm$0.4 events. We show the kinematic distributions for the background and signal expectations along with the data in Figure \[fig:xsection\]-(b). There is no distribution that hints at an excess and the data appears to be well modeled by the background prediction alone.
We show the predicted and observed cross section limits along with the NLO production cross section, which is calculated by multiplying the [pythia]{} LO cross section calculation by k-factor [@kfactors] in Figure \[fig:xsecmasslife\]. Since the number of observed events is below expectations the observed limits are slightly better than the expected limits. The ${\mbox{$\tilde{\chi}_1^0$}}$ mass reach, based on the predicted (observed) number of events is 141 GeV/$c^{2}$ (149 GeV/$c^{2}$), at a lifetime below 2 ns [@prospects]. We show the 95% C.L. NLO exclusion region as a function of mass and lifetime of ${\mbox{$\tilde{\chi}_1^0$}}$ using the fixed choice of cuts from the optimization for both for the predicted and observed number of background events in Figure \[fig:highlum\]-(a). These limits extend the reach beyond the CDF delayed photon results [@delayedPRLD] and well beyond those of DØsearches at $\tau_{{{\mbox{$\tilde{\chi}_1^0$}}}}\ll0$ [@d0search] and the limit from ALEPH/LEP [@lep], and are currently the world’s best.
Conclusions and Prospects for the future {#conclusion}
========================================
We have set limits on GMSB models using the $\gamma\gamma+{\mbox{${E\!\!\!\!/_T}$}}$ final state. Candidate events were selected based on 13 times more data, the new resolution model technique, the EMTiming system and a full optimization procedure. We found 0 events using 2.6 $fb^{-1}$ of data in run II which is consistent with the background estimate of 1.2$\pm$0.4 events from the Standard Model expectations. We showed exclusion regions and set limits on GMSB models with a ${\mbox{$\tilde{\chi}_1^0$}}$ mass reach of 149 GeV/$c^{2}$ at a ${\mbox{$\tilde{\chi}_1^0$}}$ lifetime much less than 1 ns. Our results extend the world sensitivity to these models.
To investigate the prospects of a search at higher luminosity we calculate the cross section limits assuming all backgrounds scale linearly with luminosity while their uncertainty fractions remain constant. By the end of Run II, with an integrated luminosity of 10 [$\rm fb^{-1}$]{}, we estimate a mass reach of $\simeq
160$ [GeV/c$^2$]{} at a lifetime much less than 1 ns, as shown in Figure \[fig:highlum\]-(b). For higher lifetimes (above $\sim$2 ns) the next generation delayed photon analysis will extend the sensitivity taken from Ref. [@delayedPRLD] and then will combine these results for completeness.
Eunsin Lee would like thank D. Toback, R. Culberton, A. Pronko, M. D’onofrio, T. Wright for their help on this analysis and the talk.
[99.]{} S. Dimopoulos, S. Thomas, and J. Wells, Nucl. Phys. B [**488**]{}, 39 (1997); S. Ambrosanio, G. Kribs, and S. Martin, Phys. Rev. D [**56**]{}, 1761 (1997); G. Giudice and R. Rattazzi, Phys. Rep. [**322**]{}, 419 (1999); S. Ambrosanio, G. Kane, G. Kribs, S. Martin, and S. Mrenna, Phys. Rev. D [**55**]{}, 1372 (1997). CDF Collaboration, F. Abe [*et al.*]{}, Phys. Rev. Lett. [**81**]{}, 1791 (1998); Phys. Rev. D [**59**]{}, 092002 (1999). P. Bode, J. Ostriker, and N. Turok, Astrophys. J. [**556**]{}, 93 (2001). ALEPH Collaboration, A. Heister *et. al.*, Eur. Phys. J. C** 25**, 339 (2002); A. Garcia-Bellido, Ph.D. thesis, Royal Holloway University of London (2002) (unpublished), arXiv:hep-ex/0212024. C.-H. Chen and J. F. Gunion, Phys. Rev. D [**58**]{}, 075005 (1998). CDF Collaboration, D. Acosta *et. al*., Phys. Rev. D** 71**, 031104 (2005). DØ Collaboration, V.M. Abazov *et. al*., Phys. Lett. B** 659**, 856 (2008). CDF Collaboration, A. Abdulencia *et. al.*, Phys. Rev. Lett** 99**, 121801 (2007); CDF Collaboration, T.Aaltonen *et. al.*, Phys. Rev. D** 78**, 032015 (2008). P. Wagner and D. Toback, Phys. Rev. D [**70**]{}, 114032 (2004). D. Acosta [[*et al.*]{}]{} (CDF Collaboration), Phys. Rev. D [**71**]{}, 032001 (2005). M. Goncharov *et. al.*, NIM A**565**, 543 (2006). T. Aaltonen [[*et al.*]{}]{} (CDF Collaboration), [*Search for Anomalous Production of Events with Two Photons and Additional Energetic Objects at CDF*]{}, to be submitted to Phys. Rev. D. For a discussion of the jet energy measurements, see T. Affolder [[*et al.*]{}]{}(CDF Collaboration), Phys. Rev. D. [**64**]{}, 032001 (2001). For a discussion of standard jet correction systematics, see A. Bhatti [[*et al.*]{}]{}, Nucl. Instrum. Methods, A 566, 375 (2006). We use jets with cone size $\Delta$$R$$=$$0.4$. E. Boos, A. Vologdin, D. Toback and J. Gaspard, Phys. Rev. D **66**, 013011 (2002). T. Sjöstrand [[*et al.*]{}]{}, Comput. Phys. Commun. [**135**]{}, 238 (2001). We use version 6.216. We use the standard [geant]{} based detector simulation \[R. Brun [[*et al.*]{}]{}, CERN-DD/EE/84-1 (1987)\] and add a parametrized EMTiming simulation. U. Baur, T. Han and J. Ohnemus, Phys. Rev. D [**48**]{}, 5140 (1993); U. Baur, T. Han and J. Ohnemus, [*ibid.*]{} [**57**]{}, 2823 (1998); The $W\gamma$ and $Z\gamma$ processes are simulated using the leading-order event generator and have a k-factor fixed at 1.36. Initial and final state radiation (resulting in additional jets or photons), underlying event, and additional interactions are simulated using [pythia]{} [@pythia]. P. Simeon and D. Toback, J. Undergrad. Research in Phys. 20, 1 (2007). T. Junk, Nucl. Instrum. Methods A**434**, 435-443 (1999). W. Beenakker, [[*et al.*]{}]{}, Phys. Rev. Lett. [**83**]{}, 3780 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the shifts in the gauge couplings of the top quark induced in the Littlest Higgs model with and without T parity. We find that the ILC will be able to observe the shifts throughout the natural range of model parameters.'
author:
- 'C. F. Berger'
- 'M. Perelstein'
- 'F. Petriello'
title: |
\
Top Quark Properties in Little Higgs Models
---
$\vcenter{
\hbox{\bf MADPH-05-1251}
\hbox{\bf SLAC-PUB-11589}}
$
INTRODUCTION
============
Identifying the mechanism which breaks electroweak symmetry and generates fermion masses is one of the main physics goals for both the LHC and the ILC. Studies of the top quark have the potential to illuminate this issue; since it is the heaviest of the Standard Model (SM) fermions, the top is expected to couple strongly to the symmetry-breaking sector. Consequently, the structure of that sector can have significant, potentially observable effects on the properties of the top. For example, it is well known that the vector and axial $t\bar{t}Z$ form factors receive large corrections (of order 5-10%) in certain models of dynamical electroweak symmetry breaking [@snowmass01]. At future colliders such as the LHC and the ILC, we will be able to pursue a program of precision top physics, similar to the program studying the $Z$ at LEP and SLC. In this manuscript, we study the corrections to the top quark properties in “Little Higgs” models of electroweak symmetry breaking [@review], and compare the expected deviations from the SM predictions with expected sensitivities of experiments at the LHC and the ILC.
In the Little Higgs models, electroweak symmetry is driven by the radiative effects from the top sector, including the SM-like top and its heavy counterpart, a TeV-scale “heavy top” $T$. Probing this structure experimentally is quite difficult. While the LHC should be able to discover the $T$ quark, its potential for studying its couplings is limited [@PPP; @ATLAS]. Direct production of the $T$ will likely be beyond the kinematic reach of the ILC. However, we will show below that the corrections to the gauge couplings of the SM top, induced by its mixing with the $T$, will be observable at the ILC throughout the parameter range consistent with naturalness. Measuring these corrections will provide a unique window on the top sector of the Little Higgs.
Little Higgs models contain a light Higgs boson which is a [*composite*]{} of more fundamental degrees of freedom. A generic composite Higgs model must become strongly coupled at an energy scale around 1 TeV, leading to unacceptably large corrections to precision electroweak observables. In contrast, Little Higgs models remain perturbative until a higher energy scale, around 10 TeV. The hierarchy between the Higgs mass and the strong coupling scale is natural and stable with respect to radiative corrections. Because of the special symmetry structure of the theory, the Higgs mass vanishes at tree level, as do one-loop quadratically divergent diagrams. The mass term is dominated by the logarithmically divergent one-loop contribution from the top quark, which triggers electroweak symmetry breaking.
Many Little Higgs models have been proposed in the literature. We will consider two examples in this study, the “Littlest Higgs” model [@littlest], and its variation incorporating T parity [@LHT].
THE LITTLEST HIGGS
==================
As our first example, consider the $SU(5)/SO(5)$ Littlest Higgs (LH) model [@littlest]. Since the original model turned out to be severely constrained by precision electroweak data [@pew], we focus on the version with a reduced gauge group, $SU(2)\times SU(2)\times U(1)$, which is significantly less constrained. We follow the conventions and notation of Ref. [@PPP]. The new TeV-scale states are the gauge bosons $W_H^\pm, W_H^3$, a vector-like weak-singlet quark $T$ (the “heavy top”), and a weak-triplet scalar field $\phi$. The model is parametrized by the symmetry breaking scale $f$ (assumed to be of order 1 TeV), the $SU(2)$ mixing angle $\psi$, the $htT$ coupling constant $\lambda_T$, and the triplet vacuum expectation value (vev) $v^\prime$. It can be shown that $v^\prime\sim v^2/f\ll v$, where $v=246$ GeV is the SM Higgs vev. In this analysis, we will set $v^\prime=0$, since the effects of a non-vanishing $v'$ on the observables considered here are numerically small. Instead of $f$, we will use the more physical quantity, the heavy top mass $m_T$, in our plots; these are related by $m_T/f = (\lambda_t^2+\lambda_T^2)/\lambda_T$, where $\lambda_t\approx 1$ is the SM top Yukawa. Naturalness arguments put an upper bound on this parameter, $m_T\lapproxeq 2$ TeV.
{width="60mm"} {width="60mm"}
Corrections to the gauge couplings of the top quark in the LH model arise from two sources: the mixing of the (left-handed) top with the heavy top $T$, and the mixing of the SM gauge bosons $W^\pm, Z^0$ with their heavy counterparts, $W^\pm_H$ and $W_H^3$. Using the superscripts “t” and “g” to denote the contributions from these two sources, the corrections to the $t\bar{t}Z$ coupling can be written as g\^[Z[t]{}]{}\_R = 0, & & g\^[Z[g]{}]{}\_R = g\_R\^Z,\
g\^[Z[t]{}]{}\_L = , & & g\^[Z[g]{}]{}\_L = . Here, $g_{L,R}^Z$ are the SM left- and right-handed $t\bar{t}Z$ couplings, $g_V^Z=(g_R^Z+g_L^Z)/2$ and $g_A^Z=(g_R^Z-g_L^Z)/2$ are their vector and axial combinations, $c_W,s_W$ are respectively the cosine and sine of the weak mixing angle, and $s_\psi\equiv
\sin\psi$, $c_\psi\equiv\cos\psi$. The predicted shifts in the $t\bar{t}Z$ axial and vector couplings for $m_T=0.5, 1.0,$ and 2.0 TeV, and $\lambda_T=0.5, 1, 2$, are plotted in Fig. \[fig:LH\] (left panel), along with the experimental sensitivities expected at the LHC [@LHC] and the ILC [@snowmass01]. The mixing angle $\psi$ is varied between $0$ and $\pi/2$. Note that the shifts have a definite sign. While only a rather small part of the parameter space is accessible at the LHC even with 3000 fb$^{-1}$ integrated luminosity, the ILC experiments will be able to easily observe the shifts in most of the parameter space preferred by naturalness considerations.
The corrections to the $t\bar{b}W$ coupling have the form g\^[W[t]{}]{}\_R = 0, & & g\^[W[g]{}]{}\_R = 0,\
g\_L\^[W[t]{}]{} = -, & & g\_L\^[W [g]{}]{} = g\^W s\_\^2 (c\_\^2- s\_\^2 -), where $g^W$ is the SM $t\bar{b}W$ coupling. These corrections induce a shift in the top width, $\delta\Gamma_t/\Gamma_t=2\delta g_L^W/g^W$. The induced shift, as a function of the angle $\psi$, is plotted in the right panel of Fig. \[fig:LH\], where the parameter $\lambda_T$ is varied between 0.5 and 2 for $m_T=0.5, 1.0$ and 2.0 TeV. The accuracy of the top width measurement expected at the ILC [@snowmass01] will allow to observe this effect in most of the natural parameter space.
The shift in the $t\bar{t}h$ coupling has the form = - , where the reference value is $\lambda_{t\bar{t}h} = \sqrt{2}m_t/v$, and $v$ is inferred from the experimentally measured values for $G_F$, $M_Z$ and $\alpha$.
LITTLEST HIGGS WITH T PARITY
============================
The LH model can be extended to include a discrete symmetry, T parity [@LHT], which greatly reduces the contributions to precision electroweak observables [@HMNP]. The main new feature in this model is the absence of the gauge boson mixing, since light and heavy gauge bosons have opposite charges under T parity. The top-heavy top mixing is still present, however. The resulting corrections to the $t\bar{t}Z$ and $t\bar{b}W$ vertices are identical to the corresponding shifts in the model without T parity, $\delta
g^{Z{\rm t}}_L$ and $\delta g^{W{\rm t}}_L$, given in Eqs. and . The shift in the axial $t\bar{t}Z$ coupling is plotted in the left panel of Fig. \[fig:Tpar\]. (The shift in the vector coupling is identical up to a sign.) The correction to the top width is shown in the right panel of Fig. \[fig:Tpar\]. Again, both effects should be observable at the ILC.
{width="60mm"} {width="60mm"}
The authors wish to thank Shrihari Gopalakrishna, JoAnne Hewett, Michael Peskin and Tom Rizzo for useful discussions. CFB is supported by the US Department of Energy under contract DE-AC02-76SF00515, MP is supported by the NSF grant PHY-0355005, and FP is supported by the University of Wisconsin Research Committee with funds granted by the Wisconsin Alumni Research Foundation.
[9]{} T. Abe [*et al.*]{} \[American Linear Collider Working Group\], in [*Proc. of the APS/DPF/DPB Summer Study on the Future of Particle Physics (Snowmass 2001)* ]{} ed. N. Graf, arXiv:hep-ex/0106057. For a recent review and more references, see M. Schmaltz and D. Tucker-Smith, arXiv:hep-ph/0502182. M. Perelstein, M. E. Peskin and A. Pierce, Phys. Rev. D [**69**]{}, 075002 (2004) \[arXiv:hep-ph/0310039\]. G. Azuelos [*et al.*]{}, Eur. Phys. J. C [**39S2**]{}, 13 (2005) \[arXiv:hep-ph/0402037\]. N. Arkani-Hamed, A. G. Cohen, E. Katz and A. E. Nelson, JHEP [**0207**]{}, 034 (2002) \[arXiv:hep-ph/0206021\]. H. C. Cheng and I. Low, JHEP [**0309**]{}, 051 (2003) \[arXiv:hep-ph/0308199\]; I. Low, JHEP [**0410**]{}, 067 (2004) \[arXiv:hep-ph/0409025\]. C. Csaki, J. Hubisz, G. D. Kribs, P. Meade and J. Terning, Phys. Rev. D [**67**]{}, 115002 (2003) \[arXiv:hep-ph/0211124\]; J. L. Hewett, F. J. Petriello and T. G. Rizzo, JHEP [**0310**]{}, 062 (2003) \[arXiv:hep-ph/0211218\].
U. Baur, A. Juste, L. H. Orr and D. Rainwater, Phys. Rev. D [**71**]{}, 054013 (2005) \[arXiv:hep-ph/0412021\]. J. Hubisz, P. Meade, A. Noble and M. Perelstein, arXiv:hep-ph/0506042.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.